1

Optimal Unit Commitment Solution Methods

Gonçalo Francisco Correia Relvas - Instituto Superior Técnico, University of Lisbon, Portugal
[email protected]

provide electricity to consumers with the utmost safety and

reliability obliges producer companies to plan energy supply
Abstract— The Unit Commitment (UC) problem is a typical processes at all levels. From the generation phase to the power
application of optimization methods to ensure an efficient, secure supply to the final consumer, there are many economic
and economic operation of power systems. Its main objective is to considerations to consider. Thus, the planning steps should
determine online schedules and production levels for generating allow a reliable operation of the system while being
units, based on operational costs minimization , ensuring supply
economically sustainable.
meets demand at all times. The process of finding an optimal
schedule of generating units, subject to several technical
The power system total load varies at every instant, so, the
constraints, given a planning horizon, has been solved by a diverse electric power companies must plan the power generation to
set of techniques. The formulations of the UC problem vary with meet this variable load in advance. To do this, they must decide
energy systems characteristics, as well as with other economical, between the available generators, which should be connected
technical and environmental factors. and when to synchronize them to the network, as well as the
The present article mainly addresses the deterministic, single- sequence in which the operating units should be turned off. This
objective Thermal UC problem. The formulation of the problem is decision-making process is known as Unit Commitment (UC).
discussed, and its solution is obtained by both exact and heuristic By optimizing such decisions, energy can be produced at a
methods. Are studied and developed three of the most referenced lower cost, while satisfying demand and certain constraints.
methods in energy systems optimization: Dynamic Programming
These restrictions may be of the system itself, which are used
(DP), Lagrangian Relaxation (LR) and Particle Swarm
Optimization (PSO). Both DP and LR are classical methods that
to ensure the power supply safety, or may be technological,
have been shown to be very effective in the operational scheduling reducing the freedom of choice between units, as well as the
process. PSO is a more recent population based evolutionary range of possible generation.
algorithm that has been applied to various optimization problems, The traditional Unit Commitment problem is a single objective
including UC. The effectiveness of the developed algorithms is deterministic optimization problem. Its main objective is the
tested on a 10-unit system case study. The obtained results show operational scheduling of generating units using a criterion of
the better suitability of PSO, balancing a satisfactory solution with minimization of the total operational costs during the time
a decent computing time. horizon considered, which could be several hours to a few
It is also proposed a model that integrates hydro-thermal and weeks or years. The complexity of the problem depends on the
renewable units in an attempt show a glimpse of the real UC
diversity of the technological characteristics of energy systems
problem in the present day.
and the generating units under consideration. The UC problem
Index Terms— Unit Commitment, Dynamic Programming, can be computationally challenging due to the high number of
Lagrangian Relaxation, Particle Swarm Optimization. units present in a real energy system and the various constraints
that each technology exhibits. For this reason, obtaining an
optimal solution can be challenging.
I. INTRODUCTION The main objective of the present work is the study and analysis
of some of the most popular methods applied to Unit
E nergy systems are one of the most important infrastructures
and one of the economic engines of a country, allowing its
development and providing quality of life to its citizens.
Commitment resolution: Dynamic Programming (DP),
Lagrangian Relaxation (LR) and Particle Swarm Optimization
(PSO). The theoretical principles on which these methods are
Energy is an everyday life essential asset for the modern based will be analyzed, being also developed algorithms on
societies and is of paramount importance to many companies, MATLAB® that will serve to later apply to a 10-unit system
industry and the ordinary citizen. For electricity to be always case study. An analysis and comparison of the solutions
available, the operation of the production system must be obtained by each method, will allow to infer about the
constantly planned. Worldwide demand for energy has been effectiveness of each one on the UC Problem resolution. The
increasing at a pace that the expansion, planning and comparison between algorithms will allow to evaluate the
management of energy systems have become complex and possible evolution of performance of the most recent methods
challenging problems. in relation to the older ones.
Energy systems can be divided into three main subsystems: It is also the objective of this work to formulate the UC Problem
Generation, Transmission and Distribution of energy. Each of several generation technologies, to prove the increase of its
subsystem has its own behavioral characteristics and constraints complexity
that govern the operation of the overall system. The need to
 2

II. LITERATURE REVIEW The publication [13] present the integration of different
Unit Commitment is identified as a topic of enormous economic methods on a single algorithm in a hope to benefit from the
and technological relevance in the late 1950s, early 1960s. Until advantages of each one and thus improve the effectiveness of
then, the scheduling and dispatch of thermoelectric units was the UC problem resolution. The results confirmed that the use
performed through empirical techniques, typically with priority of hybrid algorithms achieves results, in certain cases, with
lists. These methods were far from optimal in economic terms, greater efficiency than other single methods.
which translated into high production costs and high energy B. Stochastic Unit Commitment
prices practiced to the final consumer. The first studies
Article [14] present a probabilistic model to analyze the risk
produced, [1] and [2], concluded about the need of Unit resulting from the load uncertainty, which translates into the
Commitment in energy systems from an economic point of probability of committing insufficient capacity to compensate
view. They discuss several aspects and solution of the
for unit faults and not foreseen load variations. The presented
scheduling procedures to formulate the Unit Commitment
model solves the problem considering the power plant stops
problem, defining production costs functions and start-up and
random nature, problems in transmission lines as well as
shut-down costs. The defined economic operation criterion to
uncertainty in load forecasting.
minimize the cost function expected value can be obtained
through a constant periodic analysis of an equation that obtains C. Security Constrained Unit Commitment
the best combination of units to allocate. The publication [2] System safety must always be one of the most important aspects
introduces the UC problem application of DP: a major of power systems. The goal of minimizing production costs
breakthrough in power systems optimization. directly conflicts with the need to ensure the safe operation of
A. Classical Unit Commitment an energy system. Several models and methods have been
presented to solve the operational scheduling of units
Dynamic Programming was introduced by Lowery [2] and its considering safety power system restrictions. In [15], John J.
application to the UC problem has since then been studied by
Shaw proposed a new technique to solve System Constrained
several authors. Pang et al. [3] presents a DP approach for
Unit Commitment. The technique includes reserve
thermal units scheduling for a period up to 48 hours, including
requirements and power flow restrictions on transmission lines
start-up costs, spinning reserve and minimum up and down
to ensure system security.
times constraints. [4] compares several types of Dynamic
Programming, and the Priority List method. This comparison D. Profit Based Unit Commitment
showed that different versions of DP allow to achieve With the liberalization of energy markets, the goal of producing
significant reductions in costs. energy at a minimal cost has ceased to make sense, when there
Lagrangian Relaxation is proposed on [5] and [6]. The main are other competing companies on the market. On [16], the
processes involve solving the dual problem in an approximate formulation of the Profit Based UC problem is presented in a
way, ignoring load or reserve constraints. An iterative process perspective based on the new reality of the markets, where one
finds the possible dual reserve solution, properly adjusting does not try to minimize the cost, but to maximize the profit,
Lagrange multipliers. Finally, for the given viable dual reserve which makes sense on non-monopolized markets.
solution, a viable global solution is obtained by executing an
Economic Dispatch (ED) to satisfy the energy balance E. Unit Commitment with Environmental Considerations
equations. This method has been used for decades in the UC Article [17] analyze the problem with environmental
problem solving and still has a major importance in helping considerations, namely, the gas emissions from thermal units.
other more advanced methods [7]. Fuel costs and emissions minimization are contradictory
Mixed Integer Programming (MIP) is presented by John A. objectives, turning the problem into a multi-objective
Muckstadt et al. [8] for the simultaneous resolution of the optimization whose solution is represented by Pareto fronts:
thermoelectric UC and ED problems. A.I. Cohen et al. [9] compromise curves between polluting emissions and fuel costs.
present a Branch and Bound algorithm, which is a
F. Unit Commitment with Several Technologies
combinatorial optimization method that discards non-viable
solutions across lower and upper boundaries. A Mixed Integer Environmental concerns have considerably increased, a
Linear Programming (MILP) technique is proposed [10], which particularly key factor on the UC problem, due to pollutants
can be used for regulated or deregulated markets. The algorithm emissions. It was concluded that to reduce emissions of these
also provides the marginal energy price according to the substances, it would be necessary to implement new ways of
constraints of the system. introducing renewable energy technologies, which have a small
Genetic Algorithms (GA) are applied to the UC problem on the environmental impact. The article [18] present a strategy to
article [11]. GA are a class of Evolutionary Algorithms that use integrate thermal, hydro and renewable energy units.
techniques inspired by evolutionary biology such as heredity, G. Conclusions
mutation, and crossover.
The studies published during the last 50 years show a
Particle Swarm Optimization method is presented to solve the
complexity increase on the UC problem. The growing
UC problem on [12]. The method uses the particles information
complexity of the problem has been accompanied by a constant
to control the mutation operation and is similar to civil society
as a group of "leaders" influence the rest of the population in evolution of techniques used in its resolution. Nowadays, the
choosing a better decision. most used methods are Evolutionary and MIP Algorithms.
 3

III. UNIT COMMITMENT PROBLEM C. Start-up Costs

A. Problem Description Start-up costs are counted each time a generating unit is started
and are often considered constant. However, on the steam
The Unit Commitment problem consists on deciding, in a set of
turbine case, start-up costs should not be considered constant
N generation units, when each unit j ∈ N must supply energy or
because they depend on the time the unit has been switched off
not over a predefined time horizon T. In addition, it is decided,
and on the condition of the boiler, which may be hot or cold. If
for each unit in operation in the instant t ∈ to the time horizon
the boiler is kept warm during the period of inactivity, the start-
T, which is the energy Pt,j that it must produce. Therefore, the
up costs are usually modeled as a linear function over time:
problem includes two types of decisions, which are limited by
SCostj(t) = αj + γ j × HOFFj(t) (3)
load restrictions and technological constraints. For a group of N
generating units and a time horizon T, the total number of
Where HOFF is the number of hours the unit has been
possible combinations are (2N - 1)T (for 10 units and 24 hours,
continuously off until the hour t, αj [€] is the fixed start-up cost
there are 1.73x1072 possibilities). Since there often are multiple
and γj [€/h] is the cost coefficient associated with fuel
solutions that satisfy demand and certain requirements, it is
consumption to maintain the required temperature.
necessary to define a performance measure to choose the best
solution. Typically, the units scheduling is performed with the
However, if the boiler is allowed to cool, the start-up costs are
objective of minimizing the total costs of running them on
typically considered as exponentially time dependent, as in the
energy production processes.
following expression and Fig.2:
The energy demand that the units must match varies greatly
SCost(t) = α + β × (1 – eHOFF(t)/ τ) (4)
throughout the day and throughout the year. In an electrical
power system, the total system load is generally higher during Where β (€) is the cold start-up cost and τ is the cooling
the day and early evening, when industrial loads are high, and constant.
most people are awake. The load drops then during the late
evening and early morning, when most of the population is
asleep. In addition, the use of electricity has a weekly cycle,
with the load being lower on weekend days than on weekdays
and a seasonal cycle, with energy consumption being higher on
the winter than on the summer.
As already mentioned, Unit Commitment is performed with the
objective of minimizing the total costs of operation. There are
three types of costs: Cj(P), generation costs, SCost, start-up
Fig.2: Cold start cost of a steam turbine generator
costs, and shutdown costs, which are usually included in start-
up costs. The cost involved in an optimal scheduling is given On the diesel groups case, start-up costs are more difficult to
by the minimization of the total costs for all planning periods, model as they can assume intermediate levels of heating and
with Ut,j being the state (1-on and 0-off) of the generator j in fuel changes. In general, a simplified cost model is used, which
period t: can be represented by:
Min ∑𝑇𝑡=1 ∑𝑁𝑗=1[𝐶 j(Pt,j) x Ut,j + SCostt,j x (1 - Ut-1,j ) × Ut,j (1)
𝑆𝐻, 𝑖𝑓 𝐻𝑂𝐹𝐹min ≤ 𝐻𝑂𝐹𝐹(𝑡) ≤ 𝐻𝑂𝐹𝐹min + 𝐻𝑐
SCost(t)={ (5)
𝑆𝐶, 𝑖𝑓 𝐻𝑂𝐹𝐹(𝑡) > 𝐻𝑂𝐹𝐹min + 𝐻𝑐
B. Production Costs
Generally, production costs are modeled as a quadratic function Where HOFFmin is the minimum idle hours required of the
in relation to the production level. The generation cost function given unit, SH e SC are the hot and cold starting costs
illustration is given in Figure 1 and its expression is as follows: respectively, e Hc is a unit parameter such that HOFFmin+Hc
Cj(Pt,j) = aj Pt,j2 + bjPt,j + cj (2) indicates the number of hours the boiler needs to cool down.

Where aj [€/MWh2], bj [€/MWh] e cj [€/h] are the coefficients of D. Restrictions

unit j. Power systems must always meet customer demand Dt. It must
also be ensured the ability to quickly generate additional
energy, that is, a certain spinning reserve Rt should be
committed. It should be noted that each unit has a certain
minimum, PMinj, and maximum, PMaxj, production limits.

Load Demand Restrictions:

∑𝑁
𝑗=1 𝑃j × Uj = Dt (6)

Spinning Reserve Restrictions:

∑𝑁 𝑁
𝑗=1 𝑃𝑀𝑖𝑛j × Ut,j ≤ Dt + Rt ≤ ∑𝑗=1 𝑃𝑀𝑎𝑥j × Ut,j (7)
Fig.1: Quadratic Function of Thermal Production Costs
 4

Generating units impose other restrictions on their 1. Defines a small part of the problem, finding an optimal
characteristics and physical constraints. This includes unit solution for this part;
capacity and production variance (positive: ΔUpj or negative: 2. It slightly expands this small part of the problem, finding the
ΔDownj) or minimum number of hours that the unit should be optimal solution for the new problem using the optimal solution
in each state: 1-ON (HONminj) or 0-OFF (HOFFminj). previously found;
3. Continues process 2 until the expansion of the problem leads
Production Range Restrictions: to a problem that encompasses the fullness of the original
PMinj × Ut,j ≤ Pt,j ≤ PMaxj × Ut,j (8) problem. With this problem solved, the stopping conditions are
satisfied;
Ramp Restrictions: 4. The problem solution is constructed from the optimal
-ΔDownj ≤ Pt,j - Pt-1,j ≤ ΔUpj (9) solutions found for the small problems solved throughout the
process.
Minimum Uptime and Downtime Constraints:
HONj (t) ≥ HONminj and HOFFj (t) ≥ HOFFminj (10) To avoid a complete enumeration of all states, two variables can
be added to the algorithm, n and m:
• n = number of states to search in each period.
• m = number of strategies or paths to save at each step.

Fig. 3: Example of state transitions with minimum time constraints

(HONmin=3h e HOFFmin=3h) Fig. 4: Restricted search path in the Dynamic Programming algorithm
(m=3 and n=5) [19]
Further restrictions may be considered as being specific to
A. Algorithm
certain generating units or resulting from the different
production technologies that characterize energy systems.

Must Run Restriction: Some units receive mandatory execution

status during certain times of the year for voltage support on the
transmission network or other purposes such as steam supply
that can be availed for other applications.

Fuel Restrictions: Some units may have limited fuel, or

restrictions that require the burning of a specified amount of
fuel at a given time.

Hydro Restrictions: The Thermal Unit Commitment cannot be

completely separated from the scheduling of hydroelectric
units.

IV. DYNAMIC PROGRAMMING

Facing the UC problem dimensionality issues and the
inefficiency of Priority Lists, by itself, in solving the problem
for most cases, several authors have proposed the application of
Dynamic Programming to its resolution. Dynamic
Programming has many advantages over Brute Force
techniques of complete enumeration, the main one being a
possible dimensionality reduction of the problem without
sacrificing obtaining an optimal solution.
In general, Dynamic Programming is a recursive optimization
method, making a sequence of interconnected decisions: Fig. 5: Dynamic Programming algorithm flow chart [19]
 5

B. Objective Function A. Dual Optimization Process

The Dynamic Programming recursive equation applied to the In the LR method, the coupling constraints are temporarily
UC problem, at the time t for a state k, can be defined as: ignored, solving the problem as if they did not exist. This is
FCost(t,k) = minL [C(t,k) + SCost(t-1, L:t, k) + FCost(t-1 , L)] (11) done through the dual optimization procedure, which tries to
reach the optimum by maximizing the Lagrangian in relation to
Where: FCost is the lesser cumulative cost until state (t,k); the Lagrange multipliers (13), while minimizing it in relation to
SCost(t-1,L:t,k) is the start-up cost of the transition from the the other variables in the problem (14).
state (t-1 , L) to (t,k); C is the production cost of the state (t,k). q*( λ) = max λt q(λ) (13)

Given that the energy production cost function considered in q(λ) = min Ptj, Utj ℒ(P,U, λ) (14)
this work is quadratic, the lossless economic dispatch will be
computed through quadratic programming, by MATLAB The minimum of the function (14) is found for a certain optimal
function, quadprog. As the name implies, this "special" generation, Poptj, through its first derivative.
function minimizes the production costs quadratic function by There are three distinct cases to keep in mind, depending on the
analyzing different system and unit variables. The quadprog Poptj relation with the generating units limits:
function provides the optimal output that each unit must If Poptj ≤ PMinj , so:
generate to match the demand for a given combination. The
function is presented in [20]. min [Cj (Pj) - λt Pt,j ] = Cj(PMinj) - λt PMinj (15.a)
After obtaining the optimal production of each unit for a certain If PMinj ≤ Poptj ≤ PMaxj , so:
combination and hourly demand, the minimum production cost
is obtained by adding the production cost quadratic functions, min [Cj (Pj) - λt Pt,j ] = Cj(Poptj) - λt Poptj (15.b)
for Pj, of all units committed. If Poptj ≥ PMaxj , so:
For a certain hourly demand, the process of finding the
min [Cj (Pj) - λt Pt,j ] = Cj(PMaxj) - λt PMaxj (15.c)
combination with the lowest cost of production can be a
complicated task because of the many variables on which it
It should be noted that, in order to minimize q(λ) in each state
depends and the various combinations of possible states. This
and that when the state Ut,j=0 this value is zero, then the only
process will be carried out through an increasing ordering
way to get a lower value is by having:
function, with the first element being the least cost combination.
[Cj (Pj) - λt Pt,j ] < 0 (16)
To incorporate start-up costs into the DP algorithm, when a unit
changes from "0 state" to "state 1", it is sufficient to store the B. Algorithm
previous unit state in memory and compare it with the state of
the combination to be considered:
• If the difference between the current state and the previous
state is greater than zero, it means that the unit has been turned
on in the current period and the unit starting cost in question
must be added.
• If not, the unit has been switched off or it is on for at least two
consecutive hours, with no additional costs.

V. LAGRANGIAN RELAXATION
The great resulting benefits from using Lagrangian Relaxation
are the decomposition of the problem, where each unit becomes
a single entity, being optimized individually. Thus, the
commitment of each unit is done optimally, but independently
of the others. The main advantage of this method is achieved
due to the load restrictions relaxation, not requiring the
generated power to be equal to the demand at all the iterations.
A possible solution is obtained by the iterative update of
Lagrange multipliers, which approaches the relaxed solution,
called the dual solution, to the solution that respects the load
restrictions: the so-called primal solution.
The Lagrangian function is formed in the same way as in the
Economic Dispatch problem resolution [19]:
ℒ = ∑𝑇𝑡=1 ∑𝑁𝑗=1[𝐶 j(Pt,j) + SCostt,j ] × Ut,j + ∑𝑇𝑡=1 𝜆t Dt - ∑𝑇𝑡=1 ∑𝑁𝑗=1 𝜆t Pt,j Ut,j (12)

Where ℒ is the Lagrangian and λ is the Lagrange multiplier.

Fig. 6: Lagrangian Relaxation algorithm flow chart [19]

 6

C. Duality Gap 3. Compare fitness of each particle their best pbest so far. This
The difference between the primal and dual solutions is called process can be defined as follows:
duality gap, which must be minimized by approaching the dual 𝑝𝑏𝑒𝑠𝑡𝑘𝑖 = 𝑓𝑖𝑡𝑛𝑒𝑠𝑠(𝑥𝑘𝑖 )
If 𝑓𝑖𝑡𝑛𝑒𝑠𝑠(𝑥𝑘𝑖 ) < pbestik ,Then: { 𝑖 (20)
value to the primal value. 𝑥𝑘,𝑝𝑏𝑒𝑠𝑡 = 𝑥𝑘𝑖
The interval between the maximization of the dual cost
functions (q*) and the minimization of the primal cost functions 4. Compare fitness of all particles to find gbest, according to the
(J*), called the relative duality gap (Δ), can be used as a stop following expression:
criterion. Therefore, the duality gap is used as a measure of 𝑔𝑏𝑒𝑠𝑡 𝑖 = 𝑓𝑖𝑡𝑛𝑒𝑠𝑠(𝑥𝑘𝑖 )
If 𝑓𝑖𝑡𝑛𝑒𝑠𝑠(𝑥𝑘𝑖 ) <gbesti , Then: { 𝑖 (21)
convergence, being given by the following expression: 𝑥𝑔𝑏𝑒𝑠𝑡 = 𝑥𝑘𝑖
𝐽∗ − 𝑞∗
Δ= 𝑞∗
(17) 5. Update the velocity of each particle to the next iteration,
where w is the weight of inertia, c1 and c2 are constants and
D. Adjusting λ
rand is a random variable that assumes values between 0 and 1
Throughout this work, the Lagrange multipliers adjustment will in a uniformly distributed form:
be done as follows, as proposed in [19]: vk i+1= w ×vk i+c1 ×rand ×(𝑥𝑘,𝑝𝑏𝑒𝑠𝑡 - 𝑥𝑘𝑖 )+c2 ×rand ×( 𝑥𝑘,𝑝𝑏𝑒𝑠𝑡 - 𝑥𝑘𝑖 ) (22)
𝑑
λt = λt + [ 𝑑𝜆 q(λ) ] ε (18)
𝑤𝑚𝑎𝑥 − 𝑤𝑚𝑖𝑛
Where: w = 𝑤𝑚𝑎𝑥 − ×𝑖 (23)
𝑖𝑚𝑎𝑥
𝑑
ε = 0.01 When
𝑑𝜆
q(λ) > 0 (19.a)
6. Move each particle to a new position:
𝑑 0 𝑠𝑒 𝑟𝑎𝑛𝑑 ( ) ≥ 𝑠𝑖𝑔𝑚𝑜𝑖𝑑(𝑣𝑘𝑖+1 )
ε = 0.002 When q(λ) < 0 (19.b) 𝑥𝑘𝑖+1 = { (24)
𝑑𝜆 1 𝑠𝑒 𝑟𝑎𝑛𝑑 ( ) < 𝑠𝑖𝑔𝑚𝑜𝑖𝑑(𝑣𝑘𝑖+1 )
With:
VI. PARTICLE SWARM OPTIMIZATION
The particle swarm simulates a type of social optimization 7. Repeat from step 2 until desired convergence is achieved.
where each proposed solution (particle) is evaluated through an
B. Algorithm
eligibility function, with the best leading the rest through the
iterative process. This process is initiated to iteratively improve
the candidate solutions, with the particles evaluating the
suitability of these solutions, remembering the location where
they had the best success. Each particle provides this location
to adjacent particles, which also have access to the location
where their neighbors have been most successful. The
movements through the search space are guided by these
successes, with the population converging, generally, to the best
solution of the problem (or near it), at the end of the process.
Initially, the particles can randomly move in the search space,
shifting according to the velocities derived from their current
position and that of other particles in the swarm. The objective
function of the problem is solved using each particle, and its
assessed value is called fitness. Particles cross the entire search
space with their own experience and that of others, with
experience meaning their best physical value.
The best individual fitness value is stored as pbest (personal
best) and the best of all particles is stored as gbest (global best).
The social interaction between particles causes them to learn
from each other and motivate them to move to better positions.
So, the next movement of a particle, in a certain iteration, is
motivated by experience, with its new velocity being influenced
by its better physical fitness and that of its neighbors.
A. Procedures
The PSO algorithm steps based on a fully connected network
will be discussed below:
1. Randomly initialize the swarm of dimension equal to E
particles, with a particle pk on the position xk1 of the search
space, in iteration 1;
2. Compute the fitness value of each particle fitness(xi);
Fig. 7: PSO algorithm flow chart [21]
 7

VII. CASE STUDY B. Application of LR

The case study that will serve as a test for the developed The algorithm reaches the stop conditions, Δ <0.06, after 17
algorithms consists of a daily schedule (24 hours) of 10 units, iterations, with about 4 seconds of computation time and a total
the characteristics of which are found in the Appendix, Table operational cost of CTotal =459996€.
A. The demand to which the generating units must meet follows
a typical daily load profile whose values are presented in Table
1.
TABLE 1
LOAD DEMAND ON THE CASE STUDY IN [MW]
Hour t 1 2 3 4 5 6
Dt 600 500 500 700 750 850
Hour t 7 8 9 10 11 12
Dt 900 1000 1050 1100 1200 1200
Hour t 13 14 15 16 17 18
Dt 1000 900 800 700 900 1100
Hour t 19 20 21 22 23 24
Dt 1300 1500 1600 1400 1000 800

Fig.9: Hourly Production Costs and Load Demand versus Combined

A. Application of DP Maximum Power by LR
For this case study, the developed algorithm takes tcomp=1760
seconds (almost 30 minutes) to get a final solution with a total The iterative evolution of the primal J* and dual q* solutions as
operational cost of CTotal =451251€. The Fig.8 shows the hourly well as of Δ can be verified in Figure 10. The iterative evolution
costs evolution as well as the evolution of the load demand of λt is represented by Figure 11.
versus the combined maximum power available, ∑𝑁𝑗 𝑃𝑀𝑎𝑥𝑗 × 𝑈𝑗 .

Fig. 10: Iterative Evolution of the Dual Optimization

Fig.8: Hourly Production Costs and Load Demand versus Combined

Maximum Power by DP

Evaluating the obtained results, it can be verified that the

obtained combinations agree with the considered restrictions. Fig. 11: Iterative Evolution of λt
In the first hours, being the load demand minimum, only the
base units supply energy, since they have greater capacity of The final solution presents an excess of reserve in most hours
generation and a lower cost of production. Essentially, the base which contrasts with the solution obtained by Dynamic
units stay connected to the network throughout the scheduling Programming. For that reason, the total cost is much higher. As
because they have the lowest production cost and because they can be seen in Figure 10, the duality gap becomes quite small
must be connected for many hours. As demand increases, the as the dual optimization proceeds. The values of λt greatly
most economical available units will be initialized, which increase in the first 3 or 4 iterations, which translates into a
should stay on for a few hours. However, in certain cases it is rapid growth of the q* solution. λt then stabilize which causes
convenient, in a small momentary energy spike, to initialize a the stabilization of q*. Convergence is unstable near the end of
more expensive unit rather than a more economical one, if this the process, meaning that some units are "switched on" and
more expensive unit has no running time restrictions and has a "off", which causes instability in the final solution. Primal
low start-up cost. This happens at hours 11 and 12, with the solution J* is initially defined as a much higher value (1 million
algorithm opting 2 hours for the unit 8, to the detriment of €) than the expected solution value and is only minimized when
several cheaper units that were available to be initialized but the dual solution provides a schedule that meets demand at all
have more restricted conditions. times and respects all constraints.
 8

C. Application of PSO
Table 2 presents, for each method, the computation time, the
The characteristic PSO parameters used in this problem will be
total cost and its relative percentage difference in relation to the
E=50, imax=10, wmax=0.9 e wmin=0.4. The computational
optimal solution cost.
execution of the ten iterations took 50 seconds with a total
operational cost of CTotal =452510€.
TABLE 2
COMPARATIVE DATA BETWEEN ALGORITHMS
DP LR PSO
tcomp [s] 1766s 4.29 49.7
CTotal [€] 451251 459996 452510
𝑶𝒑𝒕 𝑶𝒑𝒕
(𝑪𝑻𝒐𝒕𝒂𝒍 − 𝑪𝑻𝒐𝒕𝒂𝒍 )/𝑪𝑻𝒐𝒕𝒂𝒍 [%] 0 1.94 0.28

VIII. WIND AND HYDRO-THERMAL UNIT COMMITMENT

The increased environmental considerations led to the
incorporation of renewable energy technologies in production
systems. The use of renewable energy, such as wind energy
(through wind turbines), as well as hydropower (through
Fig.12: Hourly Production Costs and Load Demand versus Combined hydroelectric plants), has the great advantage of having a small
Maximum Power by PSO environmental impact when compared with thermal energy
production technologies. Hydroelectric power has been used for
Fig. 13 shows the iterative evolution of the total cost of more than a century and is a technology that is relatively
operation, gbest. developed, with its market share stabilized (except for some
fluctuations inherent to the various levels of precipitation on
each year). The use of new renewable energy technologies on
electricity production has been increasingly implemented. In
the last two decades, these alternative sources of energy
production have gained prominence in the market in several
countries of the world, a trend that will become even more
pronounced soon.
Fig.13: iterative evolution gbest.
A. Problem Description
For a relatively complex case study like the present, obtaining The Thermal, Hydro and Wind UC problem can be solved using
a satisfactory solution is not as fast (iteratively) as it is for a a method that combines Lagrangian Relaxation, with
simple case with only a few viable solutions. In cases with many Sequential Unit Commitment (SUC) and Unit Decommitment
units (≥10 generators), there are possibly many solutions that (UD) [22]. It can be assumed that, at a certain time, wind energy
respect all problem constraints. On the random process of is not subject to dispatch, having priority access to the grid.
particle creation, there is hardly one that represents the optimal Knowing the hourly wind generation of energy, Pwindt, through
solution at the first iteration. In theory, the mechanisms that forecasting methods, it is enough to commit the hydro and
govern the PSO nature enhance an iterative improvement of the thermal units, as shown on the flowchart of Fig.15. The
solution, translating into a decrease of the gbest value until the hydrothermal units will only have to meet a load equal to the
final iteration. For this case, the solution obtained by the PSO hourly demand less the generation that the wind turbines will
algorithm, gbest, is improved up to iteration 8, as the particles supply (25). This simplification can be done because, if the
move to less expensive solutions. wind power units are able to generate energy (i.e. wind speed is
sufficient to be profitable to produce energy), all this energy
D. Comparison of Algorithms will have priority, being definitively supplied to the grid (except
Figure 14 compares the hourly costs of the final solution losses of energy). It should also be assured that the possible
obtained by each method: DP, LR and PSO. maximum hydrothermal production is greater than the demand
and reserve minus the wind generation.
∑𝑁 𝐻
𝑗=1 𝑃t,j × Ut,j + ∑ℎ=1(𝑃ℎt,h × Ut,h - Ppumpt,h) + Pwindt = Dt (25)

∑𝑁 𝐻
𝑗=1 𝑃𝑀𝑎𝑥 j×Ut,j + ∑ℎ=1(𝑃ℎ𝑀𝑎𝑥h×Ut,h -Ppumpt,h) ≥Dt+Rt- Pwindt (26)

Pwindt = φwt × PwindMax (27)

With: Phh being the production of a h hydro unit, Ppump the

energy spent on water pumping and PhMax being the maximum
production of a hydro unit. PwindMax is the maximum wind
production and φwt is the [%] of availability of wind on hour t.
Fig.14: Comparison of Hourly Costs for each method
 9

B. Algorithm of Hydro-Thermal UC [3] C. K. Pang, H. C. Chen, “Optimal short-term thermal unit

commitment”, IEEE Transactions on Power Apparatus
and Systems, Vol. PAS-95, No 4, July/August 1976.
[4] C. K. Pang, G. B. Sheble, F. Albuyeh, “Evaluation of
dynamic programming based methods and multiple area
representation for thermal unit commitments”, IEEE
Transactions on Power Apparatus and Systems, Vol. PAS-
100, No 3, March 1981.
[5] A. Merlin, P. Sandrin, “A new method for unit
commitment at Electricite de France”, IEEE Transactions
on Power Apparatus and Systems, Vol. PAS-102, No 5,
May 1983.
[6] Fulin Zhuang, F.D. Galiana, “Towards a more rigorous and
practical unit commitment by Lagrangian Relaxation”,
IEEE Transactions on Power Systems, vol.3, no.2, pp.763-
773, May 1988.
Fig.15: Hydro-Thermal UC Flowchart through LR, SUC and UD [22] [7] P. Sriyanyong, Y.H. Song, “Unit commitment using
particle swarm optimization combined with Lagrange
relaxation”, IEEE Power Engineering Society General
IX. CONCLUSIONS
Meeting, Vol. 3, pp.2752-2759, June. 2005.
It can be concluded that all developed algorithms solve the Unit
[8] John A. Muckstadt and Richard C. Wilson, “An
Commitment Problem successfully. Each of the studied
methods has advantages and disadvantages in their UC problem application of mixed-integer programming duality to
applications: scheduling thermal generating systems”, IEEE
• Dynamic Programming: It has the advantage of reaching an Transactions on Power Apparatus and Systems, vol.PAS-
optimal solution. Despite this, it presents a dimensionality 87, no.12, pp.1968-1978, Dec. 1968.
problem that manifest itself on a high computation time for real
[9] A. I. Cohen, M. Yoshimura, “A branch-and-bound
systems.
• Lagrangian Relaxation: Achieves a viable solution very algorithm for unit commitment”, IEEE Transactions on
quickly, even for complex systems. Its disadvantage is the poor Power Apparatus and Systems, Vol. PAS-102, No 2,
economical solution sometimes achieved. February 1983.
• Particle Swarm Optimization: It is the method that best [10] Chang G.W, Tsai Y.D, Lai C.Y and Chung J.S, “A
balances a solution close to the optimum, with a good
practical mixed integer linear programming based
computation time. However, it rarely achieves an optimal
solution. approach for unit commitment”, IEEE Power Engineering
Nowadays, we face many challenges on the electricity Society General Meeting, vol.1, pp. 221- 225, June 2004.
generation field, due to the growing focus on renewable [11] D. Dasgupta, D. R. McGregor, “Thermal unit commitment
energies, which have great ecological advantages. There have using genetic algorithms”, IEEE Transactions on Power
been major changes in energy systems, modifying their
Systems, Vol. 141, No 5, September 1994.
operational paradigms. The commitment to increase the
electricity production from renewable energies, characterized [12] Zwe-Lee Gaing, “Discrete Particle Swarm Optimization
by great variability (and unpredictability), must be incorporated Algorithm for Unit Commitment”, IEEE Power
in the Unit Commitment Problem, increasing, however, its Engineering Society General Meeting, Vol.1, pp. 418-424,
complexity of resolution. July 2003.
[13] A.H. Mantawy, Y. L. Abdel-Magid, S.Z. Selim,
REFERENCES
“Integrating genetic algorithms, tabu search, and simulated
[1] R. H. Kerr, J. L. Scheidt, A. J. Fontana and J. K. Wiley,
annealing for the unit commitment problem”, IEEE
“Unit Commitment”, IEEE Transactions on Power
Transactions on Power Systems, Vol. 14, No. 3, August
Apparatus and Systems. Vol. PAS-85, No. 5, pp. 417-421.
1999.
May 1966.
[2] P. G. Lowery, “Generating Unit Commitment by Dynamic [14] D. Zhai, A. M. Breipohl, F.N. Lee, R. Adapa, “The effect
Programming”, IEEE Transaction on Power Apparatus of load uncertainty on unit commitment risk”, IEEE
and Systems, Vol. PAS-85, No. 5, pp. 422-426. May 1966. Transactions on Power Systems, Vol. 9, No 1, February
1994.
 10

[15] John J. Shaw, “A direct method for security constrained [22] Chao-An Li, R. B. Johnson, A. J. Svoboda, Chung-Li
unit commitment”, IEEE Transactions on Power Systems, Tseng and E. Hsu, "A robust unit commitment algorithm
vol.10, no.3, pp. 1329-1342, Aug. 1995. for hydro-thermal optimization", Proceedings of the 20th
[16] C. W. Richter, G. B. Sheblé, “A profit-based unit International Conference on Power Industry Computer
commitment GA for the competitive environment”, IEEE Applications, Columbus, OH, 1997, pp. 186-191.
Transactions on Power Systems, Vol. 15, No. 2, May 2000.
[17] J. Catalão, S. Mariano, V. Mendes, L. Ferreira, “Unit
commitment with environmental considerations: a
practical approach”, 15th PSCC, Liege, 22-26 August
2005.
[18] Castro M, Peire J, Carpio J, Valcarcel M and Aldana F,
“Large Grid Solar Energy Integration Analysis”,
Proceedings of Annual Conference of Industrial
Electronics Society, vol.3, pp.575-580, Oct 1988.
[19] Allen J. Wood and Bruce F. Wollenberg, “Power
Generation, Operation and Control”, Wiley & Sons, USA,
1996.
[20] https://courses.csail.mit.edu/6.867/wiki/images/e/ef/Qp-
quadprog.pdf

[21] Puri, V. (2009). “Unit Commitment Using Particle Swarm

Optimization”, MSc. Thesis, Thapar University, India.

APPENDIX
TABLE A
DATA OF THE GENERATING UNITS USED ON THE CASY STUDY
Unit Pminj PMaxj aj bj cj SCostj HONminj HOFFminj Ini. State
j [MW] [MW] [€/MW2] [€/MW] [€/h] [€] [h] [h] [h]

1 150 400 0.0005 15 800 4000 7 7 8 HON

2 150 400 0.0003 16 900 3500 7 7 5 HON
3 50 200 0.002 16 500 1000 5 5 2 HOFF
4 40 150 0.002 17 650 800 5 4 5 HOFF
5 40 150 0.003 20 450 500 5 5 6 HOFF
6 35 110 0.007 23 350 200 3 3 5 HOFF
7 30 90 0.008 26 440 200 3 3 5 HOFF
8 20 60 0.005 25 660 150 1 1 5 HOFF
9 10 50 0.002 27 650 100 1 1 6 HOFF
10 10 50 0.0015 28 600 100 1 1 6 HOFF