Université de Technologie de Belfort–Montbéliard

Institut de Recherche sur les Transports, l’Énergie et la Société

Laboratoire Systèmes et Transports
École Doctorale Sciences pour l’Ingénieur et Microtechniques

Algorithmes et Architectures Multi-Agents pour la Gestion

de l’Énergie dans les Réseaux Électriques Intelligents
Application aux Centrales à Turbines à Gaz et à l’Effacement Diffus Résidentiel

Thèse no 191 présentée et soutenue publiquement en vue de l’obtention du titre de

Docteur en Génie Électrique

par

Robin Roche

le 7 décembre 2012

Membres du jury :

M. Mohamed Benbouzid IUT de Brest, LBMS Président

M. Nouredine Hadjsaid INP Grenoble, G2Elab Rapporteur
M. Lhassane Idoumghar Université de Haute Alsace, LMIA Co-encadrant
M. Emmanuel Kerrand GE Power & Water Membre invité
M. Abderrafiâa Koukam UTBM, IRTES-SET Examinateur
M. Abdellatif Miraoui UTBM, IRTES-SET Directeur de thèse
M. Stéphane Ploix INP Grenoble, GSCOP Rapporteur
M. Marcelo G. Simões Colorado School of Mines Examinateur
M. Siddharth Suryanarayanan Colorado State University Co-encadrant
 Agent-Based Architectures and Algorithms
for Energy Management in Smart Grids
Application to Smart Power Generation and Residential Demand Response

Dissertation submitted in fulfilment of the requirements for the degree of

Doctor of Philosophy (Ph.D.) in Electrical Engineering

of the

University of Technology of Belfort–Montbéliard

by

Robin Roche

December 7, 2012
In the memory of Dr. Benjamin Blunier
 On rencontre sa destinée
Souvent par des chemins qu’on prend pour l’éviter.

Jean de La Fontaine,
Livre VIII (1678–1679), L’horoscope

Please refer to this work as follows:

Robin Roche, Agent-Based Architectures and Algorithms for Energy Management in Smart
Grids: Application to Smart Power Generation and Residential Demand Response. Ph.D.
dissertation, Université de Technologie de Belfort-Montbéliard, Belfort, France, December
2012.

The author and the advisors give the authorization to consult and to copy parts of this
work for personal use only. Permission to reproduce any material contained in this work
should be obtained from the author.
 Abstract

Due to the convergence of several profound trends in the energy sector, smart grids
are emerging as the main paradigm for the modernization of the electric grid. Smart
grids hold many promises, including the ability to integrate large shares of distributed
and intermittent renewable energy sources, energy storage and electric vehicles, as well
as the promise to give consumers more control on their energy consumption. Such goals
are expected to be achieved through the use of multiple technologies, and especially of
information and communication technologies, supported by intelligent algorithms.
These changes are transforming power grids into even more complex systems, that
require suitable tools to model, simulate and control their behaviors. In this dissertation,
properties of multi-agent systems are used to enable a new systemic approach to energy
management, and allow for agent-based architectures and algorithms to be defined. This
new approach helps tackle the complexity of a cyber-physical system such as the smart grid
by enabling the simultaneous consideration of multiple aspects such as power systems, the
communication infrastructure, energy markets, and consumer behaviors. The approach is
tested in two applications: a “smart” energy management system for a gas turbine power
plant, and a residential demand response system.
An energy management system for gas turbine power plants is designed with the objec-
tive to minimize operational costs and emissions, in the smart power generation paradigm.
A gas turbine model based on actual data is proposed, and used to run simulations with
a simulator specifically developed for this problem. A metaheuristic achieves dynamic
dispatch among gas turbines according to their individual characteristics. Results show
that the system is capable of operating the system properly while reducing costs and emis-
sions. The computing and communication requirements of the system, resulting from the
selected architecture, are also evaluated.
With other demand-side management techniques, demand response enables reducing
load during a given duration, for example in case of a congestion on the transmission
system. A demand response system is proposed and relies on the use of the assets of
residential customers to curtail and shift local loads (hybrid electric vehicles, air condi-
tioning, and water heaters) so that the total system load remains under a given threshold.
Aggregators act as interfaces between grid operators and a demand response market. A
simulator is also developed to evaluate the performance of the proposed system. Results
show that the system manages to maintain the total load under a threshold by using
available resources, without compromising the steady-state stability of the distribution
system.

i
 Résumé

Avec la convergence de plusieurs tendances profondes du secteur énergétique, les

réseaux électriques intelligents (smart grids) émergent comme le paradigme principal pour
la modernisation des réseaux électriques. Les smart grids doivent notamment permettre
d’intégrer de larges proportions d’énergie renouvelable intermittente, de stockage et de
véhicules électriques, ainsi que donner aux consommateurs plus de contrôle sur leur con-
sommation énergétique. L’atteinte de ces objectifs repose sur l’adoption de nombreuses
technologies, et en particulier des technologies de l’information et de la communication.
Ces changements transforment les réseaux en des systèmes de plus en plus complexes,
nécessitant des outils adaptés pour modéliser, contrôler et simuler leur comportement.
Dans cette thèse, l’utilisation des systèmes multi-agents (SMA) permet une approche
systémique de la gestion de l’énergie, ainsi que la définition d’architectures et d’algorithmes
bénéficiant des propriétés des SMA. Cette approche permet de prendre en compte la
complexité d’un tel système cyber-physique, en intégrant de multiples aspects comme
le réseau en lui-même, les infrastructures de communication, les marchés ou encore le
comportement des utilisateurs. L’approche est mise en valeur à travers deux applications.
Dans une première application, un système de gestion de l’énergie pour centrales à
turbines à gaz est conçu avec l’objectif de minimiser les coûts de fonctionnement et les
émissions de gaz à effet de serre pour des profils de charge variables. Un modèle de turbine
à gaz basé sur des données réelles est proposé et utilisé dans un simulateur spécifiquement
développé. Une métaheuristique optimise dynamiquement le dispatching entre les tur-
bines en fonction de leurs caractéristiques propres. Les résultats montrent que le système
est capable d’atteindre ses objectifs initiaux. Les besoins en puissance de calcul et en
communication sont également évalués.
Avec d’autres mesures de gestion de la demande, l’effacement diffus permet de réduire
temporairement la charge électrique, par exemple dans la cas d’une congestion du réseau
de transport. Dans cette seconde application, un système d’effacement diffus est proposé
et utilise les ressources disponibles chez les particuliers (véhicules électriques, climatisa-
tion, chauffe-eau) pour maintenir la demande sous une valeur limite. Des aggrégateurs
de capacité de réduction de charge servent d’interface entre les opérateurs du réseau et
un marché de l’effacement. Un simulateur est également développé pour évaluer la per-
formance du système. Les résultats de simulations montrent que le système réussit à
atteindre ses objectifs sans compromettre la stabilité du réseau de distribution en régime
continu.

iii
 Acknowledgments

It is my pleasure to thank all who have helped and inspired me during the last three
years, from the day I became a Ph.D. student to the day of my defense.
At first, I would like to thank Dr. Nouredine Hadjsaid and Dr. Stéphane Ploix for
accepting to review this dissertation, despite their very busy schedules, and for their
helpful comments. I would also like to thank Dr. Mohamed Benbouzid for serving as
chair of the defense committee, and Dr. Emmanuel Kerrand, Dr. Abderrafiâa Koukam
and Dr. Marcelo G. Simões for accepting to participate in this committee.
I am very grateful to Dr. Abdellatif Miraoui for giving me the opportunity to work
under his supervision in the flourishing field of smart grids. His leadership has been a
great motivation and taught me to dare take risks and be ambitious.
I would then like to thank my co-advisors, and especially Dr. Benjamin Blunier, who
passed away in February 2012. His communicative enthusiasm for research and education
greatly inspired me. The numerous passionate and open discussions we had about a wide
variety of topics taught me a lot, and had enabled us to work very well together. He is,
and will be missed. As a co-advisor, as a colleague, and as a friend. I am thankful to
Dr. Lhassane Idoumghar for his continuous support and guidance, and for introducing
me to optimization. His precious help contributed a lot to the realization of large parts
of the work presented in this dissertation. Dr. Siddharth Suryanarayanan has been a
great motivating professor and a wonderful co-advisor. In just a few months, his extensive
knowledge in power systems, his insights, his efficient work methods, and his demand for
quality and excellence have influenced me and my work more than I could say. I am also
very thankful to him and his students for their support when I needed it most.
I would also like to thank UTBM, and especially the UTBM Foundation for funding
me throughout these years, and the Colorado State University Energy Supercluster for
funding my stay in Colorado.
Many people in Belfort have made my everyday life more productive, interesting,
and fun. My fellow doctoral students and colleagues at UTBM, with their friendship,
good temper, and amazing ability to come up with funny/weird ideas and stories, gave
me the strength to keep working toward my goal: Dr. Mohammad Kabalo, Mr. Hugues
Ostermann, Dr. Alexandre Ravey, Mr. Nicolas Watrin, and Mr. Dongdong Zhao. Without
them, these years would have been a lot more difficult, and I think our mutual support has
been an essential element in our successes. My thanks also go to, in no particular order,
Dr. El-Hassane Aglzim, M. Gillian Basso, Ms. Florence Berthold, Dr. David Bouquain, Dr.
Béatrice Bouriot, Dr. Daniela Chrenko, Mr. Sébastien Faivre, Dr. Fei Gao, Dr. Arnaud
Gaillard, Ms. Ariane Glatiny, Dr. Jérémie, Dr. Fabrice Lauri, M’Boua, Dr. Damien Paire,
and Dr. Dimitri Torregrossa, for their helpfulness and friendliness. My thanks also go to

v
vi

the numerous other people I cannot list, especially people from the EE department, from
IRTES-SET, and from GE Energy (now GE Power & Water).
I also want to thank Mr. Ayan Bhattacharyya, Mr. Manish Mohanpurkar, Mr. Sudar-
shan Natarajan, Mr. David Palchak, Mr. Mayank Panwar, Dr. Siddharth Suryanarayanan,
Prof. Daniel Zimmerle, and Mr. Adam Zipperer at Colorado State University, as well as
Dr. Marcelo G. Simões from Colorado School of Mines, for their very warm welcome. I
had a wonderful time in Colorado due to them.
Finally, I am extremely grateful to my family for their love, encouragement and con-
fidence in me. Specifically, I would like to thank my parents and my sister for their
continuous trust and support throughout the years. And last but not least, I want to
thank Aurélie for her support, especially over the last few months.
 Contents

Abstract i

Résumé iii

Acknowledgments v

Contents vii

1 Energy Management in Smart Power Systems 1

1.1 Toward Smarter Power Grids . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.1 Drivers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.2 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1.3 Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.1.4 Technologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.1.5 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.2 Modernization of Control Systems . . . . . . . . . . . . . . . . . . . . . . . 13
1.2.1 Power System Restructuring . . . . . . . . . . . . . . . . . . . . . . 13
1.2.2 Control Systems Typology . . . . . . . . . . . . . . . . . . . . . . . . 14
1.3 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.4 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2 Agent-Based Modeling, Control and Simulation 21

2.1 Multi-Agent Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.1.1 Smart Grids as Complex Adaptive Systems . . . . . . . . . . . . . . 22
2.1.2 Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.1.3 Relevance for Power Systems Applications . . . . . . . . . . . . . . . 24
2.1.4 Communication, Languages and Ontologies . . . . . . . . . . . . . . 29
2.1.5 Agent Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.1.6 Topologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.1.7 Inter-Agent Interaction . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.2 Multi-Agent Development Framework . . . . . . . . . . . . . . . . . . . . . 38
2.2.1 Development Platforms . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.2.2 The JADE Platform . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.3 Co-Simulation Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.3.1 Need for Co-Simulation Tools . . . . . . . . . . . . . . . . . . . . . . 41
2.3.2 Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.3.3 Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

vii
viii Contents

2.3.4 Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

3 Gas Turbine Power Plants for Smart Power Generation 47

3.1 Smart Power Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.1.1 Drivers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.1.2 Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.1.3 State-of-the-Art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.1.4 Application to Gas Turbine Power Plants . . . . . . . . . . . . . . . 50
3.1.5 Energy Management Systems for Smart Generating Plants . . . . . 51
3.2 Gas Turbine Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.2.1 Fuel Consumption . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.2.2 Combustion Gas Emissions . . . . . . . . . . . . . . . . . . . . . . . 53
3.2.3 Starting and Stopping Cycles . . . . . . . . . . . . . . . . . . . . . . 55
3.3 Energy Management System Architecture . . . . . . . . . . . . . . . . . . . 56
3.3.1 Agent Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.3.2 Selected MAS Architecture . . . . . . . . . . . . . . . . . . . . . . . 57
3.3.3 Agent Interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.4 Energy Management Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.4.1 Start and Stop Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 64
3.4.2 Turbine Operation Ranges . . . . . . . . . . . . . . . . . . . . . . . . 65
3.5 Economic and Environmental Dispatch . . . . . . . . . . . . . . . . . . . . . 66
3.5.1 Single and Multi-Objective Optimization . . . . . . . . . . . . . . . 66
3.5.2 Problem Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.5.3 Aggregation-Based Power Dispatching . . . . . . . . . . . . . . . . . 69
3.6 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.6.1 MAS Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.6.2 Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.6.3 Dispatching Algorithms Comparison . . . . . . . . . . . . . . . . . . 73
3.6.4 Performance Coefficients Effectiveness . . . . . . . . . . . . . . . . . 77
3.6.5 Start and Stop Algorithm Effectiveness . . . . . . . . . . . . . . . . 77
3.6.6 Energy Costs Comparison . . . . . . . . . . . . . . . . . . . . . . . . 79
3.6.7 Flexibility and Resilience Test . . . . . . . . . . . . . . . . . . . . . 79
3.6.8 Communication and Computation Requirements . . . . . . . . . . . 82
3.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

4 Aggregator-Based Residential Demand Response 85

4.1 Demand Response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.1.1 Drivers and Components . . . . . . . . . . . . . . . . . . . . . . . . 86
4.1.2 State-of-the-Art . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.1.3 Proposed Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.2 System Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.2.1 T&D Infrastructure . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.2.2 Demand Response Aggregators . . . . . . . . . . . . . . . . . . . . . 91
4.2.3 Demand Response Market . . . . . . . . . . . . . . . . . . . . . . . . 91
4.2.4 Residential Customers . . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.3 Residential Load Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
Contents ix

4.3.1 Enabling Technologies and Assumptions . . . . . . . . . . . . . . . . 93

4.3.2 Electric Water Heater Model . . . . . . . . . . . . . . . . . . . . . . 93
4.3.3 Air Conditioning Model . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.3.4 Appliances Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.3.5 PHEV Fleet Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.3.6 PV Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.4 System Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.4.1 Objective and Constraints . . . . . . . . . . . . . . . . . . . . . . . . 101
4.4.2 Metering Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
4.4.3 Demand Response Event Mode . . . . . . . . . . . . . . . . . . . . . 102
4.4.4 Rescheduling Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.5 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.5.1 Simulator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.5.2 Test Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
4.5.3 Simulation Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . 108
4.5.4 System-wide Net Load Results . . . . . . . . . . . . . . . . . . . . . 110
4.5.5 Results for Residential Customers . . . . . . . . . . . . . . . . . . . 113
4.5.6 Impact on the Distribution System . . . . . . . . . . . . . . . . . . . 113
4.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

5 Conclusions 117
5.1 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5.1.1 Conclusions on the Proposed Approach . . . . . . . . . . . . . . . . 117
5.1.2 Conclusions on the Presented Applications . . . . . . . . . . . . . . 118
5.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
5.3 Scientific Production Overview . . . . . . . . . . . . . . . . . . . . . . . . . 121

Appendices 123

A Metaheuristics for Optimal Dispatching 125

A.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
A.2 Common Metaheuristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
A.2.1 Particle Swarm Optimization (PSO) . . . . . . . . . . . . . . . . . . 125
A.2.2 Differential Evolution (DE) . . . . . . . . . . . . . . . . . . . . . . . 126
A.2.3 Genetic Algorithm (GA) . . . . . . . . . . . . . . . . . . . . . . . . . 128
A.2.4 Imperialist Competitive Algorithm (ICA) . . . . . . . . . . . . . . . 129
A.3 Hybrid Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
A.3.1 Metropolis PSO with Mutation Operation (MPSOM) . . . . . . . . 132
A.3.2 ICA-PSO Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
A.4 Performance Comparison and Analysis . . . . . . . . . . . . . . . . . . . . . 135
A.4.1 Benchmark Functions . . . . . . . . . . . . . . . . . . . . . . . . . . 135
A.4.2 Algorithms Implementation . . . . . . . . . . . . . . . . . . . . . . . 135
A.4.3 Test Results and Analysis . . . . . . . . . . . . . . . . . . . . . . . . 138

B Publications 141
x Contents

Bibliography 145

List of Acronyms 163

List of Figures 166

List of Tables 168

 1

Energy Management in Smart Power Systems

Contents
1.1 Toward Smarter Power Grids . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.1 Drivers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.2 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1.3 Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.1.4 Technologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.1.5 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.2 Modernization of Control Systems . . . . . . . . . . . . . . . . . . . . . 13
1.2.1 Power System Restructuring . . . . . . . . . . . . . . . . . . . . 13
1.2.2 Control Systems Typology . . . . . . . . . . . . . . . . . . . . . 14
1.3 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.4 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

1
2 Chapter 1. Energy Management in Smart Power Systems

1.1 Toward Smarter Power Grids

The last few years have seen the popularization of the smart grid concept, a new
paradigm for power systems. This relatively new concept, through the need for smarter
energy management systems in the electricity grid, is at the core of this dissertation.

1.1.1 Drivers
The electric power grid is over a century old in most Western countries, and has seen
relatively few breakthrough innovations since its creation. Some even argue that Thomas
Edison would recognize most of the equipment in use in today’s power grids, contrary to
Graham Bell in the telecommunication infrastructure [1]. However, due to the combination
of several profound trends, power grids will have to evolve and become much smarter than
they are today. Drivers for such changes include [2]:

– Increasing energy demand. Worldwide energy demand is expected to rise by over

150 % by 2050 [3], due to population and economic growths, and the use of new tech-
nologies such as electric vehicles (EVs). As installing new traditional and centralized
generation capacity and transmission lines is extremely expensive, alternative solu-
tions are being considered [4].

– Environmental concerns. Evidences of global warming throughout the world are

pushing legislators to foster the use of new technologies and non-pollutant energy
sources to reduce greenhouse gases emissions. The European Union has set the goal
of achieving 20 % reduction in carbon dioxide emissions, 20 % improvement in energy
efficiency and a 20 % share of renewable energy sources (RESs) in the energy mix,
by 2020 from 1990 levels. Similarly, in the United States (US), renewable portfolio
standards (RPSs) were introduced with the aim of increasing energy production from
RESs [5].

– Increasing share of intermittent renewable generation. RESs are an answer to these

environmental concerns, due to their low emission levels. However, their intermit-
tency is a source of instability for power systems, as the balance between supply
and demand is harder to maintain with such sources present in the energy mix [4,6].
Fig. 1.1 shows the solar radiation and wind speed measured in Belfort on June 6,
2011, at a height of 15 m 1 , and the total French demand observed by the transmis-
sion system operator (TSO) RTE [7]. This data shows that due to sudden variations
in solar radiation (due to clouds, for example) or wind speed, the output of RESs
can change very rapidly. Short-term generation peaks and long periods without
generation can be observed within the same day, whereas demand tends to change
rather slowly throughout the day.

– Increasing share of distributed generation. Contrary to bulk power generation, where

a few large power plants supply large areas, distributed generation (DG) relies on
small units, capable of powering from a single customer to an appropriate portion of
the grid. Such a transition requires newer control systems to operate on such scales.

1. This data is acquired using a weather station owned by UTBM.

1.1. Toward Smarter Power Grids 3

700

600
Solar radiation [W/m²]

500

400

300

200

100

0
0 500 1000 1500
Time [min]

6
Wind speed [m/s]

0
0 500 1000 1500
Time [min]

60

55
Load [GW]

50

45

40

35
0 500 1000 1500
Time [min]

Figure 1.1: Solar radiation (top) and wind speed (middle) measured in Belfort on June
6, 2011, at a height of 15 m, with a time resolution of 1 min, and total French demand
observed by RTE (bottom), with a resolution of 30 min.
4 Chapter 1. Energy Management in Smart Power Systems

– Energy costs. As most fossil fuel resources are expected to be depleted in the coming
decades or centuries, at best, energy prices prices can be expected to increase, es-
pecially because of energy demand growth in developing countries [8]. For example,
India’s electricity consumption is expected to grow by 500 % by 2050, with only a
share of 10 % for RESs [3].
– Security of energy supply. As a consequence to rising energy costs, especially with
respect to oil and gas, some countries may be at risk of supply shortage in case of
a major socio-economic or political conflict. This situation also leads to negative
balances of trade which penalize the economies of these countries. RESs and energy
efficiency may provide some locally available solutions to reduce the dependency on
foreign imports.
– Aging infrastructure. Most of the infrastructures built in the second half of the 20th
century are reaching their end of life, and due to a general trend of under-investment,
their failure rates tend to increase over the years [9].
– Information technology (IT) systems security concerns. Recent threats such as
Stuxnet [10] and Flame [11] have shown that most power systems are not as se-
cure as expected and can be targeted by malicious entities, whether they are part of
common malicious or benign hacking activities or of larger cyber-warfare plans [12].

The combination of these concerns indicates that the stability and efficiency of power
grids will be more and more at risk in the coming years, and may be compromised. The
smart grid is expected to help tackle these issues.

1.1.2 Definition
Although the term smart grid tends to be used rather as a marketing term than in
a purely technical context, its concept refers to the modernization of the electric grid 2 .
By adding an extensive communication and control infrastructure to the electric infras-
tructure, a smart power grid becomes a cyber-physical system (CPS). As defined by the
US National Science Foundation (NSF), “cyber-physical systems are engineered systems
that are built from and depend upon the synergy of computational and physical compo-
nents” [13]. For the smart grid, these components are the electrical and communication
and control infrastructures.
In 2007, the Energy Independence and Security Act [14] of the 110th US congress
defined the smart grid as follows:

It is the policy of the United States to support the modernization of the Nation’s
electricity transmission and distribution system to maintain a reliable and secure
electricity infrastructure that can meet future demand growth and to achieve each of
the following, which together characterize a Smart Grid:

2. The same term, with a similar meaning, is sometimes used for other types of networks, such as smart
water and smart gas networks.
1.1. Toward Smarter Power Grids 5

– Increased use of digital information and controls technology to improve reliabil-

ity, security, and efficiency of the electric grid.
– Dynamic optimization of grid operations and resources, with full cyber-security.
– Deployment and integration of distributed resources and generation, including
renewable resources.
– Development and incorporation of demand response, demand-side resources, and
energy-efficiency resources.
– Deployment of “smart” technologies (real-time, automated, interactive technolo-
gies that optimize the physical operation of appliances and consumer devices)
for metering, communications concerning grid operations and status, and distri-
bution automation.
– Integration of “smart” appliances and consumer devices.
– Deployment and integration of advanced electricity storage and peak-shaving
technologies, including plug-in electric and hybrid electric vehicles, and thermal-
storage air conditioning.
– Provision to consumers of timely information and control options.
– Development of standards for communication and interoperability of appliances
and equipment connected to the electric grid, including the infrastructure serving
the grid.
– Identification and lowering of unreasonable or unnecessary barriers to adoption
of smart grid technologies, practices, and services

Source: Energy Independence and Security Act of 2007 [14].

This very detailed definition focuses on the technologies and characteristics of the
smart grid from a US perspective. The European Technology Platform (ETP) for the
Electricity Networks of the Future provides another definition that focuses on the objec-
tives of the smart grid, and that shows that the smart grid impacts the entire energy
industry, from the largest nuclear power plants to the smallest consumers [15]:

A smart grid is an electricity network that can intelligently integrate the actions of
all users connected to it — generators, consumers and those that do both — in order
to efficiently deliver sustainable, economic and secure electricity supplies. A smart
grid employs innovative products and services together with intelligent monitoring,
control, communication, and self-healing technologies in order to:
– Better facilitate the connection and operation of generators of all sizes and
technologies,
6 Chapter 1. Energy Management in Smart Power Systems

– Allow consumers to play a part in optimizing the operation of the system,

– Provide consumers with more information and better options choosing their
energy supplier,
– Significantly reduce the environmental impact of the whole electricity supply
system,
– Maintain and improve the existing high levels of system reliability, quality, and
security of supply,
– Maintain and improve the existing services efficiently,
– Foster the development of an integrated European market.

Source: European Technology Platform for the Electricity Networks of the Future [15].

1.1.3 Characteristics
The complete characteristics of a smart grid heavily depend on the point-of-view of
the player: consumer, distribution and transmission system operator, generator, regulator,
equipment supplier, etc. A general consensus is that the smart grid relies on the addition
of a communication and control network to an updated electric grid.
Several levels of smart grids can be identified, and compared with legacy power grids
that were used up to a few years or decades ago, and with the current state of power grids.
Legacy power grids were only very partially automated, mostly at the transmission level,
and relied on large power plants to power customers, as pictured in Fig. 1.2.
In current power systems (Fig. 1.3), the development of information technologies and
the restructuring of the electricity sector has led to several changes, such as the separa-
tion between generation, distribution and transmission entities via unbundling, the intro-
duction of control systems based on advanced supervisory control and data acquisition
(SCADA) systems, and the integration of relatively low shares of RESs. The operation of
transmission infrastructures is also equipped with efficient communication and automatic
control capacities, while distribution infrastructures are not.
For the development of future smart grids, several steps or versions can be distin-
guished, as described by Carvallo and Cooper [16], each with an increasing degree of
complexity and automation. The first generation smart grid, called Smart grid v1, re-
quires the implementation of an advanced metering infrastructure (AMI) using smart
meters, that in turn enable basic demand response (DR) [17], of distribution automation
and of advanced energy management systems (EMSs) and distribution management sys-
tems (DMSs). Several of these technological changes are already ongoing in most parts of
the world, especially for smart meters deployment.
The advanced smart grid, or Smart grid v2, builds on the success of Smart grid v1
and includes the integration of new technologies such as EVs, large shares of DG and
RESs, and energy storage, as shown in Fig. 1.4. The communication infrastructure spans
the entire electrical infrastructure, so that each asset such as a DG source, storage unit,
1.1. Toward Smarter Power Grids 7

System operator

Power plant

Industrial
customer

Commercial
customer

Transmission Substation Distribution Substation

Residential
customer

Figure 1.2: Architecture of legacy power systems, based on [3]. Plain lines indicate elec-
trical infrastructures and dashed line communication channels.

Transmission control Distribution control

center center

Power plant

Industrial
customer

Transmission Commercial
customer

Substation Distribution
Residential
customer

Wind farm Distributed generation

Figure 1.3: Architecture of current power systems, based on [3]. Plain lines indicate
electrical infrastructures and dashed line communication channels.
8 Chapter 1. Energy Management in Smart Power Systems

customer and operator, is capable of communicating with others if required. To accommo-

date these changes, control systems are more dispersed and automated, with self-healing
capabilities [18]. Energy service providers, such as aggregators, provide services to cus-
tomers that enable them to participate in programs, such as advanced DR with the ability
to automatically control smart appliances. A Web of Things is expected to emerge from
the popularization of smart appliances and communicating electronic devices [19]. Such
changes can be expected to occur within the next couple of decades in the most advanced
Western countries. It is within this scope that this dissertation provides some original
fundamental contributions of the application of multi-agent systems, a CPS enabler, to
the energy management of smart grid elements.
Smart grid v3 is currently considered as the ultimate step of the proposed vision and
includes peer-to-peer energy trading and energy roaming, sometimes called enernet [20].
This term refers to Internet, in that as information is easily shared over the Internet,
energy could tomorrow become as easy to trade among individuals. The IT and electricity
infrastructures would then be fully integrated. The enernet concept could become a reality
by 2030 to 2050, but would require tremendous changes in the way electric energy is
managed.
Similar visions and frameworks were proposed by entities such as the French Energy
Agency (ADEME), which lists several visions for 2020 and 2050 [21], and the US National
Institute of Standards and Technology (NIST) [22].

1.1.4 Technologies
The realization of the smart grid relies on the development and maturation of numerous
technologies. According to [3], these technologies include:
– Wide-area monitoring and control technologies, such as phasor measurement units
(PMUs), are required to monitor the performance of power system components and
control their operation over large geographic areas. Wide-area monitoring systems
(WAMSs) and other technologies are required to avoid blackouts and facilitate the
integration of large RESs, such as offshore wind farms, by generating data useful to
control systems and grid operators [23].
– Information and communications technologies are required to enable the communi-
cation and IT infrastructures to adapt to the features of the smart grid. Whether
they use private utility networks or public ones, such as the Internet, data needs to
be transmitted in a reliable and efficient manner throughout the entire system. Such
transmission has to occur in real-time for some applications, such as monitoring and
control, or can be deferred in time for others, such as for billing and event logging.
Similarly, computing capabilities, both on the hardware and software sides, need to
be upgraded, especially as a lot more data is going to transit along communication
infrastructures and need to be processed.
– Renewable and distributed generation technologies, from the residential level to the
transmission level, are difficult to integrate due their intermittent output, which
does not match demand patterns. Energy storage technologies are thus required
to help mitigate these problems by providing energy buffers. For example, storage
can discharge and provide peaking power in the evening, when demand is high and
 1.1. Toward Smarter Power Grids
Transmission control Distribution control Energy service
center center provider

Electric vehicles

Power plant

Grid-scale
storage Substation Distribution

Wind and Distributed generation

solar farms and storage

Figure 1.4: Architecture of future power systems, based on [3]. Plain lines indicate electrical infrastructures and dashed line communi-
cation channels.

9
10 Chapter 1. Energy Management in Smart Power Systems

solar power decreases, and recharge at night when the wind might blow faster. DR
technologies are another response to this challenge. Tools such as EMSs, DMSs, and
geographical information systems (GISs) facilitate such integration.
– Transmission enhancement technologies, such as flexible AC transmission systems
(FACTSs), high voltage DC (HVDC), dynamic line rating (DLR) and high tem-
perature superconductors (HTSs), are useful tools for enabling a more flexible and
efficient transmission of electric power. FACTSs improve the ability to control and
transfer power flows [24], while HVDC lines enable the transmission of large amounts
of power (e.g., originating from large RESs farms) at the continental scale with low
losses [25]. DLR provides information on the carrying capability of network sections
in real-time through sensors, and enable optimizing the use of transmission assets.
HTSs use the properties of superconducting materials to reduce transmission losses,
and hereby the economic efficiency of the system.
– Distribution grid management technologies are used to improve the reliability of
distribution systems and reduce outage and repair time through advanced sensing
and automation. Distribution automation (DA) aims at processing real-time data
for applications such as fault location, automatic feeder reconfiguration, reactive
power and voltage control and DG control. The condition of distribution assets can
also be monitored through such means, for example to monitor the performance of
an aging transformer.
– The advanced metering infrastructure enables two-way information flow through
smart meters, notably between customers and utilities for consumption and pricing
data. The corresponding technologies are typically the ones that are currently being
deployed by utilities with smart meters. Such services include time-of-use pricing,
consumption profiling and diagnosis, theft detection, remote connection and discon-
nection, etc. Meter data management systems (MDMSs) are required to handle the
large amounts of data resulting from the adoption of such new devices and services.
– The electric vehicle charging infrastructure is essential in enabling the development
of plug-in hybrid electric and battery electric vehicles. It handles the billing, schedul-
ing and smart charging of vehicles, while taking into account the demand and market
context in real-time. More advanced services can also be developed, such as the par-
ticipation of vehicles in ancillary services using the vehicle-to-grid (V2G) technology.
– Customer-side systems are supposed to help users manage energy consumption, for
example by enabling them to monitor their consumption, to automate their loads
such as smart appliances, and to participate in electricity markets through the in-
termediary of aggregators. Such features can be achieved via in-home energy dash-
boards and dedicated online and smartphone applications, capable of communicating
with smart devices and controlling them adequately with an home energy manage-
ment system (HEMS).
These technologies will be used at different complementary levels given below, each
corresponding to a different power system scale.
Buildings, whether they are for residential, commercial, office or industrial use, are
expected to become smarter in their consumption of energy. Through advanced home
1.1. Toward Smarter Power Grids 11

automation, smart metering and HEMSs, the use of most loads (smart appliances, air
conditioning, water heaters, etc.) can be improved and lead to energy and electricity
bill savings, especially if systems have a learning capability. They can also integrate
DG sources such as photovoltaic (PV) panels, EV charging stations, possibly with V2G
capability, and other energy storage assets.
Microgrids are defined by the US Department of Energy (DOE) as “a group of inter-
connected loads and distributed energy resources within clearly defined electrical bound-
aries that acts as a single controllable entity with respect to the grid [and can] connect
and disconnect from the grid to enable it to operate in both grid-connected or island-
mode” [26]. The islanding capability of microgrids enables them to operate separately
from the main grid when necessary, e.g., during a blackout. Although this feature could
improve the quality of service to customers, its realization requires important advances in
control technologies, especially if large amounts of RESs and storage are to be integrated.
As large centralized power plants may become less common in the future, large networks
of interconnected microgrids could partially replace today’s large transmission systems.
The largest scale corresponds to Super grids, which are large continental or inter-
continental-scale power grids. The pan-European grid could for example be connected to
North African power grids to form a large electricity grid, where large amounts of PV and
wind energy could transit through HVDC lines, as in the Desertec project [27].
Another element of smart grids is smart power generation, or smart power plants. This
concept aims at matching electricity production with demand using multiple identical
generators which can start, stop and operate efficiently at a given load, independently
from each other, making them suitable for base load and peaking power generation [28].
Smart power generation promises to be one of the tools to help integrate intermittent
RESs, due to their ability to adapt rapidly to changing conditions.

1.1.5 Challenges
For this vision to become a reality, several challenges have to be solved. A first chal-
lenge is related to the funding and financing of smart grid projects. Most early smart
grid projects were supported by government stimulus funds initiated after several recent
blackouts, such as the 2003 blackouts in Italy [29] and in Northern America [30]. In
France, the ADEME and the National Research Agency (ANR) launched several demon-
stration projects from 2009 to 2012, as did the US DOE and other institutions [31]. In
France, frequent changes in feed-in tariffs for RESs have slowed down the adoption of these
technologies, due to the uncertainty it triggers for investors. A similar behavior could be
observed for smart grids projects in the future, unless private investors can clearly quantify
the benefits they could gain from the smart grid.
On the other hand, regulators have an essential role to play in defining policies for
the smart grid, while taking into account the constraints of all market players. Current
legislation in most countries will have to change to enable some smart grid functionalities,
especially regarding power markets and rate structures, e.g., for demand response. As
business models heavily depend on such policies, such as for energy service providers,
clear policy decisions are expected from regulating entities.
Due to the multiplicity of manufacturers and operators, standards have to be created to
enable interoperability between hardware and software from different providers, and reduce
12 Chapter 1. Energy Management in Smart Power Systems

development costs. NIST recently published a roadmap for smart grid interoperability
standards [22]. The Institute of Electrical and Electronics Engineers (IEEE) has also
been working on numerous smart grid-related standards; some of them are named in
NIST’s roadmap.
Consumer acceptance is another concern. As consumers (or prosumers) are expected
to play an increasingly important role in the smart grid, utilities have to make sure that
their customers actually feel implicated in the transition to the smart grid. Consumer
backlash in several US states has shown that rational (privacy concerns) and irrational
(radiations leading to cancers) fears can arise, simply from smart meter installation [32].
The use of intelligent HEMSs is also questionable for people who might not be comfortable
with technology. Large scale experiments must be conducted to analyze the impact of such
behaviors on operations and on the validity of business models.
And lastly, technical challenges are some of the biggest to be tackled. The largest
concerns are about the integration of RESs, storage and EVs, which introduce a lot of
stochasticity in grid operation, as opposed to today’s grid where only the load has to be
forecast. But many other challenges can be derived from these larger ones. In this work,
two technical challenges are of interest. The first one is a consequence of the development
of the metering infrastructure, which will lead to gigantic amounts of data to process in
almost real-time by utilities. The second corresponds to energy management algorithms
capable of leveraging the possibilities offered by the smart grid, namely, here, smart power
plants and demand response.

1.2 Modernization of Control Systems

1.2.1 Power System Restructuring
In addition to the change drivers listed earlier, the restructuring of power systems that
has been going on for a couple of decades in Europe and in other parts of the world is also
affecting the way control systems operate [33].
In the paradigm of vertically integrated utilities (Table 1.1), the utility owns at the
same time generation, transmission and distribution assets, and also acts as an electricity
retailer for the end-users. Such architecture enables an efficient management of costs and
benefits, as well as appropriate investment decisions for the benefit of the grid as a whole.
On the other hand, it leaves almost no space for competitors due the monopoly the utility
has. In most US states, utilities operate at the city or state scale under a structure close
to the vertically integrated utility, where generation can also be liberalized.
In unbundled electricity markets (Table 1.1), the only monopolies are for the trans-
mission and distribution infrastructures, while several generators and retailers are free to
compete on markets to sell energy to their respective customers. While this paradigm
enables multiple competitors enter the market and theoretically puts more pressure on
prices, it sometimes leads to slower technology deployments, increased complexity and
lack of investment strategies that can benefit to all activities. From the late 1990s and
until 2007, France has progressively liberalized its electricity market by switching from an
(almost) single vertically integrated utility (Électricité de France, EDF) to an unbundled
electricity market, where multiple entities (EDF, GDF-SUEZ, POWEO, etc.) compete at
1.2. Modernization of Control Systems 13

the wholesale and retail levels. RTE and ERDF have kept the monopoly for the trans-
mission and distribution (T&D) infrastructures [34].

Activities Integrated utility Unbundled market

Generation Regulated Market
Transmission Regulated
Distribution Regulated
Retail Regulated Market
Table 1.1: Comparison of vertically integrated utilities and unbundled electricity markets.

1.2.2 Control Systems Typology

Multiple control systems are at the core of electric grids, and allow these grids to
operate efficiently and safely. Four complementary objectives have to be fulfilled by control
systems:
1. Safety: The overall safety of the system and of the surrounding populations has
to be ensured, for example to avoid catastrophes such as the ones of Chernobyl or
Fukushima Dai-ichi [35], or large blackouts [29].
2. Protection: A control system must protect equipment from self-inflicted and external
damages, for example by isolating equipment from the rest of the system.
3. Stability and reliability: The system must operate correctly as long as possible while
providing its customers with electric power of satisfactory quality. Stability often
relies on maintaining electricity quantities as close as possible to their reference val-
ues, e.g., regulating frequency, voltage and reactive power, while reliability involves
the uninterrupted supply of acceptable quality of electricity supply. It also requires
operating equipment in a way that does not degrade it too rapidly, e.g., managing
storage within its proper state-of-charge limits, or not overloading transformers for
long durations.
4. Economics: All other system priorities must be fulfilled while minimizing costs for
all market players. Traditional energy management control primarily focuses on
economics, but a broader paradigm should be considered in taking into account all
technical specifications, economics and environmental concerns as well.
The work presented in this dissertation mostly focuses on the last two objectives, i.e.,
stability and economics. In order to meet these requirements, several complementary, and
sometimes overlapping, control systems are employed over several time scales, and rely on
different mechanisms depending on the market structure (Fig. 1.5).
– Capacity and operations planning aims at taking techno-economical decisions regard-
ing investments for increasing and expanding generation or transmission capacity,
as well as planning the operation of these assets over long periods, e.g., for long
maintenance duties. In vertically integrated structures, investment and planning
departments are in charge of such operations, while in liberalized market environ-
ments, capacity and forward energy contracts are used to reach agreements between
sellers and buyers.
14 Chapter 1. Energy Management in Smart Power Systems

Capacity and operations System

Operations Operations scheduling
planning balancing

1–10 years 6–12 months Day ahead

Delivery
2h–15min
< 15min
Timescales Economic
Economic dispatch
System Operation
scheduling
planning planning

Organized electricity markets

Ancillary
Capacity and forward energy Day ahead
Spot market services
contracts market
market

Mechanisms Vertically integrated utilities

Infrastructure planning and Generator Generator Operating

construction scheduling commitment reserve

Figure 1.5: Timescales and decision mechanisms for electric system operation, based
on [36].

– Operations scheduling includes two complementary processes. Day-ahead scheduling

takes place up to 24 hours before delivery, and typically aims at deciding which gen-
erators should be used for each hour of the day, using demand forecasts. Economic
dispatch uses the results of the scheduling process to derive the optimal output of
each turned-on generator to meet demand. Vertically integrated utilities and genera-
tion companies (GENCOs) use unit commitment and economic dispatch algorithms
to run such processes, while liberalized markets rely on day-ahead and real-time
markets to reach settlements.

– System balancing occurs just before delivery, and relies either on ancillary markets,
or on operational reserve. Operating reserve is used by integrated utilities for this,
while ancillary services markets are used in market environments.

Modern control systems used by utilities are based on SCADA systems. SCADAs are
information systems used for monitoring and supervising power systems or industrial pro-
cesses, but without control functionalities. However, with the recent popularization of the
term, SCADAs are now associated with extensive systems, some-times performing control
actions. With the development of communication interfaces, power systems restructuring
and distributed generation, distributed control systems (DCSs) have emerged. They have
many similarities with SCADAs, but instead of relying on a centralized intelligence, their
intelligence is rather distributed in several subsystems called intelligent electronic devices
(IEDs) [37]. They have extended capabilities such as the ability to autonomously issue
control commands, e.g., commanding a circuit breaker by detecting an abnormal voltage,
1.3. Problem Statement 15

current or frequency.
Although SCADAs only provide communication interfaces between a control system
and physical end-points, further control capabilities can be integrated: DMSs for substa-
tion automation, condition monitoring, fault location, voltage control, load flow calcula-
tions, etc.; outage management systems (OMSs) for assisting operators in system restora-
tion after a failure; GISs with geographically-referenced data, for visualizing events in a
distribution grid; additional systems for network analysis, demand response management,
business management, billing, customer information, etc.
EMSs, on which this dissertation focuses, are another category of control systems.
They include computer-aided tools that enable controlling and scheduling the operation
of a power system under given constraints (e.g., at a minimal cost), without affecting the
capability to meet their technical requirements and quality standards. EMSs are used for
multiple applications, from smart buildings to generation commitment and dispatching.

1.3 Problem Statement

The smart grid is enabling numerous new possibilities, and, at the same time, new
difficulties to tackle, such as a much higher level of complexity. Overcoming these problems
is crucial to exploit as well as possible these promises.
Energy management is considered as a major research topic for smart grids, mainly
due to the high stochasticity introduced by new generation, storage and load assets (DG,
distributed storage, DR, EVs, etc.): how can these assets be utilized as optimally as
possible and at the same time to meet their design objectives, while maintaining grid
stability, ensuring economic efficiency and customer satisfaction? The need to create
efficient and flexible EMSs arises, as well as the need to take into account the system
architecture to quantify their impact on the communication infrastructure. EMSs have
to profoundly evolve, become smarter, and make of the smart grid an evolving intelligent
system. Two requirements can be distinguished for this to happen:
– EMS architectures need to adapt to the increasing complexity of electric grids. More
decentralized architectures are favored to integrate large numbers of market players
and pro-active consumers, and solutions to evaluate the need for communication
infrastructures are required to avoid a data “deluge”. As these two objectives are in
most cases contradictory, especially if optimal or near optimal results are expected,
a compromise has to be found.
– EMSs also have to gain intelligence to become more efficient and enable new features,
such as smart power generation and advanced demand response.
This dissertation focuses on the design and development of intelligent energy manage-
ment systems, and especially on their architecture and the algorithms employed, while
trying to meet these requirements. The underlying questions are:
– How should energy management systems in a smart grid environment be designed?
– How should they be structured to operate as efficiently as possible?
– How can algorithms leverage these architectures, improve the operation of power
systems, and enable new features?
16 Chapter 1. Energy Management in Smart Power Systems

– How do the chosen architecture and algorithms impact the need for computation
and communication?

These concerns are studied through two applications, which, while providing test cases
for answering the previous issues, also aim at providing answers to the following questions:

– How can energy management systems for gas turbine power plants be improved with
respect to costs and greenhouse gases emissions?

– How can a residential demand response system be designed in order to reliably reduce
demand during peaks?

1.4 Outline
This dissertation is divided into four chapters. The current chapter has presented the
context of this research work, and especially the ongoing transition toward the smart grid
which is emerging as the main paradigm for the modernization of the electric grid. The
promises of smart grids have been listed, such as the ability to integrate large shares of dis-
tributed and intermittent renewable energy sources, energy storage and electric vehicles,
as well as the promise to give consumers more control on their energy consumption. Mul-
tiple technologies, and especially information and communication technologies supported
by intelligent algorithms, have been listed, and are expected to enable this transition to
smarter power grids. Then, by identifying architectures and algorithms for energy man-
agement systems as one of the barriers for smart grid implementation, this chapter has
given an overview of control systems and has shown that their modernization required the
ability to tackle the complexity of smart grids. The main problems this dissertation tries
to solve have been listed, both at the fundamental and at the application level.
The applications presented in chapters 3 and 4 use two specific simulators that were
designed and developed for the purpose of investigating the performance of the proposed
EMSs. Both these simulators rely on several artificial intelligence and simulation tools
that are presented in chapter 2. This chapter identifies the multi-agent systems (MAS)
concept as a facilitator for solving the problems at hand, and as a tool for designing
advanced EMSs for smart power systems. In the proposed approach, MASs enable a new
systemic and multi-disciplinary approach to energy management, and allow for agent-
based architectures and algorithms to be defined. This new approach helps tackle the
complexity smart grids by enabling the simultaneous consideration of multiple aspects
such as power systems, communication infrastructures, energy markets, and consumer
behavior. The chapter presents the multi-agent tools used to design, model, control and
simulate the EMSs. The multi-agent concept is explained, as well as its relevance for smart
grids, considered as a complex adaptive systems. Several aspects of MAS design, based
on standards and specifications, are also described. Then the multi-agent development
framework used to develop the simulators for both applications is described. The co-
simulation framework that builds on this middleware to enable evaluating the impact of
EMS decisions on steady-state grid stability is also presented. Parts of the explanations
related to this framework, which was developed in collaboration with Colorado State
University, are based on a conference publication [38].
1.4. Outline 17

Chapter 3 presents a first application in which an EMS for gas turbine power plants
is proposed. The EMS is designed with the objective to try to minimize operational costs
and emissions in the smart power generation paradigm. A gas turbine model based on
actual data from GE’s 9E turbines is proposed, and enables estimating the power plants
operational costs and NOx and CO2 emission levels. This model is then used to run
simulations with a MAS-based simulator specifically developed for this problem. The ar-
chitecture of the simulator is presented, as well as the roles and main specifications of the
constituting agents, and their interactions. Several metaheuristics are then compared on
mathematical benchmark problems. From the results of this comparison, a metaheuristic
is selected and used to achieve dynamic dispatch among gas turbines, according to their
individual characteristics. An algorithm is also designed to start and stop turbines ade-
quately according to load forecasts. Simulation results show that the system is capable
of operating the system properly while reducing costs and emissions, as expected. The
computing and communication requirements of the system, resulting from the selected ar-
chitecture, are also evaluated. Parts of this chapter are based on a publication in Applied
Energy, an Elsevier journal [39]. This work was done in collaboration with GE and with
the Université de Haute Alsace.
With other demand-side management techniques, DR enables reducing load during a
given duration, for example in case of a congestion on the transmission system. In chap-
ter 4, a second application is presented and proposes a DR system targeted at residential
customers. The system relies on the use of the assets of residential customers to curtail
and shift local loads (PHEVs, air conditioning, and water heaters), so that the total system
load remains under a given threshold. The system is integrated with the distribution and
transmission systems, as well as a DR market and aggregators serving as energy service
providers. As customers cannot directly participate in markets, aggregators act as inter-
faces between grid operators and the DR market. A MAS-based simulator is developed
to evaluate the performance of the proposed system. The architecture of this simulator is
described by presenting the roles and interactions of the agents constituting it. The algo-
rithms used to implement the load reductions are also presented. Results show that the
system manages to maintain the total load under a threshold by adequately controlling
available resources, without compromising the steady-state stability of the distribution
system. The impact for individual customers is also evaluated. This work was done in
collaboration with Colorado State University.
A general conclusion ends this dissertation, and summarizes this research work. The
advantages and drawbacks of the proposed approach are analyzed using the results ob-
tained from the two applications, and some lessons learned on the use of MASs for energy
management in smart grids are listed. Some perspectives on future work directions are
also given, especially about the further developments of the DR application.
 2

Agent-Based Modeling, Control and Simulation

Contents
2.1 Multi-Agent Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.1.1 Smart Grids as Complex Adaptive Systems . . . . . . . . . . . 22
2.1.2 Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.1.3 Relevance for Power Systems Applications . . . . . . . . . . . . 24
2.1.4 Communication, Languages and Ontologies . . . . . . . . . . . 29
2.1.5 Agent Management . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.1.6 Topologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.1.7 Inter-Agent Interaction . . . . . . . . . . . . . . . . . . . . . . . 35
2.2 Multi-Agent Development Framework . . . . . . . . . . . . . . . . . . . 38
2.2.1 Development Platforms . . . . . . . . . . . . . . . . . . . . . . 38
2.2.2 The JADE Platform . . . . . . . . . . . . . . . . . . . . . . . . 38
2.3 Co-Simulation Framework . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.3.1 Need for Co-Simulation Tools . . . . . . . . . . . . . . . . . . . 41
2.3.2 Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.3.3 Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.3.4 Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

19
20 Chapter 2. Agent-Based Modeling, Control and Simulation

2.1 Multi-Agent Systems

2.1.1 Smart Grids as Complex Adaptive Systems
The electric power grid is sometimes considered as “one of the largest and most complex
of man-made objects” [40]. Similarly, smart grids can be considered as complex adaptive
systems (CAS) [41,42], i.e., complex collections of interacting entities called agents. CASs,
as well as smart grids, are complex by their size and heterogeneity, and adaptive by their
evolving nature [43]. According to Cilliers [44], complex systems have the following main
characteristics:

– They consist of a large number of elements. Numerous types of loads, generators and
other types of components and entities impact the behavior of the grid, have their
own objectives and have specific interaction patterns. Therefore, it is impossible to
model the behavior of the grid through conventional descriptions such as differential-
algebraic equation systems alone.

– Elements have to interact with each other. In order to maintain a balance between
supply and demand, elements have to interact directly (by communicating) or indi-
rectly (by consuming or converting energy). However, not all elements are required
for the whole system to operate properly: for example, disconnecting a load may
temporarily make the grid unstable, but the system should be able to keep operating.

– Interactions are non-linear, i.e., small causes can have large results. And although
interactions have primarily a short range, their influence may extend much further.
For example, large blackouts often rely on a simple failure, such as a power line down
due to tree falling. The 2003 North American blackout is an example [30].
– Interactions can have positive or negative feedback associated. Most control systems
include such types of feedback loops, from simple proportional-integral-derivative
(PID) controllers to most elaborate systems, e.g., to maintain frequency and voltage.
– They are open systems, i.e., they interact with the environment. The size and struc-
ture of the power grid can evolve, and its constituents impact other infrastructures
such as the water and transportation infrastructures. It is also hard to draw borders.

– They need to be maintained in a state of equilibrium that is not “natural”: advanced

control systems are required to maintain the grid in operation, e.g., to maintain an
equilibrium between supply an demand. If these control systems were disabled, an
immense blackout would almost immediately occur.

– They have a history. Power grids evolve in time, not only in the way their various
constituents are used, but also in their architecture, for example when new generation
or transmission capacity is added. Their past is at least partially responsible for their
present state.

– Each element is ignorant of the behavior of the whole system. Except grid operators
that have an aggregated view of the macroscopic behavior of the grid, simple com-
ponents only have a local view of the grid, with access to little information on other
elements.
2.1. Multi-Agent Systems 21

These characteristics tend to confirm that smart grids are complex systems and also
CASs due to their advanced control systems, which justifies the choice of modeling them
using a concept called multi-agent systems. Although the first uses of this concept for the
same type of applications date back to 1997 [45], this idea has not been fully exploited by
researchers thus far.

2.1.2 Concept
Multi-agent systems (MASs) support a framework for the modeling and control of
multiple structures that can be decomposed into several interacting entities. Formal defi-
nitions for MASs have been proposed by Wooldridge and Weiss [46] and Ferber [47]. The
following definition provides a simple overview of the MAS concept: a MAS is a system
composed of a collection of interacting entities called agents, evolving in an environment
where they can autonomously perceive and act to satisfy their needs and objectives. As
shown in Fig. 2.1, agents receive data from their environment, called percepts, take deci-
sions on the basis of current and possibly past percepts, and effect actions through the
actuators they may be equipped with. The environment of an agent can be defined as the
external entities and resources the MAS can interact with.

Environment

Agent 1 Agent 2

Perception Perception

Reasoning Reasoning

Action Action

Figure 2.1: Diagram of a generic agent, from [48]. An agent perceives its environment
through sensors and acts on it with its actuators.

Agents can exhibit different behaviors and properties which give them a certain degree
of autonomy with various degrees of intelligence. Three main types of agents can be
distinguished from these characteristics [49]:

– Reactive agents only show some simple reactions to any excitation (stimuli), and
their representation of the environment is minimal. But by interacting with each
other, such agents can together lead to emerging behaviors, difficult to achieve if the
system had been modeled as a single agent [50]. They are useful when fast response
times are needed.

– Cognitive or intelligent agents have extended intellectual capabilities and can use
their resources and skills to reach their local goals. Such agents are useful to carry
out tasks that require complex decision-making.
22 Chapter 2. Agent-Based Modeling, Control and Simulation

– Learning agents can gain knowledge by analyzing the results of their actions, and
usually have a much better knowledge of the environment, which is required to take
complex decisions.
Based on such definition, many existing systems from various domains can be classified
as MASs. Examples include a society or a group, in which agents are people, and can
communicate with each other, cooperate, compete, and so on; a network of computers,
where agents are software algorithms, interacting by exchanging messages and data, for
example to solve a problem faster than a single computer; robots of a production line,
where they need to cooperate and coordinate themselves to perform a given task.
Finally, a power system, where the constituting elements of the grid (generators, loads,
distribution infrastructure, operators, etc.) are agents and interact with each other to serve
consumers while respecting given constraints can be considered as a MAS.

2.1.3 Relevance for Power Systems Applications

A shown earlier, smart power grids may be considered as complex adaptive systems,
and agent-based solutions could be a practical way to achieve efficient and reliable mod-
eling, control and simulation of such large and heterogeneous systems. Advantages of
MASs for tackling these conditions include the following properties: they are inherently
distributed, they are pro-active and they have social abilities.

[Link] Distributed Architecture

Much like power grids, MASs may be geographically dispersed, interconnected and
heterogeneous. They are distributed by design with three main attributes: (i) local knowl-
edge, (ii) flexible interactions, and (iii) bottom-up approach.
Local knowledge means that the agents view of the environment is local, and as a
consequence, their knowledge is limited to only what they can or need to know. The
perception of agents can be limited to immediate neighbors, enabling reduction in data
communications. For example, in a large power grid, an agent controlling a distributed
generator does not need to receive information about a small load, which can be several
kilometers away. Therefore, a distributed MAS architecture contributes to a scalable
distribution grid, and also enables evaluating distributed control schemes. These schemes
include market-based systems, where agents are market participants, and have a limited
knowledge of the market they can use to determine bidding or buying strategies.
A flexibly designed MAS incorporates plug-and-play, robust and fault-tolerant proce-
dures when required by changes in the environment. Control systems should be robust
enough to operate sub-optimally if needed, a property enabled by the distributed na-
ture of MASs. For example, if a generator or load agent is disconnected, i.e., turned-on
or off (regardless of whether it has been scheduled or not) or loses communication, the
MAS, if properly designed, should acknowledge this modification and take it into account
when taking decisions towards reaching its objectives (e.g., maintaining the system sta-
ble) [51]. Compared to most conventional analytical control methods where all possible
events, changes and faulty conditions, usually have to be predicted when designing the
control system so that proper corrective action can be taken, this possibility is a clear
progress. In addition, a flexible MAS can add or remove new agents and functionalities
2.1. Multi-Agent Systems 23

without requiring to completely redesign the system, which could help lower develop-
ment and maintenance costs. This characteristic is similar to how computer systems use
plug-and-play devices, and usually relies on the use of a yellow and white-pages system
described in section 2.1.5. The operation of electric vehicles with the distribution grid for
charging could for example benefit from this property, as EVs may or may not be present
at charging stations depending on the time of the day.
The previous two features, viz., local knowledge and flexible interactions, enable the
third one, i.e., a bottom-up approach. This feature is particularly well-suited for complex
and distributed problems, such as the ones related to smart grids. As a global model of the
grid does not exist, or at least does not reliably represent the real system, the adoption
of a bottom-up approach is needed. This approach relies on modeling components of
the grid separately, with their own roles, knowledge, actuators, etc., and making them
interact with other. Agents can operate autonomously (at least partially), and cooperate
or compete with each other if needed. The complexity of a control system can therefore be
reduced by distributing tasks among interacting agents. If properly modeled, a realistic
behavior, similar to the actual grid behavior, could be obtained. This property could
for example play an important role in a MAS designed for a smart grid with a high
penetration rate of distributed energy sources. The grid could be divided into several
interconnected microgrids (see section 1.1.4) containing local generators, loads and storage
devices. Intermediate layers, consisting of groups of microgrids, could also be added.

[Link] Pro-Activity

Proactive agents have goals which can be local and/or global. A single agent usually
has local goals while a group of agents may have global goals (goals the entire MAS tries to
achieve). For example, maintaining a steady voltage at a specific bus is mainly a local goal
for a power source, but maintaining balance between generation and supply is a global goal
and cannot be reached by a single agent, thus requiring cooperation. Such pro-activity
might be enabled by local intelligence with information based on knowledge about the
environment and with further information by communicating with other agents. Agents
can then take decisions based on such on-line knowledge and their goals, plan actions to
perform, and finally execute them for achieving the required actions.
The scheduling of the use of storage in a distribution system could be an example: if
the system knows (based on forecasting) that demand is going to peak and conventional
generation sources will not be sufficient to match it, a battery agent may take pre-emptive
actions by charging the battery to its maximum state-of-charge before the peak and making
it available for use during the said peak time for supplying the extra load. This strategy
would enable the system to meet demand during such peak. Other examples could be
procedures to start, synchronize and reconnect a turbine to the grid, and the planning of
required reactive power for such connection.

[Link] Social Behavior

Agents need to have a social behavior compatible with other fellow agents, expressed
under various forms. Their social organization can vary from one system to another and
with time, as well as the way they interact with each other and take decisions. Agents can
24 Chapter 2. Agent-Based Modeling, Control and Simulation

notably coordinate themselves and cooperate for reaching goals that may not be reach-
able by a single agent. Agents can influence the actions of others or act as interfaces
through negotiations, requests and contracts, or other protocols. Indirectly, this property
enables testing and comparing several interaction configurations, such as the ones listed
in section 2.1.6.
Continuing the previous example with the demand peak and storage, before the MAS
decides that the battery should absorb the peak, a “discussion” with other agents (such
as power system brokers) may occur about whether the load can be supplied by another
source, and maybe a better solution would arise from economic and/or technical points-
of-view.

[Link] Practical Implications

In addition to these aspects, the use of MASs with smart grids has several practical
advantages, that are mainly related to the ability to take into account interactions between
agents, hereby enabling proper systemic approaches.

– Communication aspects, and indirectly interaction aspects between various compo-

nents and entities, are an important feature of smart grids. MASs enable to specify
such aspects, that are rarely considered in power systems studies. For example,
this property enables specifying what information is transmitted from one agent to
another, under which form, how it is processed, when it is sent, etc.

– Due to these specifications, switching from a MAS-based simulation system to a

working real-scale prototype is easier, as communication and interaction aspects are
already taken into account. Therefore, such systems tend to be closer to implemen-
tation, as they are already distributed, and are easily deployable. Moreover, agents
can also be tested separately, as with unit tests.

– Due to the distributed nature of MASs, distributed and parallel implementations are
also facilitated. As each agent is at least partially autonomous, it can be assigned
to a particular computing system (computer, core, etc.), and exchange data with
others.

– Agent-based modeling forces modelers to respect specific constraints. Procedures

must be encapsulated in the agent, and can only be implemented using the resources
of that agent. Based on a close monitoring of the activity of each agent and on the
requirements of the user (speed, reactivity, etc.), the corresponding hardware (and
software, in some cases) specifications can be derived.

– Finally, the ability to define how and where data is structured, located and exchanged
facilitates taking into account information security aspects, for example by precisely
defining which agent has access to which information.

Due to these reasons, the MAS concept is the primary paradigm choice for this disser-
tation, and is used for designing the EMSs proposed in the following chapters.
2.1. Multi-Agent Systems 25

[Link] Application to Smart Power Systems and Energy Management

The use of MASs in smart grids, and more generally in power systems, is therefore
relevant for the multiple reasons that were just listed. Fig. 2.2 shows an example of
how MASs could be utilized for communication and intelligent decision-making in power
systems. Depending on the application, an agent may be associated to each sensor and
actuator of the grid and its control systems. Agents may then communicate and interact,
either directly by communicating with each other through a communication network, or
indirectly through measurements on the grid.

= 1 agent

Multi-agent-based
Network control system

Communication network

Sensors / actuators

Source Storage Source

Power grid
Load Load

Figure 2.2: Conceptual diagram of an example application of MASs for smart grids.

MASs have been applied to solve a variety of problems in power systems. Appli-
cations include voltage/VAR control [52, 53], restoration [54, 55], monitoring and fault
analysis [56], and energy management. State-of-the-art reviews of applications of MASs
in power systems can be found in McArthur et al. [57, 58]. Energy management solutions
using MASs have been used by several researchers over the past 10 years, including the
following applications:
– A fully decentralized MAS-based EMS for microgrids was proposed by Lagorse et
al. in [59, 60]. Each component of the microgrid (load, storage, sources and grid)
is controlled by an agent through a power electronics converter connected to a DC
bus. The agents coordinate themselves using a token mechanism, so that the DC
bus voltage is maintained to a constant value. For example, a battery may have the
token to begin with, and controls the bus voltage until it is empty. It can then give
the token to another source, storage unit or the grid, so that it takes over the bus
voltage regulation.
– Dimeas and Hatziargyriou used a partially decentralized approach to microgrid con-
trol, and proposed two energy management methodologies. In the first one, loads,
storage units and sources are modeled as agents and are grouped to form microgrids,
each with a central controller [61]. Several microgrids are interconnected and co-
ordinated by a distribution network operator. Agents then engage in market-based
26 Chapter 2. Agent-Based Modeling, Control and Simulation

negotiations to achieve dispatching [62, 63]. In the second methodology, reinforce-

ment learning is substituted to market negotiations for dispatching [64].
– Logenthiran et al. use a similar approach for multi-agent coordination of genera-
tion and supply in microgrids [65–67], for short-term generation scheduling [68], for
managing PHEVs [69], and for demand-side management [70].
– The IDAPS (Intelligent Distributed Autonomous Power Systems) concept intro-
duced by Rahman, Pipattanasomporn et al. in [71] and used in [72] also relies on a
comparable multi-agent architecture for microgrid operation.
– The PowerMatcher is another MAS and market-based approach that is tested in
field test experiments in the Netherlands. The corresponding test bed is called
PowerMatching City, and is part of several large European projects [73, 74]. The
project aims at balancing supply and demand in an economically efficient way, using
agents with the ability to competitively trade energy on a common market. Test
results showed that PowerMatcher is able to perform the technical and commercial
coordination of end-users power consumption and generation, even on a large system,
and to optimize the operation of virtual power plants (VPPs) [75–77].
– The energy market simulator MASCEM also uses a multi-agent architecture and
game theory, machine learning and optimization techniques to model power markets
and VPPs [78, 79].
– Finally, the GridAgents platform from Infotility is a commercial MAS for intelli-
gent load control, that provides a centralized decision-support solution for demand
response using a variety of agents [80].

[Link] Proposed Approach and Design Methodology

These examples show that MASs are used for multiple applications, using a wide
variety of approaches. Some of these approaches rely on distributed decision-making
mechanisms with autonomous agents, while others do not and focus on other aspects.
The approach used in this dissertation belongs to the latter category, and therefore does
not focus on distributed decision-making with autonomous agents, but rather on the other
advantages of the use of MASs.
Although this approach may seem surprising to readers used to traditional applica-
tions of MASs, this work does only consider the properties of MASs that are relevant
to the selected context. The proposed approach aims at using MASs for solving emerg-
ing engineering problems in a critical infrastructure, the power grid. MASs are chosen
as the main paradigm for EMS design in smart power systems, and enable specifying
architectures and interactions between subsystems. This in turn enables systemic and
multi-disciplinary studies.
The proposed EMSs are designed according to the following methodology:
1. Identification of the objectives and constraints of the system.
2. Definition of the architecture of the EMS/MAS.
3. Definition of the interactions between agents.
4. Modeling of the subsystems as agents.
2.1. Multi-Agent Systems 27

5. Implementation in a simulator.
6. Testing and validation.

2.1.4 Communication, Languages and Ontologies

Designing a MAS requires formalizing how the agents coordinate themselves, coop-
erate, take decisions, etc., which implies that agents and MASs need to share at least
common languages. To this end, the Foundation for Intelligent Physical Agents (FIPA)
has defined some specifications and standards for the use of MASs [81].
Languages such as ACL (Agent Communication Language) and KQML (Knowledge
Query and Manipulation Language) were created for use in MAS development, so that
agent can use a common language and vocabulary to communicate. ACL is a FIPA specifi-
cation [82] and is usually preferred. Each message is given several attributes, including the
content of the message, and information about the participants and the ownership of the
conversation. The structure of an ACL message should contain the following parameters:
– A parameter defining the type of communication, called performative; this field
indicates whether the message is a request, a reply, an information, etc.;
– The list of the participants in the conversation, with information on the sender and
the receiver(s), and reply-to fields, including the names of the corresponding agents;
– The content of the message;
– A description of the content, with the used language, encoding and vocabulary, called
ontology;
– And conversation control parameters, such as a conversation identifier and protocol.
Fig. 2.3 shows the structure of an example ACL message, sent by agent1 to agent2. The
performative is the only mandatory parameter in the specification, but others should also
be included when necessary to facilitate message sorting. The FIPA semantic language
(FIPA-SL) is a content language specification with a specific syntax and semantic for use
with ACL, that is implemented in several MAS development tools [83].

(REQUEST
:sender ( agent-identifier :name agent1@platform:1099/JADE )
:receiver (set ( agent-identifier :name agent2@platform:1099/JADE ) )
:content "Hello! How are you?"
:language FIPA-SL0
)
Figure 2.3: Sample ACL message, using the default ontology.

For complex conversations between heterogeneous agents, ontologies may be needed to

define the vocabulary agents use in their conversations. As for humans, using a common
language may not be sufficient. An ontology is a formal representation of knowledge,
under the form of a set of concepts and of relationships between these concepts. As
described in McArthur et al. [58], standards specifying data models, such as CIM (Common
Information Model), can be used as a basis to build such ontologies. A CIM-based ontology
28 Chapter 2. Agent-Based Modeling, Control and Simulation

was for example made available by the IEEE Power & Energy Society (PES) Multi-Agent
Systems Working Group [84].

2.1.5 Agent Management

FIPA specification SC00023 [85] defines two levels in how agents should be managed,
exist and operate. The first level, called agent level, corresponds to each agent itself, while
the second relates to groups of agents and how they interact with each other, called the
MAS level. At the agent level, it defines the life cycle of the agent, and at the MAS level,
it proposes agent management services and a message transport system. These services
are essential in enabling the MAS to operate in a distributed and flexible manner.
The first level defines how agents have their life cycles, from their creation to their
end. During its lifetime, an agent can be in five possible states (Fig. 2.4):
– Initiated, just after its creation. In this state, the agent executes a series of instruc-
tions defined by the user, and run once as an initialization procedure.
– Active is the normal state in which the agent operates.
– Waiting, when the agent is pooling for an external event and has not been woken
up. This state can for example be used when the agent is waiting for a message from
another agent.
– Suspended, when the agent has been halted from the active state. For example,
an agent controlling a power source may be suspended during certain maintenance
operations.
– In transit, when the agent is physically moving from one agent platform to another
(e.g., from a computer to another).
At the MAS level, each agent platform (AP) contains several entities listed below and
at least one agent. A single platform can consist of agents located in different physi-
cal locations, such as on remote computers. The platform should implement the agent
management reference model (Fig. 2.5), which defines service entities including:
– A directory facilitator (DF) that provides a yellow-pages service to agents, i.e., a list
of agents with their respective capabilities. Each agent can register its services in the
DF and then query it for knowing if other agents have a certain service registered.
Implementing a DF is optional, and multiple DFs can exist in a single AP.
– A unique agent management system (AMS), which supervises the AP and maintains
a list of all agents and their addresses in the AP. This functionality is similar to a
white-pages service, i.e., a list of agents and their name and address, to which each
agent must register.
– A message transport system (MTS), which enables messages to be transported from
one agent to the other. A message transport protocol (MTP) is used for physically
transferring messages between agents on possibly different platforms; all messages
exchanged between platforms go through the MTP.
– The agent itself, which communicates with other agents using the ACL language
and may have access to external software.
2.1. Multi-Agent Systems 29

Waiting

None Initiated Active Suspended

In transit

Figure 2.4: Life cycle of an agent, based on [85]. The cycle starts with the creation, the
transition to the initiated state and thne to the active state. The agent can then switch
to three other states, or be destroyed.

Agent platform

Agent AMS DF

MTS

Other platforms and entities

Figure 2.5: The FIPA agent management reference model, based on [85].
30 Chapter 2. Agent-Based Modeling, Control and Simulation

For enabling the previous services to operate properly, each agent is assigned a unique
Agent Identifier (AID). Each AID is composed of three parameters: a unique name, usually
consisting of a local name and the address of the AP; an address list, to which the messages
should be delivered; and a resolver, used by the AMS for resolving the transport address
of the agent.

2.1.6 Topologies
The collection of roles, authority relationships, data flow, resource allocation and co-
ordination patterns that guide the behaviors of agents is defined as an organizational
topology. The major topologies used in MASs include hierarchies, holarchies, coalitions,
teams, congregations, societies, federations, markets and matrix organizations [86]. Each
has its own strengths and weaknesses, and some topologies are more appropriate than
others, depending on the selected application (Fig. 2.6):
– A hierarchy is the earliest and the most widely used topology, in which agents are
arranged in a tree-like structure. Agents higher in the tree have a more global view
than agents below them. Lower-level agents transmit the information perceived
locally to higher-level agents, which provide directions to those below them, on the
basis of a more complete amount of information. This topology is typically used in
most current control systems.
– Holarchies consist of agents (called holons) that are constituted by several entities
and are at same time part of a larger entity. Biological species, individuals, cells and
atoms can each be viewed as holons sharing this dual characteristic.
– Coalitions are dynamic and short-lived, and emerge as soon as a goal has to be
fulfilled by a subset of an agent population. The coalition is destroyed when the
constituent agents have managed to perform the task. The structure of a coalition is
typically flat, although there can exist a leading agent that represents the coalition
as a whole. In a coalition, agents are selfish, i.e., they try to maximize their own
profit. Such topology may be suited for the restoration of supply after a blackout.
– Teams are an altruist type of organization, as opposed to coalitions; they attempt
to maximize the utility of the whole team, and coordinate their actions in order to
efficiently fulfil a common task. Team agents have an explicit representation of the
shared tasks and they know the means by which cooperation should progress.
– Congregations are generally long-lived and formed from heterogeneous agents that
have great interest to get together. Some simple examples of congregations are
clubs or academic departments. Congregating agents are expected to be rational,
by maximizing their own long-term utility. Congregations are formed if agents want
to increase information gain or decrease commitment failure.
– Societies are inherently long-lived and open. Agents living in a society may have
different goals, levels of rationality and heterogeneous capabilities. They meet and
interact according to social laws (or norms), which dictate how they should coexist.
Vehicular traffic laws are an example of social laws that minimize conflicts and
encourage efficient solutions.
2.1. Multi-Agent Systems 31

– Federations are arranged such that some agents delegate a part of their autonomy
to a single agent that represents the group. Group members interact only with this
delegate (also called facilitator, mediator or broker), which acts as an interface be-
tween the group and the outside world. The delegate typically receives undirected
messages from its group members and sends information to the delegates of other
federations. Messages from group members include skill descriptions, task require-
ments or status information, whereas messages from or to other delegates include
task requests or capability notifications. This structure could for example be used
to interface several microgrids, where the central controller of each microgrid would
be the delegate.
– Marketplaces, or market-based organizations, enable buyers and sellers to send and
receive bids for a common set of items, such as shared resources or tasks. Like for
federations, an individual or a group of individuals in a marketplace is responsible
for coordinating the actions of other agents. Unlike for a federation though, agents
within a marketplace are competitive. This topology is already used in bulk elec-
tric power markets, where independent power providers and utilities (among other
players) bid to sell and buy power.
– Matrix organizations mimic how humans influence one another, i.e., the behavior of
an agent or of an agent group may be influenced by multiple centers of authority.
How the agent perceives these influences can influence other agents as well. Simple
examples of such influences include someone receiving guidance from interacting
entities or agents.

Figure 2.6: A hierarchy (left), an holarchy (center) and a federation (right) example.

For power systems, no preferred topology is considered better than others, and actual
topologies usually consists of hybrid combinations of the previous types, as the selected
topologies often derive from the physical architecture of the systems. The MASs proposed
in the next chapters illustrate this remark. Several parameters may influence the choice
of a topology:
– The structure of the real world organization of the system or the power grid. Intu-
itively, a MAS-based grid control system would be designed to mimic the architecture
of the real grid, or of a part of it. Elaborate functionalities may however require
different configurations.
– Whether decisions require negotiations. For example, market-based mechanisms re-
quire to take into account the structure of the market, the possible coalitions that
32 Chapter 2. Agent-Based Modeling, Control and Simulation

may arise, the role of regulators, etc.

– Needs for reliability. Possible failures in the communication infrastructure may for
example require redundancy.
– The scale of the system and possible evolutions. For example, a control system
spanning a very large distribution system may need to be divided into several inter-
connected subsystems.
– The necessity of changes of topology to reach goals. In case of a localized failure
on the distribution system, the architecture of the grid and the corresponding con-
trol system may need to be reconfigured to keep the lights on for a maximum of
customers, e.g., through a self-healing mechanism.
– Constraints imposed by legislations and standards. Regulations and best practices
may also require adopting certain specific architectures, e.g., to ensure interoper-
ability.
Such choices are a trade-off between several parameters including computation, coor-
dination simplicity and organizational rules, plus the simplicity of the architecture. An
analysis according to these criteria defines the topology of the system and the way agents
interact. It should also be noted that the architecture of a MAS may change over time,
depending on what it is trying to achieve and how the environment evolves.

2.1.7 Inter-Agent Interaction

The described topologies define how agents are organized, but it is still needed to
define how they interact. Interactions are paramount notions for defining MASs [47]. An
interaction between agents can take place if they can act or communicate and if there
are situations where they can get together, such as the need to fulfil a common objective.
They are conducted under the form of discussions between at least two agents and can
occur in numerous situations: the rescue of an agent by others, a conversation between
two agents, the implicit agreement when two agents have to decide which one goes first,
the cooperation of several agents to fulfil a common task, and so on.
Interactions are usually required when agents have to satisfy a common objective while
taking into account their limited resources and individual skills. Getting together to fulfil a
common objective involves that some agents are part of a possibly emerging organization,
as described in the previous section. Therefore, every agent organization is the result of
these interactions and of the place where they take place. The dynamic characteristics of
interactions implies that new agent organizations are likely to be formed as new objectives
have to be satisfied. For example, agents can form coalitions if it helps them reach their
goals [87].
Defining interactions between agents depends on answers to the following questions:
What is the nature of their goals? What are their resources? Which skills do they possess
to fulfil these goals? Several criteria can be used to classify interactions, including the
nature of the goals pursued by the agents, their relationship to external resources and
their individual skills regarding the task at hand:
– Typically, agents are engaged in competitive tasks when their goals are contradic-
tory, whereas they cooperate or co-evolve when their individual goals are compatible
2.1. Multi-Agent Systems 33

(that is when the satisfaction of a goal by an agent does not interfere with the possi-
bility of satisfying another goal by another agent). Maintaining the balance between
electricity supply and generation is an example of problem where cooperation is
essential.
– External resources include all environment elements that agents need to satisfy a
goal, such as raw materials, energy, available amounts of space and time, etc. For
example, every energy source or power line has a limited capacity. Limited resources
can lead agents to conflicts as they will likely need the same resources at the same
time and at the same place than other agents. Such conflict situations can generally
be resolved by the coordination of agent actions.
– The last criterion relates to whether the task can be pursued by a single agent
rather than by a group of agents, and if each one of them has the appropriate skills
to perform its subtask.
To enable interoperability (i.e., the ability of several systems to exchange information
and use it), not only between agents but also between MASs, interactions are structured
and follow common rules called protocols. Basic and common types of interactions are
requests, queries, subscriptions and propositions. An example is the FIPA-Request proto-
col [88]: an agent can formulate a request, that other agents can accept or refuse (Fig. 2.7).

1 request

not- agree
2 refused
understood

3 failure inform

Figure 2.7: Diagram of the FIPA-Request protocol, based on [88]. The protocol enables
an agent to formulate a request that other agents may accept or refuse. The corresponding
conversation is divided into three consecutive steps.

Those interactions are often limited, and combinations of them are used instead. There
are several types of such complex interaction protocols. Contracting, auctions, bargaining,
voting and brokering are the most common ones.
– The contract-net protocol is an example of contracting, defined in FIPA specification
SC00029 [89]. This task-sharing protocol consists of a collection of agents forming
a contract network. Two categories of agents are distinguished: the manager, and
contractor agents. A typical round starts with a call for proposals sent by the
initiator. A deadline can be set to limit the duration spent waiting for answers.
Contractors (participants) can then submit proposals (prices, time to execute an
action, etc.) or refuse. The initiator evaluates the proposals and selects zero, one or
several agents to perform the task to be done. The selected participants are free to
accept or refuse this offer.
34 Chapter 2. Agent-Based Modeling, Control and Simulation

– Agents can interact and distribute tasks through auctions. Four main types of
auctions are commonly used: English, Dutch, sealed first-price or Vickrey, and Wal-
rasian auctions [90].
– English auctions are the most common ones. Participants bid openly against
each other, with each subsequent bid higher than the previous one. A reser-
vation price (the minimum price) may be set by the auctioneer. The auction
ends when no participant is willing to bid further. FIPA defines a specification
for English auctions in XC00031 [91].
– In Dutch auctions, the auctioneer starts with a high price which is lowered
until some participant accepts the announced price. This type of auction is
also defined by FIPA in its XC00032 specification [92].
– With sealed first-price auctions, all bidders simultaneously submit their bids,
and the winner is the one with the highest bid. Bidders can only submit a
single bid. Vickrey auctions are identical except that the winner pays the
second highest submitted price [93].
– Walrasian auctions are more complex and enable matching supply and demand
in a market of perfect competition. A market clearing price is set so that the
total demand equals the amount of sold goods, and leads to a general equi-
librium [75]. This type of auction was proposed for use in electricity markets
using locational margin pricing [94].
– Bargaining is an alternative to auctions for pricing goods, i.e., when prices are not
fixed and can be negotiated. The goods can for example be split into several parts
and be themselves subject to bargaining. As humans, agents can employ various
strategies to reach their goals.
– Interactions can also happen under the form of votes. Voting protocols, such as
Robert’s rules of order [95], define procedures for conducting votes between agents.
Voting can be used to take decisions when votes are very simple (e.g., yes or no).
Similarly to votes during elections, various rules can be adopted for selecting the
winner(s).
– Another interaction type specified by FIPA is brokering [96]. For example, an agent
can request a broker to find other agents who can answer a query; the broker would
then relay the answer back to the initiator.
As for topologies, no preferred interaction means is considered better than others for
power systems. The choice of an interaction protocol is mainly influenced by require-
ments for security, by regulations, or by existing market structures. In this dissertation,
a simplified version of the control net protocol and other basic protocols are used for the
presented applications.

2.2 Multi-Agent Development Framework

Although it is possible to develop a MAS from scratch, using a dedicated develop-
ment platform (middleware) is, in most cases, a better solution. Many different toolkits
2.2. Multi-Agent Development Framework 35

were created over the years [97], and include tools and functionalities that facilitate the
development of MASs.

2.2.1 Development Platforms

Among the numerous existing platforms, common general-purpose examples include
JADE (Java Agent DEvelopment Framework), AgentBuilder, MadKit and ZEUS. Other
platforms may be of interest for specific uses: NetLogo enables beginners to get started
with MAS programming, JANUS simplifies holonic agents building, Cougaar enables very
large scale simulations, JADEX proposes a framework for developing intelligent agents,
JACK for autonomous agents, etc. Some of them also comply with FIPA standards,
especially for messaging and agent management. A list of such platforms is available
in [98]. Most of these platforms are written in Java, but other languages can be used.

2.2.2 The JADE Platform

For power systems applications, JADE is the most popular framework used in the
literature. JADE is an open source platform for peer-to-peer agent-based applications. It
provides a runtime environment for agents and a library of classes that provide ready-made
pieces of functionality and interfaces for custom and application-dependent tasks. Its main
advantage is that it has an extensive documentation, with tutorials, books [99] and an
active community. It also includes built-in graphical tools that support the debugging
and deployment phases of MAS development. It is fully compatible with FIPA standards
and specifications, enables agent platforms to be distributed across different machines,
and supports agent mobility. As it is fully implemented in the Java language, it is also
cross-platform, and works on multiple operating systems. Many third-party plug-ins and
extensions (e.g., for mobile devices, for intelligent agents, etc.) are available for specific
uses. Due to these properties, the JADE platform is selected as a basis for building the
MAS-based simulators described in this dissertation.
Agents running in JADE are organized as follows. A platform is created, in which
at least one container containing agents is created. The first container, called the Main
Container, holds two specialized agents: the AMS and DF agents. The AMS agent is
the one that creates and destroys other agents, containers and platforms. The DF agent
implements a yellow-pages service the other agents can query to retrieve information about
services offered by other agents.
Agents created in JADE can be in any of the states shown in Fig. 2.4. Additionally,
their execution follows a model that relies on behaviors. The notion of behavior describes
how an agent reacts to an event, and enables agents to operate independently and to
be executed in parallel with other agents. This feature is made possible by assigning a
Java thread to each behavior instead of a single thread per agent, as each agent may be
involved in several tasks in parallel. Fig. 2.8 summarizes the agent execution model. In
the setup() method, agents are initialized and initial behaviors are launched. Then, as
long as the agent is alive, each active behavior b is executed until it is done. The done()
method is run at the end of each execution of the behavior and determines whether the
behavior will be run again or not. If it will not, then it is removed, otherwise, the next
36 Chapter 2. Agent-Based Modeling, Control and Simulation

active behavior is called. If the agent is taken down, the takeDown() method is run to
enable the programer to run some clean-up operations.

setup()

yes
doDelete() takeDown()
called?

no
Get the next
active behavior b

[Link]()

no
[Link]()
yes

Remove the behavior

from the list

Figure 2.8: Agent execution model of agents in JADE.

Three graphical tools enable faster and easier debugging and monitoring:
– The Introspector agent (Fig. 2.9) helps monitoring and controlling agents by provid-
ing information on their life cycle and behaviors, and on ACL messages exchanges.
– The Dummy agent can be used to create and send ACL messages to other agents
and viewing the list of messages sent and received by an agent.
– The Sniffer agent (Fig. 2.10) intercepts ACL messages and displays them similarly to
UML (Unified Modeling Language) sequence diagrams, which is useful for monitoring
how agents interact through message exchanges.
2.2. Multi-Agent Development Framework 37

Figure 2.9: Screenshot of the Introspector agent interface in JADE.

Figure 2.10: Screenshot of the Sniffer agent interface in JADE.

38 Chapter 2. Agent-Based Modeling, Control and Simulation

2.3 Co-Simulation Framework

With JADE used for modeling, controlling and simulating the studied power systems,
a dedicated co-simulation framework is developed to enable the analysis of simulation
results with a power systems analysis software, PowerWorld Simulator.

2.3.1 Need for Co-Simulation Tools

Modern technological product development requires phases of specification, develop-
ment, simulation and prototyping, before reaching the first test phases. However, as
prototyping and testing costs are often high, simulations have taken an ever increasing
importance over the last few decades, especially as computing capabilities have dramati-
cally increased and enabled new possibilities.
For the smart grid to become a reality, numerous new technological products have to
be developed, from advanced energy management systems to distributed storage, and will
have to go through extensive tests in simulation before they are actually implemented on
the grid. The risks of compromising the stability of the grid are too high to integrate new
products without ensuring they will not create problems.
In the advent of the smart grid, intelligent and distributed control systems are major
research topics, in that they are envisioned as essential elements required for making such
smart grids a reality. These smart control systems often rely on AI algorithms, such
as learning algorithms, optimization algorithms, expert systems, etc. [100]. The main
objectives justifying the use of these algorithms is their capacity to solve complex problems
efficiently. Problems to solve can range from optimal economic and environmental dispatch
and unit commitment to load forecasting and fault diagnosis and analysis. The capacity
of AI algorithms to learn from experience and explore search spaces efficiently makes them
particularly suitable for smart grid applications, although their industrial use is still rather
low.
By enabling the modeling and simulation of a large number of elements (sources, loads,
substations, market players, etc.) as well as their interaction, the MAS concept is a useful
tool for conducting smart grid-related research, as shown earlier. However, in order to
evaluate the practical validity of a developed control system, tests need to be carried out
with advanced power systems analysis tools [101]. As a decision-making algorithm in a
power system takes decisions that are physically translated by electrical components, such
as switches, generators, etc., the impact of these decisions on the power grid has to be
evaluated. An example is the effect of opening a switch on a power line, as a function of a
decision-making algorithm. This action may overload a transformer or exceed the voltage
rating of a power line. Without an analysis tool capable of running power flows (in this
case), such results could not be found. Unfortunately, AI tools do not include advanced
power systems analysis algorithms yet, and these analysis tools do not enable developing
AI-based decision-making algorithms either. Therefore, these two different kinds of tools
need to be interfaced.
Researchers have tried to resolve this issue by interfacing AI algorithms with exter-
nal physical simulation-oriented software. A brief listing of most of these co-simulation
possibilities can be found in [102]. Two main approaches have been used:
2.3. Co-Simulation Framework 39

– The first one is designed for communication analysis purposes, and software such as
OpenDSS and NS2 [103], PSLF and NS2 [104], and Modelica and NS2 [105] were
interfaced successfully.

– The second approach is designed for combined multi-agent and power systems anal-
ysis purposes. Past work has focused on interfacing Matlab/Simulink with JADE
through a TCP server [106], and JADE and Matlab’s PSAT toolbox [107]. JADE
and PowerWorld Simulator were also interfaced through a COM interface, but this
implementation is not documented [66]. Software such as GridLAB-D [108] have re-
cently attempted to integrate both simulation sides and rely on multi-agent systems,
but lack a user interface and a detailed documentation.

As current simulation tools do not enable the interaction between these two interde-
pendent fields, creating a co-simulation framework is therefore required. A framework for
interfacing a power systems analysis software like PowerWorld Simulator simultaneously
with Matlab and JADE is proposed in the following section. The objective is to enable re-
liable communication and coordinated interaction between both simulations for validating
the integration feasibility of the developed AI algorithms and systems.

2.3.2 Specifications
[Link] Requirements

Co-simulation is a simulation methodology in which several individual components are

simulated simultaneously by different tools and exchange information with each other. In
the present case, the objective of co-simulation is to run a coordinated simulation with an
AI-based (or multi-agent-based) decision-making tool, and a power systems analysis tool.
The output of the first one is the input of the second, and vice-versa. The validity of the
decision-making algorithm can then be tested, by evaluating its impact on the behavior
of the electric system.
For using the co-simulation tool in the applications presented in this dissertation,
several constraints have to be met:

– One of the simulation tools must enable easy AI-based decision-making algorithms
development, ideally multi-agent systems and advanced computational intelligence
algorithms.

– The other tool must enable easy and advanced modeling and power systems analysis,
in coordination with the first tool.

– Both simulations must be synchronized.

– Both simulators must be able to exchange information as fast as possible; one should
not dramatically slow down the other.

– Both simulations should be able to run on different computers, especially if one or

both of them are demanding in terms of computation power.
40 Chapter 2. Agent-Based Modeling, Control and Simulation

PowerWorld
JADE agents Matlab
Simulator
Agent 1

COM client
TCP server

SimAuto
Interface
Agent
··· TCP/IP COM

Agent n

Figure 2.11: Interface between JADE, Matlab and PowerWorld, from [38].

[Link] Software and Tools

A first step toward the definition of this framework, is the selection of the software
tools used for running the coordinated simulations. Three tools are used:
– PowerWorld Simulator [109] is a popular simulation software used to simulate high
voltage power systems. Using this tool, it is possible to perform power flow analysis
on a system with up to 100,000 buses. Multiple add-ons enable performing additional
analysis such as transient stability, optimal power flow, voltage stability, reserves,
transfer capacity, etc. SimAuto is an add-on used to control the simulator from
external applications. SimAuto acts as a COM object, with which other software
can communicate, by sending requests and receiving data. Any programming tool,
such as Visual Basic, Matlab or Borland Delphi, capable of accessing COM objects
can thus be interfaced with it. These add-ons, combined with the ease of use of
PowerWorld enabling fast modeling and simulation, make PowerWorld a tool well
adapted for the needs of this framework.

– MathWorks Matlab [110] is a powerful software that provides a programing envi-

ronment to perform complex numerical computations and data analysis. Numerous
toolboxes are available for many applications ranging from multi-physical systems
modeling to signal processing. Matlab can also be interfaced with various program-
ming languages such as C, C++ and Java. Matlab is commonly used for rapidly
prototyping complex numerical algorithms, especially as its toolboxes facilitate the
use of AI-based solutions. However, although the modeling and analysis of power
systems is possible through several toolboxes, their use is more complex than with
PowerWorld and requires more development time.

– The multi-agent development framework JADE, presented earlier, is also used.

2.3.3 Structure
[Link] Matlab–PowerWorld Interface

The second step in the development of this framework is the creation of an interface
between Matlab and PowerWorld. This interface is established using the COM server
offered by SimAuto, which is documented in PowerWorld’s user manual [111]. Through
this interface, PowerWorld can be requested to run instructions such as the following:
2.3. Co-Simulation Framework 41

– Open, save and close a case (network);

– List the devices of each type (buses, branches, generators, loads, etc.) present in the
opened case;
– Get the parameters (status, MW and MVAR rating, nominal voltage, etc.) of a
precise element or of all elements of a given type;
– Change the parameters of a precise element or of all elements of a given type;
– Run a power flow using a given algorithm (Newton-Raphson, Gauss-Seidel, etc.);
– Add elements to the network.
Once a connection between both software is established, functions running the above-
mentioned instructions can be used in Matlab to interact with PowerWorld and return
results.

[Link] JADE–PowerWorld Interface

In order to enable the use of MASs for power systems control while maintaining the
capability to use power systems analysis tools, the interface between Matlab and Power-
World is extended. As multi-agent frameworks such as JADE are commonly written in
Java, they cannot be directly interfaced with external power systems analysis software.
As no off-the-shelf solution exists, a custom interface must be created.
Building on the first interface between Matlab and PowerWorld, an interface between
JADE and Matlab is created. Matlab then serves as a data gateway between both soft-
wares, as in Fig. 2.11. A direct connection between JADE and PowerWorld is theoretically
possible through the SimAuto COM interface; however, contrary to Matlab, no Java doc-
umentation is readily available from PowerWorld. Moreover, the chosen structure allows
simultaneously running specific instructions in Matlab, such as for solving complex equa-
tions. Matlab embeds toolboxes and functions that enable fast prototyping, which often
take longer to develop in Java alone.
The interface between Matlab and PowerWorld/SimAuto described earlier is modified
so that it can act as a gateway for requests issued by JADE agents. A TCP connection is
established to enable communication between JADE and Matlab. The TCP/IP connection
enables running all softwares on a single computer, or using a remote computer for running
Matlab and PowerWorld. The Instrument Control Toolbox is required for Matlab to
support TCP communication (however, any other toolbox enabling TCP connection could
also be used). The connection between Matlab and PowerWorld is achieved with a COM
object through SimAuto, as mentioned earlier. On JADE’s side, a single agent handles
all communication with Matlab, and is referred to as InterfaceAgent from now on.

2.3.4 Operation
On initialization, a TCP connection is established between InterfaceAgent and Matlab,
and is maintained open throughout the entire simulation duration. Then, the following
process is used to handle each request issued by any agent (see Fig. 2.12, where messages
are numbered according to the following list):
42 Chapter 2. Agent-Based Modeling, Control and Simulation

1. The JADE agent originating the request sends a message to InterfaceAgent using
the standard MTP, with information on what the intended action is, and, if that is
the case, the data to pass to PowerWorld.
2. InterfaceAgent processes the content of the message, formats it to be sent and un-
derstood by Matlab, and sends it through TCP.
3. Matlab receives the message, processes it, and requests PowerWorld to run the appro-
priate instruction based on the content of the message it received. The instructions
are the same as the ones listed earlier for the Matlab–PowerWorld interface. In some
cases, Matlab can also run some instructions itself using the data provided by the
agent.
4. After PowerWorld (and Matlab, if that is the case) has run the selected instructions,
it returns the result to Matlab through the COM interface of SimAuto.
5. Matlab then reprocesses the answer and sends it through TCP back to InterfaceAgent.
6. Finally, InterfaceAgent processes the answer it received and sends the final answer
to the agent that issues the initial request.

Interface
Agent Agent Matlab PowerWorld

1
2
3

4
5
6

34 ms

48 ms Get bus
parameters
52 ms

9 ms

109 ms Run power

flow
174 ms

Figure 2.12: Communication flowchart of a request issued by a JADE agent, from [38].
Run times are given for each part of the framework, for two example requests with a
14-bus system.

Run times indicated in Fig. 2.12 show that these interfaces perform well for both test
cases (getting the parameters of the buses, and running a power flow). However, these
run times are given briefly as an example based on a 14-bus system, and are not to be
treated as standard. Run times may vary for other systems and other runs.
The source code for both interfaces (JADE–Matlab and Matlab–PowerWorld) are
available online at Colorado State University, with a working example [112].
 3

Gas Turbine Power Plants for Smart Power

Generation

Contents
3.1 Smart Power Generation . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.1.1 Drivers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.1.2 Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.1.3 State-of-the-Art . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.1.4 Application to Gas Turbine Power Plants . . . . . . . . . . . . 50
3.1.5 Energy Management Systems for Smart Generating Plants . . . 51
3.2 Gas Turbine Characteristics . . . . . . . . . . . . . . . . . . . . . . . . 51
3.2.1 Fuel Consumption . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.2.2 Combustion Gas Emissions . . . . . . . . . . . . . . . . . . . . 53
3.2.3 Starting and Stopping Cycles . . . . . . . . . . . . . . . . . . . 55
3.3 Energy Management System Architecture . . . . . . . . . . . . . . . . . 56
3.3.1 Agent Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.3.2 Selected MAS Architecture . . . . . . . . . . . . . . . . . . . . 57
3.3.3 Agent Interactions . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.4 Energy Management Strategy . . . . . . . . . . . . . . . . . . . . . . . 64
3.4.1 Start and Stop Algorithm . . . . . . . . . . . . . . . . . . . . . 64
3.4.2 Turbine Operation Ranges . . . . . . . . . . . . . . . . . . . . . 65
3.5 Economic and Environmental Dispatch . . . . . . . . . . . . . . . . . . 66
3.5.1 Single and Multi-Objective Optimization . . . . . . . . . . . . . 66
3.5.2 Problem Definition . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.5.3 Aggregation-Based Power Dispatching . . . . . . . . . . . . . . 69
3.6 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.6.1 MAS Implementation . . . . . . . . . . . . . . . . . . . . . . . . 70
3.6.2 Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.6.3 Dispatching Algorithms Comparison . . . . . . . . . . . . . . . 73
3.6.4 Performance Coefficients Effectiveness . . . . . . . . . . . . . . 77
3.6.5 Start and Stop Algorithm Effectiveness . . . . . . . . . . . . . . 77
3.6.6 Energy Costs Comparison . . . . . . . . . . . . . . . . . . . . . 79
3.6.7 Flexibility and Resilience Test . . . . . . . . . . . . . . . . . . . 79
3.6.8 Communication and Computation Requirements . . . . . . . . 82
3.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

43
44 Chapter 3. Gas Turbine Power Plants for Smart Power Generation

3.1 Smart Power Generation

3.1.1 Drivers
With the growing share of renewable energy sources in modern power grids and the
increasing electricity consumption, intermittency and fast power ramps are increasingly
important concerns. Short-term peaks and long periods without generation can be ob-
served within the same day (Fig. 1.1). Grid operators need to maintain a balance between
generation and supply to ensure the stability of the system, and generation patterns of
RESs do generally not match demand profiles [6].
The smart grid is expected to help tackle these problems, but the generation-side also
needs to adapt. As a consequence, the paradigm of simply increasing generation capacity
with larger generators to satisfy an increasing demand is not sufficient any more. Large
generators with long start times, slow ramp rates, and long maintenance durations are
also not well-suited for load balancing applications. A new approach relying on flexible
and fast-response generation sources is required for fast power ramps resulting from such
external factors [113]: smart power generation.

3.1.2 Concept
Smart power generation is a concept proposed by Klimstra and Hotakainen in [28],
which aims at providing maximum reliability and flexibility by matching electricity gen-
eration with demand using multiple identical generators. These generators can start, stop
and operate efficiently independently. These characteristics make these smart generating
plants (SGPs) flexible enough so that they are suitable for both base load and peaking
power generation.
This concept is useful for load balancing applications, an essential task to ensure a
stable and reliable supply of electric energy. This statement is particularly relevant when
the share of intermittent energy sources in the energy mix is high, and producers need to
adapt their output more frequently than earlier. In order to adapt to sudden variations
in intermittent generation, SGPs may need to react quickly and to operate either at
base load, at intermediate load, as peaking sources, or may even be stopped if renewable
generation is higher than demand.
Klimstra and Hotakainen identify a series of properties for SGPs [28]. The ability to
start rapidly is required to enable adapting the output of the plant to varying demand and
intermittent generation. By starting and stopping small units appropriately (a process
called cascading) instead of simply changing the output of a single larger unit, higher
efficiency and reliability levels can be achieved [28]. For the same reasons, these small units
must be able to ramp up and down rapidly, while maintaining a high efficiency over a wide
load range. As SGPs rely on multiple identical units, such plants are inherently modular
and their capacity can be adapted according to medium and long term requirements,
at a minimal cost. This specificity facilitates maintenance operations. Non-necessary
generators can also participate in ancillary services (e.g., spinning reserve). Generation
units should additionally be able to be controlled remotely, and independently from each
others. For example, the failure of a unit should not affect the operation of another one.
Fuel flexibility, low maintenance, black start capability, short building time, low spatial
3.1. Smart Power Generation 45

impact, low costs, low water use, and low sensitivity to ambient conditions are other
concerns.

3.1.3 State-of-the-Art
Due to its relative novelty, this concept has been little studied in the literature. Most
works tend to focus on precise topics such as economic and/or environmental dispatching
algorithms [114–116], unit commitment algorithms [117, 118], or turbine modeling [119,
120], but lack a systemic approach required for designing a complete EMS that takes into
account additional aspects required for practical implementation, such as interactions
between subsystems and communication [121]. Only a few studies, such as in [122, 123],
focused on optimizing gas turbine operation without altering thermodynamics. Moreover,
most papers deal with economic dispatch between power plants, each with their own
characteristics, rather than between power sources inside a single one. Tests are conducted
on well-known benchmarks but for a single data set only [114–116], which does not give
any information on how the algorithms behave in such a dynamic process. They also do
not consider the varying efficiency of each source over time, assuming it remains constant.

3.1.4 Application to Gas Turbine Power Plants

Among sources capable of meeting the requirements for SGPs, gas turbine power
plants are, with coal-fired power plants, some of the most common solutions used for
meeting such high ramps. Power plants made of simple cycle gas turbines comply with
most of the characteristics of SGPs listed earlier. Gas turbines enable a great operational
flexibility: ramp-up and ramp-down rates can exceed several tens of MWs per minute; the
turbines constituting the plant can be started and stopped frequently and independently,
depending on demand; turbines can be operated with a variety of fuels; time from planning
to completion is less than a year; capital costs are relatively low; and maintenance outage
times are short [124, 125].
However, these sources are more expensive to operate (considering the cost per gen-
erated MWh) than sources with slower dynamics such as nuclear power plants, and have
non-negligible emission levels that are only surpassed by traditional coal power plants [126].
Due to these characteristics, gas turbine power plants are often used for providing
power during demand peaks or to compensate ramps originating from intermittent sources.
Other applications include powering remote loads, where transmission capacities do not
exist or are insufficient, and powering large industrial facilities, where production processes
require large amounts of electric and thermal energy (e.g., paper mills or other chemical
processes).

3.1.5 Energy Management Systems for Smart Generating Plants

As the power grid is subject to a deep ongoing modernization, power plants also need to
become smarter, that is, in this case, more flexible and efficient, following the principles of
smart power generation. Satisfying these two requirements at the same time is however a
non-trivial objective, as designing an efficient EMS usually implies modeling very precisely
the target power system, resulting in a loss of flexibility.
46 Chapter 3. Gas Turbine Power Plants for Smart Power Generation

On the one hand, although initial investment costs for building gas turbine power
plants are relatively low, their high operation costs require efficient algorithms for manag-
ing energy flows so as to minimize costs as well as emission levels [126]. With legislations
on greenhouse gases becoming stricter in most parts of the world through cap and trade,
allowances or taxing mechanisms, limiting emissions is a growing concern to be added to
operational costs. On the other hand, an EMS with a flexible architecture enables not
only to adapt to a variety of new or existing plants, but also to make the structure of
the plant evolve over time, while requiring very little modifications. This characteristic
helps reduce commissioning costs, in addition to operation costs, and can also increase
the resilience of the system, e.g., after the failure of a component.
This chapter proposes an EMS for gas turbine power plants operated in the SGP
paradigm, addressing these concerns without altering turbine thermodynamics.

3.2 Gas Turbine Characteristics

Gas turbines are a mature technology with decades of use history, not only in the field
of power generation, but also for transportation, such as for aircraft engines. Associated
with a generator, their power output can range from a few MWs to several hundreds of
MWs, and are sometimes used as combined heat and power (CHP) or combined-cycle
sources.
Gas turbines operate upon the Brayton thermodynamic cycle. Fig. 3.1 shows the
interactions of the main parts of a simple cycle, single shaft gas turbine. Air enters
the compressor at ambient conditions, and is compressed to some higher pressure. This
compression process heats the air, which enters the combustor. In the combustor, the
compressed air is used for fuel combustion. The thermal energy generated by the com-
bustion is then converted into mechanical work, which is used to drive an alternator that
transforms mechanical energy into electricity.
Exhaust

Fuel
Combustor

Compressor Generator
Shaft

Turbine

Inlet air

Figure 3.1: Diagram of a simple cycle, single shaft gas turbine, based on [124].

General Electric (GE)’s 9E gas turbine series is an example of such a turbine, and is
used as a basis for simulation. These turbines are suitable for simple cycle peaking service,
3.2. Gas Turbine Characteristics 47

base load generation, load following and combined cycles. These gas turbines are able to
start, shut-down and handle load changes quickly, and are therefore preferred for these
applications.
Their main characteristics are summarized in Table 3.1. Gas turbine data shown in
the following sections are derived from documentation provided by GE 1 .

Characteristics Values
Output 128.1 MW
Efficiency 34.1 %
Design frequency 50 Hz
Fast start duration 16 min
Normal start duration 28 min
Ramp-up rate 12.6 MW/min
Ramp-down rate 9.5 MW/min
Emission compliant turn-down 50 % base load
Table 3.1: Typical characteristics of a simple cycle 9E gas turbine.

3.2.1 Fuel Consumption

The fuel cost of a turbine is given by the quadratic function (3.1), where f is the total
fuel cost, cfuel is the unitary fuel cost in e /kg, and p ∈ [0, 1] is the set point expressed
as a per unit of the maximum power output [127]. Fig. 3.2 shows the quadratic nature
of the fuel flow of the turbine. No valve-point effect is observable, and there is also no
prohibited operating zone.

fi (pi ) = cfuel · (ai p2i + bi pi + ci ) (3.1)

100

80
Fuel flow [%]

60

40

20
0 10 20 30 40 50 60 70 80 90 100
Load (p) [%]

Figure 3.2: Fuel flow of the turbine.

1. As some data is confidential and is covered by a non-disclosure agreement between GE and UTBM,
only their relative or p.u. value is given.
48 Chapter 3. Gas Turbine Power Plants for Smart Power Generation

Due to the wearing and maintenance history of the turbine, the performance of the
turbine can vary over time, which impacts its cost function [128]. To account for this, a
performance coefficient αi is introduced and corresponds to the ratio of the reference value
of efficiency for a new turbine, ηref , with the actual measured efficiency of the turbine,
ηmeas (3.2). It is assumed that this coefficient can be derived from performance monitoring
systems, such as the ones presented in [128]. This ratio αi is assumed constant, with values
typically included in the [1, 1.03] range, meaning that the measured efficiency of a turbine
decreases over time; although it can be temporarily re-increased with maintenance. The
final fuel cost function is then obtained with (3.3).
ηref (pi )
αi (pi ) = (3.2)
ηmeas (pi )
ctu,i (pi ) = αi (pi ) · fi (pi ) (3.3)

3.2.2 Combustion Gas Emissions

As shown in Fig. 3.3, which displays the emission curves for NOx, CO and CO2 gases,
combustion gas emissions are highly dependent on load [129]. NOx gases result from the
combustion of gases at high temperature and are therefore usually higher at base load. In
most places, air quality regulations require low NOx emissions, especially in urban areas,
due to their highly local toxic effects [130]. On the contrary, CO emissions are higher for
partial loads and result from incomplete combustion. CO2 emissions are also monitored,
as they are sometimes subject to regulations such as carbon trading.
Although technologies such as dry low NOx (DLN) help reduce NOx emissions to
low levels, it should be noted that they are trade-offs between several turbine parameters
(emissions, reliability, load range, etc.): Fig. 3.3 shows that the trends in CO and NOx
emissions tend to take opposite directions, i.e., decreasing one increases the other [129].
CO2 emissions are assumed to depend only on the amount of burned fuel, as in (3.4),
where eCO2 is the total CO2 emissions level, and d is a coefficient giving the emissions
level from the fuel flow. Consequently, reducing fuel consumption is equivalent to reducing
CO2 emissions, as reducing NOx emissions necessarily results in increasing CO emissions.
Emissions are also directly impacted by values of αi .
eCO2 ,i (pi ) = di · fi (pi ) (3.4)
Although most papers in the literature assume emissions vary according to quadratic
functions [131, 132], Fig. 3.3 shows that this assumption is clearly not verified for 9E
turbines. On the contrary, the NOx and CO curves show that 9E turbines use four
different combustion modes (Fig. 3.4):
– For 0 ≤ p < 0.225 (Mode 1): At partial loading, both NOx and CO emissions are
moderately high. This mode is typically only used for starting turbines, leading to
high levels of unburned hydrocarbons.
– For 0.225 < p < 0.50 (Mode 2): This mode exhibits a peak in CO emissions at the
beginning and a peak in NOx when nearing 50 % load.
– If the turbine stays in Mode 2 longer than 5 minutes, it then enters in an extended
version of Mode 2, called Mode 2 Extended, even if the load increases above 50 %,
and exits this extended mode only after going back to Mode 1.
3.2. Gas Turbine Characteristics 49

CO [ppmvd]
300

Mode 1 Mode 2 Mode 3

CO [ppmvd]

200

100

0
0 10 20 30 40 50 60 70 80 90 100

NOx @ 15% O2 [ppmvd]

100

80
NOx [ppmvd]

60

40

20

0
0 10 20 30 40 50 60 70 80 90 100
Load (p) [%]

Figure 3.3: NOx and CO emission curves of 9E gas turbines, based on [129]. Emissions
are heavily dependent on the load and combustion mode of the turbine.

– For p ≥ 0.50 (Mode 3): This last mode enables reaching the lowest emissions (ex-
cept for CO2 ), by maintaining them to stable levels around 9 ppmvd for NOx and
25 ppmvd for CO, due to DLN combustion, even for high load values.

3.2.3 Starting and Stopping Cycles

In order to be able to account for the costs of starting a turbine, the fuel consumption
during the starting cycle is measured and is shown in Fig. 3.5. The corresponding emissions
are obtained using a similar procedure. The cycle lasts 16 minutes [133], after which the
loading cycle can begin and the load increases. Regarding the stopping cycle, the turbine
simply decreases its output to zero and disconnects, without any special procedure that
would have an impact on fuel consumption or emissions.
50 Chapter 3. Gas Turbine Power Plants for Smart Power Generation

Mode 2
extended

p < 0.225

t1
p ≥ 0.225 t2

Mode 1 Mode 2 Mode 3

p < 0.25 p < 0.50

Figure 3.4: Finite state machine describing the turbines, from [39]. Condition t1 is verified
if the turbine stays in Mode 2 more than 5 minutes, and t2 if p ≥ 0.5 without staying
more than 5 minutes in Mode 2.

50

40 Starting cycle
Fuel flow [p.u.],
Load [p.u.]

30

20
Fuel flow
10 Load

0
0 2 4 6 8 10 12 14 16 18 20
Time [min]

Figure 3.5: Fuel consumption of a turbine during its starting cycle, from [39]. During this
cycle, the turbine is gradually brought to its operational state, which consumes fuel.
3.3. Energy Management System Architecture 51

3.3 Energy Management System Architecture

One of the main objectives of this EMS is to reach a high level of flexibility, not
only during its operation, but also during partial outages, and over the entire life-cycle:
the system is designed to adapt to most changes in the structure of the power plant,
whether they are intentional (e.g., increasing the generation capacity by adding a turbine,
or stopping one for maintenance) or not (e.g., after an outage). MASs enable this flexibility
due to their inherent characteristics.

3.3.1 Agent Structure

All agents in this application are based on a similar structure (Fig. 3.6). As the agent
is created, it runs a set of initialization instructions. These instructions enable it to gather
information on the component it is connected to (e.g., the turbine), and to register with
the other agents interacting with it. The agent then reaches a state in which it waits for a
message from another agent. This message can be a request or can contain an information
the agent needs. Depending on the content of the message, the agent runs a given sequence
of instructions, that can include running measurements or interacting with other agents.
At the end of this sequence, the agent goes back to waiting for a message, and can run
background tasks until a new message is received. If the message is a request for the
agent to delete itself, the agent runs some clean-up instructions and deletes itself after un-
registering from other agents. In case of a communication problem, the agent also has the
capability to switch a degraded mode, in which the agent can choose to progressively stop
to ensure the safety of the system, if it does not manage to re-establish communication
on its own [51].

3.3.2 Selected MAS Architecture

The selected MAS architecture consists of seven agent types, divided into three cate-
gories. A first category corresponds to agents controlling or representing physical compo-
nents in the power plant: turbine and load agents.

– A turbine agent is a reactive agent created for each turbine in the power plant,
and that resides in the turbine control system. Through its integration with the
turbine control system, the agent is capable of retrieving information about the
current status of the turbine, for example through measurements, and to impact its
operation, by changing the set point and the state of the turbine. Low level control
actions (fuel flow regulation, air compression, temperature monitoring, etc.) are
not controlled by the agent but by the control system itself; however, the agent has
access to all available data on the turbine. Depending on interactions with other
agents, the turbine agent executes different processes (Fig. 3.7). For example, it can
communicate with the SCADA agent by sending data about the actual turbine state,
and receive new set points in return. Similarly, if requested to change of state (e.g.,
start or stop), or to be deleted (e.g., when the turbine is un-installed or shut-down
for long maintenance), the agent takes appropriate measures to ensure the correct
operation of the system.
52 Chapter 3. Gas Turbine Power Plants for Smart Power Generation

Start-up sequence

Registration

Background tasks
Message
received? No

Sequence 1 Sequence 2 De-registration

Shut-down sequence

Figure 3.6: Generic flowchart of an agent life cycle, from creation to deletion, from [39].
Instruction sequences are selected according to interactions with other agents.

– The load agent is a reactive agent in charge of communicating with the other agents
to provide them information on the total load. It converses with the SCADA agent
and provides it with the current load value.

The second category contains agents fulfilling energy management-related roles: SCADA,
dispatching, forecasting and yellow/white pages agents.

– The SCADA agent plays a role similar to commercial SCADA systems, and acts
as a communication interface between the different agent types of the system, as
well as the power plant operator through the control interface. It gathers data
from the turbine and load agents, and forwards them the set points issued by the
dispatching agent. It also provides other agents with information they require to
operate, whenever they request it, and if they are allowed to do so, and stores logs
of events occurring in the plant. This agent makes the system ready for future
evolutions including additional functionalities, such as advanced scheduling.

– The dispatch agent (Fig. 3.8) is an intelligent agent in charge of controlling the oper-
ation of the turbines with respect to the forecast and actual load, while minimizing
costs and emission levels. At first, based on load forecasts, it decides whether each
turbine needs to be switched on or off. It then computes the optimal set points for
the switched-on turbines, based on data provided by these turbines and the load
agent and forwarded by the SCADA agent. It also maintains a database containing
information on the turbines, dispatch results and events.
 3.3. Energy Management System Architecture
Initialization

Registration

Message
received? No

Yes
Check request type

Set point
State update State change Deletion
change
request? request? request?
request?

Get updated data Check current state Check current state Un-register

Answer requester Extract set point Extract new state Save results
with updated data from message from message
Delete agent
Update set point Update state
and mode

Figure 3.7: Simplified turbine agent flowchart.

53
54 Chapter 3. Gas Turbine Power Plants for Smart Power Generation

– The forecasting agent is an intelligent agent capable of forecasting the load several
hours ahead. It is in charge of computing forecasts of the expected peak load at a
given frequency for the dispatch agent. Forecasts can be obtained either through a
learning algorithm, or through the results of power market bids. This behavior is
not implemented in this application, and the agent simply provides the future load.
– The yellow/white pages agent serves as a name and service server, providing re-
questing agents with the names and addresses of the agents corresponding to their
needs. All agents can communicate with this particular agent. Upon being created
and deleted, each agent in the power plant registers with the pages agent so that it
can be found by the other agents who might need its services. This agent is essen-
tial in enabling the system to evolve when its structure changes, by facilitating the
discovery of new services.
The third and last category corresponds to the human-machine interface agent. The
GUI agent receives updates on the current status of the system provided by the SCADA
agent. This information enables human operators to monitor it, and to take appropriate
decisions when required. In a real implementation, the operators would have the ability
to override the decisions taken automatically by the system, and to operate in a manual
mode. The commands issued by the operators are also sent to the SCADA agent.
Fig. 3.9 provides an overview of the EMS and of the basic interactions between these
agents. This figure indicates that the selected MAS topology has some similarities with
the federations described in section 2.1.6, as the SCADA agent acts as a delegate to other
agents.

3.3.3 Agent Interactions

The way the agents of the MAS interact depends on the situation at hand. Two
situations are described here: architecture change and normal system operation. In the
following figures and simulations, measures taken to ensure data is properly transferred
and received are not considered. However, in a real test case, these should be implemented
to guarantee the proper operation of the system.

[Link] MAS Architecture Change

When the architecture of the power plant changes, e.g., when it is started for the first
time, or when a turbine is added, the MAS structure needs to adapt. In the first case, the
GUI and the energy management-related agents are started, initialized and parameterized
by the operator. The system then waits for power plant agents to register. The operation
of the system can then begin. For a turbine agent, the following process is followed
(Fig. 3.11, where numbers given at the beginning of each line correspond to the numbers
given in the following enumeration):
1. The agent of the turbine is created as its control system starts.
2. As it initializes, it loads or measures the characteristics of the turbine (model, max-
imum output, type of fuel, etc.).
3. It registers with the yellow/white pages agent, so that other agents can find it when
they look for turbines.
3.3. Energy Management System Architecture 55

Initialization

Register with pages

Get SCADA address

from pages

Register with SCADA

Start request
received?

Yes
Extract run parameters

Message
received? No

Yes
Check request type

Dispatching Deletion
request? request?

Update components Un-register

list and characteristics

Save results
Update turbines
operation bonds
Delete agent

Run start/stop
algorithm

Compute optimal
turbine set points

Send results to
SCADA

Store and analyze

results

Figure 3.8: Simplified dispatch agent flowchart.

56 Chapter 3. Gas Turbine Power Plants for Smart Power Generation

Energy management agents

Forecasting agent Dispatch agent

Plant operator

SCADA agent GUI agent

Pages agent

Power plant agents

Turbine agent Load agent

Turbine agent Turbine agent

Figure 3.9: Overview of the power plant EMS architecture, from [39]. Each arrow corre-
sponds to a usual communication channel between agents in normal operation mode.

4. It then registers with the SCADA agent, and waits for a request to transmit infor-
mation about the current state. The agent has then reached the normal operational
state.
5. The SCADA agent finally registers the new agent with the dispatch agent. The
turbine is then fully operational.
Similarly, when a component is required to disconnect, either by the operator or by
the system, the corresponding agent starts by un-registering from the pages and from the
other databases the agents have, so that they do not try to communicate with it any more.
A series of instructions related to the deletion of the agent, e.g., saving measured data,
are then run before the actual deletion from the system. In cases where the agent only
needs to be temporarily disconnected, the agent can also be suspended, by switching to a
standby mode. When requested, the agent can then switch back to normal mode.
This ability of the system to accommodate structural changes enables it to adapt to a
wide variety of power plants, with variable numbers and types of turbines and loads.

[Link] Normal System Operation

During normal system operation, i.e., when the structure of the system does not
change, the system runs the same instructions continuously at a given frequency chosen
by the operator or imposed by regulators (Fig. 3.11):
1. The dispatch agent asks the SCADA agent to send measurements on the current
state of the whole system.
2. The SCADA agent requests each registered agent to send the latest information it
has, e.g., by running measurements.
3.3. Energy Management System Architecture 57

Turbine Pages SCADA Dispatch

Figure 3.10: Operation and interactions of agents when a new turbine is connected.

Turbine Load SCADA Dispatch GUI

10

Figure 3.11: Interactions between agents during normal operation of the EMS. Each arrow
describes the direction of a message exchange between two agents. This example is limited
to two components, one load and one turbine, for ease of understanding.

3. Each agent, based on these measurements, updates its characteristics, such as the
operation zone.
4. The updated data is then sent to the SCADA agent, which centralizes the measure-
ments and updates the database.
5. The SCADA agent sends only the required data to the dispatch agent.
6. The dispatch agent computes the optimal set points.
58 Chapter 3. Gas Turbine Power Plants for Smart Power Generation

7. These set points are then sent to the SCADA agent.

8. The results are forwarded to the respective recipients.
9. The turbines update their status, after they have either accepted or rejected the
requested set point changes.
10. The results are sent to the GUI agent for display.

3.4 Energy Management Strategy

In order to properly manage power flows so as to meet demand in the most efficient
and flexible way, an energy management strategy is defined and is implemented in three
complementary steps:
1. The first step, called start and stop algorithm (SSA), defines whether and when each
turbine should be turned-on or off.
2. The second step is the definition of the operation range of each turbine according to
a set of criteria.
3. The last step is the dispatching algorithm presented in details in section 3.5.

3.4.1 Start and Stop Algorithm

The SSA handles decisions regarding whether and when each turbine should be started
or stopped. Although such decisions are traditionally taken by advanced unit commit-
ment algorithms using day-ahead load forecasting, such algorithms are not required or
applicable here. The first reason is that as all generators are identical except for the value
of their performance coefficient, there is no reason to privilege using a specific turbine
over the others. The second reason is that due to the uncertainties related to intermittent
energy sources, the load forecasting error increases with the forecasting horizon [134,135].
Consequently, very short term (one hour-ahead or less) forecasts are more reliable than
day-ahead predictions and can be used to take the required decisions, even if they some-
times simply consist in persistence models [136].
The decisions taken by the SSA are therefore based on one hour-ahead forecasts of
peak load provided by the forecasting agent, and are used to maintain a minimum spinning
reserve required to meet demand. Reserve is defined here as the amount of generation
capacity that is not used by the generators, i.e., the difference between the rated power
output and the actual output. As reserve is computed when the algorithm takes the
decision and not when the turbine actually starts (in the case of a turbine starting), the
actual reserve value may be lower than its reference value until the turbine is operational.
The algorithm derives the number of turbines to start (nstart ) or stop (nstop ) from the
total forecast peak load (Plpeak ), the minimum amount of reserve Rstart or Rstop given as a
percentage of the total load, the maximum output of a turbine P tu,max and the number of
turbines currently running nrun . The turbines with the best performance coefficients are
started first, similarly to a priority list-based algorithm, and using the fast start procedure
shown in Fig. 3.5.
If the spinning reserve is too low, i.e., if condition (3.5) is verified, nstart is derived
from (3.6):
3.4. Energy Management Strategy 59

Plpeak · (1 + Rstart ) > P tu,max · nrun (3.5)

" # !
Plpeak
nstart = max ceil − nrun , 0 (3.6)
(1 − Rstart ) · P tu,max

On the contrary, if the spinning reserve is too high, the turbine with the highest (worst)
performance coefficient is requested to stop, and the same process is repeated as long as
condition (3.7) is verified:

Plpeak · (1 + Rstop ) < P tu,max · (nrun − 1) (3.7)

3.4.2 Turbine Operation Ranges

The second step is distributed in each turbine agent. As each turbine agent has access
to all information available on the turbine it controls, it can autonomously define its
operation range. The rest of the system has no influence on such decisions. The agents
define the operation ranges of the turbines through these criteria:
– The turbines cannot be started or stopped as long as the minimum up and down time
constraints, (3.8) and (3.9), respectively, are not verified. These constraints avoid
the turbines to be switched-on or off for just a few minutes, which would degrade
them faster than expected:
up
Tiup (t) ≥ Ti,min (3.8)
Tidown (t) ≥ Ti,min
down
(3.9)
up down the minimum down-time.
where Ti,min is the minimum up-time of turbine i and Ti,min
– The physically possible operation ranges of the turbine are bounded with their min-
imum and maximum limits:

Pitu,min ≤ Pitu (t) ≤ Pitu,max (3.10)

where Pitu is the power output of turbine i, and Pitu,min and Pitu,max are the minimum
and maximum values, respectively.
– While its starting cycle is not over, or if it is in standby mode, the turbine cannot
produce any output. Its operation range is thus restricted to a nil output.
– In normal operation mode, the positive and negative ramp rates (Ru , Rd > 0) are
also enforced:

Pitu (tk ) − Rd · (tk+1 − tk ) ≤ Pitu (tk+1 ) ≤ Pitu (tk ) + Ru · (tk+1 − tk ) (3.11)

– Finally, if the turbine is being stopped, its output is forced to decrease according
to (3.12):
Pitu (tk+1 ) = Pitu (tk ) − Rd · (tk+1 − tk ) (3.12)
60 Chapter 3. Gas Turbine Power Plants for Smart Power Generation

Based on these criteria, each turbine is capable of providing the dispatching system
with its operation range for the next period. In a real implementation of the EMS, the
turbine agent could also take decisions regarding the safety of the turbine, for example
by turning it off if a problem is detected. As a consequence, the turbine agent has the
highest priority among agents and can override requests by other agents if required, for
example to ensure the safety of the equipment. A degraded mode, used if communication
channels fail, would also be implemented at this level, and would base its decisions on
direct measurements.

3.5 Economic and Environmental Dispatch

After the energy management strategy has defined which component should be used
and how, the economic and environmental dispatch algorithm performs optimal control for
the power plant. From this strategy, an optimization problem is dynamically defined and
solved through a metaheuristic technique (see Appendix A) with one or more objectives.
As gas turbines use carbon-based products as fuels, their emissions contribute to a large
share of greenhouse gases resulting from power generation. Minimizing these emissions is
therefore one of the objectives of power plant operators, especially in a context of growing
environmental legislation. The DLN combustion system is an example of a technology
helping such minimization, while another leverage is to optimize the load dispatch while
taking into account emissions in addition to fuel costs.
However, these two objectives are sometimes conflicting: reducing fuel costs by resort-
ing to uncleaner options can increase emissions, and vice-versa, for the same total amount
of energy generated [137, 138]. For example, only optimizing costs would enable a turbine
to operate in Mode 2 during long periods (see Fig. 3.3), whereas also taking emissions into
account would favor exiting this zone to Mode 3 to reduce NOx emissions. The operator
is required to take a decision resulting from a trade-off between emissions and costs, as
no solution is entirely better than all others. This problem can therefore be seen as a
multi-objective optimization problem, where the first objective is to minimize operation
costs, resulting from fuel consumption, and the other is to minimize gas emissions.

3.5.1 Single and Multi-Objective Optimization

An optimization problem aims at minimizing a real function, called objective, cost,
utility or fitness function, by systematically choosing input values from a specific set and
evaluating the corresponding fitness. Mathematically, for a minimization problem, this
proposition is equivalent to searching for an element x0 in a search space A such that
f (x0 ) ≤ f (x) for all x ∈ A.
Another formulation is used for most numerical optimization problems. An optimiza-
tion problem with m objectives can be formulated as follows:

min [f1 (x), f2 (x), . . . , fm (x)]T (3.13)

x
s.t. gi (x) ≤ 0 i = 1, . . . , nineq (3.14)
hi (x) = 0 i = 1, . . . , neq (3.15)
xli ≤ xi ≤ xui i = 1, . . . , n (3.16)
3.5. Economic and Environmental Dispatch 61

where x is the vector of n decision variables, fi is the i-th objective function, gi is the
i-th of nineq inequality constraints, hi is the i-th of neq equality constraints, and xli and
xui are the lower and upper bounds for the i-th decision variable of x. If m = 1, then the
problem has a single objective, and if m > 1, then this is a multi-objective problem.
Solving single-objective problems returns a single solution, that is the best solution
found by the algorithm. However, for multi-objective problems, two main approaches can
be used, depending on the objectives and constraints of the problem: aggregating the
objective functions to create a composite objective function to minimize, or looking for a
set of equivalent solutions, called the Pareto front, from which the most appropriate can
be chosen.

[Link] Aggregation-Based Multi-Objective Optimization

Aggregation relies on combining the m objective functions fi to create a single com-

posite objective function Φ, as in (3.17). This function can then be minimized by a
single-objective optimization algorithm, and return a single solution. The main advan-
tage of the aggregation technique is its simplicity; however, tuning the weight αi to give
to each objective may be problematic, especially if the objectives correspond to different
quantities. For example, combining costs and emissions requires to give emissions a cost,
so that the total cost of the solution can be evaluated.
m
X
Φ(x) = (αi · fi (x)) (3.17)
i=1

[Link] Pareto Multi-Objective Optimization

The second approach is based on the search for a set of non-dominated (equivalent)
Pareto-efficient solutions, forming a m-dimensional surface called Pareto front obtained
by a multi-objective optimization algorithm. These solutions are not dominated by any
other solution for all objectives at the same time, and require an additional algorithm to
select the most appropriate solution that will be used.
Formally, a vector x∗ is Pareto-optimal (efficient) if, for all vectors x from the feasible
solutions set Ω, and all m objective functions f , both the following conditions are verified
for I = {1, . . . , m} [139]:

∀ i ∈ I : fi (x∗ ) ≤ fi (x) (3.18)

∃ j ∈ I : fj (x∗ ) < fj (x) (3.19)

The set of Pareto-optimal solutions P ∗ includes all found x∗ vectors belonging to Ω.

Consequently, the Pareto front F (Fig. 3.12) is defined as:

F = {f (x) ∈ Rm | x ∈ P ∗ } (3.20)

This Pareto dominance approach has the advantage of not requiring any experimental
tuning of parameters for running the optimization, as in the previous one, but as it provides
a set of solutions, another algorithm has to be used to select the most appropriate solution.
62 Chapter 3. Gas Turbine Power Plants for Smart Power Generation

f2

A C

f1

Figure 3.12: A Pareto front example. Here, C is dominated by A on f1 and B on f2 .

Contrary to A, B is better than A on f2 but not on f1 . A and B are not Pareto-dominated
by any other solution and thus belong to P ∗ .

[Link] Method Selection

For the EMS application, the aggregation method is selected for several reasons:
– It is simpler to implement;
– The simple-objective algorithms it requires are simpler to implement than multi-
objective ones;
– Aggregation does not require selecting a particular solution from a set of equivalent
solution, as such selection criteria are difficult to define. Such selection requirement
is also not well-suited for a dynamic problem.
As a consequence, only single-objective metaheuristics are considered in the following
sections.

3.5.2 Problem Definition

Equations (3.13) to (3.16) describe the general formulation of a multi-objective op-
timization problem. In the present case, vector x, i.e., the vector of decision variables,
includes a list of set points and powers corresponding to the turbines:
h iT
x = P0tu (t), . . . , Pntutu (t) (3.21)

where ntu is the number of turbines. Their respective power outputs Pitu use the same
indexes.
The first objective is the total fuel cost (3.22), which is obtained by computing the
sum of the operation costs of all connected components.
h ntu h
X ii
min ctot (t) = ctu,i (Pitu (t)) + ∆Pitu (t) (3.22)
i=0

where ctot (t) is the total generation cost, and ci are the cost functions using the same
indexes as in (3.21). ctu is defined by (3.3).
3.6. Simulation Results 63

The term ∆P represents the power output change for turbine i from the last set point,
where β0 is a real positive constant, and helps the algorithm avoid fluctuating around
optimal solutions over time:

∆Pitu (t) = β0 Pitu (t) − Pitu (t − 1) (3.23)

The second and third objectives correspond to the emissions of gas g (NOx and CO2 )
from the gas turbines as expressed in (3.24).
" ntu
#
X
min eg,tot (t) = eg,tu,i (Pitu (t)) (3.24)
i=0

where eg,tot is the total emissions amount, and etu,g is the emission function of the turbines
for this specific gas.
The only equality constraint is the required balance between the total generation from
the turbines and the total load PL (3.25).
ntu
X
Pimb (t) = Pitu (t) − PL (t) = 0 (3.25)
i=0

The inequalities for this problem are defined in section 3.4.2 and are summarized in
the optimization boundaries provided by the turbine agents.

3.5.3 Aggregation-Based Power Dispatching

As described earlier in section 3.5.1, the aggregation technique is selected for this
application, and the objective functions are therefore combined to create a single composite
objective function Φ. This function is then minimized by the single-objective optimization
algorithm MPSOM (see Appendix A for the description and comparison of MPSOM and
other metaheuristics). However, combining costs and emissions is not straightforward,
and tuning the weight to give to each objective may be problematic. A solution is to give
emissions an equivalent cost to add to fuel costs: the total cost is then to be minimized.
For example, CO2 emissions can be subject to a carbon tax [140], as well as NOx emissions.
This penalty method is also used to handle the power balance constraint: a high penalty
is added to prevent any imbalance. The composite function becomes:

Φ(t) =β1 · [γc · ctot (t) + γNOX · eNOX ,tot (t) + γCO2 · eCO2 ,tot (t)] + β2 · |Pimb (t)| (3.26)

where γc is a binary variable that controls whether costs are included in the optimiza-
tion or not, γNOX and γCO2 are respectively the costs of NOx and CO2 emissions, and β1
and β2 are positive real values.

3.6 Simulation Results

The performance of the proposed system is evaluated through several tests. Three
main aspects of the EMS are tested in simulations: its efficiency with respect to costs
and emissions, its flexibility, and its requirements in communication and computational
power.
64 Chapter 3. Gas Turbine Power Plants for Smart Power Generation

3.6.1 MAS Implementation

In order for the system to be tested, a simulator called MAEMS (Multi-Agent EMS)
is implemented in JADE. A graphical user interface (Fig. 3.13) is also developed, and
enables the user to define test cases and to monitor simulation results.

3.6.2 Parameters
Efficiency tests are run using a custom load profile shown in Fig. 3.14. A first load
curve, spanning the four first days of year 2012, is extracted from data available on the
website of the French transmission system operator RTE [7]. This profile corresponds
to the total load for France during four days, starting with two weekend days. Data is
scaled-down so that the maximum load is slightly lower than the maximum output of four
9E turbines (i.e., approximately 480 MW).
As one of the objectives of the EMS is to enable the operation of a SGP with a
medium to high penetration of renewable energy sources, a wind power profile is added.
This profile is based on a wind speed profile measured by UTBM’s weather station, also
during the first days of 2012. This profile is then scaled-up, so that the output of a large
wind farm with a peak output of about 150 MW is obtained, i.e., a 33 % penetration in
terms of rated power. It is assumed that the output of the wind farm can be smoothed,
e.g., with storage [141, 142], so that sudden variations do not have to be absorbed by the
power plant.
This wind farm output profile is then subtracted to the previous load profile in order
to obtain the final net load profile. A ramp is added at the beginning of the profile to
enable the turbines to start, as each simulation starts with all turbines turned off.
Table 3.2 lists the parameters used in the simulations. The following tests are run for
a power plant with four 9E series turbines, each with a different performance coefficient
ranging from 1.0 to 1.03, i.e., a 3 % change at maximum. Data related to taxes on
emissions were extracted from Norway’s legislation, one of the only countries in the world
to have such a system, and where taxes on NOx (16.34 NOK/kg – 2,149 e/tonne) and
CO2 (342 NOK/tonne – 44.46 e/tonne) emissions have been in use for years [143, 144].
This price for NOx can be considered as high compared to the 300 $/ton price currently
set for NOx emission allowances by the US FERC [145]. Set points are computed every
60 s throughout the test profile. Values of β0 , β1 and β2 are obtained empirically by the
authors.

3.6.3 Dispatching Algorithms Comparison

In this first series of tests, several dispatching algorithms are used and compared, while
the turbines are not enabled to start and stop. Results are summarized in Table 3.3. The
objective here is to analyze the behavior of the algorithms and the decisions they take, by
comparing the resulting costs and emissions. The tested algorithms are:
– Algorithm A – Sequential dispatch: This rule-base algorithm states that if the total
load is lower than the maximum output of one turbine, only one (selected randomly)
is used, otherwise this turbine is kept at its maximum output, and the others share
the rest of the load. This algorithm is currently used in some power plants, and
 3.6. Simulation Results
Figure 3.13: Screenshot of the developed GUI.

65
66 Chapter 3. Gas Turbine Power Plants for Smart Power Generation

8
x 10
5
Adapted RTE load [W]

4.5

3.5

2.5

2
0 0.5 1 1.5 2 2.5 3 3.5
5
x 10
7
x 10
15
Wind farm output [W]

10

0
0 0.5 1 1.5 2 2.5 3 3.5
5
x 10
8
x 10
5
Total net load [W]

1
0 0.5 1 1.5 2 2.5 3 3.5
Time [s] x 10
5

Figure 3.14: Total French load from RTE (top), synthetic wind farms output (middle),
and load profile used in the simulations (bottom).
3.6. Simulation Results 67

Parameter Value
αi {1.0, 1.01, 1.02, 1.03}
cfuel 0.3748 e/kg
γc 1
γNOx 2,149 e/tonne
γCO2 44.46 e/tonne
∆t 60 s
β0 0.5
β1 105
β2 1
niter 3,000
Table 3.2: Selected simulation parameters.

does not consider the individual performance of each turbine. Results show that
the performance of this algorithm is average for fuel costs; however, the lowest NOx
emissions are achieved. Fig. 3.15 shows that the turbines are in Modes 2 and 2
Extended during most of the simulation, except for turbine T3 that rapidly switches
to Mode 3, which enables drastically reducing NOx emissions.

7000

6000

5000
Duration [min]

4000

3000

2000

1000

0
1 2 3 4
Turbine
Mode 1 Mode 2 Mode 2 Extended Mode 3

Figure 3.15: Duration spent by each turbine in each mode during the simulation for
Algorithm A.

– Algorithm B – Equal dispatch: In this algorithm, the total load is dispatched equally
between the turbines, regardless of their respective performance. This algorithm
returns relatively low fuel costs, but also the highest NOx emissions because, as
shown in Fig. 3.16, all turbines switch to Mode 2 Extended where NOx emissions
are the highest.
– Algorithm C – Fuel costs optimization (DE with γNOx = γCO2 = 0): Fuel consump-
tion is the only objective, and as expected, the total fuel cost is reduced by 2.3 %
68 Chapter 3. Gas Turbine Power Plants for Smart Power Generation

7000

6000

5000
Duration [min]

4000

3000

2000

1000

0
1 2 3 4
Turbine
Mode 1 Mode 2 Mode 2 Extended Mode 3

Figure 3.16: Duration spent by each turbine in each mode during the simulation for
Algorithm B.

compared to Algorithm A, and is close to the cost returned by Algorithm B. If the

same profile was run continuously, savings of about 5.1 million e could be achieved
compared to Algorithm A over a year. In terms of operating modes (Fig. 3.17),
results for this algorithm are similar to the ones observed for Algorithm B, except
that the influence of the performance coefficients is observable: the higher the per-
formance of a turbine, the more it is used.

7000

6000

5000
Duration [min]

4000

3000

2000

1000

0
1 2 3 4
Turbine
Mode 1 Mode 2 Mode 2 Extended Mode 3

Figure 3.17: Duration spent by each turbine in each mode during the simulation for
Algorithm C.

– Algorithm D – NOx optimization (DE with γc = γCO2 = 0): NOx emissions are
3.6. Simulation Results 69

reduced by 34.8 % compared to Algorithm B, but fuel costs increase by 3.4 %. This
algorithm achieves lower NOx emissions by minimizing the use of turbine T3, which
spends more time in Modes 1 and 2 than the others (Fig. 3.18). However, contrary
to what could be expected, this algorithm does not return the lowest NOx emissions
and is outperformed by algorithm A. Additionally, this algorithm has the highest
costs.

7000

6000

5000
Duration [min]

4000

3000

2000

1000

0
1 2 3 4
Turbine
Mode 1 Mode 2 Mode 2 Extended Mode 3

Figure 3.18: Duration spent by each turbine in each mode during the simulation for
Algorithm D.

– Algorithm E – Fuel and emissions optimization (DE): The costs of fuel and of emis-
sions are combined to create a composite objective function, as in (3.26). Results
show that the obtained fuel costs are close to the ones obtained for Algorithms B and
C, while average NOx emission levels are obtained. Fig. 3.19 shows results similar
to the ones obtained for Algorithm C.

For all algorithms, the maximum total imbalance remains low, with less than 0.05 %,
which means that the balance constraint is verified. As expected, the lowest costs are
obtained for Algorithm C, but fuel costs for Algorithms B and E are also very close.
However, the counter-performance of Algorithm D on NOx emissions is more surpriz-
ing, although the total emission levels are close to the ones obtained by Algorithm A,
which has the best performance on this criterion. This result can be partially explained
by the fact that the algorithm, without the SSA, has only limited means to lower NOx
emissions, especially as it is not capable of taking into account load forecasts. The SSA
should however help improve these results.
Finally, Algorithm E manages to obtain rather low costs and average NOx emissions
at the same time; the impact of the high price for NOx emissions is clearly observable.
This last algorithm thus proposes an interesting compromise solution, taking both fuel
costs and emissions into account.
70 Chapter 3. Gas Turbine Power Plants for Smart Power Generation

7000

6000

5000
Duration [min]

4000

3000

2000

1000

0
1 2 3 4
Turbine
Mode 1 Mode 2 Mode 2 Extended Mode 3

Figure 3.19: Duration spent by each turbine in each mode during the simulation for
Algorithm E.

Algorithm Fuel costs (ke) NOx (kg) CO2 (ktonne)

(A) Sequential dispatch 3,069.7 14,452 20,379
(B) Equal dispatch 2,999.1 22,682 19,910
(C) Fuel costs optimization 2,998.8 22,240 19,908
(D) NOx optimization 3,099.8 14,799 20,579
(E) Cost and emissions optimization 3,000.4 19,150 19,919
Table 3.3: Results for each dispatching algorithm without the SSA.

3.6.4 Performance Coefficients Effectiveness

As the efficiency of turbines varies over time and with maintenance, performance co-
efficients were introduced in (3.2). Three elements are used to measure the impact of
these coefficients, using results for Algorithms B and C: (1) by comparing the total fuel
costs and emissions; (2) by comparing the total energy generated by each turbine to meet
the same demand; and (3) by comparing the time spent by each turbine in the first two
modes.
As shown earlier in Table 3.3, the coefficients enable reducing costs and emissions. In
addition to these results, Table 3.4 and Figs. 3.16 and 3.17 show how coefficients impact
the behavior of each turbine. For example, turbine T1, which has the best performance,
produces more energy (i.e., is more used) than the other ones, and the higher the coeffi-
cient, the less the turbine is used. As a consequence, the best performing turbines spend
less time in Modes 1 and 2.

3.6.5 Start and Stop Algorithm Effectiveness

In addition the optimal dispatch algorithm, the SSA is another way to further cut
costs and emissions. Moreover, as shown earlier with NOx emissions, the system needs to
3.6. Simulation Results 71

Metric Unit Algorithm T1 T2 T3 T4

Coefficient – – 1.0 1.01 1.02 1.03
B 6.8759 6.8759 6.8759 6.8759
Energy GWh
C 7.3153 7.0248 6.7207 6.4418
B 30 30 30 30
Mode 1 min
C 28 30 32 56
B 2,810 2,810 2,810 2,810
Mode 2 min
C 2,100 2,523 3,140 3,656
Table 3.4: Impact of the performance coefficients on dispatching results.

be guided because it does not have a knowledge of the full emissions curve of the turbines,
but only of a portion of it due to ramp rates. Consequently, the decisions of the algorithm
at time t depend on the decisions it took earlier. Additionally, the algorithm has no
memory, learning capability, or forecasting ability for periods longer than an hour. The
SSA is therefore expected to increase the time spent by each turbine in Mode 3, where
NOx emissions are the lowest.
In order to tune the parameters of the SSA so that it operates properly and efficiently,
several simulations were run for fuel costs optimization with various values of Rstart and
Rstop . The system is allowed to choose which turbines to turn off or on and when to do
it. Reserve values ranging from 0 to 30 % (with increments of 10 %) were selected and
tested. Results showed that any combination including 0 leads to unserved energy, as the
turbines do not have enough time to start. Also, choosing values of Rstart and Rstop such
that Rstart > Rstop leads to a chaotic behavior where turbines are started and stopped for
short periods of time. Results with the lowest costs are obtained for reserve values equal
to 10 %. This value is therefore selected for running the tests. Higher values result in
higher costs and emissions, as the algorithm is more conservative in the decisions it takes.
Simulations with Algorithms C, D and E are re-run with the SSA enabled and the
selected reserve values, and turbines T1 and T2 are turned-on by default. Results in
Table 3.5 show that the SSA can reduce fuel costs by 8.7 % for fuel costs optimization.
NOx emissions are dramatically cut by all algorithms, especially for Algorithm D (-21.9 %
compared to without the SSA), which now achieves the lowest NOx emissions, and for
Algorithm C (-28.6 %). Similarly to without the SSA, Algorithm E returns a fuel cost close
to the one obtained by Algorithm C, with moderately low NOx emissions. Compared to
Algorithm A, the SSA and the dispatching algorithms enable reducing fuel costs and CO2
emissions by 10.8 % (Algorithm C), and NOx emissions by 20.0 % (Algorithm D).

Algorithm Fuel costs (ke) NOx (kg) CO2 (ktonne)

(C) Fuel costs optimization 2,739.1 15,885 18,184
(D) NOx optimization 2,780.7 11,557 18,461
(E) Cost and emissions optimization 2,739.9 13,794 18,186
Table 3.5: Results for each dispatching algorithm with the SSA.

Figs. 3.20 to 3.22 provide additional details on the behavior of the algorithms with
respect to turbine operation modes. The duration spent by each turbine in each mode
72 Chapter 3. Gas Turbine Power Plants for Smart Power Generation

show that the SSA enables one to two turbines to operate in Mode 3, which was never
used earlier except for Algorithm A, and strongly impacts the amount of NOx emitted. At
the same time, turbines T3 and T4 are temporarily stopped when they are not required,
which enables achieving lower fuel costs and emissions.
These results also show that in a best case scenario (i.e., switching from the worst case
to the best), fuel costs can be reduced by as much as 360 ke, and NOx emissions by as
much as 10.6 metric tonnes, simply by using the SSA and the optimization algorithms,
and for the selected load profile.

7000

6000

5000
Duration [min]

4000

3000

2000

1000

0
1 2 3 4
Turbine
Mode 1 Mode 2 Mode 2 Extended Mode 3

Figure 3.20: Duration spent by each turbine in each mode during the simulation for
Algorithm C with the SSA enabled.

7000

6000

5000
Duration [min]

4000

3000

2000

1000

0
1 2 3 4
Turbine
Mode 1 Mode 2 Mode 2 Extended Mode 3

Figure 3.21: Duration spent by each turbine in each mode during the simulation for
Algorithm D with the SSA enabled.
3.6. Simulation Results 73

7000

6000

5000
Duration [min]

4000

3000

2000

1000

0
1 2 3 4
Turbine
Mode 1 Mode 2 Mode 2 Extended Mode 3

Figure 3.22: Duration spent by each turbine in each mode during the simulation for
Algorithm E with the SSA enabled.

3.6.6 Energy Costs Comparison

In order to compare the cost efficiency of the algorithms, the instantaneous cost per
MWh generated is plotted in Fig. 3.23. Results show that, as observed earlier with total
costs results, the optimization algorithms and the SSA enable significantly improving the
efficiency of the system. The peaks present in Fig. 3.23 for Algorithm C with SSA are the
result of turbines stopping, which mathematically increases the cost per MWh generated.

170
Algorithm A
160 Algorithm C without SSA
Algorithm C with SSA
150
MWh cost [EUR/MWh]

140

130

120

110

100

90
0.5 1 1.5 2 2.5 3 3.5
Time [s] x 10
5

Figure 3.23: Comparison of the MWh costs obtained by the algorithms.

74 Chapter 3. Gas Turbine Power Plants for Smart Power Generation

3.6.7 Flexibility and Resilience Test

The claimed flexibility and resilience of the system relies on its architecture, but needs
to be tested on an example. The outage of a turbine is simulated to analyze the behavior
of the algorithms. Fig. 3.24 shows that the EMS is capable of responding to the situation
quickly and to minimize its impact, by reducing as much as possible the unserved energy.
At mark 1, turbine T1 fails and is immediately disconnected. Turbine T2 increases its
output to its maximum value. At mark 2, right after the failure has been detected, turbine
T3 starts and increases its output as fast as possible to replace the power lost from T1.
The simulation then goes on as it would have without the failure, at mark 3. The system
has thus managed to keep operating despite the unexpected loss of a turbine.

8
x 10
1.5

1
1
Power [W]

0.5

2 3
0

−0.5

−1
2.5 3 3.5 4 4.5 5 5.5
Time [s] 4
x 10
T1 T2 T3 T4 Imbalance

Figure 3.24: Results of the flexibility test, based on [39].

In addition to this test, the ones run in the following section with various numbers of
turbines also show that the system can be adapted to different power plant sizes, and could
also include various types of turbines as long as their characteristics, costs and emission
curves are inputted in the system.

3.6.8 Communication and Computation Requirements

The last series of tests concerns the communication and computational power require-
ments of the system. Intuitively, the more agents there are, the higher their needs, but the
increase rate has to be empirically quantified. Simulations are run on a dedicated Sony
S12V9E laptop with a 2.40 GHz CPU.
From the results in Table 3.6 and Fig. 3.25, empirical expression giving the number of
messages exchanged and the optimization duration as a function of the number of turbines
are obtained. The mean number of messages nmess exchanged between the agents for one
round increases linearly with the number of turbines in the system nturb (3.27). The mean
3.7. Conclusion 75

optimization duration topt increases at least polynomially with nturb (3.28).

nmess = 5.0 nturb + 18 (3.27)

topt = 0.44 n2turb + 9.5 nturb + 39 (3.28)

These empirical results indicate that the requirements in communication and compu-
tation power are limited, and that managing up to a few dozens of turbines is possible.
However, controlling a large number of sources would increase these needs rapidly (es-
pecially as the number of iterations in the optimization would have to be increased). A
more decentralized control system would become more appropriate, although the algo-
rithms would perform less well than in the proposed centralized architecture.

nturb nmess topt (ms)

1 23.0 43.7
5 42.7 95.1
10 67.4 197.2
20 117.8 389.2
30 166.9 726.6
Table 3.6: Number of messages exchanged and optimization duration as a function of the
number of turbines to control.

800
Optimization duration [ms]

2
y = 0.44*x + 9.5*x + 39
600

400

200
Data
Fit
0
0 5 10 15 20 25 30
Number of turbines [−]

Figure 3.25: Plot of the mean optimization duration as a function of the number of turbines
to control.

3.7 Conclusion
This chapter has presented a flexible, resilient and efficient gas power plant EMS based
on a multi-agent architecture, which enables the EMS to adapt to a large variety of power
plant structures, as well as to their temporary or definitive evolutions, and facilitates
its transfer to the real system mostly by taking into account communication aspects by
defining how agents interact. Another advantage of the EMS is its ability to run with
several dispatching algorithms according to the objectives and constraints of the operator.
76 Chapter 3. Gas Turbine Power Plants for Smart Power Generation

Simulations based on actual data of GE 9E turbines have shown that the algorithms are
capable of reducing costs and NOx emission levels, and that further reductions can be
achieved using the SSA.
 4

Aggregator-Based Residential Demand Response

Contents
4.1 Demand Response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.1.1 Drivers and Components . . . . . . . . . . . . . . . . . . . . . . 86
4.1.2 State-of-the-Art . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4.1.3 Proposed Approach . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.2 System Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.2.1 T&D Infrastructure . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.2.2 Demand Response Aggregators . . . . . . . . . . . . . . . . . . 91
4.2.3 Demand Response Market . . . . . . . . . . . . . . . . . . . . . 91
4.2.4 Residential Customers . . . . . . . . . . . . . . . . . . . . . . . 92
4.3 Residential Load Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.3.1 Enabling Technologies and Assumptions . . . . . . . . . . . . . 93
4.3.2 Electric Water Heater Model . . . . . . . . . . . . . . . . . . . 93
4.3.3 Air Conditioning Model . . . . . . . . . . . . . . . . . . . . . . 95
4.3.4 Appliances Model . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.3.5 PHEV Fleet Model . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.3.6 PV Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.4 System Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.4.1 Objective and Constraints . . . . . . . . . . . . . . . . . . . . . 101
4.4.2 Metering Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
4.4.3 Demand Response Event Mode . . . . . . . . . . . . . . . . . . 102
4.4.4 Rescheduling Algorithm . . . . . . . . . . . . . . . . . . . . . . 104
4.5 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.5.1 Simulator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.5.2 Test Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
4.5.3 Simulation Parameters . . . . . . . . . . . . . . . . . . . . . . . 108
4.5.4 System-wide Net Load Results . . . . . . . . . . . . . . . . . . 110
4.5.5 Results for Residential Customers . . . . . . . . . . . . . . . . . 113
4.5.6 Impact on the Distribution System . . . . . . . . . . . . . . . . 113
4.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

77
78 Chapter 4. Aggregator-Based Residential Demand Response

4.1 Demand Response

4.1.1 Drivers and Components
Electricity demand for the residential sector in the US is expected to grow by 18 % from
2010 to 2035, not only due to population growth but also due to changes in consumption
patterns and habits [146]. The deployment of EVs transfers a part of the total energy con-
sumption from the transportation infrastructure to the electricity infrastructure, and may
cause increases in peak load on distribution systems [147]. In order to meet these demand
peaks, utilities usually resort to increasing generation capacity and distribution assets,
e.g., by building new power plants, and installing larger transformers, transmission lines,
and distribution cables. Such changes require expensive investments, and the operational
costs of peaking power plants (gas or coal-based) are also generally more expensive than
their base load counterparts [148]. Instead of trying to increase generation, transmission,
and distribution capacity, other approaches, labelled as demand-side management (DSM)
programs, try to mitigate demand in various ways (Fig. 4.1).
A first solution is to try to permanently reduce the energy consumption by resorting
to energy efficiency solutions. The main objective of energy efficiency is to reduce the
energy consumption for a same or better level of service for the consumer. This can be
achieved through various means. An example is the improvement of housing insulation
to reduce the energy consumed for space heating or cooling, through the use of better
materials. Another common example is the use of compact fluorescent lights instead of
traditional incandescent light bulbs — the latter consume more energy for a similar level
of illumination than the former [149]. Although such approaches can result in potentially
significant energy savings [150], the required capital investments are often prohibitive and
tend to slow down the adoption of such technologies.
Another category of solutions is known as demand response (DR). The programs using
this concept aim at reducing demand peaks by shifting or shedding loads directly or
indirectly, in response to supply conditions. This concept has been in use in Europe and
in the US under various forms since the 1980s [151], and has been gaining momentum
again over the last few years. A study by the US FERC has shown that, in a best case
scenario, a reduction of up to 20 % in peak demand could be achieved by 2019 using
DR programs [152]. As supply costs can exceed 1000 $/MWh during extreme demand
peaks (while costs typically remain under 50 $/MWh most the time), DR could provide a
solution to drastically reduce electricity costs during such peaks. Studies have shown that
a 5 % reduction in peak demand during the California energy crisis of 2000–2001 could
have reduced highest prices by 50 % [153]. In addition to 40 % lower capital investment
costs compared to building new power plants, DR can also be deployed as fast as in a few
minutes [36, 153], which enables DR to participate in day-ahead dispatch and ancillary
services. Examples of existing programs and benefits can be found in [154] and [155].
Two main categories of DR programs can be distinguished [157] (Fig. 4.1). A first
category of solutions corresponds to non-dispatchable (non-event-based) programs based
on custom pricing schemes, that rely on the assumption that customers tend to reduce
their consumption when prices increase. Examples include:
– Time-of-use (TOU) schedules, in which the price paid for a kWh of electricity de-
pends on the time window when it is consumed, and tends to be higher when demand
 4.1. Demand Response
Demand-side management

Demand response Energy efficiency

Dispatchable Non-dispatchable

Reliability-based Economics-based Time-based pricing:

– Time-of-use (TOU)
– Critical peak pricing (CPP)
– Real-time pricing (RTP)
Capacity Ancillary services – Demand bidding
– Direct load control

– Direct load control – Spinning reserves

– Interruptible demand – Non-spinning reserves
– Regulation

Figure 4.1: Main components of demand-side management, based on [156].

79
80 Chapter 4. Aggregator-Based Residential Demand Response

increases. On and off-peak periods are typically defined within each day, while the
on-peak period may be different in winter and summer.
– Real-time pricing (RTP) schedules, in which prices vary in real-time with supply
and demand, in order to better reflect generation costs.
– Critical peak pricing schedules, in which prices are higher during peaks, and espe-
cially when the peak load is much higher than the average load. This scheme may
be combined with TOU schedules.
Another category of DR programs corresponds to dispatchable (event-based) DR sys-
tems, which use dedicated control systems to control loads in response to a signal sent
by the utility to improve or maintain the reliability of the system, or to market price
conditions. Two main types of dispatchable DR are used:
– Interruptible demand corresponds to loads that can be reduced or stopped, usually
manually, with little or no notice. These programs are mostly suited for entities that
may need low energy prices to be profitable, i.e., mainly industrial and commercial
customers.
– Direct load control (DLC) programs are able to directly control specific loads, such
as air conditioning and electric water heaters. These programs are best suited for
residential and commercial customers, and are therefore favored for these applica-
tions.
Dispatchable programs are considered to allow a high degree of certainty compared to
non-dispatchable programs, as they are capable of predicting the amount of load reduction
with a higher reliability, and are also capable of reaching their goal faster than their price-
based counterparts [156]. However, as such programs are event-based, they cannot be used
more than a few times, perhaps a dozen times a year, which makes them only suitable to
mitigate the largest peaks for short periods of time.

4.1.2 State-of-the-Art
Additionally, although such DR systems have been used for industrial and commercial
customers for years, most DR programs offered by utilities for residential customers rely
either on pricing schemes or on basic DLC, using only one type of load (typically air condi-
tioning or water heaters), without taking into account the potential of newer technologies
that are penetrating the distribution grid such as electric vehicles.
Existing research work has focused on the participation in DR of: PHEVs [158, 159];
thermal loads such as air conditioning (AC) and electric water heaters (EWHs) [160–162];
and other appliances [163]. Others have looked at incentives for DR participation [164] and
at control architectures [165]. Thomas et. al [166] studied the impact of price-responsive
residential demand on power market operations, but do not consider the need to achieve
a given load reduction with a reasonable degree of confidence. Also, the focus is set on
the impact of DR on markets, but not on how residential loads are used for DR.
As 38 % of the US total demand is from residential customers [167], this essentially
untapped potential provides opportunities for new residential DR systems [168]. Exper-
iments have shown that load reductions of up to 1.5 kW per residential customer can be
reached [154], which does not enable them to participate in power markets. Aggregators
4.2. System Architecture 81

are thus required, and are able to bid on markets using the aggregated capacity from their
customers [169]. Depending on the local market structure, utilities themselves can serve
as aggregators, which can also be independent commercial entities.
To the knowledge of the authors, no work in the literature examines how residential
customers with diverse assets could participate in large-scale DR through aggregators, and
with the ability to achieve a given load reduction with a reasonable degree of confidence.
Communication aspects are also rarely considered.

4.1.3 Proposed Approach

This chapter addresses this concern and proposes an architecture for a DR system
targeted at residential customers. Independent aggregators serve as interfaces between
end-users and a DR market, where buyers and sellers of DR services can meet. Participat-
ing customers can have their PHEV, AC unit and EWH use controlled by the aggregator
though a DLC program. This strategy is preferred to other solutions to enable the system
operator to have a high level of confidence on the load reduction that is achieved by the
DR system. It is also assumed that higher participation levels could be reached using such
method, and may be better accepted by consumers than price-based programs.
Section 4.2 describes the architecture of the system and the roles of each element. In
section 4.3, the models used for residential customers are presented. Section 4.4 details how
the system operates using these architecture and models, by specifying how the various
elements interact with each other to reach the DR goal. Finally, section 4.5 presents an
analysis of the simulation results at both the system and customer level. Commercial and
financial aspects for aggregators and customers are not discussed.

4.2 System Architecture

The architecture of the proposed system, as shown in Fig. 4.2, contains several el-
ements: a T&D infrastructure with physical assets as well as market entities, such as
aggregators, a DR marketplace, and residential customers. Commercial and industrial
customers are not considered in this study.

4.2.1 T&D Infrastructure

The T&D infrastructure used in the proposed system reproduces a simplified version of
the infrastructure found in North American power systems. This infrastructure contains
physical and market assets. Physical assets include residential customers, load points,
feeders, and substations, ordered by increasing voltage level. Customers are connected to
the distribution system through load points and feeders, which are in turn connected to
substations, themselves connected to the transmission system. In this study, substations
are “transparent”, in that their role is not modeled or considered. Feeders and load points
are only considered in the power system simulation described in section 4.5, and are not
equipped any autonomous intelligence.
Two other entities are also included: DSOs and ISOs. DSOs are in charge of operating
and maintaining the distribution infrastructure, while ISOs are non-profit entities in charge
of maintaining the balance between supply and demand in the grid in the selected area,
82 Chapter 4. Aggregator-Based Residential Demand Response

DRX ISO

DSO 1 DSO 2

Substation 1 Substation 2

Aggregator 1 Aggregator 2

Feeder 1 Feeder 2

Load point 1 Load point 2

Customer 1 Customer 2 Customer 3

Figure 4.2: Architecture of the DR system. Short-dashed lines indicate that multiple
avatars of entities are not represented, e.g., only three customers are shown while there
are many. Dashed-dotted lines indicate communication channels, while plain lines indicate
power and communication flows.

sometimes encompassing multiple US states, by providing non-discriminatory access to the

transmission infrastructure and by facilitating competition among wholesale suppliers.
Most ISO operations are regulated by the FERC. Based on load forecasts and on the
knowledge of the available generation capacity, ISOs are capable of determining whether
the stability of the regional grid is at risk, and are in charge of finding proper solutions, e.g.,
through the intermediary of markets. In case a transmission congestion or a temporary
insufficient generation capacity or available transmission capacity is detected, the ISO
may need to leverage available DR capacity to avoid making the grid unstable.

4.2.2 Demand Response Aggregators

Although large customers can sign up for DR programs directly with utilities, it is
often simpler for them to outsource this activity to another entity, due to the regulatory,
administrative and technical difficulties. Utilities may also resort to such entities for the
same reasons. Aggregators are energy service providers that are capable of providing DR
services. Companies such as EnerNOC and Comverge in the US [170] operate on this
business model, and provide DR capacity to utilities, either directly or through markets.
Contrary to large customers, smaller customers (e.g., residential customers) do generally
4.2. System Architecture 83

not qualify for direct DR programs with utilities, which are primarily targeted at customers
with loads of at least a few MWs. For the same reason, smaller customers are also not able
to participate in markets. A residential DR aggregator is therefore proposed to enable
residential customers to participate in DR programs and markets through the intermediary
of aggregators.
Each customer can have a specific contract with an aggregator, and make some of its
loads available for curtailment or DLC under given conditions. By aggregating the capacity
of thousands of small customers, aggregators are capable to participate in power markets
where minimum capacities are usually of several MWs, something only large industrial
customers could achieve on their own. For end users, the main benefit is the possibility
to save on electricity bills, estimated to about $40 per year per residential customers for a
few DR events a year [153]. Several business model for aggregators are possible, and are
not developed in this dissertation. However, it is assumed that several aggregators use
different business models so that they are able to compete for residential DR capacity.

4.2.3 Demand Response Market

A DR market, or DR exchange (DRX), is implemented in the system and serves as a
marketplace where DR is treated as a public good, and sellers and buyers of DR capacity
meet. The DRX presented here is inspired by Nguyen et al.’s work [171], in which a
pool-based DRX is proposed. Such market enables multiple players to benefit from DR,
including ISOs, DSOs, retailers and aggregators. ISOs, and DSOs may for example need
to buy DR capacity to improve the reliability of their system, while aggregators have
some capacity to offer. Several applications of DR, such as transmission and distribution
congestion management or peak shaving, may be achieved using the aggregator and DRX
model.
A DRX operator collects the bids and offers from market players (buyers and sellers
of DR capacity). The market is then cleared a day ahead according to specific objectives
and constraints. In the present case, the players place bids for capacity, and aggregators
submit offers based on the capacity they have at their disposal, the bid price, and the
profit margin they intend to achieve. This enables multiple aggregators with different
business models to compete for DR markets and customers. As market aspects are not
developed in this work, it is also assumed that only a single ISO is capable of buying DR
capacity.

4.2.4 Residential Customers

As the proposed system focuses on residential DR, other types of customers such as
industrial, commercial and offices, are not considered. Residential customers are assumed
to have typical loads (Fig. 4.3), such as air conditioning, electric water heating and other
appliances listed in section 4.3.4. Some of the customers may also have DGs such as PV
panels, and/or a PHEV. The V2G capability of PHEVs is not considered. The models
used for these loads and sources are presented in the following section. The residential
building is assumed to be equipped with a smart meter and an autonomous HEMS.
84 Chapter 4. Aggregator-Based Residential Demand Response

Load point

Residential building
Smart meter

Aggregator HEMS User

DG Appliances Thermal PHEV

loads

Figure 4.3: Diagram representing the interactions between residential loads, assets, smart
meter, HEMS and the rest of the system.

4.3 Residential Load Model

Probabilistic residential load models rely on several methodologies based on statistical
usage data and are used to generate synthetic load curves, corresponding to thermal loads
such as air conditioning and electric water heaters, typical appliances, PHEVs and PVs.
This bottom-up approach enables achieving DR operation with high granularity, where a
specific load or asset from a given customer can be controlled separately from the rest.

4.3.1 Enabling Technologies and Assumptions

For the proposed system to operate properly, this paper assumes that several tech-
nologies are available and adopted by all considered residential customers, and are used
according to the following principles:

– Each residence is assumed to be equipped with a smart meter owned by the utility.
It is also assumed that the smart meter is only used for metering purposes.

– An intelligent HEMS capable of forecasting the load of the customer using weather
and historical data (especially for thermal loads) [172], as well as schedule the
recharge of a PHEV (if present) using user-supplied information [173], is also as-
sumed to be present at each residence. In addition, the HEMS could also serve as an
advanced home automation device, and would communicate with the smart meter.

– The customers who have a PHEV are assumed to have a charger with a rating of
1.8 kW (120 V / 15 A) or 7.2 kW (240 V / 30 A) [174,175]. The former does not require
any special modification in the power installation of the user, while the second one
does and would therefore be more expensive, although it would reduce the charging
duration.
4.3. Residential Load Model 85

– The customers who have a PV installation on their roof are assumed to sell all the
energy they produce to the grid, without storing it.
– Selected loads such as air conditioning (AC) units and electric water heaters (EWH)
are assumed to be directly controllable by the HEMS, using signals sent by aggre-
gators. These loads are all assumed to be electric.

4.3.2 Electric Water Heater Model

The selected EWH model is based on the model proposed in [176]. The water temper-
ature is given by (4.1), where Tw is the water temperature, Ti is the indoor temperature,
Cw is the tank thermal capacity, Rw is the thermal resistance of tank walls, Kw is the
status of the EWH (1 for on and 0 for off), Pw is the maximum power output, cp is the
specific heat constant for water, q is the hot water flow, Tref,w is the desired temperature
set by the user, and Tcold is the temperature of the inlet water. Parameter values are
given in Table 4.1, and are distributed according to a normal distribution with standard
deviations arbitrarily chosen as 10 % of the mean value.
dTw (t) 1
Cw =− (Tw (t) − Ti (t)) + Kw · Pw − cp · q · (Tref,w − Tcold ) (4.1)
dt Rw
The hot water flow is obtained assuming a daily consumption of 266 l with 17 events of
15 minutes a day [176], each with the same amount of hot water consumed, and following
a linear event distribution from 6am to 12am. Solving (4.1) for Tw yields the current
temperature, which is then used as an input for the on-off controller of the EWH. This
thermostat from [176] is given by (4.2), where ∆Tw is the deadband value (Fig. 4.4).
(
0 → 1, if Tw (t − 1) ≤ Tref,w
Kw (t) = (4.2)
1 → 0, if Tw (t − 1) ≥ Tref,w + ∆Tw

Thermostat
status

On

Off

Temperature

Tref,w Tref,w + ∆Tw

Figure 4.4: Operation principle of the EWH thermostat.

86 Chapter 4. Aggregator-Based Residential Demand Response

Variable Mean value

Cw 0.351 kWh/◦ C
Rw 370 ◦ C/kW
Pw 4.50 kW
cp 0.00117 kWh/(◦ C·l)
Tref,w 60.0 ◦ C
Tcold 15.0 ◦ C
∆Tw 5.00 ◦ C
Table 4.1: EWH model parameters, with data from [176].

4.3.3 Air Conditioning Model

The AC model is based on a household temperature model given in (4.3) and proposed
in [177], where Ti is the inside temperature, To is the outside temperature,  is the system
inertia, η is the coefficient of performance (COP) of the AC system, Kac is its status (1
for cooling, -1 for heating, and 0 for off), Pac is its maximum power output, and A is the
thermal conductivity of the building. Parameter values from [177] are given in Table 4.2,
and are distributed according to a normal distribution with standard deviations arbitrarily
chosen equal to 10 % of the mean value. As parameters are given in ◦ F, the temperature
is converted to ◦ C at the end of the simulation.
 
Pac
Ti (t) =  · Ti (t − 1) + (1 − ) To (t) − η · Kac (t) · (4.3)
A

An hysteresis controller is adapted from (4.2), and uses the temperature model in (4.3)
and user settings to decide when to turn-on or off the AC, as in (4.4), where Tref,i is
the optimal comfort temperature set by the user and ∆Ti is the deadband temperature
(Fig. 4.5). The AC is assumed to operate for cooling or heating, depending on the outdoor
temperature.


0 → −1, if Ti (t − 1) < Tref,i − ∆Ti [Heating]


−1 → 0, if Ti (t − 1) ≥ Tref,i
Kac (t) = (4.4)
0 → 1,

 if Ti (t − 1) > Tref,i + ∆Ti [Cooling]


1 → 0, if Ti (t − 1) ≤ Tref,i

Variable Mean value

 0.93
η 2.5
Pac 3.5 kW
A 0.14 kW/◦ F
Tref,i 70.0 ◦ F
∆Ti 2.00 ◦ F
Table 4.2: AC model parameters, with data from [177].
4.3. Residential Load Model 87

Thermostat
status

Heating

Off

Cooling

Temperature

Tref,i − ∆Ti Tref,i Tref,i + ∆Ti

Figure 4.5: Operation principle of the AC thermostat.

4.3.4 Appliances Model

The probabilistic load curve model proposed by Dickert and Schegner in [178] is used to
generate synthetic load profiles for appliances. Appliances include kitchen, laundry and
household appliances, consumer goods (TV, computer, etc.) and lighting. A four step
methodology adapted from [178] is used to generate the load profile of each appliance:
1. At first, the statistical penetration level for the appliance is used to generate whether
or not the appliance is present in the household.
2. If it is present, the frequency of use, turn-on time and operation duration are gen-
erated according to random distributions (linear or normal) that depend on the
appliance.
3. Similarly, the power consumption of the appliance is generated randomly according
to a normal distribution.
4. Using the parameters obtained in the two previous steps, the load profile of the
appliance is generated.
Parameters for appliances use are partially derived from [178] and are given in Ta-
ble 4.3. Parameters not given in [178] were arbitrarily selected by the authors from their
own use habits. For each daily occurrence of each appliance, a start time is obtained from
one of the distributions listed in the start time column of Table 4.3. After the appliance
has started, its power consumption is considered constant for a duration given in 15 min
blocks.
Figure 4.6 shows the load curve obtained for a single customer over a day. The load
curve includes not only appliances but also the load for the AC and the EWH (PHEVs
88 Chapter 4. Aggregator-Based Residential Demand Response

and PVs are not included). Figure 4.7, on the other hand, shows the aggregated (i.e.,
summed) load for 1000 customers, and is obtained using the same methodology as for a
single customer. This aggregation of load assumes the same usage characteristics for all
customers. The aggregated load curve is much smoother than the load curve of a single
customer, and is similar to load curves generally observed on most distribution systems: a
peak is observed in the late afternoon, between 4pm and 8pm; the load is lower at night;
and increases again in the morning till late afternoon.

4
Power [kW]

0
0 5 10 15 20
Time [h]

Figure 4.6: Load profile example for one customer.

4500

4000
Power [kW]

3500

3000

2500

2000
0 5 10 15 20
Time [h]

Figure 4.7: Aggregated load profile example for 1000 customers.

4.3.5 PHEV Fleet Model

The PHEV fleet generation methodology is based on the one presented in [179], and
adapted and used in [174] and [175], that gives a realistic model of a PHEV fleet charac-
teristics and usage data. PHEVs are identified as one of the means for deploying and inte-
grating advanced electricity storage and peak-shaving technologies. Additionally, contrary
to battery electric vehicles (BEVs), PHEVs have high driving ranges and can be used even
when the battery is discharged. Due to this property and to range anxiety [180], PHEVs
 4.3. Residential Load Model
Appliance Penetr. [%] Rating [kW] Occur. Start time [h] Duration [15 min blocks]
Electric cooker 100 N(3.00,1.00) 2 N(12,1), N(18,1) 3
Water kettle 95 N(2.00,0.20) 2 N(8,1), L(14-18) 1
Dishwasher 95 N(1.60,0.10) 1 N(10,2), N(14,2), N(20,2) 4
Coffee maker 80 N(1.00,0.20) 2 N(7,1), L((14-18) 2
Microwave 85 N(1.00,0.10) 3 N(8,1), N(12,2), N(18,2) 1
Toaster 92 N(1.00,0.10) 1 N(9,1) 2
Fridge-freezer 100 N(0.13,0.03) – – Continuous
Freezer 45 N(0.10,0.03) – – Continuous
Other kitchen appl. 95 N(0.50,0.10) 3 N(8,1), N(13,2), N(19,2) 1
Washing machine 99 N(1.60,0.20) 1 L(8-20) 4
Laundry dryer 80 N(2.50,0.25) 1 L(10-22)) 5
Vacuum cleaner 97 N(1.50,0.50) 1 L(10-20)) 2
Hair dryer 80 N(2.00,0.10) 2 N(8,1), N(18,2) 1
Circulation pump 100 N(0.05,0.01) – – Continuous
TV 1 and 2 99 and 80 N(0.15,0.03) 2 N(13,5), N(19,2)) and N(13,5), N(21,2) 16 and 8
PC 1 and 2 84 and 40 N(0.15,0.03) 1 N(14,7) 16 and 8
Hi-fi system 81 N(0.03,0.01) 1 N(19,3) 4
Other goods 100 N(0.05,0.03) – – Continuous
Lighting 100 N(0.20,0.05) 2 N(7,1), N(19,3) 12
Table 4.3: Appliances model parameters, partially based on [178]. N(A,B) indicates that the value is obtained from a normal distribution
with mean A and standard deviation B, and L(A-B), that the value is obtained from a linear distribution between hours A and B.

89
90 Chapter 4. Aggregator-Based Residential Demand Response

are expected to penetrate the market faster than BEVs, and are therefore considered in
this study.
The selected methodology is separated in four main parts. In the first part, the pen-
etration rate of PHEVs pphev is used to determine whether the PHEV is present at the
considered household. A random number r is generated from the uniform distribution,
such that r ∼ U (0, 1). If r ≥ pphev , then it is assumed that one PHEV is present is the
household. Otherwise, the household is assumed to possess no PHEV, and only conven-
tional internal combustion engine (ICE) vehicles.
If a PHEV is present in a household, the characteristics of the vehicle are generated
according to the following procedure:
1. At first, the class of each PHEV is selected randomly from four vehicle classes, which
are arbitrarily defined to provide a diverse representation of a future US vehicle
fleet [174, 179]. The following vehicles are used as inspiration for each class: Honda
Civic and Ford Taurus for class 1 (compacts), Honda Accord and Ford Taurus for
class 2 (sedans), Ford Explorer and Ford F-150 for class 3 (medium sport utility
vehicles (SUVs)), and Chevrolet Suburban and Chevrolet Silverado for class 4 (large
SUVs).
2. Two PHEV control strategies are distinguished. In charge-depleting mode, the
PHEV only uses power drawn from the battery for driving. When the charge de-
pleting distance Ddep is reached, i.e., when the state-of-charge (SOC) of the battery
reaches its lower limit, the ICE is switched on. In charge-sustaining mode, the ICE is
used to maintain the SOC of the battery around a given average value. The variable
k ∈ [0, 1] represents the share of electric driving power for each vehicle, and is equal
to 0 for an ICE vehicle and to 1 for a full electric one.
3. For each PHEV class, the used battery capacity Bcap and a value for k are selected
according to a specific method that uses the ranges given in Table 4.4 for each PHEV
class.

Class Share [%] Bcap range [kWh] k range [-] ae [kWh/mi] be [-]
Class 1 20 [8,12] [0.2447,0.5976] 0.3790 0.4541
Class 2 30 [10,14] [0.2750,0.6151] 0.4288 0.4179
Class 3 30 [17,21] [0.3217,0.5428] 0.6720 0.4040
Class 4 20 [19,23] [0.3224,0.4800] 0.8180 0.4802
Table 4.4: Main characteristics of the four PHEV classes, based on data from [179].

Bcap and k are assumed to be distributed according to a bivariate normal distri-

bution, with a correlation parameter ρ of 0.8 as suggested in [179]. This type of
distribution enables the possibility to have a correlation between Bcap and k, which
should intuitively vary similarly. The parameters of the distribution are defined as
follows, for each class. At first, the mean vector µ is obtained using (4.5).
 
kmin + kmax

µ= B 2 
(4.5)
cap,min + Bcap,max

2
4.3. Residential Load Model 91

Then, the covariance matrix Σ is obtained using (4.6) to (4.8).

kmax − kmin
σk = (4.6)
4

Bcap,max − Bcap,min
σB = (4.7)
4
 
σk2 ρ · σk · σB
Σ= 2 (4.8)
ρ · σk · σB σB
Using the Cholesky decomposition, the covariance matrix is decomposed into a lower
triangular matrix S, such that Σ = S · ST . Then, a vector of two standard normal
values N is generated using the Box-Müller method, where Nbm is a standard normal
value, and r1 and r2 are distributed from a uniform distribution U (0, 1):
p
Nbm = −2 · ln(r1 ) · cos(2 π · r2 ) (4.9)

This vector N is then used to obtain the desired multivariate normal distribution:
 
k
=µ+S·N (4.10)
Bcap

4. The required energy per mile driven Emil is a performance metric of the vehicle, and
is generated using (4.11) where ae and be are given in Table 4.4 for each class.

Emil = ae · kbe (4.11)

5. The charge-depleting distance Ddep is then obtained with (4.12).

Bcap
Ddep = (4.12)
Emil
In the third part of the methodology, the daily usage characteristics of the PHEVs are
obtained. These characteristics include the daily driven distance Ddri , the energy required
to fully recharge the battery of the PHEV Erec , and the departure and arrival time from
and to the residence.
1. Ddri is derived from a log normal distribution, with mean µm and standard deviation
σm , as in (4.13) where r is a standard normal random number. In this case, µm and
σm are assumed to be equal to 3.37 and 0.5, respectively, as in [179].

Ddri = exp (µm + σm · r) (4.13)

2. The daily energy to recharge the PHEV required from the grid Erec is equal to
the battery capacity if the driven distance is equal to or larger than the depleting
distance, or to the consumed driving energy if not, as in (4.14).
(
Bcap if Ddri ≥ Ddep
Erec = (4.14)
Ddri · Emil otherwise.
92 Chapter 4. Aggregator-Based Residential Demand Response

3. The departure and arrival times tdep and tarr , respectively, are obtained using (4.15)
and (4.16), where N1 and N2 are standard normal random numbers obtained with (4.9),
µdep and µarr are the departure and arrival mean times, and σdep and σarr are the
corresponding standard deviations. Values for these parameters from [179] are given
in Table 4.5 for a typical weekday.

tdep = µdep + σdep · N1 (4.15)

tarr = µarr + σarr · N2 (4.16)

Parameter Departure Arrival

Mean 7.0 18.0
Standard deviation 1.732 1.732
Table 4.5: Departure and arrival time distribution parameters of PHEVs for a typical
weekday, based on data from [179].

In the last part of the methodology, the battery charger present in the household is
selected and determines the charging rate. As in [179], the rating of the charger is obtained
from a random distribution, where each rating (1.8 or 7.2 kW) has a 50 % probability to
be picked.
From these results, the initial charging schedule of the PHEV, i.e., its load curve, can
be established, as in Fig. 4.19. The initial charging schedule is based on a simple strategy
in which the vehicle starts charging as fast as possible as soon as it arrives at the residence,
and until the battery is fully charged. In the following, the charging efficiency is assumed
to be equal to 0.88, as in [181].

4.3.6 PV Model
The output of the photovoltaic panels is obtained using a methodology similar to the
ones employed for the appliances and the PHEVs: if a PV is present at the considered
household, its rating Pr is obtained from a normal distribution with mean 5.9 kW and a
standard deviation of 3 kW [182]. The output of the system is simply obtained using an
irradiation profile E given in W/m2 as input, as in (4.17), which is then normalized with
respect to the “one sun” (1000 W/m2 ) method [183].

E(t)
PP V (t) = · Pr (4.17)
1000

4.4 System Operation

4.4.1 Objective and Constraints
The objective of the proposed system is to shape the load so that it remains under a
given threshold for a given duration, typically a few hours, using curtailable and shiftable
capacity made available by residential customers through aggregators. In the following,
the term event is used to refer to the time during which the proposed DR system operates
4.4. System Operation 93

to maintain the load under the threshold. The selected constraints set for system operation
require that, after the DR system has been run:
– Thermal loads (AC and EWH) can be controlled by the aggregator as long as critical
temperatures are not reached.
– PHEVs should be fully charged by the time the user has specified.
– Users can manually override settings until one hour before the event begins. This
duration is arbitrarily selected so that aggregators have enough time to adapt their
strategy to meet the total bid capacity, based on the number of customer overrides.
The proposed system can operate in two modes: for metering, and for solving a DR event.

4.4.2 Metering Mode

The first mode is used for basic metering purposes, i.e., for obtaining the forecast and
actual load in each location of the system. In this mode, the T&D infrastructure described
earlier is used to obtain the load at each node of the system. A request is issued by the
ISO and is transmitted in the hierarchy down to the end customers. Each customer then
measures its net load and sends it to the element that originally sent the request, i.e., the
feeder. The feeders aggregate the answers of the customers that are connected to it, and
send the results to the substation they are connected. The same process is repeated along
the infrastructure hierarchy. At the end of this process, the net load in each node of the
system is known. This mode is used primarily to obtain the actual net load in the system
after the DR action has been taken. It is also used to obtain a baseline load curve before
the event; in reality, this would be achieved using load forecasting algorithms.

4.4.3 Demand Response Event Mode

The second mode is more elaborate and uses the elements listed earlier to implement
a DR action. In this mode, the elements of the system interact and cooperate to solve
a DR scenario such as a congestion [166]. Fig. 4.8 shows the main chronological steps
of a DR event, seen from the point-of-view of a residential customer: the customer is
noticed in advance that a DR event is going to happen, and receives information about its
characteristics, e.g., start time, duration, etc.; then the DR is deployed, until the end of
the planned event when the loads affected by DR are released. The recovery period may
last a few hours, depending on the strategy used by the aggregators, and after this period
is over, the system is back to normal operation.
DR event Recovery period

Time

Advance Deployment Release Normal

notice operation

Figure 4.8: Main chronological steps of a DR event, from the point-of-view of a customer.
94 Chapter 4. Aggregator-Based Residential Demand Response

However, many more steps have to be followed before the DR capacity is actually
deployed (Fig. 4.9). These steps follow a process that starts a day before the DR event:
1. The area ISO detects a potential transmission congestion issue for the next day
using load and generation forecasting. It computes the required DR event duration
and capacity, i.e., the load to curtail and shift. The ISO then submits a request
for this capacity to the DRX containing the event start time, duration, the required
capacity, and the MWh price. According to FERC order 745, this price is equal to
the price offered to generators [184].
2. The DRX informs all registered aggregators that a request has been received, and
provides them with its characteristics.
3. Aggregators request the customers that have a contract with them to submit the
capacity they are willing to curtail during the event, using the provided event char-
acteristics.
4. The HEMS of each customer computes the average capacity that is predicted for
curtailment or shifting during the event, and reschedules the recharge of the PHEV
accordingly, if required. The bid is then submitted to the aggregator. Bids contain
the average value of the predicted load reduction over the event duration. In this
study, three categories of loads are used, in the order of decreasing curtailment prior-
ity: PHEVs, thermal loads (AC, EWH), and others loads (which are not considered
for DR). These loads are selected because they have a limited impact on customers if
their use is scheduled properly, and because the corresponding load reduction during
the DR event can be quantified.
5. Aggregators centralize the bids from the customers, and decide on a bid for the DRX
according to their respective internal criteria and business model.
6. Using the bids of the aggregators and the offer of the ISO, the DRX dispatches the
capacity among aggregators and informs the ISO of whether the request can be fully
met or not. As the ISO is neutral and non-partisan, it is assumed in this study that
each aggregator is allocated a portion of the total required DR capacity proportional
to its bid.
7. Aggregators randomly select customers to commit to their bid capacity, until 200 %
of the bid capacity aggregators are expected to achieve is reached. The value of
this adjustment coefficient α is obtained empirically in this study, and accounts for
several phenomenon: 1) customers bid their average load reduction, but the goal is
to mitigate the peak, which requires to curtail more than the average amount; 2)
new loads may be switched-on during the event; 3) other loads may increase their
consumptions; and 4) customers may later override the committed capacity, which
needs to be compensated. Lower values would result in the total net load exceeding
the threshold value during the event.
After the commitment has occurred, and until one hour before the event begins:
9. Each customer that has been requested to curtail or shift load by an aggregator has
the ability to manually override this automated commitment, using the HEMS inter-
face. This could also be achieved through a dedicated smartphone application [185].
Choosing to override a DR event would reduce the financial benefit for the customer.
4.4. System Operation 95

Right before, during, and after the event:

10. Aggregators send a signal to the customers that have committed to curtail or shift
their loads to implement the plan they committed to. From this moment on, the
customer is assumed not to be able to override the curtailment request.
11. Customer HEMSs implement the committed measures, until the end of the scheduled
event: the controls of the AC and EWH units are changed, and the charge of the
PHEV, if present, is rescheduled.
12. At the end of the event, each customer HEMS returns to its normal (i.e., non-DR)
operation mode. The end of the event is scheduled at different times for customers
in order to avoid rebound effects, as explained in sections 4.4.4 and 4.5.4.

The steps required to check that customers have actually implemented the load reduc-
tion they committed to are not considered in this study. However, in a real system, this
would be an essential feature, especially for fair compensation.

Customer
ISO DRX Aggregator Customer
HEMS

Figure 4.9: Interactions between agents during a DR event. Interactions are numbered
according to the list given in section 4.4.3.

4.4.4 Rescheduling Algorithm

In order to decrease the local residential load, several strategies may be used by the
customer. These strategies rely on temporary load reduction or load shifting achieved by
rescheduling PHEV charging and changing the control parameters of thermal loads.
96 Chapter 4. Aggregator-Based Residential Demand Response

[Link] PHEV Rescheduling Strategy

The PHEV charging schedule is modified to decrease the load as much as possible
during the event. As shown in Figs. 4.10 and 4.11, several cases are distinguished.

– If the PHEV cannot be fully recharged during the time it is plugged-in, then the
initial schedule is not altered, and the PHEV does not participate in the event.

– If the battery can be fully recharged without charging at all during the event, then a
first strategy is used: the maximum charging rate is used before the event; charging
stops during the event; and restarts as late as possible so that the battery is full
when required by the user (Strategy 1). This strategy helps avoiding all PHEVs to
start charging at the same time right after the event.

– If there is not enough time to recharge the vehicle without charging during the event,
and the vehicle arrives before or after the event, then the maximum charging rate is
used before and after the event, and the charging rate during the event is chosen so
that the remaining energy to recharge is equally split at each time (Strategy 2).

– If the vehicle is plugged-in during the event and there is not enough time to recharge
the vehicle without charging during the event, the same process as in Strategy 2 is
used, except that the charging starts when the vehicle is plugged-in (Strategy 3).

start

no
Full recharge Do not participate
feasible?

yes

Enough time yes

to recharge Use strategy 1
with event?

no

Vehicle no
plugged-in Use strategy 2
during event?

yes

Use strategy 3

Figure 4.10: Flowchart of the PHEV charging rescheduling algorithm.

4.4. System Operation 97

DR event

Strategy 1

Arriving Leaving

Strategy 2

Arriving Leaving

Strategy 3

Arriving Leaving

Fast charging allowed

Slowest charging allowed

No charging allowed

Figure 4.11: Diagram of the available PHEV rescheduling strategies.

[Link] Thermal Loads Rescheduling Strategy

For thermal loads, the rescheduling strategy simply overrides the temperature settings
of the user and temporarily modifies them. The settings are changed so that the indoor
temperature may become warmer and cooler than it would with the original settings,
but without altering the user comfort significantly, i.e., the room temperature should be
bearable. The same principle is used for the EWH. The following settings are selected:

– For AC units, Tref,i is set to 80 ◦ F (i.e., +10 ◦ F compared to the inital set point) for
cooling, and to 65 ◦ F (-5 ◦ F) for heating. These values are chosen arbitrarily by the
authors, according to their own habits and opinion. However, these settings could
be easily changed to more conservative values.

– Similarly, for EWHs, Tref,w is set to 45 ◦ C (-5 ◦ C) [176].

– In order to avoid all thermal loads restarting at the same time and resulting in
a rebound effect, the end of the event is locally postponed by a random duration
shorter than 2 hours.
98 Chapter 4. Aggregator-Based Residential Demand Response

4.5 Simulation Results

4.5.1 Simulator
In order to evaluate the performance of the proposed system, a simulator was devel-
oped. The architecture of the system is implemented using JADE. The models described
earlier are implemented in Java, are embedded in the corresponding agents, and are set
up so that they can communicate with each other. The physical agents (feeder and sub-
station) are also interfaced with PowerWorld using the co-simulation framework described
in section 2.3.
Three consecutive steps are used to obtain the results shown in the following pages:

1. The system is first run in metering mode, so as to obtain the initial load profiles,
where no DR action is taken.

2. The DR system is then run, and acts on load profiles through the process described
earlier in section 4.4.4.

3. A second metering pass is run in order to obtain the load curves after the loads and
PHEVs have been rescheduled, i.e., to verify the effectiveness of the system.

4.5.2 Test Case

The proposed methodology is tested with a portion of the RBTS, a large test system
with transmission and distribution subsystems, a peak load of 85 MW, and about 15,000
customers [186]. Bus 5 of the RBTS is modeled in PowerWorld Simulator (Fig. 4.12) and
is interfaced with the simulator using the co-simulation. The peak load for this bus is
20 MW, with an average load of 11.29 MW, and a total of 2,858 customers of all types
(residential, commercial, industrial, and offices).
As shown in Table 4.6 and according to data given in [186], each load point contains
1 to 250 customers, which can be of residential, industrial, commercial or office type.
For this study, all non-residential customers are converted to residential customers, by
assuming a 4 kW average peak load. For example, an industrial customer with a 400 kW
peak load is replaced by 100 residential customers of 4 kW rating each. A total of 5,555
residential customers is thus simulated.

Load points Customer type Peak load [MW] Avg. load [MW] No. of cust.
1-2,20,21 Residential 0.7625 0.4269 210
4,6,15,25 Residential 0.7450 0.4171 240
26,9-11,13 Residential 0.5740 0.3213 195
3,5,8,17,23 Gov. and inst. 1.1100 0.6247 1
7,14,18,22,24 Commercial 0.7400 0.4089 15
12,16,19 Office buildings 0.6167 0.3786 1
Table 4.6: Customer data for load points at bus 5 of the RBTS, based on data from [186].
 4.5. Simulation Results
Figure 4.12: Oneline diagram of bus 5 of the RBTS in PowerWorld Simulator. Bus numbers are different from the ones given in Table 4.6.

99
100 Chapter 4. Aggregator-Based Residential Demand Response

4.5.3 Simulation Parameters

The time resolution of operation is set to 15 minutes, to coincide with that of bulk
power markets [187]. The penetration rates for PV and PHEVs are chosen as 15 % and
30 % of the customers, respectively. Regarding aggregators, 70 % of the customers are
assumed to have a contract with one (and only one) of the aggregators, while 10 % of
these contracted customers override either the rescheduling of their loads (thermal loads)
or the charging of their PHEV. The number of aggregators is arbitrarily chosen to be
three in this study, with market shares of 40 %, 40 % and 20 % of the customers having a
DR contract. However, the number of aggregators and their respective market shares can
be changed. Customers are required to set the time by which they expect the PHEV to
be charged the next day. The objective of the DR system is to maintain the system load
under a threshold, assumed to be constant during the DR event, and arbitrarily equal to
90 % of the maximum forecast load (i.e., demand during the peak should be decreased by
10 %).

4.5.4 System-wide Net Load Results

A first run in metering mode acquires the forecast load profile, and detects the maxi-
mum capacity threshold shown as an horizontal line in Fig. 4.13. The corresponding event,
where the system load exceeds 90 % of the maximum forecast value, starts at 4:30pm and
ends at 8:00pm.
Figure 4.13 shows the net load curve obtained for the system with the first pass of the
system with metering, i.e., the baseline or forecast load curves, and the net load curve
obtained after the DR system is used, i.e., the actual load. This load curve shows that the
DR system decreases the load so that it remains below the threshold during the event, by
curtailing and shifting load for use later in the day or night.
Figure 4.14 shows the load reduction, which corresponds to the difference between the
baseline and the actual load curve. It shows that the load is reduced on average by about
1.96 MW during the event, and fluctuates for a few hours after the event has ended. This
fluctuation is due to the shifted loads restarting. The average of this curve is 0.0642 MW,
which indicates that the total energy is is almost exactly maintained by shifting the use
of the loads, except a small part that corresponds to the energy not used by the ACs
and EWHs during the event. The total energy consumed is reduced by 0.43 %, which is
considered negligible in the following.
Tables 4.7 and 4.8 indicate that the average load reduction during the event is of
1.96 MW (10.0 %), with a maximum of 2.36 MW, or 12.3 % of the total baseline load, and
a minimum of 0.0 (i.e., the actual load does not exceed the threshold). Considering the
target load reduction of 10 %, these results indicate that the objective is met. Consequent
to the DR event, the load is increased by as much as 1.03 MW (+6.81 %, peak value) after
the event to compensate for the energy shifted during the peak. Additionally, 6.86 MWh,
or 0.92 %, of the on-peak energy consumption is shifted to off-peak hours. An average of
0.73 kW is is curtailed or shifted by each participating customer during the event.
The rebound effect is a phenomenon where the load increases above its expected value
after it was temporarily reduced [188]. The rebound effect indicates that the load is
essentially shifted later in time: PHEVs still need to be charged, and the indoor and
4.5. Simulation Results 101

22
Baseline
20 Actual use
Threshold
18
Power [MW]

16

14

12

10
10 15 20 25 30 35
Time [h]

Figure 4.13: Baseline and actual net load curves.

3
Load reduction
2 Average
Power [MW]

−1

−2
10 15 20 25 30 35
Time [h]

Figure 4.14: Total load reduction achieved with the DR system.

Metric Value [MW] Value [% of total load]

Average load reduction -1.96 -10.0
Max. load reduction -2.26 -12.3
Max. load increase +1.03 +6.81
Table 4.7: Metrics for the DR event (Part 1).

Metric Baseline value Actual value

On peak energy 68.1 MWh 61.2 MWh
On peak energy share 9.52 % 8.60 %
Table 4.8: Metrics for the DR event (Part 2).
102 Chapter 4. Aggregator-Based Residential Demand Response

water temperatures need to be brought back to their reference value after the event. The
rescheduling algorithm described earlier contains two measures to mitigate this effect:
postponing PHEV charging as late as possible, and restarting the use of thermal loads
randomly a few hours after the event. Fig. 4.15 shows the load curve that would be
obtained without these measures: as soon as the event is over (here, at 7pm), the loads and
PHEVs that were curtailed all restart at the same time, which leads to a large peak that
largely exceeds the threshold capacity by about 3 MW. The rebound mitigation measures
are thus needed to properly maintain the net load below the capacity threshold. As a load
profile is generated randomly at each run, the load curve show in Fig. 4.15 corresponds to
a different run than the previous one; however, a similar profile would be obtained if the
same run could have been used.

Baseline
20
Actual use
Threshold
Power [MW]

15

10

10 12 14 16 18 20 22 24 26 28 30
Time [h]

Figure 4.15: Rebound effect example, observed after the DR event if no measure is taken
to mitigate it.

Taking a closer look at how each category of loads is affected, Fig. 4.16 shows that
the load of PHEVs is decreased during the event, and increases after it, so that the total
energy remains unchanged (18.8 MWh). A small proportion of the PHEVs (2.8 % of them)
cannot by fully charged when leaving (the DR event has no impact on it), which is not
problematic as, in a worst case scenario, PHEVs can operate using fuel instead of the
battery. The curves also show that PHEVs account for about a third of the total load
reduction, with a maximum of almost 1 MW load reduction during the event.
Similarly, for thermal loads (ACs and EWHs), the load is reduced during the event
and consequently increases after it, leading to a difference between the baseline and the
actual load curve for several hours. The rebound mitigation strategy manages to limit
the load increase after the event to a reasonable value. The use of these loads enables
reducing the total net load by more than 1 MW during the event, which means that these
loads account for about half of the total load reduction.

4.5.5 Results for Residential Customers

Figure 4.18 shows the impact on indoor and water temperature of rescheduling for
a single typical residential customer. During the event, the indoor temperature clearly
increases (as the outdoor temperature is warmer than the setting temperature) and the
4.5. Simulation Results 103

4
Baseline
Actual use
3
Power [MW]

0
10 15 20 25 30 35
Time [h]

Figure 4.16: Baseline and actual load curves for PHEV charging.

10

9
Power [MW]

7
Baseline
Actual use
6
10 15 20 25 30 35
Time [h]

Figure 4.17: Baseline and actual load curves for thermal loads.

water temperature decreases. These two temperatures are brought back to their set values
after the event.
Regarding the charging schedule of the PHEV, Fig. 4.19 shows that the PHEV uses
recharging strategy 1 after having its charging scheduled modified: charging starts before
the event, and stops during its entire duration. Charging restarts only a few hours after
the event has ended, so that the vehicle is fully charged when leaving. Fig. 4.20 shows a
PHEV using recharging strategy 2, where the charging rate is decreased but not stopped
during the DR event, and Fig. 4.21 shows the result of strategy 3, where the event starts
before the PHEV is plugged-in.

4.5.6 Impact on the Distribution System

The impact of the DR system on the distribution is evaluated using the interface with
PowerWorld. For each time period, a power flow algorithm is run and checks whether
the voltage on every bus of the system Vi meets condition (4.18). The power flow solved
by the PowerWorld setup is a balanced three-phase power flow; its use is enabled by the
selected approach of load aggregation. Results show that before and after the DR event,
104 Chapter 4. Aggregator-Based Residential Demand Response

26
Temperature [°C]

24

22

20

18
10 15 20 25 30 35
Time [h]
Forecast indoor temp. Actual indoor temp. Outdoor temp.

70
Temperature [°C]

65

60

55
10 15 20 25 30 35
Time [h]
Forecast water temperature Actual water temperature

Figure 4.18: Impact of the DR system on indoor and water temperature for a typical
residential customer.

2.5
Baseline
2 Actual use
Power [kW]

1.5

0.5

0
10 15 20 25 30 35 40
Time [h]

Figure 4.19: Rescheduling of the charge of a PHEV using strategy 1.

4.5. Simulation Results 105

2
Baseline
Actual use
1.5
Power [kW]

0.5

0
10 15 20 25 30 35 40
Time [h]

Figure 4.20: Rescheduling of the charge of a PHEV using strategy 2.

2
Baseline
Actual use
1.5
Power [kW]

0.5

0
10 15 20 25 30 35 40
Time [h]

Figure 4.21: Rescheduling of the charge of a PHEV using strategy 3.

this condition is verified, which means that the system does not lead to steady-state
instabilities. The minimum and maximum voltage deviations for all buses are 0.9963 and
0.9995 p.u., respectively. Fig. 4.22 shows the voltage profile at Feeder 4, and indicates
that the DR system enables slightly improving the voltage profile. As the same customer
model is used at each load point, similar results are obtained for other feeders.

0.95 ≤ Vi ≤ 1.05 p.u. (4.18)

106 Chapter 4. Aggregator-Based Residential Demand Response

0.999

0.9985
Voltage [p.u.]

0.998

0.9975

0.997 Baseline
Actual use
0.9965
10 15 20 25 30 35
Time [h]

Figure 4.22: Voltage profile at Feeder 4 of bus 5 in RBTS.

4.6 Conclusion
This chapter has shown that an agent-based DR architecture and system can use small
residential assets such as AC units, EWHs and PHEVs to temporarily reduce load under
a given threshold for participation in a DR event. By rescheduling the use of these assets
at a limited cost and impact on comfort for the end user, aggregators can coordinate their
actions through a DRX and meet a request for capacity issued by the ISO. The resulting
rebound effect can also be mitigated by carefully planning the end of the DR event in
time, while limiting the impact on customer comfort. The system does not cause any
steady-state instability when tested on a 5000+ customers test system.
 5

Conclusions

5.1 Concluding Remarks

5.1.1 Conclusions on the Proposed Approach
In the advent of the smart grid, the need for energy management solutions capable
of handling the increasing complexity of electric power systems has become even more
important. This dissertation has proposed an new and unique approach for the design
and development of EMSs for smart power systems, based on the use of MASs and other
artificial intelligence techniques to model, control and simulate such systems.
This approach relies on the combination of several disciplines: it tackles issues in the
field of electrical engineering, using computer engineering and applied mathematics tools,
while also taking into account economic and sociological parameters. It therefore facilitates
multi-disciplinary studies of CPSs including physical power systems, the communication
and control infrastructure, energy markets, and their users. It also enables truly systemic
studies of smart grids, which is a requirement to handle the inherent complexity of these
complex adaptive systems.
This work has also shown that MASs do have a potential application of high impact
in power systems, even for problems where decision-making cannot be fully distributed.
Traditional approaches generally tend to focus on decision-making algorithms, but do
not discuss how system entities interact with each other. The selected approach helps
simultaneously taking into account communication and interaction aspects, that are at
the core of smart grid technologies, and facilitates the specification of EMS architectures
and of the corresponding hardware and communication infrastructure, ultimately enabling
a faster transition from simulation to real-scale prototypes.
The approach has been applied to two applications of energy management in smart
grids. Due to the specificity of each application, the approach has been adapted to define
the most appropriate architecture and interactions to meet design objectives. However,
it should be re-stated that the definition for an agent adopted in this work is generally
broader than the one used in “traditional” MAS works, especially for non-power systems
applications, in that not all agents are fully autonomous and may depend on each other.
The dispatching agent in the power plant application is an example. However, for such an
application, a distributed dispatching algorithm would return sub-optimal results [189],
and would therefore not make sense compared to centralized ones considering economical

107
108 Chapter 5. Conclusions

design constraints. Nevertheless, the other advantages of the agent concept are utilized,
for the reasons and advantages described in chapter 2. The operational constraints and
objectives are thus the main drivers in deciding on the centralized or decentralized nature
of decision-making in an EMS.
In a broader perspective, it is expected that the adoption of the proposed approach
could facilitate the design and development of EMSs with a truly systemic point-of-view.
It holds promises for a variety of applications, also outside the fields that were presented in
this work. Examples include the comparison of centralized and distributed architectures
and algorithms, the study of various market structures and regulations, the analysis of
various planning studies, e.g., for the integration of storage and DG, etc.

5.1.2 Conclusions on the Presented Applications

The proposed applications have shown that the two presented systems could provide
practical solutions for energy management in smart grids, and especially for the integra-
tion of intermittent RESs, through two main means: by making generation capable of
handling high demand ramps with a high economic and environmental performance, and
by controlling residential assets through market mechanisms.
In the first application, the proposed unique gas turbine power plant system has shown
that agent-based modeling and simulation enables to design and test a flexible, resilient
and efficient power plant EMS, capable of handling a large penetration of intermittent
generation. GE’s 9E gas turbines were modeled using real fuel consumption and NOx
and CO2 emissions data. A specific MAS architecture was designed to enable a flexible
and resilient power plant operation, and was combined with advanced decision-making
algorithms to reduce operational costs and emissions. The agent-based approach has en-
abled specifying how each element in the system interacts with the others, as well as
evaluating the computational and communication requirements, which may facilitate a
faster transfer to real-world applications. Simulation results on a dynamic load profile
with high intermittent RESs penetration also showed that significant cost savings and
emissions reductions could be achieved without altering turbine thermodynamics. Dy-
namic dispatching optimization was done by a metaheuristic selected from a comparison
of several, and a rule-based automatic start and stop algorithm in charge of controlling
the operation of the gas turbines further improved results.
In the second application, the proposed unique and scalable DR system has shown that
agent-based approaches enable to model, simulate and control large power systems. An
agent-based architecture was developed to model a distribution grid based on the RBTS
with numerous residential customers, each with their own characteristics (housing insu-
lation, PHEV, DG, etc.). Probabilistic models of residential customers with several load
categories were implemented and used with an advanced load control algorithm that in-
cludes PHEV charging rescheduling and DLC for AC and EWH units, i.e., small residential
loads. The role of each entity in the system, from end-users to a regional market operator,
was specified, as well as their respective interactions in part through a DR market and
aggregators. Simulation results showed that the system is capable of reducing residential
load under a given threshold, with a limited impact on customer comfort. A steady-state
stability analysis was performed using a specific co-simulation framework that was made
5.2. Future Work 109

available to the research community, and showed that the operation of the system did not
compromise stability.

5.2 Future Work

Future research work will focus on several aspects of the proposed approach and appli-
cations. As the studied field is an almost virgin territory, future avenues are very fertile.
At first, the MAS design methodology proposed in this dissertation could be enhanced
with more detailed steps, in order to obtain a full design methodology suited for power
systems applications. Another aspect would be to focus on distributed decision-making
algorithms, that were not considered in this work, and are particularly relevant given the
selected MAS framework.
Then, solutions to overcome the soft spots of the applications could be developed.
Additional topics related to the proposed approach and framework could be explored.
A solution to take into account forecasting inaccuracies could be included in both ap-
plications. For the power plant EMS, this could simply consist in including the forecasting
error estimation in the computations for the SSA, which could result in a more conserva-
tive but also more reliable behavior. For the DR system, a simple solution could be to
increase the value of the adjustment coefficient α, so that more capacity is committed.
However, in both cases, a detailed study would be required to find a compromise between
costs and reliability.
Enhanced forecasting capabilities could also enable reducing the probability of non-
optimal choices being made by the power plant EMS. Equipping the EMS with the capa-
bility to establish strategies in advance, e.g., through a unit commitment algorithm, while
taking into account the entire emissions curve could also result in even lower costs and
emissions, especially for NOx. The difficulty for metaheuristics-based dispatching algo-
rithms to return results that do not exhibit unnecessary variations (which mostly depend
on the optimization objective) could be tackled by combining these stochastic algorithms
with a deterministic algorithm.
The integration of the power plant EMS with markets would also enable new studies,
e.g., related to the participation of power plants in ancillary services, or in day-ahead
markets. Another example would be the possibility for the power plant to sell the surplus
energy it does not need to power a local load (such as a paper mill), depending on electricity
prices.
In addition to these aspects, many other topics related to the residential DR system
may also be investigated in the future. A first objective would be to evaluate the financial
impact of the DR system on customer billing and on aggregator business models. This
aspect is currently not accounted for, and is expected to give an insight on how residen-
tial customers could benefit from such DR measures, but also on which business models
would benefit the most to aggregators. A better integration with power markets, e.g., by
interfacing it with test beds such as the AMES wholesale power market test bed [190]
from Iowa State University, USA, would enable detailed studies on the impact of DR on
markets, and of markets on the behavior of customers with respect to DR.
Enhancing customers behavior models, e.g., with respect to overrides, could also pro-
vide interesting results, and could improve the reliability of the results (i.e., more realistic
110 Chapter 5. Conclusions

results may be obtained after such as study). Through a collaboration with researchers
from Colorado State University (CSU), and with the RECITS laboratory at UTBM, which
has specialists in consumer acceptance toward technological changes, the modeling of the
behaviors of customers could be greatly improved by taking into account the various cat-
egories of customers and their different reactions toward new technologies. For example,
a teenager and an elderly person could have different comfort zones with respect to room
temperatures, as well as have completely different uses of energy management interfaces.
Improving residential load models could also increase the accuracy of the results. In
the future, more accurate load models, separating different dwellings sizes and types could
be introduced to reflect the diversity of housing in a distribution system. GridLAB-D,
developed by the US Pacific Northwest National Laboratory, may be an alternative to the
models that are currently used [108]. Additionally, a sensitivity analysis on the penetration
level of the loads and DGs (PHEVs, PV, etc.) could provide interesting perspectives on
the impact and potential of such resources on the grid.
Using such detailed load models, a learning HEMS, such as the one assumed to be
present in the DR application, could be developed. Using learning techniques such as
reinforcement learning or artificial neural networks, the behavior of customers could be
learned automatically, which would enable the HEMS to take decisions based on informa-
tion extracted from users habits. This capability could enable estimating the amount of
capacity that could be shed of shifted, and to take action automatically so as to maximize
the benefit for the user. For example, savings could probably be achieved if the system
was able to control heating and cooling based on users schedules and price forecasts.
Implementing dynamic, real-time or time-of-use pricing, and elastic load models is
another aspect. As customers would react differently if prices vary according to the balance
between supply and demand, various pricing schemes may be used. This alternative DR
method could be compared with the proposed one, with the main disadvantage that the
DR capacity can only be roughly estimated using historical empirical data, as it depends
on the behavior of each customer, which is by nature hard to predict. Although the use of
smart appliances was not considered in the proposed study, it could be included with such
price-based DR methods. Another approach based on heuristic optimization proposed
in [191] could also be tested and compared with current results.
Integrating additional elements in the grid is another aspect that could provide inter-
esting results. On the one hand, integrating distributed energy storage at the distribution
level would provide additional flexibility, and could contribute to temporarily reduce the
net load during demand peaks, by serving as a buffer and a complement — or competi-
tor — to DR. On the other hand, the integration of larger shares of DG resources could
enable microgrid islanding. Although distribution PV resources are currently considered,
larger DG sources could be added, and their impact on the operation of the system eval-
uated. This would ultimately enable parts of the distribution system to be islanded and
to operate autonomously, using local generation, storage and DR resources.
Finally, two last aspects of the DR simulator could be modified and improved. Firstly,
the system could be implemented (fully or partially) on a real-time simulator. The IRTES-
SET laboratory recently acquired two real-time OPAL-RT simulators that could be used
to simulate the DR system. Secondly, replacing PowerWorld Simulator with DIgSILENT
PowerFactory would give more accurate results regarding the impact of the DR system
5.3. Scientific Production Overview 111

on the distribution grid. PowerWorld assumes three-phase balanced systems only, which
does not fully represent real distribution system conditions. PowerFactory has the ability
to model and simulate unbalanced power flows, and would fill this gap [192].

5.3 Scientific Production Overview

The work presented in this dissertation was also an opportunity for collaborations
with the industry and with local and international academic partners. It led to multiple
publications, as shown in Appendix B:
– The development of the gas turbine power plant EMS was realized in collaboration
with the company GE established in Belfort, France, which provided data on the gas
turbines, and with the Université de Haute Alsace (UHA) from Mulhouse, France,
for the study of metaheuristics. It resulted in an international journal publication
in Applied Energy [39], and a French and international patent [193]. Additionally,
this work was ranked second of the international GE Energy Innovation Awards of
2011.
– The design and development of the residential DR system was done in collaboration
with CSU, Fort Collins, USA, and was initiated during a visiting period of four
months in 2012. Its results were submitted to an international journal, in collabo-
ration with a researcher from the Lawrence Berkeley National Laboratory (LBNL),
Berkeley, USA.
– Several other publications were derived from these works, including a co-simulation
framework [38], other applications of metaheuristics for energy management [194,
195], a state-of-the-art of MASs [48], a book chapter on MAS design for power
systems [196], a comparison of smart grid developments in Europe and in the US [2,
197], and a hybrid metaheursitic algorithm [198].
In addition to these works, several other topics, mostly related to PHEVs, were in-
vestigated in collaboration with other researchers and are not presented in this disserta-
tion. The impact of PHEVs with V2G capability on distribution systems was investigated
in [175], in collaboration with researchers from CSU and the US National Renewable
Energy Laboratory (NREL). A study of the combined optimal energy management and
sizing of hybrid electric vehicles was conducted at UTBM [199], as well as an experimental
analysis of the performance of a lithium-ion battery multi-physical model and its appli-
cation to Kalman filter-based SOC estimation [200]. Finally, a short-term distribution
load forecasting methodology was recently developed in collaboration with GE Power &
Water [201].
Additionally, some of the source code of the developed algorithms were made available
to the research community [112, 202], with the hope to facilitate the re-use of elements of
the proposed approach by other researchers.
Appendices

113
 A

Metaheuristics for Optimal Dispatching

The dispatching algorithm, which is at the heart of the EMS proposed in chapter 3,
relies on optimization algorithms called metaheuristics. This appendix describes and com-
pares several metaheuristics in order to select one for the EMS.

A.1 Definition
Metaheuristics are computational methods that optimize a problem by iteratively try-
ing to improve candidate solutions with regards to a given measure of quality called fit-
ness. These stochastic optimization algorithms require very few or no assumption about
the problem and can explore almost any kind of search space, but do not guarantee that
the best solution found is optimal.
For real-valued search spaces, classical optimization algorithms generally derive the
gradient (i.e., compute the Hessian matrix) of the fitness function to be optimized and
then employ gradient descent or a quasi-Newton methods to find local maxima or minima.
Metaheuristics do not use this process and can therefore be employed with problems where
the fitness function may not be continuous or differentiable. In the following sections, it
is assumed that the goal of the optimization process is to find the global minimum.

A.2 Common Metaheuristics

Since the invention of genetic algorithms in the early 1970’s that made metaheuris-
tics popular for solving complex optimization problems, dozens of new algorithms have
been invented [203]. In the following, particle swarm optimization (PSO), genetic algo-
rithms (GAs), differential evolution (DE) and imperialist competitive algorithms (ICAs)
are presented.

A.2.1 Particle Swarm Optimization (PSO)

The original PSO algorithm (Algorithm 1) is inspired from the behavior and movement
of bird and fish swarms [204] and enables exploring a multi-dimensional search space
based on this principle. In this population-based stochastic optimization technique, each
individual, called particle, tries to improve itself by observing other group members and
imitating the best ones. To do that, each particle keeps track of its coordinates in the

115
116 Appendix A. Metaheuristics for Optimal Dispatching

solution space using two values: the best solution it has achieved so far, called personal
best pbest , and the best value obtained so far by any particle in the nneighb particles
constituting the neighborhood of that particle, called group best gbest .
The movement of each particle depends on its velocity v, computed according to (A.1).
The new velocity depends not only on the particle’s previous velocity and position, but
also on pbest and gbest . The relative importance of each term of the equation can be tuned
by modifying the cognitive and social coefficients c1 and c2 , respectively, as well as the
inertia weight wt which determines the influence of the previous velocity. To enable the
algorithm to search the solution space and to avoid particles getting stuck in local minima,
two random real values r1 , r2 ∈ [0, 1] are also generated.

vt+1 = wt · vt + c1 · r1 · (pbest − xt ) + c2 · r2 · (gbest − xt ) (A.1)

| {z } | {z }
cognitive component social component

The particle’s position x is then obtained from its previous position and its current
velocity:
xt+1 = xt + vt+1 (A.2)

Each particle moves in the search space based on the previous movements, until a
stopping condition is met, e.g., a maximum number of function evaluations or generations.
xt+1

gbest

vt+1
vt gbest − xt

pbest
xt pbest − xt

Figure A.1: Diagram of the movement of a particle with the PSO algorithm, from [195].

Algorithm 1 Pseudo-code of the PSO algorithm.

Initialize the population of npop particles
while the stopping condition is not satisfied do
for each particle xi do
Update the particle’s velocity vi
Update the particle’s position xi
Evaluate the fitness f (xi (t)) of the particle
Update the values of pbest and gbest
end for
end while
Return the best found solution.

The parameter settings used for PSO are given in Table A.1.
A.2. Common Metaheuristics 117

Parameter Value
c1 2.0
c2 2.0
ω 0.5
vmax 1.0
npop 50
nneighb 10
Table A.1: Parameter settings for the PSO algorithm.

A.2.2 Differential Evolution (DE)

DE is a stochastic population-based metaheuristic developed by Storn and Price [205],
that is particularly well-suited for continuous problems. Similarly to other evolutionary
algorithms like genetic algorithms, it uses mutation, crossover and selection operators
to make its solutions population explore the search space while looking for the best fit-
ness value. Algorithm 2 describes its operation principle. The main advantages of this
algorithm are its fast operation, and its low number of parameters.
The mutation operator generates a trial vector ui (t) for each individual xi (t) by com-
bining an individual xi1 (t) of the population with the weighted difference between two
other individuals, as in (A.3) where β is a positive real value and i 6= i1 6= i2 6= i3.

ui (t) = xi1 (t) + β · (xi2 (t) − xi3 (t)) (A.3)

The crossover operator uses this trial vector to produce an offspring x0i (t) with (A.4),
where xi,j (t) refers to the j-th element of individual i, r is a random number between 0
and 1, and γ is a real positive value.
(
ui,j (t) if r < γ
x0i,j (t) = (A.4)
xi,j (t) otherwise.

Finally, the selection operator is used to decide which solution to keep in the population
between xi (t) and its offspring x0i (t). The offspring replaces its parent only if its fitness
f (x0i (t)) is better than the one of its parent f (xi (t)).

Algorithm 2 Pseudo-code of the DE algorithm.

Initialize the population of npop individuals
while the stopping condition is not satisfied do
for each individual xi (t) do
Evaluate the fitness f (xi (t)) of the individual
Create a trial vector ui (t) (mutation)
Create an offspring x0i (t) (crossover)
Keep only the best between xi (t) and x0i (t) (selection)
end for
end while
Return the best found solution.

The parameter settings used for DE are given in Table A.2.

118 Appendix A. Metaheuristics for Optimal Dispatching

Parameter Value
β 0.8
γ 0.8
npop 50
Table A.2: Parameter settings for the DE algorithm.

A.2.3 Genetic Algorithm (GA)

GAs model genetic evolution, where the characteristics of individuals are expressed
using genotypes [206]. Individuals called chromosomes consist of several genes, one for each
dimension. The algorithm (Algorithm 3) uses the selection, recombination and mutation
operators. Numerous variations of each operator are possible, and the following explains
the operation principle of the ones that were selected.
– The selection operator models the survival of the fittest. Tournament selection in-
volves running tournaments among chromosomes randomly chosen from a genotype.
The winner of each tournament is selected for crossover. Parameter ntourn control
the number of chromosomes in the tournament.
– The crossover operator models reproduction. New individuals can be obtained either
by generating an offspring from a parent, by using two parents to generate one or
two offspring, or by using multiple parents to generate one or more offspring. The
uniform crossover operator belongs to the second category, uses a fixed mixing ratio,
and enables parent chromosomes to contribute at the gene level rather than at the
segment level. For example, if a mixing ratio of 0.5 is chosen, each gene has a 50 %
probability to be part of the first offspring, and 50 % to be part of the second. These
probabilities are exclusive, i.e., each gene can only be part of a single offspring. As
a consequence, the offspring statistically has half the genes from its first parent and
the second half from its other parent. The probability of a crossover happening is
set by pcross < 1.
– The mutation operator aims at introducing new genetic material to maintain genetic
diversity in the genotype. This is achieved by attributing random values to some
genes. Parameters pmut and σ control the mutation probability.
– The use of elitism can help improve convergence speed by enabling the selection of
the best chromosomes to form the new population.
The parameter settings used for GA are given in Table A.3.

Parameter Value
pcross 0.85
pmut 0.05
ntourn 2
σ 0.01
npop 50
Table A.3: Parameter settings for the GA algorithm.
A.2. Common Metaheuristics 119

Algorithm 3 Pseudo-code of the GA algorithm.

Initialize the population of npop chromosomes
while the stopping condition is not satisfied do
for each chromosome xi do
Evaluate the fitness f (xi (t)) of the chromosome
Select the next generation
Perform reproduction using crossover
Perform mutation
end for
end while
Return the best found solution.

A.2.4 Imperialist Competitive Algorithm (ICA)

The ICA is a recent evolutionary optimization approach introduced in 2007 by E.
Atashpaz-Gargari [207]. It is inspired by the imperialistic competition processes of human
societies. The algorithm can be seen as a social counterpart of genetic algorithms. Several
of its steps are indeed similar: countries can undergo revolutions as chromosomes can
mutate, for example. Historical events such as the competition between France and Britain
during the 18th century for taking control of India are used by the author of [207] to
illustrate the concept of the algorithm.
This algorithm uses a precise terminology, in which a solution is called a country.
There are two types of countries: imperialist countries, and colonies which depend on
these imperialists. An imperialist and its colonies form a group of countries called empire.

Algorithm 4 Pseudo-code of the ICA algorithm.

1: Initialize the countries and form the empires
2: while the stopping condition is not satisfied do
3: for all empires do
4: Move the colonies toward the imperialist (assimilation)
5: Make some colonies undergo a revolution
6: if a colony is more powerful than the imperialist then
7: The colony becomes the imperialist and vice versa (overthrow)
8: end if
9: end for
10: if two empires are too close then
11: Merge them (unification)
12: end if
13: Make imperialistic competition occur
14: if there is an empire with no colony then
15: Eliminate this empire
16: end if
17: end while
18: Return the best found solution.

ICA works as illustrated in Algorithm 4, where the following processes are used:
– Initialization and empire formation: Like other evolutionary algorithms, ICA starts
120 Appendix A. Metaheuristics for Optimal Dispatching

with an initial population of solutions called countries, of size Npop . Among them,
the Nimp best countries (the most powerful) are selected to be imperialists. The re-
maining Ncol countries form the colonies of these imperialists. The n initial empires
are formed by dividing the colonies among imperialists according to their normalized
power P derived from their cost (fitness) c (A.5).

cn − maxi ci
Pn = PN (A.5)
imp
i=1 (cn − maxi ci )

The number of colonies Ncol,n attached to empire n is computed according to (A.6):

Ncol,n = round (Pn · Ncol ) (A.6)

– Assimilation: Imperialist countries attract colonies using the assimilation policy

illustrated in Fig. A.2. To update its position x, each colony moves toward its
imperialist by updating its position using (A.7).

xt+1 = xt + β · γ · r · d (A.7)

where β > 1 causes the colonies to get closer to the imperialist, γ < 1 corresponds
to an assimilation coefficient, r is a random number chosen from the uniform dis-
tribution U (−θ, θ), θ adjusts the deviation from the original direction and enables
searching around the imperialist, and d is the distance between the colony and the
imperialist.

Imperialist
d

θ
Colony xt+1
xt

Figure A.2: Movement of a colony toward its imperialist, from [194].

– Revolution: The revolution process introduces sudden random changes in the po-
sition of some countries. It plays the same role as the mutation operator in GAs.
Parameter Rr can be tuned for change the revolution probability.
– Overthrow: After assimilation and revolution, a colony might reach a better position
than the imperialist of the empire. In this case, the colony can become the imperialist
and vice versa.
– Unification: If two empires are too close to each other, they can unite and become
a single empire, with the sum of the colonies of the two initial empires. Parameter
Ut can be tuned for change the minimal unification distance.
– Imperialistic competition: Each empire tries to take possession of colonies of other
empires and to control them. This imperialistic competition is modeled by selecting
A.3. Hybrid Algorithms 121

the weakest colonies of the weakest empire, and giving them to the empire that has
the highest likelihood to possess them.
The total power Ptot,n of each empire n is defined by the power of its imperialist plus
its average colonies power, as defined in (A.8) where ζ  1, I refers to the empire’s
imperialist, and C to its colonies.

Ptot,n = P (In ) + ζ · mean(P (Cn )) (A.8)

The likelihood pn , called possession probability, is then derived from each empire’s
power (A.9).
ctot,norm,n
p n = PN (A.9)
imp
i=1 ctot,norm,i

where ctot,norm,n = ctot,n − maxi ctot,i is the total normalized cost of empire n and
ctot,n its total cost.
In order to divide the colonies among empires based on their possession probability,
a vector A is built (A.10), where Pn is the power of empire n, and rn is a random
value between 0 and 1. The selected colonies are then assigned to the empire whose
relevant index in A is the highest.

A = [P1 − r1 , P2 − r2 , ..., PNimp − rNimp ] (A.10)

The parameters for ICA are empirically determined by running iterative trials using
the mathematical functions described in Table A.7, and starting with the parameters given
by the authors of the algorithm in [207]. The tuned parameters are listed in Table A.4.

Parameter Value
Npop 50
Nimp 6
Rr 0.1
β 2
π
θ 6
ζ 0.02
Ut 0.02
Table A.4: Parameter settings for the ICA approach.

A.3 Hybrid Algorithms

In order to improve the performance of standard metaheuristics, hybrid metaheuristics
were introduced. These algorithms are skilled combinations of two optimization methods,
that can enable the resulting algorithm to benefit from the strengths of each constituting
algorithm [208].
122 Appendix A. Metaheuristics for Optimal Dispatching

A.3.1 Metropolis PSO with Mutation Operation (MPSOM)

The PSO algorithm has been hybridized multiple times, and the MPSOM variant was
introduced in [209] to avoid a premature convergence of PSO. This hybrid algorithm uses
a combination of PSO and simulated annealing with the Metropolis rule [210]. When a
local optimal solution is reached with a PSO algorithm, all particles gather around it, and
escaping from this local optimum may become difficult. MPSOM introduces the following
changes:
– The social component of each particle in (A.1) is computed as a weighted average
of all solutions which are better than its own. This allows the particles to prefer
following a group of particles rather than a single one. In (A.11), Nt is a set of
solutions which are better than the particle, and vmax is the maximum velocity.
 
|Nt |
X pbest − xt 
vt+1 = wt · vt + c1 · r1 · (pbest − xt ) + min vmax , (A.11)
i
i=1

– The value of pbest is updated according to the Metropolis rule which gives the proba-
bility of accepting the position x of a particle according to its fitness f . Rule (A.12)
uses a random variable r3 and a variable T representing a temperature, similarly to
temperature in the physical annealing process [210]. This modification improves the
convergence capability of the algorithm.
  
1 if f (x) ≤ f (p ) or r · exp f (x) − f (pbest ) < 1
best 3
p= T (A.12)

0 otherwise.

– If gbest has not been improved after a given number of iterations (here 60), a mutation
operator gives the particles the maximum allowed velocity to help them escape a local
optimum.
The parameter settings used for MPSOM are given in Table A.5.

Parameter Value
c1 1.7
c2 1.3
wt 0.9
vmax 1.0
npop 50
nneighb 10
T 2.5
Table A.5: Parameter settings for the MPSOM algorithm.

A.3.2 ICA-PSO Algorithm

The ICA-PSO algorithm combines the ICA and PSO approaches in order to improve
the exploration capacity of ICA for single and multi-objective problems. The ICA-PSO
A.3. Hybrid Algorithms 123

approach (Algorithm 5) proposed by the authors introduces the following main changes
to the original ICA algorithm [198]:
– Two archives are used to keep a memory of the best solutions found by the individ-
uals. A global archive stores the non-dominated solutions found by the algorithm,
and local archives store the best solutions found by each individual. A crowding
distance operator is used to determine which solutions should be kept in the archive.
The crowding distance value of a solution provides an evaluation of the density of
solutions surrounding that solution [211].
– Equation (A.7), which defines the movement of the colonies toward the imperialists,
is replaced by an adapted version of the equation defining the movement of particles
in the PSO algorithm. The colonies now have a memory of the best solutions they
could find (similarly to pbest ), and adapt their position according to the position
of the imperialist and their best solution in memory. The speed of each colony is
updated according to (A.13):

vk+1 = wt · vk + c1 · r1 · (pk − xk ) + c2 · r2 · (ek − xk ) (A.13)

Similarly, the speed of the imperialists is updated according to (A.14):

vk+1 = wt · vk + c1 · r1 · (pk − xk ) + c2 · r2 · (gk − xk ) (A.14)

The respective positions of the colonies and of the imperialists are then updated
using (A.2). In these equations, pk is the best position of the colony, ek is the best
position of its imperialist, and gk is the best position in the global archive.
– A crossover operator is also introduced to improve the solutions contained in the
local archive. The crossover mixes a solution of the archive of a colony with a
solution of the local archive of an imperialist.
Although this algorithm was primarily designed for multi-objective problems, it can
also be used for single-objective problems, and, in this case, returns a single solution. The
parameters for ICA-PSO are given in Table A.6.

Parameter Value
Npop 50
Nimp 5
ω 0.65
c1 1.0
c2 1.5
vmax 18
ζ 0.01
Ut 0.02
Table A.6: Parameter settings for the ICA-PSO approach.
124 Appendix A. Metaheuristics for Optimal Dispatching

Algorithm 5 Pseudo-code of the ICA-PSO algorithm.

1: Initialize and evaluate the empires
2: Initialize the archive
3: Initialize the particles memory
4: while the stopping condition is not satisfied do
5: for each empire do
6: Compute imperialist and colonies speed
7: Move each particle according to its speed
8: Evaluate each particle
9: if a colony dominates its imperialist then
10: This colony becomes the new imperialist of its empire
11: end if
12: end for
13: Crossover operations
14: Update the memory of the particles
15: Update the archive
16: Compute the total cost of the empires
17: Imperialistic competition
18: if there is an empire with no colony then
19: Eliminate the empire
20: end if
21: end while
22: return the external archive.

A.4 Performance Comparison and Analysis

In order to compare the algorithms and to select the most appropriate one, a series of
tests is run. The performance of each algorithm is tested with mathematical functions, so
that their proper operation can be verified and their respective performance compared.

A.4.1 Benchmark Functions

Several benchmark functions, described in Table A.7 where n is the dimension of the
problem, are used to test the performance of the algorithms [212]. As the focus is pri-
marily on the dispatching problem, the number of functions is limited to four, although
very rigorous tests in the optimization field generally rely on much more advanced tests,
which are not relevant in this study. These functions provide a good start for testing the
credibility of an optimization algorithm.
Each of these functions (except the Sphere function) has many local optima in its
solution space. The amount of local optima increases with their dimension, which is set
to 20, as in [213]. For each algorithm, the maximum number of evaluations is set to
200, 000. The optimal minimal value of these functions is 0, as show in Figs. A.3 to A.6
where the dimension is set to 2 to enable display. For each algorithm, a total of 20 runs
are conducted and the average fitnesses and standard deviations of the best solutions are
recorded.
A.4. Performance Comparison and Analysis 125

Function Problem Range

Xn
Sphere x2i [-5.12;5.12]
i=1
n
X
Rastrigin (x2i − 10 cos(2πxi ) + 10) [-5.12;5.12]
i=1
n−1
X
Rosenbrock (100(xi+1 − x2i )2 + (xi − 1)2 ) [-2.048;2.048]
i=1
1 Pn 1 1 Pn
x2i ) 2
Ackley 20 + e − 20 e−0.2 (n i=1 − en i=1 cos(2πxi )
[−32.0; 32.0]
Table A.7: Standard mathematical benchmark functions definition.

Figure A.3: Plot of the Sphere function for n = 2.

126 Appendix A. Metaheuristics for Optimal Dispatching

Figure A.4: Plot of the Rastrigin function for n = 2.

4000

3500

3000

2500

2000

1500

1000

500

0
2
1 2
0 1
−1 0
−1
−2
−2
−3 −3

Figure A.5: Plot of the Rosenbrock function for n = 2.

A.4. Performance Comparison and Analysis 127

Figure A.6: Plot of the Ackley function for n = 2.

A.4.2 Algorithms Implementation

The algorithms described earlier are implemented in the Java language as follows. The
PSO, DE and GA algorithms are directly used from their implementation in MetaHeuris-
ticDemo, a framework for single objective metaheuristic optimization [214]. The ICA
algorithm is implemented in Java by the author, based on a Matlab implementation pro-
posed by E. Atashpaz-Gargari in 2008 [215]. This implementation is available under the
LGPL licence at Github [202]. The MPSOM algorithm was implemented in MetaHeuris-
ticDemo by L. Idoumghar [209]. The ICA-PSO algorithm was implemented by N. Chérin
and L. Idoumghar in jMetal, a framework for multi-objective optimization [216].

A.4.3 Test Results and Analysis

The mean solutions and the corresponding standard deviations obtained for the algo-
rithms are listed in Table A.8. MPSOM obtains the best results for the Sphere, Rastrigin
and Ackley functions, while DE obtains the best solutions the Rosenbrock function. From
these results, MPSOM clearly stands out as the best metaheuristic for these kinds of
mathematical problems, followed by DE. On the other hand, PSO, GA and ICA return
average to low performance results, comparatively. More surprisingly, ICA-PSO has an
average performance, even on a simple problem such as the Sphere. As a consequence of
these results, the MPSOM algorithm is selected for running the EMS simulations.
 128
Function PSO DE GA ICA MPSOM ICA-PSO
22.31 7.744 × 10−12 6.000 × 10−06 1.311 × 10−20 0 4.754 × 10−01
Sphere
±8.583 ±5.395 × 10−12 ±3.513 × 10−06 ±3.631 × 10−20 ±0 ±1.729 × 10−01

Appendix A. Metaheuristics for Optimal Dispatching

138.5 89.61 40.69 67.13 0 76.84
Rastrigin
±30.61 ±12.66 ±11.12 ±4.513 ±0 ±37.99
437.3 1.606 13.01 17.29 18.79 114.6
Rosenbrock
±201.0 ±0.4734 ±4.367 ±22.79 ±0.1129 ±51.20
16.00 2.056 × 10−05 1.037 × 10−02 1.718 4.441 × 10−16 1.718
Ackley
±1.327 ±6.623 × 10−06 ±2.896 × 10−03 ±4.556 × 10−16 ±5.059 × 10−32 ±4.556 × 10−16
Table A.8: Comparison of the solutions obtained by the selected metaheuritics.
 B

Publications

Publications list updated on December 7, 2012.

International Journals
1. [Submitted] Roche R., Suryanarayanan S., Kiliccote S. and Miraoui A., An Aggregator-
Based Residential Demand Response System.

2. Roche R., Idoumghar L., Suryanarayanan S., Daggag M., Solacolu C.-A. and Miraoui
A., A Flexible and Efficient Gas Power Plant Operation System With Economic and
Environmental Constraints. Applied Energy, vol. 101, pp. 644–654, January 2013.

3. Watrin N., Roche R., Ostermann H., Blunier B. and Miraoui A., Multi-physical
lithium-based battery model for use in state-of-charge determination. IEEE Trans-
actions on Vehicular Technology, vol. 61, no. 8, pp. 3420–3429, October 2012.

4. Simões M.G., Roche R., Kyriakides E., Suryanarayanan S., Blunier B., McBee K.,
Nguyen P., Ribeiro P. and Miraoui A., A Comparison of Smart Grid Technologies
and Progresses in Europe and the U.S. IEEE Transactions on Industry Applications,
vol. 48, no. 4., pp.1154–1162, July-August 2012.

Book Chapter
1. [In press] Roche R., Lauri F., Blunier B., Miraoui A. and Koukam A., Multi-Agent
Technology in Power Systems, in Chakraborty S., Simões M.G., Kramer W.E. (eds.),
Power Electronics for Renewable and Distributed Energy Systems, Springer.

Patent
1. Daggag M., Roche R., Idoumghar L., Blunier B., Miraoui A. and Koukam A., Sys-
tem and method for controlling an electrical energy production installation. Patent
no. WO/2012/143424 (international) / FR1153369 (France).

129
130 Appendix B. Publications

International Conferences and Workshops

1. [Submitted] Lauri F., Roche R., Basso G., Zhu J., and Hilaire V., Distributed
Microgrid Control Using Multi-Agent Reinforcement Learning.
2. [Submitted] Couraud B., Roche R., and Traill B., A Distribution Load Forecasting
Methodology Based on Primary Substations SCADA Data.
3. [Submitted] Giràldez J., Roche R., Suryanarayanan S., and Zimmerle D., A Linear
Programming Methodology to Quantify the Impact of PHEVs with V2G Capabilities
on Distribution Systems.
4. Roche R., Idoumghar L., Blunier B., and Miraoui A., Imperialist Competitive Algo-
rithm for Dynamic Optimization of Economic Dispatch in Power Systems, in Hao
J.-K. et al. (Eds.): Artificial Evolution, Lecture Notes in Computer Science (LNCS),
vol. 7401, pp. 217–228. Springer, Heidelberg. 2012.
5. Hansen T., Roche R., Suryanarayanan S., Siegel H.J., Zimmerle D., Young P.M.,
and Maciejewski A.A., A Proposed Framework for Heuristic Approaches to Resource
Allocation in the Emerging Smart Grid. IEEE PES International Conference on
Power Systems Technology (POWERCON 2012). pp. 1–6. October 2012.
6. Roche R., Natarajan S., Bhattacharyya A., and Suryanarayanan S., A Framework
for Co-simulation of AI Tools with a Power Systems Analysis Software. 1st In-
ternational Workshop on Intelligent Agent Technology, Power Systems and Energy
Markets (IATEM 2012). pp. 350–354. September 2012.
7. Ravey A., Roche R., Blunier B. and Miraoui A., Combined Optimal Sizing and En-
ergy Management of Hybrid Electric Vehicles. IEEE Transportation Electrification
Conference and Expo (ITEC 2012). pp. 1–6. June 2012.
8. Basso G., Hilaire V., Lauri F., Roche R., Cossentino M., A MAS-based simulator for
the prototyping of Smart Grids. 9th European Workshop on Multi-Agent Systems
(EUMAS 2011). November 2011.
9. Roche R., Idoumghar L., Blunier B. and Miraoui A., Imperialist Competitive Al-
gorithm for Dynamic Optimization of Economic Dispatch in Power Systems. In-
ternational Conference on Artificial Evolution (EA 2011). pp. 375–386. October
2011.
10. Roche R., Idoumghar L., Blunier B. and Miraoui A., Optimized Fuel Cell Array
Energy Management Using Multi-Agent Systems. 46th IEEE Industry Applications
Annual Meeting (IAS 2011). pp. 1–8. October 2011.
11. Simões M.G., Roche R., Kyriakides E., Miraoui A., Blunier B., McBee K., Surya-
narayanan S., Nguyen P. and Ribeiro P., Smart-Grid Technologies and Progress
in Europe and the USA. IEEE Energy Conversion Congress & Exposition (ECCE
2011). pp. 383–390. September 2011.
12. Roche R., Blunier B. and Miraoui A., Multi-Agent Systems For Grid Energy Manage-
ment: A Short Review. 36th Annual Conference of the IEEE Industrial Electronics
Society (IECON 2010). pp. 3341–3346. November 2010.
 131

National Conferences and Symposia

1. Roche R., Idoumghar L., Blunier B. and Miraoui A., Algorithmes hybrides pour
la gestion intelligente de l’énergie dans les smart grids. 7èmes Journées Franco-
phones Planification, Décision, et Apprentissage pour la conduite de systèmes (JF-
PDA 2012). May 2012.
2. Chérin N., Idoumghar L., Siarry P., Roche R. and Blunier B., Métaheuristique hy-
bride pour les problèmes d’optimisation continue. 13e congrès annuel de la Société
Française de Recherche Opérationnelle et d’Aide à la Décision (ROADEF 2012).
April 2012.
3. Roche R., Application de métaheuristiques pour la gestion optimale de l’énergie dans
les réseaux électriques intelligents. Conférence des Jeunes Chercheurs en Génie Elec-
trique (JCGE 2011). December 2011.
4. Roche R., Blunier B., Miraoui A., Gestion intelligente de l’énergie dans les smart
grids : combiner flexibilité et efficacité. IngéDoc seminar for young researchers,
UTBM, Belfort, France. December 2011.

Oral Communications
1. Roche R., An aggregator-based architecture and simulator for residential demand
response. UTBM, Belfort, France. July 2012.
2. Roche R., A Framework for a Multi-Agent Distribution Management System, Clean
Energy Supercluster & Cenergy Expo 2012. Colorado State University, Fort Collins,
USA. April 2012.
3. Roche R., Design and Development of an Intelligent, Flexible and Integrated Energy
Management System for Smart Microgrids. Colorado State University, Fort Collins,
USA. February 2012.
4. Roche R., Blunier B., Miraoui A., Smart grids et smart meters : Vers des réseaux
électriques décentralisés et intelligents. Vers une ville post-carbone symposium, with
the French Academy of Technology, UTBM, Belfort, France. April 2010.
 Bibliography

[1] J. Miller, “A modest proposal for kickstarting smartgrid with smarter devices:
Lessons from wireless spectrum regulatory and policy models for procotols and
devices necessary for realizing the smarter grid,” 2012, available at SSRN:
[Link]

[2] M. Simões, R. Roche, E. Kyriakides, S. Suryanarayanan, B. Blunier, K. McBee,

P. Nguyen, P. Ribeiro, and A. Miraoui, “A comparison of smart grid technologies
and progresses in europe and the U.S.” IEEE Transactions on Industry Applications,
vol. 48, no. 4, pp. 1154–1162, 2012.

[3] International Energy Agency, “Technology roadmap smart grids,” 2011, accessed
September 30, 2012. [Online]. Available: [Link]
freepublications/publication/smartgrids_roadmap.pdf

[4] S. Suryanarayanan, F. Mancilla-David, J. Mitra, and Y. Li, “Achieving the smart

grid through customer-driven microgrids supported by energy storage,” in IEEE
International Conference on Industrial Technology (ICIT), 2010, pp. 884–890.

[5] K. Cory and B. Swezey, “Renewable portfolio standards in the states: Balancing
goals and implementation strategies,” 2007, NREL/TP-670-41409. [Online].
Available: [Link]

[6] European Renwable Energy Council, Greenpeace International, “Renewables 24/7:

Infrastructure needed to save the climate,” November 2009, accessed October 6,
2012. [Online]. Available: [Link]
Publications/global%20energy%20grid%[Link]

[7] RTE France - Réseau de Transport d’Électricité. (2012) Customers portal.

Accessed June 25, 2012. [Online]. Available: [Link]
an/visiteurs/vie/[Link]

[8] U.S. Energy Information Administration, “International energy outlook 2011,”

2011, DOE/EIA-0484(2011). [Online]. Available: [Link]
ieo/pdf/0484(2011).pdf

[9] D. Coll-Mayor, M. Paget, and E. Lightner, “Future intelligent power grids: Analysis
of the vision in the European Union and the United States,” Energy Policy, vol. 35,
no. 4, pp. 2453–2465, 2007.

[10] T. Chen, “Stuxnet, the real start of cyber warfare? [editor’s note],” IEEE Network,
vol. 24, no. 6, pp. 2–3, 2010.

133
134 Bibliography

[11] W. Boothby, W. von Heinegg, J. Michael, M. Schmitt, and T. Wingfield, “When is

a cyberattack a use of force or an armed attack?” Computer, pp. 82–84, 2012.
[12] A. Stefanov and C.-C. Liu, “Cyber-power system security in a smart grid environ-
ment,” in IEEE PES Innovative Smart Grid Technologies (ISGT), jan. 2012, pp. 1
–3.
[13] National Science Foundation. (2012) Cyber-physical systems (CPS). Accessed
September 10, 2012. [Online]. Available: [Link]
jsp?pims_id=503286
[14] 110th U.S. Congress, “Energy independence and security act of 2007,”
2007, publication 110-140. [Online]. Available: [Link]
BILLS-110hr6enr/pdf/[Link]
[15] SmartGrids European Technology Platform. (2012) The SmartGrids European
Technology Platform. Accessed August 15, 2012. [Online]. Available: [Link]
[Link]/web/node/81
[16] A. Carvallo and J. Cooper, The Advanced Smart Grid: Edge Power Driving Sus-
tainability. Artech House, 2010.
[17] F. Rahimi and A. Ipakchi, “Demand response as a market resource under the smart
grid paradigm,” IEEE Transactions on Smart Grid, vol. 1, no. 1, pp. 82–88, 2010.
[18] M. Amin, “Challenges in reliability, security, efficiency, and resilience of energy in-
frastructure: Toward smart self-healing electric power grid,” in IEEE Power and
Energy Society General Meeting - Conversion and Delivery of Electrical Energy in
the 21st Century, 2008, pp. 1–5.
[19] S. Duquennoy, G. Grimaud, and J. Vandewalle, “The web of things: interconnecting
devices with high usability and performance,” in IEEE International Conference on
Embedded Software and Systems (ICESS), 2009, pp. 323–330.
[20] E. Palm. (2009, March) Enernet - A smart-grid vision from a net tycoon. CNET
News. Accessed September 10, 2012. [Online]. Available: [Link]
8301-11128_3-[Link]
[21] Agence de l’Environnement et de la Maitrise de l’Energie (ADEME),
“Roadmap for smart grids and electricity systems integrating renewable energy
sources,” 2009, accessedSeptember 10, 2012. [Online]. Available: [Link]
[Link]/url?sa=t&rct=j&q=&esrc=s&source=web&cd=1&ved=0CGsQFjAA&
url=http%3A%2F%[Link]%2Fservlet%2FgetBin%3Fname
%[Link]&ei=d_
wrUKvvKMn8igL2tIG4Dw&usg=AFQjCNHx5_0-gSeDufpzZuj71Ry-ahHUpg
[22] NIST, “Framework and roadmap for smart grid interoperability standards, release
2.0,” NIST, 2012, Special Publication 1108R2.
[23] V. Terzija, G. Valverde, D. Cai, P. Regulski, V. Madani, J. Fitch, S. Skok, M. Be-
govic, and A. Phadke, “Wide-area monitoring, protection, and control of future
electric power networks,” Proceedings of the IEEE, vol. 99, no. 1, pp. 80–93, 2011.
Bibliography 135

[24] N. Hingorani, L. Gyugyi, and M. El-Hawary, Understanding FACTS: concepts and

technology of flexible AC transmission systems. IEEE Press New York, 2000.
[25] N. Kirby, X. Lie, M. Luckett, and W. Siepmann, “HVDC transmission for large
offshore wind farms,” Power Engineering Journal, vol. 16, no. 3, pp. 135–141, 2002.
[26] U.S. Department of Energy – Office of Electricity Deliver and Energy Reliability
– Smart Grid R&D Program, “Microgrid workshop report,” 2011. [Online].
Available: [Link]
%20August%[Link]
[27] F. Trieb and H. Müller-Steinhagen, “The DESERTEC Concept-Sustainable Elec-
tricity and Water for Europe, Middle East and North Africa,” Whitebook of TREC
and Club of Rome—Clean Power from Deserts, pp. 23–43, 2007.
[28] J. Klimstra and M. Hotakainen, Smart power generation: the future of electricity
production. Avain, 2012.
[29] A. Berizzi, “The Italian 2003 blackout,” in IEEE Power Engineering Society General
Meeting, 2004, pp. 1673–1679.
[30] D. White, A. Roschelle, P. Peterson, D. Schlissel, B. Biewald, and W. Steinhurst,
“The 2003 blackout: solutions that won’t cost a fortune,” The Electricity Journal,
vol. 16, no. 9, pp. 43–53, 2003.
[31] [Link]. (2012) Project information and locations. Accessed September
10, 2012. [Online]. Available: [Link]
information
[32] G. Zachary, “Saving smart meters from a backlash,” IEEE Spectrum, vol. 48, no. 8,
p. 8, august 2011.
[33] N. Hadjsaïd and J. Sabonnadière, Power Systems and Restructuring. ISTE, 2009.
[34] N. Hadjsaid, La distribution d’énergie électrique en présence de production décen-
tralisée (série Génie électrique). Hermès – Lavoisier, 2010.
[35] M. Aoki and G. Rothwell, “A comparative institutional analysis of the Fukushima
nuclear disaster: Lessons and policy implications,” 2012, accessed September 10,
2012. [Online]. Available: [Link]
[36] U.S. Department of Energy, “Benefits of demand response in electricity markets
and recommendations for achieving them,” 2006, Report to the U.S. Congress
persuant to Section 1252 of the Energy Policy Act of 2005. [Online]. Available:
[Link]
[37] T. Sezi and B. Duncan, “New intelligent electronic devices change the structure of
power distribution systems,” in IEEE IAS Annual Meeting, vol. 2, 1999, pp. 944–952.
[38] R. Roche, S. Natarajan, A. Bhattacharyya, and S. Suryanarayanan, “A framework
for co-simulation of AI tools with a power systems analysis software,” in 1st In-
ternational Workshop on Intelligent Agent Technology, Power Systems and Energy
Markets (IATEM 2012), 2012, pp. 350–354.
136 Bibliography

[39] R. Roche, L. Idoumghar, S. Suryanarayanan, M. Daggag, C. Solacolu, and A. Mi-

raoui, “A flexible and efficient multi-agent gas turbine power plant energy manage-
ment system with economic and environmental constraints,” Applied Energy, vol.
101, pp. 644–654, 2013.
[40] T. Overbye and J. Weber, “Visualizing the electric grid,” IEEE Spectrum, vol. 38,
no. 2, pp. 52 –58, 2001.
[41] M. Shahidehpour and Y. Wang, Communication and control in electric power sys-
tems: applications of parallel and distributed processing. Wiley-IEEE Press, 2003.
[42] C. Petermann, S. Ben Amor, and A. Bui, “A pretopological multi-agents based model
for an efficient and reliable smart grid simulation,” in International Conference on
Artificial Intelligence (ICAI), 2012.
[43] J. Miller and S. Page, Complex adaptive systems: An introduction to computational
models of social life. Princeton Univ Press, 2007.
[44] P. Cilliers, Complexity and Postmodernism: Understanding Complex Systems.
Routledge, 1998.
[45] A. Wildberger, “Autonomous adaptive agents for distributed control of the electric
power grid in a competitive electric power industry,” in IEEE International Con-
ference on Knowledge-Based Intelligent Electronic Systems (KES), vol. 1, 1997, pp.
2–11.
[46] M. Wooldridge and G. Weiss, Multi-Agent Systems. The MIT Press, 1999.
[47] J. Ferber, Multi-Agent Systems: An Introduction to Artificial Intelligence. Addison-
Wesley, 1999.
[48] R. Roche, B. Blunier, A. Miraoui, V. Hilaire, and A. Koukam, “Multi-agent systems
for grid energy management: A short review,” in 36th Annual Conference on IEEE
Industrial Electronics Society (IECON), 2010, pp. 3341–3346.
[49] S. Russell and P. Norvig, Artificial Intelligence, A modern approach. Prentice-Hall,
1995.
[50] G. Serugendo, “Self-organisation and emergence in multi-agent systems,” The
Knowledge Engineering Review, vol. 20, no. 2, pp. 165–189, 2005.
[51] S. Srivastava, S. Suryanarayanan, P. Ribeiro, D. Cartes, and M. Stcurer, “A con-
ceptual power quality monitoring technique based on multi-agent systems,” in Pro-
ceedings of the 37th Annual North American Power Symposium (NAPS), 2005, pp.
358–363.
[52] M. Baran and I. El-Markabi, “A multiagent-based dispatching scheme for distributed
generators for voltage support on distribution feeders,” IEEE Transactions on Power
Systems, vol. 22, no. 1, pp. 52–59, 2007.
[53] H. F. Wang, H. Li, and H. Chen, “Coordinated secondary voltage control to eliminate
voltage violations in power system contingencies,” IEEE Transactions on Power
Systems, vol. 18, no. 2, pp. 588–595, 2003.
Bibliography 137

[54] T. Nagata and H. Sasaki, “A multi-agent approach to power system restoration,”

IEEE Transactions on Power Systems, vol. 17, no. 2, pp. 457–462, 2002.
[55] J. Solanki, S. Khushalani, and N. Schulz, “A multi-agent solution to distribution
systems restoration,” IEEE Transactions on Power Systems, vol. 22, no. 3, pp.
1026–1034, 2007.
[56] E. Davidson, S. McArthur, J. McDonald, T. Cumming, and I. Watt, “Applying
multi-agent system technology in practice: automated management and analysis of
SCADA and digital fault recorder data,” IEEE Transactions on Power Systems,
vol. 21, no. 2, pp. 559–567, 2006.
[57] S. McArthur, E. Davidson, V. Catterson, A. Dimeas, N. Hatziargyriou, F. Ponci,
and T. Funabashi, “Multi-Agent Systems for Power Engineering Applications - Part
I: Concepts, Approaches, and Technical Challenges,” IEEE Transactions on Power
Systems, vol. 22, no. 4, pp. 1743–1752, 2007.
[58] ——, “Multi-Agent Systems for Power Engineering Applications - Part II: Technolo-
gies, Standards, and Tools for Building Multi-agent Systems,” IEEE Transactions
on Power Systems, vol. 22, no. 4, pp. 1753–1759, nov. 2007.
[59] J. Lagorse, D. Paire, and A. Miraoui, “A multi-agent system for energy management
of distributed power sources,” Renewable Energy, vol. 35, no. 1, pp. 174–182, 2010.
[60] J. Lagorse, M. Simoes, and A. Miraoui, “A multiagent fuzzy-logic-based energy man-
agement of hybrid systems,” IEEE Transactions on Industry Applications, vol. 45,
no. 6, pp. 2123–2129, 2009.
[61] A. Dimeas and N. Hatziargyriou, “Operation of a multiagent system for microgrid
control,” IEEE Transactions on Power Systems, vol. 20, no. 3, pp. 1447–1455, 2005.
[62] ——, “A multiagent system for microgrids,” in IEEE Power Engineering Society
General Meeting, vol. 1, 2004, pp. 55–58.
[63] N. Hatziargyriou, A. Dimeas, A. Tsikalakis, J. Lopes, G. Karniotakis, and J. Oyarz-
abal, “Management of microgrids in market environment,” in International Confer-
ence on Future Power Systems, 2005, p. 7.
[64] A. Dimeas and N. Hatziargyriou, “Multi-agent reinforcement learning for micro-
grids,” in IEEE Power and Energy Society General Meeting, 2010, pp. 1–8.
[65] T. Logenthiran, D. Srinivasan, A. Khambadkone, and H. N. Aung, “Multiagent
system for real-time operation of a microgrid in real-time digital simulator,” IEEE
Transactions on Smart Grid, vol. 3, no. 2, pp. 925–933, 2012.
[66] T. Logenthiran, D. Srinivasan, and D. Wong, “Multi-agent coordination for DER in
microgrid,” in IEEE International Conference on Sustainable Energy Technologies
(ICSET 2008), 2008, pp. 77–82.
[67] T. Logenthiran, D. Srinivasan, A. Khambadkone, and H. Aung, “Scalable multi-
agent system (MAS) for operation of a microgrid in islanded mode,” in Joint Inter-
national Conference on Power Electronics, Drives and Energy Systems (PEDES),
2010, pp. 1–6.
138 Bibliography

[68] ——, “Multi-agent system (MAS) for short-term generation scheduling of a micro-
grid,” in IEEE International Conference on Sustainable Energy Technologies (IC-
SET), 2010, pp. 1–6.
[69] T. Logenthiran and D. Srinivasan, “Multi-agent system for managing a power dis-
tribution system with plug-in hybrid electrical vehicles in smart grid,” in IEEE PES
Innovative Smart Grid Technologies - India (ISGT India), 2011, pp. 346–351.
[70] T. Logenthiran, D. Srinivasan, and T. Z. Shun, “Multi-agent system for demand side
management in smart grid,” in IEEE International Conference on Power Electronics
and Drive Systems (PEDS), 2011, pp. 424–429.
[71] S. Rahman, M. Pipattanasomporn, and Y. Teklu, “Intelligent distributed au-
tonomous power systems (IDAPS),” in IEEE Power Engineering Society General
Meeting, 2007, pp. 1–8.
[72] M. Pipattanasomporn, H. Feroze, and S. Rahman, “Multi-agent systems in a dis-
tributed smart grid: Design and implementation,” in IEEE/PES Power Systems
Conference and Exposition (PSCE), 2009, pp. 1–8.
[73] J. Kok, B. Roossien, P. MacDougall, O. Pruissen, G. Venekamp, I. Kamphuis,
J. Laarakkers, and C. Warmer, “Dynamic pricing by scalable energy management
systems - Field experiences and simulation results using PowerMatcher,” in IEEE
Power and Energy Society General Meeting, 2012.
[74] F. Bliek, A. van den Noort, B. Roossien, R. Kamphuis, J. de Wit, J. van der Velde,
and M. Eijgelaar, “PowerMatching City, a living lab smart grid demonstration,” in
IEEE PES Innovative Smart Grid Technologies Conference - Europe (ISGT Europe),
2010, pp. 1–8.
[75] J. Kok, M. Scheepers, and I. Kamphuis, “Intelligence in electricity networks for
embedding renewables and distributed generation,” Intelligent Infrastructures, pp.
179–209, 2010.
[76] J. K. Kok, C. J. Warmer, and I. G. Kamphuis, “PowerMatcher: multiagent control
in the electricity infrastructure,” in Proceedings of the international joint conference
on autonomous agents and multiagent systems (AAMAS), 2005, pp. 75–82.
[77] M. Hommelberg, C. Warmer, I. Kamphuis, J. Kok, and G. Schaeffer, “Distributed
control concepts using multi-agent technology and automatic markets: An indis-
pensable feature of smart power grids,” in IEEE Power Engineering Society General
Meeting, 2007, pp. 1–7.
[78] I. Praca, C. Ramos, Z. Vale, and M. Cordeiro, “MASCEM: a multiagent system that
simulates competitive electricity markets,” IEEE Intelligent Systems, vol. 18, no. 6,
pp. 54–60, nov-dec 2003.
[79] Z. Vale, T. Pinto, H. Morais, I. Praca, and P. Faria, “VPP’s multi-level negotiation
in smart grids and competitive electricity markets,” in IEEE Power and Energy
Society General Meeting, 2011, pp. 1–8.
[80] Infotility. (2012) Infotility products. Accessed October 4, 2012. [Online]. Available:
[Link]
Bibliography 139

[81] Foundation for Intelligent Physical Agents (FIPA). (2012) Accessed September 10,
2012. [Online]. Available: [Link]
[82] FIPA ACL Message Structure Specification, Foundation For Intelligent Physical
Agents Std. SC00 061G, 2002.
[83] FIPA SL Content Language Specification, Foundation For Intelligent Physical
Agents Std. SC00 008I, 2002.
[84] V. Catterson, P. Baker, E. Davidson, and S. McArthur. (2010) An upper ontology
for power engineering applications. Accessed July 1, 2011. [Online]. Available:
[Link]
[85] FIPA Agent Management Specification, Foundation For Intelligent Physical Agents
Std. SC00 023K, 2004.
[86] B. Horling and V. Lesser, “A survey of multi-agent organizational paradigms,” The
Knowledge Engineering Review, vol. 19, no. 4, pp. 281–316, 2004.
[87] J. Yen, Y. Yan, B. Wang, P. Sin, and F. Wu, “Multi-agent coalition formation in
power transmission planning,” in Proceedings of the Thirty-First Hawaii Interna-
tional Conference on System Sciences, vol. 4, 1998, pp. 433–443.
[88] FIPA Request Interaction Protocol Specification, Foundation For Intelligent Physical
Agents Std. XC00 026F, 2001.
[89] FIPA Contract Net Interaction Protocol Specification, Foundation For Intelligent
Physical Agents Std. SC00 029H, 2002.
[90] G. Sheble, Computational auction mechanisms for restructured power industry op-
eration. Springer, 1999.
[91] FIPA English Auction Interaction Protocol Specification, Foundation For Intelligent
Physical Agents Std. XC00 031F, 2001.
[92] FIPA Dutch Auction Interaction Protocol Specification, Foundation For Intelligent
Physical Agents Std. XC00 032F, 2001.
[93] L. Ausubel and P. Milgrom, “The lovely but lonely Vickrey auction,” Combinatorial
Auctions, pp. 17–40, 2006.
[94] A. Motto, F. Galiana, A. Conejo, and M. Huneault, “On walrasian equilibrium for
pool-based electricity markets,” IEEE Transactions on Power Systems, vol. 17, no. 3,
pp. 774–781, 2002.
[95] J. Pitt, L. Kamara, M. Sergot, and A. Artikis, “Voting in multi-agent systems,” The
Computer Journal, vol. 49, no. 2, pp. 156–170, 2006.
[96] FIPA Brokering Interaction Protocol Specification, Foundation For Intelligent Phys-
ical Agents Std. SC00 033H, 2002.
[97] Wikipedia, “Comparison of agent-based modeling software — Wikipedia,
The Free Encyclopedia,” 2012, accessed August 7, 2012. [Online]. Avail-
able: [Link]
modeling_software&oldid=505030846
140 Bibliography

[98] Foundation for Intelligent Physical Agents. (2003) Publicly available agent platform
implementations. Accessed August 7, 2012. [Online]. Available: [Link]
org/resources/[Link]

[99] F. Bellifemine, G. Caire, and D. Greenwood, Developing multi-agent systems with

JADE. Wiley, 2007.

[100] K. Warwick, A. Ekwue, and R. Aggarwal, Artificial intelligence techniques in power

systems. IET, 1997.

[101] G. Venayagamoorthy, “Potentials and promises of computational intelligence for

smart grids,” in IEEE Power & Energy Society General Meeting (PES’09), 2009,
pp. 1–6.

[102] J. de Haan, P. Nguyen, W. Kling, and P. Ribeiro, “Social interaction interface for
performance analysis of smart grids,” in IEEE First International Workshop on Grid
Modeling and Simulation (SGMS), 2011, pp. 79–83.

[103] T. Godfrey, S. Mullen, R. Dugan, C. Rodine, D. Griffith, and N. Golmie, “Modeling

smart grid applications with co-simulation,” in First IEEE International Conference
on Smart Grid Communications (SmartGridComm), 2010, pp. 291–296.

[104] H. Lin, S. Sambamoorthy, S. Shukla, J. Thorp, and L. Mili, “Power system and
communication network co-simulation for smart grid applications,” in IEEE PES
Innovative Smart Grid Technologies (ISGT), 2011, pp. 1–6.

[105] V. Liberatore and A. Al-Hammouri, “Smart grid communication and co-simulation,”

in IEEE Energytech, 2011, pp. 1–5.

[106] J. Gomez-Gualdron and M. Velez-Reyes, “Simulating a multi-agent based self-

reconfigurable electric power distribution system,” in IEEE Workshops on Com-
puters in Power Electronics (COMPEL’06.), 2006, pp. 1–7.

[107] C. J. Bankier, “GridIQ - A Test Bed for Smart Grid Agents,” Master’s thesis,
University of Queensland, 2010. [Online]. Available: [Link]

[108] D. Chassin, K. Schneider, and C. Gerkensmeyer, “GridLAB-D: An open-source

power systems modeling and simulation environment,” in IEEE/PES Transmission
and Distribution Conference and Exposition (T&D 2008), 2008, pp. 1–5.

[109] PowerWorld Corporation. (2012) PowerWorld, the visual approach to electric

power systems. Accessed September 21, 2012. [Online]. Available: [Link]
[Link]/

[110] MathWorks. (2012) MATLAB and Simulink for technical computing. Accessed
September 21, 2012. [Online]. Available: [Link]

[111] PowerWorld Corporation, “User’s Guide - Simulator Version 16,” 2011. [Online].
Available: [Link]

[112] R. Roche. (2012) Co-simulation framework for JADE-Matlab-PowerWorld

Simulator. [Online]. Available: [Link]
Bibliography 141

[113] C. De Jonghe, E. Delarue, R. Belmans, and W. D’haeseleer, “Determining opti-

mal electricity technology mix with high level of wind power penetration,” Applied
Energy, vol. 88, no. 6, pp. 2231–2238, 2011.
[114] N. Sinha, R. Chakrabarti, and P. Chattopadhyay, “Evolutionary programming tech-
niques for economic load dispatch,” IEEE Transactions on Evolutionary Computa-
tion, vol. 7, no. 1, pp. 83–94, 2003.
[115] M. Abido, “Environmental/economic power dispatch using multiobjective evolution-
ary algorithms,” IEEE Transactions on Power Systems, vol. 18, no. 4, pp. 1529–1537,
2003.
[116] ——, “Multiobjective evolutionary algorithms for electric power dispatch problem,”
IEEE Transactions on Evolutionary Computation, vol. 10, no. 3, pp. 315–329, 2006.
[117] S. Kazarlis, A. Bakirtzis, and V. Petridis, “A genetic algorithm solution to the unit
commitment problem,” IEEE Transactions on Power Systems, vol. 11, no. 1, pp.
83–92, 1996.
[118] N. Padhy, “Unit commitment - A bibliographical survey,” IEEE Transactions on
Power Systems, vol. 19, no. 2, pp. 1196–1205, 2004.
[119] S. K. Yee, J. Milanovic, and F. Hughes, “Overview and comparative analysis of gas
turbine models for system stability studies,” IEEE Transactions on Power Systems,
vol. 23, no. 1, pp. 108–118, 2008.
[120] S. Simani, “Identification and fault diagnosis of a simulated model of an industrial
gas turbine,” IEEE Transactions on Industrial Informatics, vol. 1, no. 3, pp. 202–
216, 2005.
[121] E.-H. T. El-Shirbeeny and M. K. Kadum, “Communication schemes for electric
energy management,” Applied Energy, vol. 24, no. 4, pp. 277–286, 1986.
[122] V. V. Silva, W. Khatib, and P. J. Fleming, “Performance optimization of gas turbine
engine,” Engineering Applications of Artificial Intelligence, vol. 18, no. 5, pp. 575–
583, 2005.
[123] A. K. Kralj and P. Glavic, “Optimization of a gas turbine in the methanol process,
using the NLP model,” Applied Thermal Engineering, vol. 27, no. 11-12, pp. 1799–
1805, 2007.
[124] F. Brooks, “GE gas turbine performance characteristics,” GE Power Systems,
Schenectady, NY, Tech. Rep. GER3567h, 2000. [Online]. Available: [Link]
[Link]/prod_serv/products/tech_docs/en/downloads/[Link]
[125] M. Boyce, Gas turbine engineering handbook. Butterworth-Heinemann, 2011.
[126] L. Freris and D. Infield, Renewable energy in power systems. Wiley, 2008.
[127] A. Wood and B. Wollenberg, Power generation, operation, and control. Wiley,
1996.
[128] J. Petek and P. Hamilton, “Performance monitoring for gas turbines,” ORBIT,
vol. 25, no. 1, pp. 64–74, 2005.
142 Bibliography

[129] L. Davis and S. Black, “Dry Low NOx Combustion Systems for GE Heavy-Duty
Gas Turbines,” GE Power Systems, Tech. Rep. GER-3568G, 2000.
[130] U.S. Environmental Protection Agency, “Alternative Control Techniques Document
– NOx Emissions from Stationary Gas Turbines,” 1993, EPA-453/R-93-007.
[Online]. Available: [Link]
[131] R. Hooshmand, M. Parastegari, and M. Morshed, “Emission, reserve and economic
load dispatch problem with non-smooth and non-convex cost functions using the
hybrid bacterial foraging-Nelder–Mead algorithm,” Applied Energy, vol. 89, no. 1,
pp. 443–453, 2012.
[132] V. Vahidinasab and S. Jadid, “Multiobjective environmental/techno-economic ap-
proach for strategic bidding in energy markets,” Applied Energy, vol. 86, no. 4, pp.
496–504, 2009.
[133] GE Energy. (2012) 9E Heavy Duty Gas Turbine. Accessed March 7, 2012.
[Online]. Available: [Link]
gas_turbines_heavy_duty/9e_heavy_duty_gas_turbine.jsp
[134] M. Albadi and E. El-Saadany, “Overview of wind power intermittency impacts on
power systems,” Electric Power Systems Research, vol. 80, no. 6, pp. 627–632, 2010.
[135] A. Botterud, J. Wang, V. Miranda, and R. J. Bessa, “Wind Power Forecasting in
U.S. Electricity Markets,” The Electricity Journal, vol. 23, no. 3, pp. 71–82, 2010.
[136] M. Lei, L. Shiyan, J. Chuanwen, L. Hongling, and Z. Yan, “A review on the forecast-
ing of wind speed and generated power,” Renewable and Sustainable Energy Reviews,
vol. 13, no. 4, pp. 915–920, 2009.
[137] J. Talaq, F. El-Hawary, and M. El-Hawary, “A summary of environmental/economic
dispatch algorithms,” IEEE Transactions on Power Systems, vol. 9, no. 3, pp. 1508–
1516, 1994.
[138] C.-M. Huang and Y.-C. Huang, “A novel approach to real-time economic emission
power dispatch,” IEEE Transactions on Power Systems, vol. 18, no. 1, pp. 288–294,
2003.
[139] C. Coello, “A comprehensive survey of evolutionary-based multiobjective optimiza-
tion techniques,” Knowledge and Information systems, vol. 1, no. 3, pp. 129–156,
1999.
[140] M. Shao and W. Jewell, “CO2 emission-incorporated ac optimal power flow and its
primary impacts on power system dispatch and operations,” in IEEE Power and
Energy Society General Meeting, 2010, pp. 1–8.
[141] M. Khalid and A. Savkin, “A model predictive control approach to the problem of
wind power smoothing with controlled battery storage,” Renewable Energy, vol. 35,
no. 7, pp. 1520–1526, 2010.
[142] S. Teleke, M. Baran, A. Huang, S. Bhattacharya, and L. Anderson, “Control strate-
gies for battery energy storage for wind farm dispatching,” IEEE Transactions on
Energy Conversion, vol. 24, no. 3, pp. 725–732, 2009.
Bibliography 143

[143] OECD, “OECD Environmental Performance Reviews: Norway 2011,” Tech. Rep.,
2011, accessed September 10, 2012. [Online]. Available: [Link]
norway/[Link]

[144] J. Sumner, L. Bird, and H. Smith, “Carbon taxes: A review of experience and policy
design considerations,” NREL, Tech. Rep. NREL/TP-6A2-47312, 2009.

[145] Federal Energy Regulatory Commission. (2012) Emissions Allowances - GHG,

SOx & NOx. Accessed June 25, 2012. [Online]. Available: [Link]
market-oversight/othr-mkts/[Link]

[146] U.S. Energy Information Administration, “Annual energy outlook 2012,” 2012.
[Online]. Available: [Link]

[147] R. Green II, L. Wang, and M. Alam, “The impact of plug-in hybrid electric vehicles
on distribution networks: A review and outlook,” Renewable and Sustainable Energy
Reviews, vol. 15, no. 1, pp. 544–553, 2011.

[148] S. Kaplan, “Power plants: Characteristics and costs,” 2008, Congressional Research
Service RL34746. [Online]. Available: [Link]
pdf

[149] A. Kumar, S. Jain, and N. Bansal, “Disseminating energy-efficient technologies: a

case study of compact fluorescent lamps (CFLs) in India,” Energy Policy, vol. 31,
no. 3, pp. 259–272, 2003.

[150] R. Brown, “U.S. building-sector energy efficiency potential,” Lawrence Berkeley

National Laboratory, Tech. Rep. LBNL-1096E, 2008. [Online]. Available: http://
[Link]/info/[Link]

[151] A. Rosenfeld, D. Bulleit, and R. Peddie, “Smart meters and spot pricing: exper-
iments and potential,” IEEE Technology and Society Magazine, vol. 5, no. 1, pp.
23–28, 1986.

[152] Federal Energy Regulatory Commission, “A national assessment of demand

response potential,” 2009. [Online]. Available: [Link]
staff-reports/[Link]

[153] J. Osborne and D. Warrier, “A Primer On Demand Response – The Power

Grid: Evolving from a Dumb Network to a Smart Grid,” Thomas Weisel
Partners, Tech. Rep., 2007. [Online]. Available: [Link]
com/internetevolution/Thomas_Weisel_Demand_Response.pdf

[154] Federal Energy Regulatory Commission, “Assessment of demand response and

advanced metering,” 2006, AD06-2-000. [Online]. Available: [Link]
legal/staff-reports/[Link]

[155] E. McKenna, K. Ghosh, and M. Thomson, “Demand response in low-carbon

power systems: a review of residential electrical demand response projects,” in
2nd International Conference on Microgeneration and Related Technologies, 2011.
[Online]. Available: [Link]
144 Bibliography

[156] AEIC Research Load Committee, “Demand response measurement & ver-
ification,” 2009. [Online]. Available: [Link]
[Link]
[157] M. Albadi and E. El-Saadany, “A summary of demand response in electricity mar-
kets,” Electric Power Systems Research, vol. 78, no. 11, pp. 1989–1996, 2008.
[158] S. Shao, M. Pipattanasomporn, and S. Rahman, “Demand response as a load shaping
tool in an intelligent grid with electric vehicles,” IEEE Transactions on Smart Grid,
2012, to appear.
[159] ——, “Grid integration of electric vehicles and demand response with customer
choice,” IEEE Transactions on Smart Grid, vol. 3, no. 1, pp. 543–550, 2012.
[160] K. Schneider, J. Fuller, and D. Chassin, “Analysis of distribution level residential de-
mand response,” in IEEE/PES Power Systems Conference and Exposition (PSCE),
2011, pp. 1–6.
[161] H. Sæle and O. Grande, “Demand response from household customers: Experiences
from a pilot study in norway,” IEEE Transactions on Smart Grid, vol. 2, no. 1, pp.
102–109, 2011.
[162] A. Thomas, P. Jahangiri, D. Wu, C. Cai, H. Zhao, D. Aliprantis, and T. L., “In-
telligent residential air-conditioning system with smart-grid functionality,” IEEE
Transactions on Smart Grid, 2012, to appear.
[163] M. Pipattanasomporn, M. Kuzlu, and S. Rahman, “An algorithm for intelligent
home energy management and demand response analysis,” IEEE Transactions on
Smart Grid, to appear.
[164] M. Mallette and G. Venkataramanan, “Financial incentives to encourage demand
response participation by plug-in hybrid electric vehicle owners,” in IEEE Energy
Conversion Congress and Exposition (ECCE), 2010, pp. 4278–4284.
[165] S. Lu, N. Samaan, R. Diao, M. Elizondo, C. Jin, E. Mayhorn, Y. Zhang, and
H. Kirkham, “Centralized and decentralized control for demand response,” in IEEE
PES Innovative Smart Grid Technologies (ISGT), 2011, pp. 1–8.
[166] A. Thomas, C. Cai, D. Aliprantis, and T. L., “Effects of price-responsive residential
demand on retail and wholesale power market operations,” in IEEE Power and
Energy Society General Meeting, 2012, to appear.
[167] T. Lui, W. Stirling, and H. Marcy, “Get smart,” IEEE Power and Energy Magazine,
vol. 8, no. 3, pp. 66–78, 2010.
[168] C. O’Dwyer, R. Duignan, and M. O’Malley, “Modeling demand response in the
residential sector for the provision of reserves,” in IEEE Power and Energy Society
General Meeting, 2012, to appear.
[169] C. Quinn, D. Zimmerle, and T. Bradley, “The effect of communication architecture
on the availability, reliability, and economics of plug-in hybrid electric vehicle-to-grid
ancillary services,” Journal of Power Sources, vol. 195, no. 5, pp. 1500–1509, 2010.
Bibliography 145

[170] N. Hopper, C. Goldman, R. Bharvirkar, and D. Engel, “The summer of 2006: A

milestone in the ongoing maturation of demand response,” The Electricity Journal,
vol. 20, no. 5, pp. 62–75, 2007.
[171] D. Nguyen, M. Negnevitsky, and M. de Groot, “Pool-based demand response ex-
change – concept and modeling,” IEEE Transactions on Power Systems, no. 99, pp.
1677–1685, 2011.
[172] F. Lai, F. Magoulès, and F. Lherminier, “Vapnik’s learning theory applied to energy
consumption forecasts in residential buildings,” International Journal of Computer
Mathematics, vol. 85, no. 10, pp. 1563–1588, 2008.
[173] M. Pedrasa, T. Spooner, and I. MacGill, “Coordinated scheduling of residential
distributed energy resources to optimize smart home energy services,” IEEE Trans-
actions on Smart Grid, vol. 1, no. 2, pp. 134 –143, 2010.
[174] J. Giráldez, A. Jaiantilal, J. Walz, S. Suryanarayanan, S. Sankaranarayanan,
H. Brown, and E. Chang, “An evolutionary algorithm and acceleration approach
for topological design of distributed resource islands,” in IEEE PowerTech, 2011,
pp. 1–8.
[175] J. Giràldez, R. Roche, S. Suryanarayanan, and D. Zimmerle, “A Linear Program-
ming Methodology to Quantify the Impact of PHEVs with V2G Capabilities on
Distribution Systems,” submitted, in review.
[176] J. Laurent, G. Desaulniers, R. Malhame, and F. Soumis, “A column generation
method for optimal load management via control of electric water heaters,” IEEE
Transactions on Power Systems, vol. 10, no. 3, pp. 1389–1400, 1995.
[177] M. Ilic, J. Black, and J. Watz, “Potential benefits of implementing load control,” in
IEEE Power Engineering Society Winter Meeting, 2002, pp. 177–182.
[178] J. Dickert and P. Schegner, “A time series probabilistic synthetic load curve model
for residential customers,” in IEEE PowerTech, 2011, pp. 1–6.
[179] S. Meliopoulos, J. Meisel, G. Cokkinides, and T. Overbye, “Power sys-
tem level impacts of plug-in hybrid vehicles,” Power Systems Engineer-
ing Research Center (PSERC), Tech. Rep. 09-12, 2009. [Online]. Avail-
able: [Link]
meliopoulos_phev_pserc_report_t-34_2009.pdf
[180] E. Graham-Rowe, B. Gardner, C. Abraham, S. Skippon, H. Dittmar, R. Hutchins,
and J. Stannard, “Mainstream consumers driving plug-in battery-electric and plug-
in hybrid electric cars: A qualitative analysis of responses and evaluations,” Trans-
portation Research Part A: Policy and Practice, vol. 46, no. 1, pp. 140–153, 2012.
[181] C.-S. N. Shiau, C. Samaras, R. Hauffe, and J. J. Michalek, “Impact of battery weight
and charging patterns on the economic and environmental benefits of plug-in hybrid
vehicles,” Energy Policy, vol. 37, no. 7, pp. 2653–2663, 2009.
[182] G. Barbose, N. Dargouth, R. Wiser, and J. Seel, “Tracking the sun IV: An
historical summary of the installed cost of photovoltaics in the U.S. from 1998
146 Bibliography

to 2010,” Lawrence Berkeley National Laboratory, Tech. Rep. LBNL-5047E, 2011.

[Online]. Available: [Link]

[183] G. Masters, Renewable and efficient electric power systems. Wiley - IEEE Press,
2004.

[184] P. Davis, “Energy Efficiency & Demand Response – Smart Grid, Evolution of
DR and the Impact of FERC 745,” Electric Light and Power, vol. 89, no. 3,
p. 46, 2011, accessed October 4, 2012. [Online]. Available: [Link]
index/display/article-display/8190235016/articles/electric-light-power/volume-89/
issue-3/sections/[Link]

[185] L. Huang, J. Walrand, and K. Ramchandran, “Optimal smart grid tariffs,” in Infor-
mation Theory and Applications Workshop (ITA), 2012, pp. 212–220.

[186] R. Billinton and S. Jonnavithula, “A test system for teaching overall power system
reliability assessment,” IEEE Transactions on Power Systems, vol. 11, no. 4, pp.
1670–1676, 1996.

[187] X. Ma, D. Sun, and K. Cheung, “Evolution toward standardized market design,”
IEEE Transactions on Power Systems, vol. 18, no. 2, pp. 460–469, 2003.

[188] L. Greening, D. Greene, and C. Difiglio, “Energy efficiency and consumption – The
rebound effect – A survey,” Energy policy, vol. 28, no. 6, pp. 389–401, 2000.

[189] P. Werbos, “Computational intelligence for the smart grid – history, challenges, and
opportunities,” IEEE Computational Intelligence Magazine, vol. 6, no. 3, pp. 14–21,
2011.

[190] L. Tesfatsion. (2012) The AMES Wholesale Power Market Test Bed. [Online].
Available: [Link]

[191] T. Hansen, R. Roche, S. Suryanarayanan, H. Siegel, D. Zimmerle, P. Young, and

A. Maciejewski, “A proposed framework for heuristic approaches to resource alloca-
tion in the emerging smart grid,” in IEEE PES International Conference on Power
Systems Technology (POWERCON 2012), 2012, pp. 1–6.

[192] DIgSILENT GmbH. (2012) PowerFactory – DIgSILENT Germany. Accessed

October 4, 2012. [Online]. Available: [Link]
[Link]

[193] M. Daggag, R. Roche, L. Idoumghar, B. Blunier, A. Miraoui, and A. Koukam,

“System and method for controlling an electrical energy production installation,”
2012, patent no. WO/2012/143424 (international) / FR1153369 (France).

[194] R. Roche, L. Idoumghar, B. Blunier, and A. Miraoui, “Imperialist competitive algo-

rithm for dynamic optimization of economic dispatch in power systems,” in Inter-
national Conference on Artificial Evolution (EA 2011), 2011, pp. 375–390.

[195] ——, “Optimized fuel cell array energy management using multi-agent systems,” in
IEEE Industry Applications Society Annual Meeting (IAS), 2011, pp. 1–8.
Bibliography 147

[196] R. Roche, F. Lauri, B. Blunier, A. Miraoui, and A. Koukam, Power Electronics

for Renewable and Distributed Energy Systems. Springer, 2012, ch. Multi-Agent
Technology in Power Systems, in press.
[197] M. Simões, R. Roche, E. Kyriakides, A. Miraoui, B. Blunier, K. McBee, S. Surya-
narayanan, P. Nguyen, and P. Ribeiro, “Smart-grid technologies and progress in Eu-
rope and the USA,” in IEEE Energy Conversion Congress and Exposition (ECCE),
2011, pp. 383–390.
[198] N. Chérin, L. Idoumghar, P. Siarry, R. Roche, and B. Blunier, “Métaheuristique
hybride pour les problèmes d’optimisation continue,” in 13e congrès annuel de la
Société française de Recherche Opérationnelle et d’Aide à la Décision (ROADEF
2012), 2012.
[199] A. Ravey, R. Roche, B. Blunier, and A. Miraoui, “Combined optimal sizing and en-
ergy management of hybrid electric vehicles,” in IEEE Transportation Electrification
Conference and Expo (ITEC), 2012, pp. 1–6.
[200] N. Watrin, R. Roche, H. Ostermann, B. Blunier, and A. Miraoui, “Multi-physical
lithium-based battery model for use in state-of-charge determination,” IEEE Trans-
actions on Vehicular Technology, vol. 61, no. 8, pp. 3420–3429, 2012.
[201] B. Couraud, R. Roche, and B. Traill, “A Distribution Load Forecasting Methodology
Based on Primary Substations SCADA Data,” 2013, submitted, in review.
[202] R. Roche. (2011) JICA: A Java Implementation of the Imperialist Competitive
Algorithm for real variables. Accessed September 8, 2012. [Online]. Available:
[Link]
[203] G. Kochenberger, Handbook of metaheuristics. Springer, 2003.
[204] J. Kennedy and R. C. Eberhart, Swarm Intelligence. Morgan Kaufmann Academic
Press, 2001.
[205] R. Storn and K. Price, “Differential evolution – A simple and efficient heuristic for
global optimization over continuous spaces,” Journal of global optimization, vol. 11,
no. 4, pp. 341–359, 1997.
[206] J. Holland, “Genetic algorithms,” Scientific american, vol. 267, no. 1, pp. 66–72,
1992.
[207] E. Atashpaz-Gargari and C. Lucas, “Imperialist competitive algorithm: An algo-
rithm for optimization inspired by imperialistic competition,” in IEEE Congress on
Evolutionary Computation, 2007, pp. 4661–4667.
[208] C. Blum and A. Roli, “Hybrid metaheuristics: An introduction,” Hybrid Metaheuris-
tics, pp. 1–30, 2008.
[209] L. Idoumghar, M. Idrissi-Aouad, M. Melkemi, and R. Schott, “Metropolis particle
swarm optimization algorithm with mutation operator for global optimization prob-
lems,” in 22nd IEEE International Conference on Tools with Artificial Intelligence
(ICTAI), 2010, pp. 35–42.
148 Bibliography

[210] M. Locatelli, Handbook Of Global Optimization. Kluwer Academic Pub, 2002, ch.
Simulated annealing algorithms for continuous global optimization, pp. 179–229.
[211] C. Raquel and P. Naval Jr, “An effective use of crowding distance in multiobjective
particle swarm optimization,” in Proceedings of the 2005 ACM conference on Genetic
and evolutionary computation, 2005, pp. 257–264.
[212] P. N. Suganthan, N. Hansen, J. J. Liang, K. Deb, Y.-P. Chen, A. Auger, and S. Ti-
wari, “Problem definitions and evaluation criteria for the CEC 2005 special session
on real-parameter optimization,” Nanyang Technological University, Singapore and
IIT Kanpur, India, Tech. Rep. 2005005, 2005.
[213] M. Pant, R. Thangaraj, and A. Abraham, “Particle swarm based meta-heuristics for
function optimization and engineering applications,” in Conference on Computer
Information Systems and Industrial Management Applications, vol. 7, 2008, pp. 84–
90.
[214] L. Idoumghar. (2012) MetaHeuristicDemo. Accessed September 9, 2012. [Online].
Available: [Link]
[215] E. Atashpaz Gargari. (2008) Imperialist competitive algorithm (ica). Ac-
cessed September 8, 2012. [Online]. Available: [Link]
matlabcentral/fileexchange/22046-imperialist-competitive-algorithm-ica
[216] J. J. Durillo and A. J. Nebro, “jMetal: A Java framework for multi-objective opti-
mization,” Advances in Engineering Software, vol. 42, pp. 760–771, 2011.
 List of Acronyms

AC Air conditioning
ACL Agent communication language
AI Artificial intelligence
AID Agent identifier
AMS Agent management system
AP Agent platform
AMI Advanced metering infrastructure
BEV Battery electric vehicle
CAS Complex adaptive system
CHP Combined heat and power
CIM Common information model
COM Component object model
COP Coefficient of performance
CPP Critical peak pricing
CPS Cyber-physical system
DA Distribution automation
DCS Distributed control system
DE Differential evolution
DF Directory facilitator
DG Distributed generation
DLC Direct load control
DLN Dry low NOx
DLR Dynamic line rating
DMS Distribution management system
DR Demand response
DRX Demand response exchange
DSM Demand-side management
DSO Distribution system operator
ED Economic dispatch
EMS Energy management system
EV Electric vehicle
EWH Electric water heater
FACTS Flexible AC transmission system

149
150 List of Acronyms

FIPA Foundation for intelligent physical agents

GA Genetic algorithm
GENCO Generation company
GIS Geographical information system
GUI Graphical user interface
HEMS Home energy management system
HTS High temperature superconductor
HVDC High voltage DC
ICA Imperialist competitive algorithm
ICE Internal combustion engine
IED Intelligent electronic device
ISO Independent system operator
IT Information technology
JADE Java agent development framework
KQML Knowledge query and manipulation language
MAS Multi-agent system
MDMS Meter data management system
MPSOM Metropolis particle swarm optimization with mutation operation
MTP Message transport protocol
MTS Message transport system
OMS Outage management system
PHEV Plug-in hybrid electric vehicle
PID Proportional-integral-derivative
PMU Phasor measurement unit
PSO Particle swarm optimization
PV Photovoltaic
RBTS Roy Billinton test system
RES Renewable energy source
RTP Real-time pricing
SCADA Supervisory control and data acquisition
SGP Smart generating plant
SOC State-of-charge
SSA Start and stop algorithm
T&D Transmission and distribution
TCP Transmission control protocol
TOU Time-of-use
TSO Transmission system operator
UML Unified modeling language
US United States
V2G Vehicle-to-grid
VPP Virtual power plant
List of Acronyms 151

WAMS Wide area measurement system

 List of Figures

1.1 Solar radiation and wind speed measured in Belfort, and total French load on
June 6, 2011 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Architecture of legacy power systems . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 Architecture of current power systems . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Architecture of future power systems . . . . . . . . . . . . . . . . . . . . . . . . 9
1.5 Timescales and decision mechanisms for electric system operation . . . . . . . . 15

2.1 Diagram of a generic agent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

2.2 Conceptual diagram of an example application of MASs for smart grids . . . . 28
2.3 Sample ACL message . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.4 Life cycle of an agent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.5 The FIPA agent management reference model . . . . . . . . . . . . . . . . . . . 32
2.6 Diagrams of hierarchy, holarchy and federation topologies . . . . . . . . . . . . 34
2.7 Diagram of the FIPA-Request protocol . . . . . . . . . . . . . . . . . . . . . . . 37
2.8 Agent execution model of agents in JADE . . . . . . . . . . . . . . . . . . . . . 40
2.9 Screenshot of the Introspector agent interface in JADE . . . . . . . . . . . . . . 40
2.10 Screenshot of the Sniffer agent interface in JADE . . . . . . . . . . . . . . . . . 41
2.11 Interface between JADE, Matlab and PowerWorld . . . . . . . . . . . . . . . . 43
2.12 Communication flowchart of a request issued by a JADE agent . . . . . . . . . 46

3.1 Diagram of a simple cycle, single shaft gas turbine . . . . . . . . . . . . . . . . 52

3.2 Fuel flow of the turbine. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.3 NOx and CO emission curves of 9E gas turbines . . . . . . . . . . . . . . . . . 54
3.4 Finite state machine describing the turbines combustion modes . . . . . . . . . 55
3.5 Fuel consumption of a turbine during its starting cycle . . . . . . . . . . . . . . 56
3.6 Generic flowchart of an agent life cycle . . . . . . . . . . . . . . . . . . . . . . . 57
3.7 Simplified turbine agent flowchart . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.8 Simplified dispatch agent flowchart . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.9 Overview of the power plant EMS architecture . . . . . . . . . . . . . . . . . . 61
3.10 Operation and interactions of agents when a new turbine is connected . . . . . 62
3.11 Interactions between agents during normal operation of the EMS . . . . . . . . 63
3.12 Pareto front example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.13 Screenshot of the developed GUI for the power plant EMS . . . . . . . . . . . . 71
3.14 Load profile used in the simulations . . . . . . . . . . . . . . . . . . . . . . . . 72
3.15 Duration spent by each turbine in each mode for Algorithm A . . . . . . . . . 74
3.16 Duration spent by each turbine in each mode for Algorithm B . . . . . . . . . . 74

152
List of Figures 153

3.17 Duration spent by each turbine in each mode for Algorithm C . . . . . . . . . . 75

3.18 Duration spent by each turbine in each mode for Algorithm D . . . . . . . . . 76
3.19 Duration spent by each turbine in each mode for Algorithm E . . . . . . . . . . 76
3.20 Duration spent by each turbine in each mode for Algorithm C with the SSA . . 79
3.21 Duration spent by each turbine in each mode for Algorithm D with the SSA . 80
3.22 Duration spent by each turbine in each mode for Algorithm E with the SSA . . 80
3.23 Comparison of the MWh costs obtained by the algorithms . . . . . . . . . . . . 81
3.24 Results of the EMS flexibility test . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.25 Plot of the mean optimization duration as a function of the number of turbines
to control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

4.1 Main components of demand-side management . . . . . . . . . . . . . . . . . . 87

4.2 Architecture of the DR system . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.3 Diagram representing the interactions between residential loads, assets, smart
meter, HEMS and the rest of the system . . . . . . . . . . . . . . . . . . . . . . 92
4.4 Operation principle of the EWH thermostat . . . . . . . . . . . . . . . . . . . . 94
4.5 Operation principle of the AC thermostat . . . . . . . . . . . . . . . . . . . . . 96
4.6 Load profile example for one customer . . . . . . . . . . . . . . . . . . . . . . . 98
4.7 Aggregated load profile example for 1000 customers . . . . . . . . . . . . . . . 98
4.8 Main chronological steps of a DR event . . . . . . . . . . . . . . . . . . . . . . 103
4.9 Interactions between agents during a DR event . . . . . . . . . . . . . . . . . . 105
4.10 Flowchart of the PHEV charging rescheduling algorithm . . . . . . . . . . . . . 106
4.11 Diagram of the available PHEV rescheduling strategies . . . . . . . . . . . . . . 107
4.12 Oneline diagram of bus 5 of the RBTS in PowerWorld Simulator . . . . . . . . 109
4.13 Baseline and actual net load curves . . . . . . . . . . . . . . . . . . . . . . . . . 110
4.14 Total load reduction achieved with the DR system . . . . . . . . . . . . . . . . 111
4.15 Rebound effect example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
4.16 Baseline and actual load curves for PHEV charging . . . . . . . . . . . . . . . . 112
4.17 Baseline and actual load curves for thermal loads . . . . . . . . . . . . . . . . . 113
4.18 Impact of the DR system on indoor and water temperature . . . . . . . . . . . 114
4.19 Rescheduling of the charge of a PHEV using strategy 1 . . . . . . . . . . . . . 114
4.20 Rescheduling of the charge of a PHEV using strategy 2 . . . . . . . . . . . . . 115
4.21 Rescheduling of the charge of a PHEV using strategy 3 . . . . . . . . . . . . . 115
4.22 Voltage profile at Feeder 4 of bus 5 in RBTS . . . . . . . . . . . . . . . . . . . 116

A.1 Diagram of the movement of a particle with the PSO algorithm . . . . . . . . . 126
A.2 Movement of a colony toward its imperialist in the ICA algorithm . . . . . . . 130
A.3 Plot of the Sphere function for n = 2 . . . . . . . . . . . . . . . . . . . . . . . . 136
A.4 Plot of the Rastrigin function for n = 2 . . . . . . . . . . . . . . . . . . . . . . 136
A.5 Plot of the Rosenbrock function for n = 2 . . . . . . . . . . . . . . . . . . . . . 137
A.6 Plot of the Ackley function for n = 2 . . . . . . . . . . . . . . . . . . . . . . . . 137
 List of Tables

1.1 Comparison of vertically integrated utilities and unbundled electricity markets 14

3.1 Typical characteristics of a simple cycle 9E gas turbine . . . . . . . . . . . . . . 52

3.2 Selected parameters for the power plant EMS simulation . . . . . . . . . . . . . 73
3.3 Results for each dispatching algorithm without the SSA . . . . . . . . . . . . . 77
3.4 Impact of the performance coefficients on dispatching results . . . . . . . . . . 78
3.5 Results for each dispatching algorithm with the SSA . . . . . . . . . . . . . . . 78
3.6 Number of messages exchanged and average optimization duration as a function
of the number of turbines to control . . . . . . . . . . . . . . . . . . . . . . . . 82

4.1 Electric water heater model parameters . . . . . . . . . . . . . . . . . . . . . . 94

4.2 Air conditioning model parameters . . . . . . . . . . . . . . . . . . . . . . . . . 95
4.3 Appliances model parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.4 Characteristics of the four PHEV classes . . . . . . . . . . . . . . . . . . . . . . 99
4.5 Departure and arrival time distribution parameters of PHEVs for a typical
weekday . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.6 Customer data for load points at bus 5 of the RBTS . . . . . . . . . . . . . . . 108
4.7 Metrics for the DR event (Part 1) . . . . . . . . . . . . . . . . . . . . . . . . . 111
4.8 Metrics for the DR event (Part 2) . . . . . . . . . . . . . . . . . . . . . . . . . 111

A.1 Parameter settings for the PSO algorithm . . . . . . . . . . . . . . . . . . . . . 127

A.2 Parameter settings for the DE algorithm . . . . . . . . . . . . . . . . . . . . . . 127
A.3 Parameter settings for the GA algorithm . . . . . . . . . . . . . . . . . . . . . . 129
A.4 Parameter settings for the ICA approach . . . . . . . . . . . . . . . . . . . . . 132
A.5 Parameter settings for the MPSOM algorithm . . . . . . . . . . . . . . . . . . . 133
A.6 Parameter settings for the ICA-PSO approach . . . . . . . . . . . . . . . . . . 134
A.7 Standard mathematical benchmark functions definition . . . . . . . . . . . . . 135
A.8 Comparison of the solutions to the benchmark functions obtained by the se-
lected metaheuritics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

154
 List of Algorithms

1 Pseudo-code of the PSO algorithm. . . . . . . . . . . . . . . . . . . . . . . . 127

2 Pseudo-code of the DE algorithm. . . . . . . . . . . . . . . . . . . . . . . . 128
3 Pseudo-code of the GA algorithm. . . . . . . . . . . . . . . . . . . . . . . . 129
4 Pseudo-code of the ICA algorithm. . . . . . . . . . . . . . . . . . . . . . . . 130
5 Pseudo-code of the ICA-PSO algorithm. . . . . . . . . . . . . . . . . . . . . 134

155