Mathematical Optimization: A Powerful Tool for the Energy Industry

Although artificial intelligence and machine learning have captured the

world’s imaginations, a powerful but lesser-known technology is at work
all around us, touching our lives every day.

It’s the unseen force behind daily conveniences — calculating the quickest
route on your mobile phone, ensuring timely package deliveries, and
streamlining your airline and hotel bookings.

It’s at work in the energy industry, too — addressing challenges

associated with battery storage, the variability of wind and solar power
generation, and the design and operation of complex energy industry
processes.

This technology is called “mathematical optimization,” and although it’s

not a household name, it has been transforming the way enterprises make
decisions for decades.

What Is Mathematical Optimization?

Some may define it as “a powerful decision-making technology” or “a set

of algorithms for solving complex problems.” But at its heart, it involves
three key elements: a goal you want to achieve, the constraints you’re
facing and the questions you’re asking.

For example, in the electric power industry, providers must produce

electricity as efficiently as possible. That’s a goal.

But they face constraints involving customer demand; the physics of

electrical power flow; available raw sources of energy (e.g., coal, water,
sunlight and wind); and operational resources.

They also face questions about which power sources to use, when to use
them, and at what usage level. These are the decision variables.

How can you solve a complex problem like this? With mathematical
optimization.

With mathematical optimization, you input your goals, constraints and

decision variables into software that identifies the optimal answer to your
questions. In other words, it tells you exactly how to achieve your goals,
given your constraints and the decision variables.

Is Mathematical Optimization the Same as Machine Learning?

Machine learning and mathematical optimization are both data analytics

technologies. But they’re fundamentally different. Machine learning can
predict the future based on the past. Mathematical optimization can
identify your best path forward, regardless of what’s happened before or
what happens next.

Data scientists are beginning to combine these two technologies. They are
using mathematical optimization to identify the best path forward, given
machine learning’s predicted future.

For the electric-power example, machine learning can be used to predict

fluctuating customer demand and energy produced by more variable
sources, such as wind and solar. Given this predicted data, mathematical
optimization can then identify the most cost-effective way to meet
customer demand while satisfying operational constraints.

Mathematical optimization supports different types of models, but they all

share a notion of a goal or objective to be minimized or maximized. For
example, the overall cost of power generation could be minimized. The
total power generated could be maximized. The models also include
constraints imposed by limited resources, often associated with a physical
system.

Taking a Peek Under the Hood

If you’re interested in the science behind mathematical optimization,

here’s a quick lesson.

The simplest type of mathematical-optimization model is the linear

program. It includes a linear expression of the decision variables to be
minimized, as well as constraints involving other linear expressions of
decision variables associated with the limitations of the physical system.

Consider a very simple electric-power model. The linear objective could be

minimizing the total cost — the sum of the costs per kWh of each type of
power source multiplied by the amounts of power that were generated by
each source.

A linear constraint could involve meeting overall energy-demand

requirements. Specifically, the sum of the amounts of each type of
available power source must match or exceed the total demand.

Another example of a linear constraint is ensuring that the power flowing

into a node on the grid equals the power flowing out.

With the introduction of discrete decisions, nonlinear objectives and

nonlinear constraints, the models can quickly become elaborate and
computationally challenging. For example, complexity increases when you
include a yes/no decision about placing a solar panel in a particular
location on a rooftop1 or deciding how many turbines to build on a wind
farm.2
In these examples, a fractional value of the decision variable has no
practical meaning. After all, you can’t build a fourth of a turbine or place a
tenth of a solar panel.

The decision variables in a linear program can sometimes take on

fractional values. That’s essential when dealing with continuous variables
like the usage level of power sources in an electric-power model — or the
amount of fluid flowing through the system of a concentrated solar power
plant.

Mathematical Optimization’s History

In the 1960s, the oil industry was an early adopter of mathematical-

optimization software. At the time, software and computers were far less
evolved than today — so practitioners often used linear programming, the
simplest mathematical-optimization algorithm.

The heavy use of linear programming by the oil industry in the 1960s and
1970s led to significant improvements in linear-programming software for
all industries.

The electric-power industry knew about mathematical-optimization

software in its early stages, but the software couldn’t meet its needs until
later.

For example, unit commitment problems — such as scheduling generators

to meet demand at minimum cost3 — involve discrete variables such as
yes/no decisions to access a power source. This type of problem is a
mixed-integer program, not a simple linear one.

By the late 1980s, state-of-the-art mixed-integer programming was still

unable to solve practical unit commitment problems effectively. But the
1990s saw huge improvements, with the software surpassing the
performance threshold needed to solve these models.

By 2010, most Independent System Operators and other organizations

that manage grids and electricity markets used mixed-integer
programming software to solve real-time, practical unit commitment
problems for day-ahead and real-time markets.

In the last decade, mathematical-optimization software has continued to

improve and can now solve extremely complex problems. Today, the
energy industry can model and solve more precise representations of
physical systems instead of approximations. The AC Optimal Power Flow
Problem, which minimizes the cost of power production while adhering to
the physics of the generators and transmission lines, is a noteworthy
example.4, 5
Furthermore, the tried-and-true industrial applications of linear
programming remain relevant as well. For example, since the transition to
electric vehicles will not be immediate, a need remains for efficient
blending and pooling of crude oil to produce gasoline. 6, 7 Blending is a
linear problem and pooling is a nonlinear problem.

Optimization is a mathematical modeling methodology used to find the

best solutions to complex problems subject to user-specified criteria and
system constraints. This article provides a basic technical background on
the application of optimization modeling techniques in energy market
analysis, trying to demystify the technical terminology and make it
accessible to practitioners.

Key Takeaways

 Mathematical optimization is vital in energy market analysis as it

helps navigate multi-dimensional decisions and trade-offs between
objectives.

 Optimization models are built on key components: decision

variables, parameters, objective functions, and constraints, which
collaboratively guide decision-making processes toward optimal
solutions.

 There are various categories of optimization models that serve

distinct purposes in energy market applications, facilitate various
degrees of complexity, and address uncertainties in data.

Introduction

As the previous articles in this series have demonstrated, mathematical

optimization is an indispensable tool for the analysis of energy markets
and, specifically, electric power markets. Its value stems from its ability to
consider multi-dimensional decisions and account for tradeoffs between
goals and requirements. But how does it work? With the ubiquity of AI, any
computer-generated analytical solution might seem like “AI magic” to a
novice user, making optimization appear like a black box. Although
optimization serves as a building block for some AI methods and is an AI
methodology itself, its unique structure and algorithms set it apart.

Optimization is a prescriptive methodology, offering outputs that aid in

decision-making. Grasping the technical background of this framework
enhances appreciation for the recommendations produced by optimization
models. This knowledge is beneficial for consumers of modeling results
and essential for analysts to derive insights from the model’s solutions.

While this article deals with the technical aspects of optimization, it

provides practical examples from applications to electric power markets
and is accessible to a wide audience.

Building Blocks of Optimization Models

Optimization is useful in cases where there is some set of

available courses of action to choose from in order to achieve
an objective subject to a set of constraints. Accordingly, models consist of
several elements that work together to represent the problem, explore
possible solutions, and achieve specific objectives.

These components, which are standard across all models, are described
below and include decision variables, parameters, an objective function,
and constraints. Each plays a crucial role in shaping the model and
guiding it towards an optimal solution.

Decision Variables

Decision variables represent the available options or actions that the

model seeks to optimize. In other words, they represent the choices that
are under the decision maker’s control. When optimization algorithms run
to solve optimization problems, they search for the optimal values of
these decision variables. Here are some examples of decision variables in
the realm of power markets:

o The capacity to build of a specific resource type, such as a

solar resource or a combined cycle, in a given year

o The optimal timing of a coal plant retirement

o Dispatch levels for each generator in a utility’s fleet in a given

hour

o Whether a generator is synchronized to the grid or offline

o The amount of energy to store or release from energy storage

in a given hour

These variables are crucial for decision making as they directly influence
the outcomes of the model. Decision variables are usually “free” variables
chosen by the model in trying to optimize some objective which depends
on these variables, but some of the variables can also be fixed at specific
values. By manipulating these variables, one can simulate various
strategies and identify the most effective approach for achieving the
desired objectives.

Defining and understanding decision variables is essential because they

effectively represent specific levers that modelers and stakeholders can
pull to explore the trade-offs and gauge their influence on overall profits
or costs.

Parameters

Parameters are fixed and exogenous inputs to the model that describe the
environment or constraints within which decisions are made. These values
are typically derived from physical system characteristics, historical data,
forecasts, or policy requirements. They are assumed to be beyond the
control of the decision maker. In electric power markets, common
parameters include:

o Natural gas, coal, uranium, and other generation fuel prices

o Hourly wind and solar generation availability as a percentage

of maximum rated capacity

o Electricity demand forecasts, their hourly shape, and their

year-to-year growth

o Capital investment costs for a new power plant

o Physical constraints such as generation capacity limits,

ramping abilities, or transmission line ratings

o Regulatory requirements like emissions caps or renewable

portfolio standards

The role of parameters is to set the stage for the optimization process.
They provide the context within which decision variables operate, and
their values significantly influence the outcomes of the model. By
accurately defining parameter values, modelers ensure that the
optimization models reflect real-world conditions.

Objective Function

The objective function defines the model’s goal, whether it is minimizing

costs, maximizing profits, or achieving an optimal balance of multiple
goals. This function is the driving force behind the optimization process,
guiding the model towards its desired outcome.

In electric power markets, common objective functions include:

 

o Cost minimization, which focuses on reducing the total cost

of electricity generation and delivery

o Profit maximization, which aims to maximize revenues for

power producers while considering market prices and costs of
production

o Environmental goals, such as minimizing greenhouse gas

emissions or renewable curtailment

Formulating the objective function is a critical step in mathematical

optimization. It requires a deep understanding of the problem at hand and
the ability to translate complex goals into mathematical terms. This
process may involve balancing multiple objectives and making tradeoffs to
achieve the best possible outcome.

Constraints

Constraints are the rules, limits, or requirements that the solution must
adhere to. They ensure the feasibility and practical applicability of the
model’s outputs. Constraints are represented by mathematical equations
that define the feasible space of the optimization problem. Following are
some examples of constraints in electric power markets:

o Physical constraints include power plants’ capacity and

ramping limits (e.g., generating up to but not above the rated
capacity) and energy limits (e.g., storage systems being able
to discharge only the energy they have stored). Transmission
line ratings are another example of physical constraints as
they are used to specify limits on transmission flows between
regions.

o Market constraints involve adherence to market rules, such

as ensuring that demand is met by generation and net imports
or that the frequency regulation requirements are met at
every time step.

o Policy constraints include compliance with emissions limits

or renewable energy mandates. In a capacity expansion
context, planning reserve margins or expected peak energy
requirements expressed in terms of effective load carrying
capability or a similar metric are also policy constraints.
 o Logical constraints are used in cases where certain
decisions depend on others. They enforce logical relationships,
such as an “if-then” or an “either-or” dependency. For
example, a requirement to build a storage resource if more
than a certain amount of solar capacity is added to a system
can be encoded as a logical constraint. Another example could
be a choice between two mutually exclusive options, such as
building either a new peaker or a new combined cycle.

Defining the decision variables, identifying parameters, and formulating

an optimization model by writing out the objective function and
constraints are crucial steps of the optimization framework. A skilled
modeler can translate the important aspects of a real-world process into
mathematical equations which allows for detailed analysis of the process
and robust decision-making.

Categories of Optimization Models in Energy Market Applications

Optimization models come in various forms, each tailored to specific types

of problems and challenges. Some of the aspects that affect the choice of
model are the assumptions that the stakeholders make, the complexity of
the problem, and the level of precision required. The classification will
affect the ease with which algorithms are able to solve the problem.

We will go over three common categories of optimization models: linear

optimization, integer optimization, and stochastic optimization. Each of
these types of models offers unique capabilities and is suited to different
types of problems.

Linear Optimization

Linear programming (LP) models are used when the relationships between
variables are linear. In the context of an economic dispatch problem, this
means that, for example, we assume the rates of increase of variable
O&M costs and fuel costs are constant with respect to generation. LP
models also assume that decision variables are continuous, which means
that a decision variable, such as the amount of generation or the capacity
to build, can take on any value between its minimum and maximum
bounds.

One important property of LP models is that they can be solved to global

optimality. In other words, an LP solution algorithm is able to locate a
solution such that there is no alternative set of values that satisfy all
constraints and that can produce a better outcome. This is often very
insightful for decision makers as the optimal course of action can be
something they never even considered. This property sets LPs apart from
non-linear programming (NLP) models where global optimality of solutions
cannot be guaranteed.

Applications of LP models in the energy market include real-time market

clearing, determining least-cost generation dispatch, and optimizing fuel
procurement decisions. The ease of implementation and the ability to
solve large-scale problems efficiently make LP models widely used in
these applications. LP models also have some limitations, chief of which is
their inability to capture non-linear relationships, such as elasticity of
demand or fixed costs. In spite of this, LP models remain a powerful tool
for optimization in energy markets, providing optimal solutions when
assumptions hold.

Integer Optimization

What sets integer programming (IP) models apart is that the decision
variables are restricted to take on discrete integer values. In some
generation capacity expansion models, capacity can only be added in
blocks. For example, if there is a standard size for a peaker, the integer
variable can represent the number of peaker units to build. It can take on
values of 0, 1, 2, and so forth, but not fractional values such as 1.7.

Further, a special case of an integer variable is a binary variable that is

restricted to take values of 0 or 1, which is useful for handling logical
constraints. In a unit commitment model, binary variables are used to
represent whether or not a specific generator is synchronized to the grid
and to track startup and shutdown decisions and account for the
associated costs. These variables are essential for decisions that involve
discrete choices.

While IP models offer flexibility to model complex operational constraints

and capture real-world discrete decisions, they are computationally
intensive for large-scale problems. The increased computational demands
can be a significant drawback, but the ability to represent discrete
decisions makes IP models invaluable for certain types of optimization
tasks.

Stochastic Optimization

Stochastic optimization models incorporate uncertainty into the

optimization process, making them ideal for problems with variable inputs
such as fuel prices or weather-dependent generation availability. Unlike
scenario analysis, which solves multiple deterministic models separately,
stochastic programming endogenizes the uncertainty and produces a
single solution that accounts for the fact that the exact values of some
high impact parameters are not known with certainty at the time
stakeholders have to make the decision.
A useful feature of stochastic models is that their objective can be
expressed in terms of a risk metric (e.g., conditional value at risk).
Optimizing with regards to expected value of costs or revenues assumes
that the stakeholders are risk neutral. With stochastic models, modelers
are able to optimize the decisions with regards to a risk metric, which
aligns the solution and the recommendations with the stakeholders’ risk
preferences.

Applications of stochastic optimization in the energy market include

capacity expansion under uncertain demand growth, fuel costs, or
regulatory requirements or scheduling of maintenance outages facing
uncertain demand and renewable resource availability profiles.

The increased complexity and computational demands of stochastic

optimization models are notable limitations. However, the robustness of
solutions and the ability to account for uncertainty make stochastic
optimization a powerful tool for strategic planning in energy markets.

Optimization Solvers

Optimization solvers are the engines that drive the optimization process.
They transform mathematical models into actionable insights.
Optimization solvers use algorithms that are designed to find the best
possible solution to an optimization problem. Given a mathematical
model, an optimization solver systematically explores the feasible solution
space defined by the constraints to identify an optimal solution for the
specified objective. If you are in a room, think of the feasible solution
space as that room. An optimal solution to the optimization problem can
be anywhere in the room but not outside of it. The floor, ceiling, and walls
are the constraints. A solver explores the room to find an optimal solution.
It does so in a methodical way and efficiently, so that it does not have to
test every single point in the room.

When the model is feasible and bounded (i.e., when there exists at least
one solution that satisfies all the constraints and there is at least one
constraint that prevents the objective function from improving
indefinitely), the algorithm will yield an optimal solution. A solution to an
optimization problem consists of a set of values for the decision variables
that produces the best objective function value and can be used to guide
the stakeholders in their decision-making process.