Machine Learning for Ranking Day-ahead Decisions

in the Context of Short-term Operation Planning

Laurine Duchesne, Efthymios Karangelos, Antonio Sutera, Louis Wehenkel
Montefiore Institute - Department of EE&CS
University of Liège, Liège, Belgium
{[Link], [Link], [Link], [Link]}@[Link]

Abstract—In operation planning, probabilistic reliability as- making simulator using the next-day information state to
sessment consists in evaluating, for various candidate planning compute next-day decisions and the different terms of next-day
decisions, the induced probability of meeting a reliability target operation cost.
and the expected operating cost over a certain future time period.
In this paper, we propose to exploit Monte-Carlo simulation and To evaluate the expected cost of operation, a possible
machine learning to predict operation costs for various day-ahead solution would exploit a crude Monte-Carlo (CMC) approach
unit commitment and economic dispatch decisions and a range using the real-time decision making simulator, for many next-
of realisations of uncertain loads and renewable generations over day scenarios drawn by using the probabilistic model of
the next day. We describe how to generate a database, how to exogenous uncertainties. However, this approach is most of
apply supervised machine learning to it, and how to use the
learnt proxies to rank candidate day-ahead decisions in terms of the time impractical. Indeed, simulating real-time operation
the expected operating cost they induce over the next day. We consists often in running Security Constrained Optimal Power
illustrate the approach on the IEEE-RTS96 benchmark where we Flow (SCOPF) types of problems with a very large number
use the DC power-flow approximation and the N-1 criterion to of variables [2] especially for large-scale power systems. In
simulate real-time operation and to generate generation schedules this regard, given that in day-ahead the operator has only a
in the day-ahead operation planning stage.
few hours to select a day-ahead decision, running a very large
Index Terms—Machine learning, Monte Carlo simulation, number of SCOPFs for each candidate day-ahead decision may
Operation planning, Probabilistic reliability assessment not be realistic from a practical point of view.
To reduce the computational burden of the CMC approach,
I. I NTRODUCTION we proposed in [3] to replace the real-time decision making
In operation planning, reliability management aims at taking simulator by a much faster proxy built with machine learning
decisions ahead in time so that the system can be later algorithms, and we also leveraged variance reduction tech-
on operated according to its reliability targets. For example, niques, such as control variates, to yield unbiased estimates
day-ahead decisions may be taken to adjust market clearing of the expected operating cost. In that work, the problem was
outcomes, to postpone planned outages, etc, so as to enable the tackled in a set-up of assessing only one single already selected
operator to meet his reliability targets over the next day while day-ahead decision.
minimising the cost of operating the system. In general, this In the present paper, we propose a generalisation of the
reliability management problem is decomposed into reliability approach presented in [3], to further help choosing among day-
assessment and reliability control [1]. ahead decisions. We thus address the problem of building a
In this paper we consider as ‘template’ the day-ahead proxy predicting operation costs with acceptable performances
operation planning context, and we address the problem of for several candidate day-ahead decisions, even unseen ones.
ranking various candidate day-ahead planning decisions in This proxy could replace the real-time decision-making simu-
terms of the expected next-day operating cost they induce, lator in the CMC approach, allowing one to much more rapidly
while considering exogenous uncertainties such as load or assess the expected cost of next-day operation for various
renewable generation. The final purpose of this would be candidate day-ahead decisions. We propose a methodology to
to help selecting a day-ahead decision among a given set automatically build a database combining different candidate
of candidate such decisions, while considering the impact of day-ahead decisions with a sample of scenarios representing
uncertainties on operation. More specifically, we propose a the expected range of possible next-day conditions, and show
method to evaluate, for a list of candidate day-ahead decisions, how to exploit such a database to learn and validate a proxy
their corresponding expected cost of real-time operation and of real-time operation. In particular we investigate the use of
then use this evaluation to rank the decisions to, in fine, ease neural networks and multitask learning [4] to assess different
the identification of a ‘good’ day-ahead decision. day-ahead decisions and to rank them in terms of their induced
To state our problem, we model in a probabilistic way the next-day operating cost. We test this approach on the three-
exogenous uncertainties about load and renewable generation area IEEE-RTS96 system, while using the DC power flow
that would be faced by the operation planner. We also model model and the N-1 security criterion in order to simulate
the behaviour of the real-time operator via a real-time decision real-time operation as the resolution of a sequence of DC-

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
SCOPF problems. In our case study, we consider as candidate day, and subject to uncertainty on renewable power generation
day-ahead decisions several unit commitments and market and demand. In particular, the mission of the operation planner
dispatches, as well as provisional wind curtailment, and we is to select in advance (i) the commitment and dispatch of
focus on the estimation of the expected value of the real-time generating units, and, (ii) provisional curtailment of wind
preventive control costs. power generation. To help in his decision-making, we assume
The idea of exploiting machine learning in power system that he has access to a simulator of real-time operation along
security and reliability assessment dates back to the 1970s the next day (nd), modelling both the physical behaviour of
[5]. More recently, due to the advances in the field of ma- the power system and the decision-making of control room
chine learning, an increasing number of papers exploiting this operators in response to the resolution of uncertainties in real-
approach have been published (see, for example [6]–[9]). We time. We use a sequence of 24 SCOPF computations following
refer the interested reader to [10] for a comprehensive survey the N-1 criterion while minimising the costs of real-time
of recent works applying machine learning in the context of operation to this end.
reliability management. Closer to the subject of the present We also assume that the operation planner has at his disposal
work, several papers propose to use machine learning to learn a generative model of day-ahead uncertainties, allowing him to
the outputs of a SCOPF. For instance, [11]–[13] exploit deep sample scenarios of next-day load and renewable generations,
1 2 i
learning to predict the generators set-points. Finally, several denoted {ξnd , ξnd , ...}. For a given candidate decision δda and
j
other papers propose machine learning approaches to build a given scenario ξnd , the application of the real-time operation
proxies of shorter-term decision-making contexts to be used simulator allows the planner to anticipate real-time operation
when solving longer-term reliability assessment problems, but along the next day and in particular to evaluate the costs
in other contexts than short-term operation. For example, y i,j of real-time operation. Furthermore, we suppose that the
in [14], [15], the authors propose, in the context of outage operation planner is interested in evaluating, for any given
scheduling, to build proxies of short-term operation based on day-ahead decision, the consequence over next-day operation
the nearest neighbor algorithm. by the expected value of the cost of operating the system the
Our approach also shares strong similarities with analytical next day.
two-stage optimisation such as stochastic unit commitment The problem addressed in this paper then amounts to
[16]. However, we see several differences. First, our approach screening candidate day-ahead decisions to select a good
can be used to rank under uncertainties various sets of day-ahead decision in terms of its expected impact on next-
candidate first-stage decisions, for which we do not need day operating costs, while exploiting the available generative
to know how they have been chosen. In contrast, solving a model and real-time operation simulator.
two-stage stochastic program returns a (locally or globally)
optimal candidate decision but no quantitative information on B. Generating a database of day-ahead decisions and next-
the merit of this decision with respect to the other available day scenarios
alternatives. Furthermore, our methodology relies essentially We extend the methodology described in [3] to generate
only on massively parallel simulations of the operator’s real- both a set of day-ahead decisions and a set of next-day scenar-
time behaviour and induced costs along different scenarios. On ios. In this approach, m next-day scenarios {ξnd 1 2
, ξnd m
, ..., ξnd }
the other hand, stochastic unit commitment approaches require are sampled with the generative model of day-ahead uncertain-
that the real-time operation strategy of the operator is explicitly ties available to the operation planner. They are then combined
modelled in the form of a second-stage optimisation problem, with k day-ahead decisions {δda 1 2
, δda k
, ..., δda } generated as
and typically impose strong restrictions on such second-stage described in the following subsection.
models in order to yield computationally tractable two-stage 1) Generating k day-ahead decisions: Here we assume
stochastic formulations. that the planner has a day-ahead decision-making support
The rest of this paper is organised as follows. Section software, e.g. in the form of a deterministic multi-period Unit
II states the problem studied and presents methodologies Commitment and Economic Dispatch (UCED) program.
to generate automatically a database of real-time operation To generate the k day-ahead decisions with such a tool, we
trajectories, to build proxies with supervised learning and to first generate a large sample (n  k) of next-day scenarios by
rank candidate day-ahead decisions with the help of these using our generative model of uncertainties, and then apply to
proxies. Section III reports the case-study on the IEEE-RTS96 them the k-means clustering algorithm [17].1 We then compute
system, where we analyse the generalisation capability of the k day-ahead decisions {δda 1 2
, δda k
, ..., δda } from these k next-
proxies and we exploit them to rank candidate day-ahead 1 2 k
day scenarios {ξnd , ξnd , ..., ξnd } by applying to each scenario
decisions. Section IV concludes and suggests future works. the available day-ahead decision-making support software.
II. P ROBLEM STATEMENT AND PROPOSALS 2) Simulating real-time operation: For a given pair
i j
A. Problem statement (δda , ξnd ) combining a day-ahead decision and a next-day
scenario (we call this combination the trajectory τ i,j ), we
In this work, we consider the standard setting wherein in
day-ahead (da) the operation planner seeks to ensure N-1 1 If n is sufficiently large, the resulting k real-time scenarios will cover as
secure operation over every time period of the forthcoming well as possible the uncertainty space, for a given budget k.

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
apply the real-time operation simulator to compute the cor- good day-ahead decision. To do so, we propose to first select
responding next-day operation costs, denoted by y i,j . This (randomly) a subset ∆s of candidate day-ahead decisions
gives us a database {(τ i,j , y i,j )}N of size N = k × m, with that will be used to build a proxy of the real-time operation
features describing day-ahead decisions and next-day scenarios simulator. This subset should be large enough for the proxy
as inputs and real-time operation costs as outputs, that we can to be able to generalise well to unseen decisions but small
exploit to learn proxies of the real-time operation simulator. enough to avoid as much as possible the computational burden
stemming from the use of the heavy real-time simulator. For
C. Learning and generalising the proxies i j 1 2 m
all δda ∈ ∆s and ξnd ∈ {ξnd , ξnd , ..., ξnd }, we compute the
A proxy is a simplified model of real-time operation, i,j
next-day operation costs y with the real-time simulator and
allowing to predict the real-time operation costs yi,j for a given then we build the proxy as described in section II-C. Then
i j
day-ahead decision δda and a given next-day scenario ξnd . We we apply the proxy to predict the next-day operation cost y i,j
propose to build proxies with supervised learning, which is a i j 1 2 l
for all δda ∈ ∆u and ξnd ∈ {ξnd , ξnd , ..., ξnd }2 . After that we
branch of machine learning that consists in finding a function average the predictions over the l scenarios to have an estimate
h(·) that maps some vector of inputs x to (a vector) of outputs ȳpi of the expected next-day operation cost for each decision in
y, given a database {(xi , y i )}N
i=1 of input-output pairs [17]. i
δda ∈ ∆u and we average the y i,j over the m scenarios to have
1) Splitting the database in training and test sets: We an estimate ȳ i for each decision in δda i
∈ ∆s . Finally, these
are interested in evaluating the accuracy of our proxies to estimates are used to rank the candidate decisions according
unseen day-ahead decisions, to unseen next-day scenarios, to their expected next-day operation cost.
and to combinations of both. For that, we need to split the
database into a learning set L and a validation set V used both III. C ASE - STUDY ON THE IEEE-RTS96 BENCHMARK
to train the proxies, and three test sets T used to evaluate In order to test our proposed methodology, we place our-
their generation capabilities. Figure 1 shows this database selves in the context of day-ahead operation planning, when
decomposition in a graphical way. the operation planner has to select a unit commitment and
economic dispatch for the next day.
day-ahead decisions δ Our real-time operation simulator as well as the scenario
generator are implemented in JULIA. We use Python for
scenarios
unseen

T∆s ,Ξu T∆u ,Ξu learning: scikit-learn for the clustering algorithms and Pytorch
ξu

as the deep learning framework.

next-day scenarios ξ

A. Test system, uncertainties, day-ahead decision-making and

real-time operation simulator
1) Test system: We consider as test system the 3-area IEEE-
scenarios

RTS96 benchmark [18], where 19 wind farms have been

seen

V∆s ,Ξs T∆u ,Ξs

ξs

added, as was proposed in [19]. We consider as initial scenario

L∆s ,Ξs the demand and wind generation of the first day of the year,
for a peak demand of 3135 MW per area, and favourable wind.
2) Day-ahead decision-making: A day-ahead decision cor-
responds to the commitment status and economic dispatch
seen unseen of all dispatchable generators as well as the provisional
decisions δs decisions δu curtailment of wind power generation. To simulate day-ahead
decision-making, we apply to the forecasted scenarios a multi-
Fig. 1. Schematic representation of the database partition between train (L, in
white), validation (V, in green) and test (T , in blue) sets. Rows and columns
period SCOPF, under the DC-approximation. We consider that
respectively represent scenarios and decisions. the market dispatch should satisfy the N-1 security criterion on
transmission system elements, using provisional wind curtail-
2) Assessing the accuracy of proxies: In order to assess the
ment as a measure of last resort. We also impose a minimum
accuracy of a proxy hp (·) based on a sample S of size |S|,
upward and downward spinning reserve capacity constraint of
we consider the square-loss `(ŷ, y) = (ŷ − y)2 and compute
300 MW in each area of the system. The full formulation can
the empirical loss by
be found in the appendix of [3].
1 X
2
3) Uncertainties and next-day scenarios: We place our-
L(hp , S) = hp (xi ) − y i .
|S| i i selves at noon, the day ahead. We consider uncertainties
(x ,y )∈S
coming from errors in demand and wind generation forecast,
From there, we say that hp (·) generalises well to unseen day- modelled by Gaussian distributions. In particular, for each
ahead decisions if L(hp , T∆u ,Ξu ) ≈ L(hp , T∆s ,Ξu ). element (load or wind farm) we consider two terms of error,
D. Using the proxies for ranking day-ahead decisions a local one and a global one (see [3]). We exploit this model
Once one has shown that the proxies generalise well to 2 Note that that these l scenarios can be different from the m scenarios
unseen decisions, one could use them in order to identify a generated for learning the proxy.

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
coupled with Monte-Carlo simulation to generate plausible
next-day scenarios of realized demand and wind generation.
Since we chose a one-hour time step, each next-day scenario
corresponds to 24 snapshots of one day. The precise method-
ology is described in our previous work [20].
4) Real-time operation simulator: The real-time opera-
tion simulator is modelled as a sequence of 24 SCOPF
computations (using as well the N-1 criterion and the DC-
approximation), to adapt to the realization of uncertainties via
preventive and corrective generator redispatch, load shedding (a) Day-ahead unit commitment (b) Day-ahead economic dispatch
Fig. 2. Hamming distance between each candidate day-ahead decision for
and wind curtailment. These recourse decisions are chosen so respectively the unit commitment and day-ahead dispatch.
that it is always possible to come back to the market dispatch
decided for the next hour, given the ramping constraints of TABLE I
generating units, and with an emphasis on the preventive M EAN , STANDARD ERROR , MINIMUM AND MAXIMUM VALUE OF THE
TOTAL PREVENTIVE COST FOR THE k = 20 STUDIED DAY- AHEAD
control cost so as to favour corrective actions whenever they DECISIONS COMPUTED OVER m = 600 NEXT- DAY SCENARIOS .
can help. We refer the reader to the appendix of [3] for the S ORTED IN INCREASING ORDER OF MEAN .
mathematical formulation of this SCOPF. i Mean Standard err Min Max
δda
(×106 ) (×104 ) (×105 ) (×106 )
B. Description of the database 2 1.641 3.439 5.278 6.339
19 1.656 3.163 4.912 7.024
For our study, we generated k = 20 candidate day-ahead 1 1.661 3.350 5.072 6.672
3 1.676 3.263 5.382 7.404
decisions, applying the methodology of Section II-B with n = 11 1.689 3.384 5.136 7.466
20, 000 next-day scenarios. We combined further m = 600 8 1.691 3.411 5.970 7.091
next-day scenarios with each one of these k = 20 day-ahead 6 1.692 3.150 5.546 8.139
7 1.719 3.741 5.151 7.680
decisions to yield 12, 000 trajectories of 24 hourly time-steps. 10 1.728 3.615 5.153 7.664
In our database, the input-features per hourly snapshot of 9 1.731 3.705 4.788 7.182
each trajectory are the following: 4 1.732 3.264 6.359 6.215
20 1.740 3.560 5.835 7.119
• demand realisations for each load, 16 1.746 3.412 5.270 6.781
• wind generation realisations for each wind farm,
14 1.748 3.518 5.944 7.608
18 1.788 3.700 5.061 7.287
• difference between the real-time scenario and the forecast 12 1.797 4.046 5.248 8.192
scenario used to generate the corresponding day-ahead 5 1.798 3.485 6.345 6.317
13 1.855 3.631 5.928 6.293
decision (in MW and in %), 15 1.872 3.793 5.412 7.547
• total demand, total wind generation and net load, 17 1.917 3.962 6.546 7.912
• maximum and minimum generation capacity,
• hour of the day. commitments is equal to 3% of the vector components. This
Concerning the outputs of the database, we focus here proportion rises to 12% for the economic dispatches.
on the total cost of preventive actions, which is the sum of Another analysis we can make to compare candidate day-
the preventive generation redispatch cost, the preventive load ahead decisions is to look at their impact on next-day operation
shedding cost and the preventive wind curtailment cost along costs. Since we used the same 600 next-day scenarios with
a next-day scenario. each candidate day-ahead decision, the results are directly
comparable. We look at some statistics of the total preventive
C. Analysis of the 20 candidate day-ahead decisions cost over the 600 scenarios, such as the average value ȳ i , the
We begin our analysis with the comparison of the unit com- standard error which is defined as √σm , with m = 600 and
mitments (on-off statuses) and economic dispatches (active σ the standard deviation of y i,j over the 600 scenarios, and
power levels) of the 20 candidate day-ahead decisions. For the minimum and maximum values of y i,j . The results can
that, we represent each decision as a vector containing as be seen in Table I, which is sorted in increasing order of the
elements either the unit commitment or day-ahead dispatch mean total preventive cost.
of each one of the 96 generating units for each one of the 24 When analysing the mean total preventive cost, we see
hours of the day. To compare the vectors pairwise, we use the that there is clear difference between decisions. For instance
Hamming distance, expressed as the number of components decision 17, the most costly decision, is in average 300,000e
for which the two vectors differ. The results are presented in more expensive in real-time than decision 2. If one analyses
the form of heat maps and can be seen in Figure 2. the components of the total preventive cost, it can be seen
We observe that the decisions are different, and that the that this decision leads to the largest load shedding cost in
distances between them are of same order of magnitude, average over next day. Note that given the standard error
both in the unit-commitment space and the space of eco- values, we cannot guarantee that decision 2 is effectively
nomic dispatches. On average, the Hamming distance for unit the candidate decision with the smallest expected next-day

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
 T∆ku ,Ξu with only the left-out decision δ k
preventive control cost, but the less expensive decision on
average should nevertheless be among the first few decisions
of Table I.
D. Machine learning protocol
L1∆s ,Ξs L2∆s ,Ξs ··· Lk∆s ,Ξs
To predict the total preventive cost along a next-day trajec-
tory we divide this trajectory in 24 hourly snapshots, predict
the total preventive cost for each hour (denoted as hourly
prediction in [3]) and then sum the 24 predictions. Note that 1 2 k
we provide information about the hour of the day in two ways, k folds
one using one single input with values ranging from 1 to Fig. 3. Principle of the leave-one-decision-out experiment (unused data are
in grey).
24 and another one using 24 binary inputs with a one-hot-
encoding. This leaves us with 240 hourly input features and
one single output variable. TABLE II
S TATISTICS OF THE HOURLY (H.) AND TRAJECTORY- WISE (T.)
As supervised learning algorithm, we chose the neural R2 - SCORES OBTAINED OVER THE 20 FOLDS OF THE
network algorithm [17]. A neural network has many meta- LEAVE - ONE - DECISION - OUT EXPERIMENT.
parameters that need to be selected. We used a gridsearch H. train score H. test score H. test score H. test score
to find a suitable configuration of meta-parameters among Seen dec. Seen dec. Unseen dec. Unseen dec.
Seen scen. Unseen scen. Unseen scen. Seen scen.
the following candidate values: 3, 4, 5 or 6 layers and 50, Mean 0.9876 0.9262 0.9207 0.9820
100 and 200 neurons per layer. Furthermore, we tested 5 Std 0.0004 0.0074 0.0144 0.0062
different initializations of the networks’ weights. The other Min 0.9863 0.9070 0.8984 0.9648
Max 0.9883 0.9360 0.9434 0.9889
meta-parameters were kept constant. In particular, we used a T. train score T. test score T. test score T. test score
batch size of 200, a learning rate of 10−3 , the Adam optimizer Seen dec. Seen dec. Unseen dec. Unseen dec.
[21] and a weight decay of 10−4 . The maximum number of Seen scen. Unseen scen. Unseen scen. Seen scen.
epochs is 200, and we keep the model corresponding to the Mean 0.9852 0.9069 0.8987 0.9791
Std 0.0015 0.0101 0.0196 0.0057
epoch minimising the loss on the validation set. Min 0.9813 0.8798 0.8703 0.9614
In reference to Figure 1, 100 next-day scenarios are always Max 0.9870 0.9200 0.9395 0.9866
kept out in order to yield the test sets. As concerns the splitting
along the day-ahead decisions, we investigate different settings To select the meta-parameters, we performed a gridsearch
in the sequel. In any case, the validation set V corresponds to and kept for each fold the model with the combination of meta-
5% of the trajectories not used in the test sets; it is used to parameters leading to the maximum score on the validation
select the meta-parameters of the learning algorithms3 , while set. We thus have 20 proxies, each with a different day-ahead
the remaining 95% provide the learning set L used only to decision left out. The statistics of the test scores of these 20
tune the parameters of the neural network predictors. proxies on the different test sets can be seen in Table II.
To evaluate the proxies, we always use the R2 -score, which We observe that the scores of unseen decisions are almost
is computed on the basis of a sample S by [22]: equal to those of seen decisions. Therefore the proxy is able
P i i 2 to generalise to unseen decisions. Furthermore, the scores are
2 (xi ,y i )∈S (y − hp (x ))
R (hp , S) = 1 − P i 2
, quite good, meaning that the proxy is able to predict the total
(xi ,y i )∈S (y − ȳS ) preventive cost with acceptable performances. The trajectory
where ȳS is the mean of the targets y i over S. The best scores are generally a bit smaller than the hourly scores but are
possible R2 -score is 1 and corresponds to a proxy hp that still good. We also see that when the scenarios have already
perfectly predicts the target values y i over S. The best constant been seen by the neural network, the score is close to 1, even
model would systematically predict the sample-mean ȳS and for unseen decisions.
obtain an R2 -score of 0. 2) Machine learning improvement: multitask learning:
Instead of predicting only the target output with a proxy,
E. Generalisation over day-ahead decisions
we can try to simultaneously predict a vector of outputs;
1) Leave-one-decision-out: For this experiment, we use this corresponds to multitask learning [4]. In our case, this
samples from k − 1 decisions to learn a proxy and we test means predicting at the same time the target (total preventive
it with data from the k th (unseen) decision applied on unseen cost) and auxiliary outputs as the preventive redispatch cost,
scenarios. We redo this experiment k times, with each time a the preventive load shedding cost and the preventive wind
different unseen decision in the test set. We then average the curtailment cost. The main advantage of this method is that
k test scores obtained. If this score is high, the proxy is able the model can benefit from extra knowledge brought by the
to well generalise to unseen decisions. Figure 3 illustrates the additional auxiliary outputs, so as to improve the performances
principle of the leave-one-decision-out experiment. of the proxy on the main target output.
3 We keep the combination of meta-parameters minimising the loss on the We repeated the leave-one-decision-out experiment while
validation set. exploiting this multitask learning approach. We performed

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
 TABLE III to generalise well with only 5 decisions and that the scores
S TATISTICS OF THE HOURLY (H.) AND TRAJECTORY- WISE (T.) are already quite good. With 10 decisions we get scores close
R2 - SCORES OBTAINED OVER THE 20 FOLDS OF THE
LEAVE - ONE - DECISION - OUT EXPERIMENT, WITH MULTITASK LEARNING . to those reported for 19 decisions in Table II.
H. train score H. test score H. test score H. test score
Seen dec. Seen dec. Unseen dec. Unseen dec. F. Using the proxies for ranking day-ahead decisions
Seen scen. Unseen scen. Unseen scen. Seen scen.
Mean 0.9868 0.9444 0.9387 0.9817
In this subsection we consider the use of the learnt proxies
Std 0.0005 0.0057 0.0134 0.0052 in order to rank a set of unseen candidate decisions according
Min 0.9860 0.9286 0.9049 0.9704 to the expected next-day operating cost they would induce.
Max 0.9874 0.9523 0.9595 0.9884
T. train score T. test score T. test score T. test score
Since we have seen that the proxies generalise well to unseen
Seen dec. Seen dec. Unseen dec. Unseen dec. decisions, we conjecture that they could be used in order to
Seen scen. Unseen scen. Unseen scen. Seen scen. identify a good day-ahead decision, while avoiding as much
Mean 0.9845 0.9257 0.9157 0.9784
Std 0.0017 0.0080 0.0240 0.0066
as possible to resort to heavy SCOPF computations over large
Min 0.9797 0.9021 0.8634 0.9620 samples of next-day scenarios combined with each candidate
Max 0.9861 0.9351 0.9519 0.9864 day-ahead decision.
We compare two methods exploiting the proxies to estimate
again a gridsearch analysis with the same meta-parameters as the expected total preventive cost ȳ i associated to a day-ahead
before, but this time predicting a vector of outputs. We keep decision δda i
. For each method, we first select a subset ∆s
the network configuration maximising the validation score for of candidate day-ahead decisions that we exploit to build a
the total preventive cost, since it is the target of interest, proxy. Since we noticed in Table IV that only 5 decisions are
for each fold. Looking only at the predictions of the total needed in the learning set for the proxies to generalise well
preventive cost, we obtain the results presented in Table III. on unseen decisions, we selected randomly 5 decisions and
We see that with multitask learning, we can improve the test assigned them to the learning set. We performed this operation
scores of the proxy by 2%, which is quite interesting given 5 times, each time with a different set ∆s of size 5. We also
that the scores were already close to the maximum score. realised this experiment with 10 decisions in the learning set,
3) Impact of the number of training day-ahead decisions: for comparison.
In this experiment, we first selected randomly l < k candidate
1) Method 1: For this method, we first generate 2000
decisions that we consider as our test decisions. Then we learn
next-day scenarios. We then apply the proxy to predict the
k −l times a proxy, each time adding a new decision (different j
total preventive cost y i,j for all δda i
∈ ∆u and ξnd ∈
from the l test decisions) in the training set. At the end, we 1 2 2000
{ξnd , ξnd , ..., ξnd }. Finally we average the predictions over
compare the k − l test scores obtained both on the test set
the 2000 scenarios to have an estimate ȳpi of the expected total
with seen decisions and unseen scenarios and the test set with i
preventive cost for each decision in δda ∈ ∆u .
unseen decisions and unseen scenarios and check when the
2) Method 2: This method extends the previous one by
test score corresponding to unseen decisions is similar to the
using the control variates approach presented in [3] to correct
one corresponding to seen decisions. This would indicate that
a possible bias in the estimation ȳpi performed with the proxies.
enough decisions have been taken into account for a proxy i
For that, we compute the estimated total preventive cost ȳcv
to be able to generalise well to unseen decisions. Figure 4 i
of decision δda ∈ ∆u as follows:
illustrates the principle of this experiment.
100
unused decisions train decisions test decisions i
X
ȳcv = ȳpi + (y i,j − ypi,j ),
j=1

where the 100 scenarios belongs to the set Ξu .

··· 3) Results: We consider as ground truth the average value
of y presented in Table I and computed with 600 scenarios per
i
decision. Note that with both methods, for each δda ∈ ∆s , the
estimated expected value of total preventive cost is the one
k − l cases presented in Table I, given that the SCOPF calls had to be
Fig. 4. Principle of the gradual increase of the learning set L size experiment.
made to build the proxy.
To analyse the quality of the ranking, we use the Kendall’s
Here we report the results obtained with l = 5. We use tau coefficient and the Spearman’s rank correlation coefficient.
the best configuration of neural network on average on the The Kendall’s tau coefficient τ is defined as τ = C−D
C+D , where
validation score from the previous experiment (5 layers and 50 C is the number of concordant pairs and D the number of
neurons per layer) and we repeated each experiment 5 times, discordant pairs. The Spearman’s rank correlation coefficient
with a different initialisation of the neural network weights. ρ is defined as ρ = cov(rg X ,rgY )
σrgX σrgY , where cov(rgX , rgY ) is the
The results for the best initialisation (based on the validation covariance of the rank variables and σrgX and σrgY are the
score) are presented in Table IV. We see that the proxy is able standard deviations of the rank variables.

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
 TABLE IV
M EAN HOURLY (H.) AND TRAJECTORY (T.) R2 - SCORES AS A FUNCTION OF THE NUMBER OF DAY- AHEAD (DA) DECISIONS IN THE LEARNING SET.
Nb of DA decisions H. test score H. test score H. test score T. test score T. test score T. test score
in the training set Seen decision Unseen decision Unseen decision Seen decision Unseen decision Unseen decision
Unseen scenario Unseen scenario Seen scenario Unseen scenario Unseen scenario Seen scenario
1 0.7329 0.6621 0.8234 0.6413 0.6272 0.7961
2 0.7888 0.7140 0.9001 0.7323 0.6840 0.8860
3 0.8241 0.8045 0.9426 0.7903 0.7582 0.9366
4 0.8611 0.8447 0.9603 0.8352 0.8131 0.9546
5 0.8620 0.8668 0.9703 0.8321 0.8384 0.9651
6 0.8861 0.8824 0.9739 0.8636 0.8581 0.9703
7 0.8969 0.8814 0.9745 0.8720 0.8544 0.9709
8 0.9046 0.8964 0.9776 0.8870 0.8737 0.9741
9 0.9103 0.9045 0.9797 0.8993 0.8829 0.9782
10 0.9120 0.9091 0.9801 0.8942 0.8881 0.9779
11 0.9239 0.9243 0.9805 0.9045 0.9015 0.9782
12 0.9241 0.9225 0.9820 0.9100 0.9073 0.9798
13 0.9265 0.9307 0.9821 0.9094 0.9049 0.9807
14 0.9257 0.9245 0.9824 0.9057 0.8963 0.9816
15 0.9257 0.9244 0.9818 0.9143 0.9099 0.9804

TABLE V TABLE VI
M INIMUM AND MAXIMUM K ENDAL’ S TAU COEFFICIENT AND T RUE AND ESTIMATED EXPECTED TOTAL PREVENTIVE COST PER
S PEARMAN ’ S CORRELATION COEFFICIENT FOR BOTH ESTIMATION DECISION AND THE ASSOCIATED RANKING r(·). T HE DECISIONS USED TO
METHODS , WITH 5 OR 10 DECISIONS IN THE LEARNING SET. LEARN THE PROXIES ARE COLORED IN RED .

ȳp ȳcv ȳ i ȳpi i

ȳcv
δi r(ȳ i ) r(ȳpi ) i )
r(ȳcv
Min Max Min Max (×106 ) (×106 ) (×106 )
τ 0.5158 0.8421 0.8211 0.9263 2 1.642 1.639 − 1.635 − 2 19 ∧1 19 ∧1
5 decisions 19 1.656 1.626 − 1.626 − 19 2 ∨1 2 ∨1
ρ 0.7038 0.9534 0.9338 0.9835
τ 0.6211 0.8632 0.8211 0.9159 1 1.661 1.681 + 1.646 − 1 11 ∧2 1=
10 decisions 3 1.676 1.664 − 1.664 − 3 8 ∧2 3=
ρ 0.7624 0.9654 0.9353 0.9820
11 1.689 1.643 − 1.672 − 11 3 ∨1 11 =
8 1.691 1.663 − 1.706 + 8 1 ∨3 6 ∧1
Both metrics minimum and maximum values can be found 6 1.692 1.692 ◦ 1.692 ◦ 6 6= 8 ∨1
7 1.719 1.700 − 1.716 − 7 7= 7=
in Table V for the different estimations. One can see that with 10 1.728 1.725 − 1.742 + 10 4 ∧2 9 ∧1
the control variates approach, these metrics are closer to 1, 9 1.731 1.731 ◦ 1.731 ◦ 9 14 ∧4 10 ∨1
meaning that there is a stronger relationship between the true 4 1.732 1.711 − 1.761 + 4 10 ∨2 4=
20 1.740 1.728 − 1.788 + 20 20 = 20 =
ranking and the estimated one. One can also notice that the 16 1.746 1.728 − 1.796 + 16 16 = 16 =
ranking is better when there is 10 decisions in the learning 14 1.748 1.716 − 1.807 + 14 9 ∨4 12 ∧2
set, but at the cost of more SCOPF calls. 18 1.788 1.750 − 1.810 + 18 18 = 5 ∧2
12 1.797 1.797 ◦ 1.797 ◦ 12 12 = 14 ∨2
Table VI presents the different estimations of the real- 5 1.798 1.798 ◦ 1.798 ◦ 5 5= 18 ∨2
time operating costs associated to a decision as well as the 13 1.855 1.855 ◦ 1.855 ◦ 13 15 ∧1 13 =
corresponding ranking for the best case experiment when 15 1.872 1.816 − 1.871 − 15 13 ∨1 15 =
17 1.918 1.857 − 1.905 − 17 17 = 17 =
there is only 5 decisions in the learning set. One can directly
notice that the control variates approach allows to improve the
estimation of the preventive total cost and thus the ranking, We showed with a case-study on the three-area IEEE-RTS96
but it has a larger computational burden (50% more SCOPF benchmark that our proxies of real-time operation, predicting
calls) than method 1. Note that, even if the first decision in the next-day preventive costs, are able to generalise well to
both estimated rankings is not the correct decision, one can unseen decisions. We also showed that they can be exploited
see that the good decisions (small expected total preventive to identify candidate decisions with smallest expected induced
cost) are identified with both methods. costs in real-time operation.
There are many possible future directions of research fol-
IV. C ONCLUSIONS AND FUTURE WORK lowing this work. We highlight the following ones:
With respect to the methodology proposed in [3], we took • how to improve the ranking methodology;
here the additional steps of (i) studying how to generalise over • how to adapt the proposed methodology to evaluate,
‘unseen’ day-ahead decisions, and (ii) testing the usefulness of instead of the expected real-time operation costs for a
the learnt proxies for ranking candidate day-ahead decisions. given decision, the probability to meet the reliability
To do so, we proposed a methodology to build automatically target in real-time and in particular be able to evaluate
a database of candidate day-ahead decisions and next-day the probability of rare events;
scenarios, to each combination of which we can apply a real- • how to exploit the presented methodology to reverse the
time operation simulator in order to compute the resulting problem and find hints for the operation planner about
next-day operating costs. We also proposed a methodology what forecast scenarios could lead him to a ‘good’ day-
to build and validate proxies, exploiting such a database. ahead decision.

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
 ACKNOWLEDGEMENT [18] P. Wong, P. Albrecht, R. Allan, R. Billinton, Q. Chen, C. Fong,
S. Haddad, W. Li, R. Mukerji, D. Patton et al., “The ieee reliability
The authors thank the Isaac Newton Institute for Mathe- test system-1996. a report prepared by the reliability test system task
matical Sciences, University of Cambridge, for support and force of the application of probability methods subcommittee,” Power
Systems, IEEE Transactions on, vol. 14, no. 3, pp. 1010–1020, 1999.
hospitality during the programme Mathematics of Energy Sys- [19] H. Pandzic, Y. Dvorkin, T. Qiu, Y. Wang, and D. Kirschen, “Unit
tems where work on this paper was initiated. This work was Commitment under Uncertainty - GAMS Models,” Library of the
supported by EPSRC grant no EP/R014604/1. They warmly Renewable Energy Analysis Lab (REAL), University of Washington,
Seattle, USA, [Online]. Available at: [Link]
thank the support of RTE-France (the French Transmission research/real/gams [Link].
System Operator) and Consortium des Équipements de Calcul [20] L. Duchesne, “Machine learning of proxies for power systems reliability
Intensif funded by the Fonds de la Recherche Scientifique de management,” Master’s thesis, Université de Liège, Liège, Belgique,
2016.
Belgique (F.R.S.-FNRS) under Grant No. 2.5020.11 and the [21] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
Walloon Region for providing the computational resources. arXiv preprint arXiv:1412.6980, 2014.
LW acknowledges the support of F.R.S.-FNRS for his sabbat- [22] Scikit-learn developers, “Model evaluation: quantifying the quality
of predictions,” [Online], Available at [Link]
ical year and the Simons Foundation for his fellowship during modules/model [Link], accessed on 2016-05-11.
his stay at the Isaac Newton Institute.
A PPENDIX
R EFERENCES C OMPUTING TIMES
[1] GARPUR Consortium, “D2.2 - guidelines for implementing the new The average computing times of the experiments presented
reliability assessment and optimization methodology,” 2016, available in this paper, with a MacBookPro (2.2GHz Intel Core i7, 16GB
at: [Link] RAM), are presented in Table VII.
[2] F. Capitanescu, “Critical review of recent advances and further devel-
opments needed in AC optimal power flow,” Electric Power Systems In the upper part of the Table, we highlight the CPU times
Research, vol. 136, pp. 57–68, 2016. needed to assess a single scenario and day-ahead decision,
[3] L. Duchesne, E. Karangelos, and L. Wehenkel, “Using machine learning composed of first sampling the scenario, and then computing
to enable probabilistic reliability assessment in operation planning,” in
2018 Power Systems Computation Conference (PSCC). IEEE, 2018, the costs induced by the decision either via the detailed
pp. 1–8. SCOPF-wise simulation or via applying the learnt proxy. We
[4] R. Caruana, “Multitask learning,” Machine learning, vol. 28, no. 1, pp. observe a gain of a factor 10,000 with the ML proxy.
41–75, 1997.
[5] L. A. Wehenkel, Automatic learning techniques in power systems.
Springer Science & Business Media, 2012. TABLE VII
[6] J.-M. H. Arteaga, F. Hancharou, F. Thams, and S. Chatzivasileiadis, AVERAGE COMPUTING TIMES
“Deep learning for power system security assessment,” in 13th IEEE Average time (s)
PowerTech 2019. IEEE, 2019.
[7] B. Donnot, I. Guyon, A. Marot, M. Schoenauer, and P. Panciatici, “Opti- Sampling of one scenario 0.002
mization of computational budget for power system risk assessment,” in Real-time SCOPF simulation for one scenario 303.600
2018 IEEE PES Innovative Smart Grid Technologies Conference Europe Using the ML proxy for one scenario 0.027
(ISGT-Europe). IEEE, 2018, pp. 1–6. Learning/validation/test dataset generation 3,643,200.000
[8] J. L. Cremer, I. Konstantelos, S. H. Tindemans, and G. Strbac, “Data- Learning one proxy 5,660.000
driven power system operation: Exploring the balance between cost and
Optimising the meta-parameters of the proxy 169,813.000
risk,” IEEE Transactions on Power Systems, vol. 34, no. 1, pp. 791–801,
2018.
[9] F. Thams, L. Halilbasic, P. Pinson, S. Chatzivasileiadis, and R. Eriks- The CPU times in the lower part of the Table correspond to
son, “Data-driven security-constrained opf,” in X Bulk Power Systems the off-line learning stage based on the leave-one-decision-out
Dynamics and Control Symposium, 2017.
[10] L. Duchesne, E. Karangelos, and L. Wehenkel, “Recent developments experiment. We observe that the bulk of the computations are
in machine learning for energy systems reliability management,” Pro- about the dataset generation (corresponding in our simulations
ceedings of the IEEE, in press. to 12,000 trajectories (i.e. 600 scenarios combined with 20
[11] F. Fioretto, T. W. Mak, and P. Van Hentenryck, “Predicting AC optimal
power flows: Combining deep learning and lagrangian dual methods,” day-ahead decisions) times 24 hours, while optimising the
arXiv preprint arXiv:1909.10461, 2019. meta-parameters of the machine learning method corresponded
[12] X. Pan, T. Zhao, and M. Chen, “DeepOPF: Deep neural network for in our case to tuning 30 different proxy models. It is important
DC optimal power flow,” in 2019 IEEE International Conference on
Communications, Control, and Computing Technologies for Smart Grids to realise that the computing times of both the dataset gener-
(SmartGridComm). IEEE, 2019, pp. 1–6. ation and the meta-parameters’ optimisation parts could be
[13] A. Zamzam and K. Baker, “Learning optimal solutions for extremely reduced by using massive parallel computations, respectively
fast AC optimal power flow,” arXiv preprint arXiv:1910.01213, 2019.
[14] G. Dalal, E. Gilboa, S. Mannor, and L. Wehenkel, “Unit commitment
to about 303 seconds and 5,660 seconds.
using nearest neighbor as a short-term proxy,” in 2018 Power Systems
Computation Conference (PSCC). IEEE, 2018, pp. 1–7.
[15] ——, “Chance-constrained outage scheduling using a machine learning
proxy,” IEEE Transactions on Power Systems, 2019.
[16] M. Håberg, “Fundamentals and recent developments in stochastic unit
commitment,” International Journal of Electrical Power & Energy
Systems, vol. 109, pp. 38–48, 2019.
[17] T. Hastie, R. Tibshirani, and J. Friedman, The Elements of Statistical
Learning: Data Mining, Inference, and Prediction, 2nd ed. Springer
Series in Statistics, 2009.

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020