2030 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 22, NO.

4, NOVEMBER 2007

Hybrid Particle Swarm Optimization Approach for

Solving the Discrete OPF Problem Considering
the Valve Loading Effects
M. R. AlRashidi, Student Member, IEEE, and M. E. El-Hawary, Fellow, IEEE

Abstract—This paper presents a hybrid particle swarm opti- backbone tool that has been extensively researched since its
mization algorithm (HPSO) as a modern optimization tool to solve first introduction by Carpentier in 1962 [1].
the discrete optimal power flow (OPF) problem that has both The goal of OPF is to find the optimal settings of a given
discrete and continuous optimization variables. The problem is
classified as constrained mixed integer nonlinear programming
power system network that optimize a certain objective function
with multimodal characteristics. The objective functions consid- while satisfying its power flow equations, system security, and
ered are the system real power losses, fuel cost, and the gaseous equipment operating limits. Different control variables, some of
emissions of the generating units. Two different types of fuel cost which are generators’ real power outputs and voltages, trans-
functions are considered in this study, namely the conventional former tap changing settings, phase shifters, switched capaci-
quadratic function and the augmented quadratic function to tors, and reactors, are manipulated to achieve an optimal net-
introduce more accurate modeling that incorporates the valve
loading effects. The latter model presents nondifferentiable and work setting based on the problem formulation. A major dif-
nonconvex regions that challenge most gradient-based optimiza- ficulty of the OPF problem is the nature of the control vari-
tion algorithms. The proposed algorithm makes use of the PSO, ables since some of them are continuous (real power outputs
known for its global searching capabilities, to allocate the optimal and voltages) and others are discrete (transformer tap setting,
control settings while Newton–Raphson algorithm minimizes phase shifters, and reactive injections). The most commonly
the mismatch of the power flow equations. A hybrid inequality
constraint handling mechanism that preserves only feasible solu-
used objective is the minimization of the overall fuel cost func-
tions without the need to augment the original objective function tion. However, other traditional objectives are minimization of
is incorporated in the proposed approach. To demonstrate its active power loss, bus voltage deviation, emission of generating
robustness, the proposed algorithm was tested on the IEEE 30-bus units, number of control actions, and load shedding. Deregu-
system with six generating units. Several cases were investigated lation of the electric power industry has also introduced new
to test and validate the consistency of detecting optimal or near objectives to the OPF problem like maximization of the social
optimal solution for each objective. Results are compared to solu-
tions obtained of MATPOWER software outcomes that employs welfare and individual supplier’s profit [2], [3].
sequential quadratic programming algorithm to solve the OPF. Researchers proposed different mathematical formulations of
The impact of the proposed inequality constraint handling method the OPF problem. It can be broadly classified as follows.
in improving the HPSO performance is illustrated. Furthermore, 1) Linear problem in which objectives and constraints are
a study of HPSO parameters tuning with regard to the OPF given in linear forms with continuous control variables.
problem is presented and analyzed.
2) Nonlinear problem where either objectives or constraints or
Index Terms—Emission, optimal power flow (OPF), particle both combined are linear with continuous control variables.
swarm optimization, power system operation. 3) Mixed-integer linear and linear problems when control
variables are both discrete and continuous.
Various traditional optimization techniques were developed
I. INTRODUCTION to solve the OPF problem but the most popular ones are linear
programming, sequential quadratic programming, generalized

E LECTRIC power grids are considered the most complex

man-made systems mainly due to their wide geographical
coverage, various transactions among different utilities, and
reduced gradient method, and Newton method. The reader is
referred to [4] and [5] for a complete list of the most commonly
used conventional optimization algorithms with regard to the
diversity in individual electric power companies’ layouts, size, OPF. Despite the fact that some of these techniques have excel-
and equipment used. Engineers need special tools to optimally lent convergence characteristics and some are widely used in the
analyze, monitor, and control different aspects of such sophis- industry, some of their drawbacks are as follows.
ticated system. Some of these tools are economic dispatch, unit 1) They are local optimizers in nature, i.e., they might con-
commitment, state estimation, automatic generation control, verge to local solutions instead of global ones if the initial
and optimal power flow (OPF). The latter is regarded as the guess happens to be in the neighborhood of a local solu-
tion. This occurs as a result of using Kuhn–Tucker condi-
Manuscript received February 16, 2007; revised April 21, 2007. Paper no.
TPWRS-00136-2007.
tions as termination criteria to detect stationary points. This
The authors are with the Department of Electrical and Computer Engineering, practice is commonly used in most commercial nonlinear
Dalhousie University, Halifax, NS, B3J 2X4 Canada (e-mail: [email protected]; optimization programs [6].
[email protected]).
Color versions of one or more of the figures in this paper are available online
2) Each technique is tailored to suit a specific OPF optimiza-
at http://ieeexplore.ieee.org. tion problem based on the mathematical nature of the ob-
Digital Object Identifier 10.1109/TPWRS.2007.907375 jectives and/or constraints.
0885-8950/$25.00 © 2007 IEEE
ALRASHIDI AND EL-HAWARY: HYBRID PARTICLE SWARM OPTIMIZATION APPROACH FOR SOLVING THE DISCRETE OPF PROBLEM 2031

3) The theoretical assumptions behind these developed al- subject to

gorithms that may not be suitable for the OPF formula-
tion like convexity, differentiability, and continuity among (2)
other things. (3)
A new category of classical optimization tools has emerged
to cope with some of the traditional optimization algorithms’ where vector denotes the state variables of a power system
shortcomings. The main modern optimization techniques are network that contains the slack bus real power output ,
genetic algorithm (GA), evolutionary programming (EP), ar- voltage magnitudes and phase angles of the load buses
tificial neural network (ANN), simulated annealing (SA), ant , and generator reactive power outputs . Vector
colony optimization (ACO), and particle swarm optimization represents both integer and continuous control variables that
(PSO). Most of these relatively new developed tools mimic a consist of real power generation levels and voltage
certain natural phenomenon in its search for an optimal solu- magnitudes , transformer tap setting , and reactive
tion like species evolution (GA and EP), human neural system power injections due to volt-amperes reactive (VAR)
(ANN), thermal dynamics of a metal cooling process (SA), or compensations; i.e.,
social behavior (ACO and PSO). They have been successfully
applied to a wide range of optimization problems in which
global solutions are more preferred than local ones.
PSO is one of the new optimizers that have been investigated (4)
to solve the OPF problem. Researchers in [7]–[9] have attempted
A. Objective Function
to utilize the PSO to solve the OPF problem considering different
objective functions. In their mathematical formulation, only In this study, minimization of different objectives is consid-
continuous control settings were considered as optimization ered to examine the performance of the proposed algorithm. The
variables which restricts its applicability to real power systems. objective functions taken into considerations are fuel emission,
In a different formulation, authors in [10]–[12] employed the fuel cost, and the network real power losses. Each objective is
PSO to solve the OPF with the inclusion of both discrete and briefly described as follows:
continuous optimization variables. Then, they augmented the 1) Fuel Emission: Fossil-based thermal plants are consid-
OPF objective function by adding penalty terms to transform ered a major player in the pollution crisis that we are facing
the constrained OPF into an unconstrained one. This approach nowadays. The industrial growth led to greater demands to
usually encounters a major difficulty in how to properly select generate more electricity. Consequently, the emission of these
penalty factor values. If the penalty factors selected are high, generating units is gradually building up in the atmosphere
the optimization algorithm will get trapped in local solutions. which is having a severe impact on our environment. One way
On the other hand, the algorithm may not be able to detect to cope with this problem is to dispatch electric power with
a feasible solution if the penalty factors are low [9]. emission considerations. The objective of the fuel emission dis-
In this paper, a discrete HPSO algorithm that handles both patch problem is to minimize the total emission of all thermal
discrete and continuous optimization variables is proposed to units by allocating optimal control settings while satisfying var-
tackle the OPF problem. It presents an example of how HPSO ious network operation constraints. Fuel emission of a number
can easily incorporate other optimization methods, in this case of generating units can be modeled mathematically as follows:
Newton–Raphson method, to solve part of the overall problem.
Newton–Raphson iterative technique is incorporated within the
ton/h (5)
HPSO algorithm to minimize the power mismatch (i.e., the
equality constraints) in the power flow equations. In the pro-
posed algorithm, a hybrid mechanism for handling inequality 2) Fuel Cost: The aim of the fuel cost dispatch problem is to
constraints is developed to reinforce constraints avoiding the allocate the best network settings that minimize the overall fuel
use of penalty factors or augmenting the objective function. cost function while imposing all network constraints. Conven-
This mechanism makes use of the memory element embedded tionally, the overall fuel cost function for a number of thermal
in the PSO theory. The objectives considered in this study are generating units can be modeled by a quadratic function (convex
the fuel cost (convex and convex) and gaseous emission of the and differentiable) as follows:
generating units, and the total network real power losses. A
study of HPSO parameters tuning with regard to the OPF is $ h (6)
also presented and analyzed.

However, this model ignores the valve point loading that intro-
II. PROBLEM FORMULATION duces rippling effects to the actual input–output curve. Equation
The OPF goal is to optimize a certain objective subject to (6) is modified by adding an additional sine term to account for
several equality and inequality constraints. The problem can be the valve effects in this manner [13]
mathematically modeled as follows:
$ h

(1) (7)
2032 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 22, NO. 4, NOVEMBER 2007

3) The discrete transformer tap settings

(14)

4) The discrete reactive power injections due to capacitor

banks

(15)

Note that , and are continuous variables while

and are discrete ones.
5) The transmission lines loading
Fig. 1. Generator input–output curve considering the valve point effects.
(16)

This more accurate modeling adds more challenges to most III. PROPOSED HYBRID ALGORITHM
derivative-based optimization algorithms in finding the global
PSO is a population based stochastic global optimizer that
solution since the objective is no longer convex nor differen-
was first introduced primitively by Eberhart and Kennedy in
tiable everywhere. Fig. 1 shows the shape of the fuel cost func-
1995 [14], [15]. The original PSO model was intended to handle
tion with the valve loading effects included.
only continuous nonlinear optimization problems. However,
3) Real Power Losses: With this objective, all control set-
further enhancements elevated the PSO capabilities to solve a
tings are adjusted such that the total real power losses are mini-
wider class of problems. The reader is directed to [16]–[18]
mized. Power losses can be modeled as follows:
for more information on PSO recent developments. The idea
behind this optimizer came from early attempts conducted by
(8) Eberhart and Kennedy to model the flocking behavior of many
species, like birds or school of fish, in their food hunting. They
realized that such species try to approach their target in an
where is the number of transmission lines in the system, optimal manner which resembles finding the optimal solution
is the conductance of the line connecting buses and , and to any mathematical optimization problem.
the bus voltage is represented in polar form by and . In PSO, a number of particles form a swarm that evolve or
fly throughout the problem hyperspace to search for optimal or
B. Constraints near optimal solution. The coordinates of each particle represent
The OPF problem has two categories of constraints: a possible solution with two vectors associated with it, the po-
1) Equality Constraints: These are the sets of nonlinear sition and velocity vectors. In -dimensional search
power flow equations that govern the power system, i.e., space, and
are the two vectors associated with each particle . During their
search, particles interact with each others in a certain way to
(9) optimize their search experience. There are different variants
of the particle swarm paradigms but the most general one is
the gbest model where the whole population is considered as a
(10)
single neighborhood throughout the optimization process [19].
In each iteration, the particle with the best solution shares its
where and are the real and reactive power outputs in- position coordinates (gbest) information with the rest of the
jected at bus respectively, the load demand at the same bus is swarm. Then, each particle updates its coordinates based on its
represented by and , and elements of the bus admit- own best search experience (pbest) and gbest according to the
tance matrix are represented by and . following equations:
2) Inequality Constraints: These are the set of continuous
and discrete constraints that represent the system operational
and security limits like the bounds on the following. (17)
1) The generators real and reactive power outputs (18)

(11) — and are two positive acceleration constants, they keep

(12) balance between the particle’s individual and social be-
havior when they are set equal.
2) Voltage magnitudes at each bus in the network — and are two randomly generated numbers with a
range of [0, 1] added in the model to introduce stochastic
(13) nature.
ALRASHIDI AND EL-HAWARY: HYBRID PARTICLE SWARM OPTIMIZATION APPROACH FOR SOLVING THE DISCRETE OPF PROBLEM 2033

— is the inertia weight and it keeps a balance between ex-

ploration and exploitation. It is a linearly decreasing func-
tion of the iteration index

(19)

— is the iteration index.

The most important factor that governs the PSO performance
in its search for optimal solution is to maintain a balance
between exploration and exploitation. Exploration is the PSO
ability to cover and explore different areas in the feasible search
space while exploitation is the ability to concentrate only on
promising areas in the search space and to enhance the quality
of the potential solution in the promising region. Exploration
requires bigger step sizes at the beginning of the optimization
process to determine the most promising areas then the step
size is reduced to focus only on that area.
Recently, PSO developments and applications have been
widely explored in engineering and science mainly due to
its distinct favorable characteristics. Just like in the case of
other evolutionary algorithms, PSO has many key features that
attracted many researchers to employ it in different applications
in which conventional optimization algorithms might fail such
as follows.
• It only requires a fitness function to measure the “quality”
of a solution instead of complex mathematical operations Fig. 2. Flow chart diagram of the proposed algorithm.
like gradient, Hessian, or matrix inversion. This reduces the
computational complexity and relieves some of the restric-
tions that are usually imposed on the objective function like subroutine. This mechanism of multiple initial solutions can
differentiability, continuity, or convexity. provide better probability of detecting an optimal solution to
• It is less sensitive to a good initial solution since it is a the power flow equations that would globally minimize a given
population-based method. objective function. The importance of such hybridization is
• It can be easily incorporated with other optimization tools signified by realizing the fact that in a transmission system, the
to form hybrid ones. solution to the power flow equation is not unique, i.e., multiple
• It has the ability to escape local minima since it follows solutions within the stability margins may exist and only one
probabilistic transition rules. can globally optimize a certain objective [20]. The flow chart
More interesting PSO advantages can be emphasized when of the proposed algorithm is shown in Fig. 2.
compared to other members of evolutionary algorithms like the
following. A. Constraints Handling Methods
• It can be easily programmed and modified with basic math-
ematical and logic operations. There are different ways to handle constraints in evolutionary
• It is inexpensive in terms of computation time and memory. computation optimization algorithms just like in the case of the
• It requires less parameter tuning. PSO. The following constraint handling methods are the most
• It works with direct real-valued numbers that eliminates the commonly used ones [21].
need to do binary conversion of classical canonical genetic 1) Preserving feasible solution method: In this method, solu-
algorithm. tions are initially placed in the feasible search space and
The proposed hybrid approach combines PSO technique remain within by adapting an update mechanism that gen-
with Newton–Raphson based power flow program in which erates only feasible solutions.
the former technique is used as a global optimizer to find the 2) Infeasible solution rejection method: This approach rejects
best combinations of the mixed type control variables while the any solution that violates the feasible search space.
latter serves as a minimizer to reduce the nonlinear power flow 3) Penalty function method: in which a penalty factor is added
equations mismatch. Newton–Raphson method used in this to the objective once any constraint violation occurs.
implementation is the one with the full Jacobian evaluated and 4) Solution repair method: This approach converts the infea-
updated at each iteration. The HPSO utilizes a population of sible solution to a feasible one by performing special op-
particles or possible solutions to explore the feasible solution erations.
hyperspace in its search for an optimal solution. Each particle’s Selecting the proper constraints handling method is highly re-
position is used as a feasible initial guess for the power flow liant on the problem’s nature. Reference [21] indicates that, in
2034 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 22, NO. 4, NOVEMBER 2007

the solution repair method, the process of reinstating the infea- require special initialization and treatment of the position
sible solution to a feasible one can be as challenging as solving vector throughout the optimization process. The continuous
the original problem. variables are initialized with uniformly distributed pseudo-
Within the context of PSO applications to the OPF and eco- random numbers that take the range of these variables, e.g.,
nomic dispatch, inequality constraints that represent the permis- and .
sible operating range of each optimization variable are typically However, in the case of the discrete variables, an additional
handled in the following two ways: operator was needed to account for the distinct nature of these
— Set to limit approach (SLA): If any optimization variable variables. A rounding operator was included to ensure that
exceeds its upper or lower bound, the value of the variable each discrete variable is rounded to its nearest decimal integer
is set to the violated limit. This resembles the idea found value that represents the physical operating constraint of a
in operating all generating units at equal incremental pro- given variable. Each transformer tap setting was rounded to its
duction cost to reach optimal power dispatch. Once any nearest decimal integer value of 0.01 by utilizing the rounding
unit violates its operating range, the real power is set to the operator as follows . The
violated limit. It is important to note that PSO has some same principle applies to the discrete reactive injection due
randomness in the update equation that might cause sev- capacitor banks with the difference being the step size, i.e.,
eral variables to exceed their limits during the optimization . This ensures that the fitness
process. Thus, this approach may fix multiple optimization of each solution is measured only when all elements of the
variables to their operating limits for which global solution solution vector are properly represented to reflect the real world
may not be reached. Also this approach fails to utilize the nature of each variable. Since the particle update equations
memory element that each particle has once it exceeds its have some uniformly distributed random operators built into it
boundaries. References [22]–[25] employ this approach. and because of the addition of two different types of vectors,
— Penalty factor: The other approach is to use penalty factors the rounding operator is called again after each update to act
to incorporate the inequality constraints with the objective. only on the discrete variables as follows: and
The main problem with this approach is introducing new . Once the rounding process is over, all solution
parameters that need to be properly selected in order to elements go through a feasibility check. This simple rounding
reach acceptable PSO performance. Values of the penalty method guarantees that power flow calculations and fitness
factors are problem dependant; thus, this approach requires measurements are obtained only when all problem variables
proper adjustments of the penalty factors in addition to are properly addressed and their nature types are accounted for.
tuning the PSO parameters. References [8], [11], and [12]
make use of this approach.
1) Hybrid Inequality Constraints Handling Mechanism: IV. SIMULATION RESULTS AND DISCUSSION
The inequality constraints are handled by a hybrid constraints The proposed algorithm was implemented in Matlab com-
handling mechanism that brings together the concepts of two puting environment and the standard IEEE 30-bus test system
common methods currently used in evolutionary computation was used to validate its potential. The test system consists of six
optimization methods to deal with constraints. It combines the generating units interconnected with 41 branches of a transmis-
ideas of preserving feasible solution and infeasible solution sion network to serve a total load of 189.2 MW and 107.2 Mvar
rejection methods to retain only feasible solutions throughout as shown in Fig. 3. Detailed description of the system’s data is
the optimization process without the need to introduce penalty presented in the Appendix, [26], [27]. The emission data used in
factors in the objective function. In most of the evolutionary this study are for nitrogen oxides . Note that the original
computation optimization methods that employ the infeasible system has two capacitors banks installed at bus 5 and 24 with
solution rejection method to handle constraints, any solution ratings of 19 and 4 Mvar, respectively. A series of experiments
candidate among the population is randomly re-initialized once were conducted to properly tune the HPSO parameters to suit
it crosses the boundaries of the feasible region. The majority the targeted OPF problem. Considering the quadratic fuel cost
of methods do not have memory elements associated with each function as an objective, Table I shows the HPSO outcomes as
candidate in the population. However, in the case of HPSO, a result of varying its parameters. To quantify the results, 50 in-
each particle has a memory element (pbest) that recalls the best dependent runs were executed for each parameter variation. The
visited location through its own flying experience to search most noticeable observation from this groundwork is that the op-
for the optimal solution and may use this information once timal settings for and are found to be 1.0. These values are
it violates the problem boundaries. Thus, this hybridization relatively small since most of the values reported in the previ-
makes use of the memory element that each particle has to ously related work are in the range of 1.4–2 [12], [22]–[25]. The
maintain its feasibility status. This restoration operation keeps best settings for number of particles and particle’s maximum ve-
the infeasible particle alive as a possible candidate that could locity are 20 and 0.1 respectively.
locate the optimal solution instead of a complete rejection that In our implementation, the is the upper limit used to
eliminates its potential in the swarm. randomly initialize and enforce the velocity vector throughout
the optimization process. A fixed value is used for all
B. Control Variables Treatment problem variables. This means that the velocity vector of all
The problem variables considered in our formulation are optimization problems is randomly initialized with a range of
of two types: continuous and discrete as shown in (4) that . It appears from the results shown in Table I
ALRASHIDI AND EL-HAWARY: HYBRID PARTICLE SWARM OPTIMIZATION APPROACH FOR SOLVING THE DISCRETE OPF PROBLEM 2035

TABLE II
COMPARISON BETWEEN HPSO AND SQP FOR CASE 1

the velocity vector update equation. This parameter directly im-

pacts the velocity vector with larger values in the early stage of
the search process and lower values toward the end of the itera-
tions. In the presence of the inertia weight parameter, properly
setting of would take the least priority when compared to
setting the acceleration constants or swarm’s size. Note that in-
creasing the number of particles beyond than 20 will improve
the solution accuracy slightly at the expense of increasing the
computation time significantly.
Fig. 3. Single line diagram of IEEE 30-bus standard test system. The inertia weight is kept fixed throughout the simulation
process between the upper and lower bounds of 0.9 and 0.4 re-
spectively since changing its values did not have a great impact
TABLE I in improving the convergence characteristics. The same param-
STUDY OF TUNING HPSO PARAMETERS
eters were suitable for cases 1 and 2 as follows..
Once the HPSO best parameters were set, the following cases
were considered to test the proposed approach.
Case 1: The quadratic emission and fuel cost functions in (5)
and (6) were minimized considering only the continuous control
variables, i.e., real power outputs and voltages at voltage-con-
trolled buses. A comparison of results obtained using the HPSO
to those obtained using MATPOWER, MATLAB-based soft-
ware that uses sequential quadratic programming (SQP) to solve
the OPF, are shown in Table II. MATPOWER is capable of
solving the OPF when the objective is represented in polynomial
form and is only capable of handling continuous variables. SQP
took 28 evaluations of fuel cost and emission functions to con-
verge to its answer while in HPSO each particle is evaluated 30
times. Results clearly indicate that HPSO achieved better solu-
tion in both cases when only continuous optimization variables
are used.
Case 2: The test system is modified by introducing four
tap-changing transformers between buses 6–9, 6–10, 4–12, and
27–28. The operating range of all transformers is set between
that HPSO performed better with lower values of . In addi- 0.9–1.05 with a discrete step size of 0.01. The capacitor banks
tion, our proposed algorithm incorporates an inertia weight with at buses 5 and 24 are also considered as new discrete control
decreasing function as shown in (17) and (19) to provide better variables with a range of 0–40 Mvar and a step size of 1. With
control on balancing exploration and exploitation by controlling this modification, the problem now has both continuous and
2036 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 22, NO. 4, NOVEMBER 2007

TABLE III
RESULTS OF DIFFERENT OBJECTIVE MINIMIZATION WHEN BOTH DISCRETE
AND CONTINUOUS VARIABLES ARE CONSIDERED (CASE 2)

Fig. 4. High degree of convexity in the shape of the objective once valve
loading effects are included.

TABLE IV
RESULTS OF CASE 2 UNDER DIFFERENT SWARM’S SIZE

discrete control variables that can be troublesome to most

conventional optimization methods. In addition to the objec-
tives considered in case 1, the total real power losses are also
introduced as new objective in this case. Table III summarizes
the results of each minimization problem along with the best
solution vector achieved. The inclusion of discrete variables in
this case further improved the results of minimizing the fuel
cost and emission functions when compared to case 1.
Case 3: Since HPSO is capable of handling optimization
problems in which the objective is not required to be convex or
differentiable, the fuel cost function is augmented with an addi-
tional sine term as in (7). This addition increases the degree of
smoothness in the shape of the objective function significantly.
Note that the fuel cost coefficients are modified to create more To demonstrate the consistency and robustness of the devel-
challenging objective function shape within the permissible oped algorithm, 20 independent runs were conducted for each
operating range. The two coefficients and are made such case to measure the frequency of reaching the optimal or near
that the generating units experience at least four ripples to give optimal solution while maintaining the same stopping criteria
more realistic representation of real-world power unit response (maximum iterations of 30). Results and computation time are
[28]. The number of ripples is proportional to the degree of shown in Table V that reflect the better performance of HPSO
convexity in the shape of the objective function and it increases in solving the OPF problem.
the difficulty of detecting the global solution. Fig. 4 shows the It is evident that in case 1, even the HPSO worse performance
shape of the fuel cost function of two generation units, which outperformed SQP in both fuel cost and emission minimization.
represents part of the overall problem, when the valve point However, it is noted here that in case 3, there was a noticeable
effects are accounted for and in the absence of any constraints. deviation between best and worst run when incorporating the
Note that even when considering only two units, the shape of valve loading effects into the fuel cost function. This is due to the
the objective is highly convex with multiple peaks and differen- highly smooth feasible region as a result of adding sine terms to
tiable valleys. In this case, more particles are needed to explore the quadratic functions. Similar deviations were noted in earlier
this complex solution hyperspace efficiently. Table IV tabulates work conducted in [29] and [30] when considering the valve
the results obtained using different swarm’s size. Increasing the loading effects.
swarm’s size improved the HPSO performance in achieving Case 4: For a better understanding of the effectiveness of the
better results at the expense of computational time. proposed hybrid inequality constraints treatment in enhancing
ALRASHIDI AND EL-HAWARY: HYBRID PARTICLE SWARM OPTIMIZATION APPROACH FOR SOLVING THE DISCRETE OPF PROBLEM 2037

TABLE V TABLE VIII

STATISTICAL DATA FOR CASES 1–3 BUS DATA FOR IEEE 30-BUS SYSTEM

TABLE VI
STATISTICAL DATA FOR CASE 4

TABLE IX
BRANCH DATA FOR IEEE 30-BUS SYSTEM

TABLE VII
CHARACTERISTICS OF THE GENERATION UNITS

the flying experience of the swarm in the search for optimal so-
lution, the algorithm was modified such that the SLA approach
was incorporated to deal with the inequality constraints. Cases
2 and 3 were repeated with the same HPSO parameters pre-
viously used. The performance results of this version of PSO
are tabulated in Table VI. This type of inequality constraints
handling strategy degraded the PSO performance in solving the
OPF in both cases and for all the objectives considered when
compared to results shown in Table V. As anticipated, multiple objectives. Second, the equality constraints, i.e., the power
optimization variables (in some cases up to eight variables) were equations mismatch is minimized using the Newton–Raphson
set to their limits. This signifies the importance of utilizing the method in a separate subroutine. Third, the inequality con-
memory element of each particle to recall its pbest position once straints are handled by the proposed hybrid constraints handling
it flies outside the feasible solution hyperspace. mechanism to preserve feasibility of solutions. Results are
compared to the outcomes of other optimization techniques,
V. CONCLUSION namely SQP and modified PSO with conventional inequality
The work presented in this paper investigates the applica- constraints handling strategy. Comparison shows that HPSO
bility of HPSO in solving the OPF problem under different consistently outperforms both techniques once its parameters
formulations and considering different objectives. The HPSO are properly tuned. A study of HPSO parameters tuning is
algorithm treats the OPF problem in three major steps. First, the presented to enhance the optimizer performance in finding
optimal combinations of the mixed-type control variables are the minimum of several objectives. The PSO is combined
handled by the discrete HPSO to optimize convex and convex with Newton–Raphson algorithm to form a hybrid optimizer
2038 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 22, NO. 4, NOVEMBER 2007

that can solve the discrete OPF problem with the inclusion of [13] D. C. Walters and G. B. Sheble, “Genetic algorithm solution of eco-
valve loading effects. This emphasizes the HPSO capability of nomic dispatch with valve point loading,” IEEE Trans. Power Syst.,
vol. 8, no. 3, pp. 1325–1332, Aug. 1993.
handling optimization problems with more complex modeling [14] R. Eberhart and J. Kennedy, “A new optimizer using particle swarm
of system objectives and/or constraints. In contrast, most gra- theory,” in Proc. 6th Int. Symp. Micro Machine and Human Science,
Nagoya, Japan, 1995, pp. 39–43.
dient-based optimization methods add more restrictions on the [15] J. Kennedy and R. Eberhart, “Particle swarm optimization,” in Proc.
nature of the problem formulation like continuity, convexity, IEEE Int. Conf. Neural Networks, Perth, Australia, 1995, vol. 4, pp.
and differentiability to ensure finding the optimal solution. 1942–1948.
[16] X. Hu, Y. Shi, and R. Eberhart, “Recent advances in particle swarm,”
Sometimes, such restrictions may lead to intolerable modeling in Proc. Congr. Evolutionary Computation, Portland, OR, 2004, vol.
simplifications that can produce gross error in computing the 1, pp. 90–97.
[17] R. C. Eberhart and Y. Shi, “Guest editorial special issue on particle
minimum of the objective at hand. swarm optimization,” IEEE Trans. Evol. Comput., vol. 8, no. 3, pp.
HPSO performance and robustness in its search for optimal 201–203, Mar. 2004.
[18] J. Kennedy and R. C. Eberhart, Swarm Intelligence. San Francisco,
solution is highly dependant on its parameters tuning and the CA: Morgan Kaufmann, 2001.
shape of the objective function. Objective functions with smooth [19] Y. Shi and R. Eberhart, “A modified particle swarm optimizer,” in Proc.
shapes tend to require less number of particles and iterations IEEE World Congr. Computational Intelligence, 1998, pp. 69–73.
[20] A. J. Korsak, “On the question of uniqueness of stable load-flow
to converge to the optimal solution while the ones with rough solutions,” IEEE Trans. Power App. Syst., vol. PAS-91, no. 3, pp.
surface would require more number of particles and iterations 1093–1100, May 1972.
[21] G. Coath and S. K. Halgamuge, “A comparison of constraint-han-
to reach the same quality of solution. dling methods for the application of particle swarm optimization
to constrained nonlinear optimization problems,” in Proc. Congr.
Evolutionary Computation, Canberra, Australia, 2003, vol. 4, pp.
APPENDIX 2419–2425.
[22] J. G. Vlachogiannis and K. Y. Lee, “A comparative study on particle
Tables VII–IX show the standard IEEE 30-bus test system swarm optimization for optimal steady-state performance of power sys-
tems,” IEEE Trans. Power Syst., vol. 21, no. 4, pp. 1718–1728, Nov.
data used to validate the proposed approach. 2006.
[23] J. B. Park, K. S. Lee, J. R. Shin, and K. Y. Lee, “A particle swarm
optimization for economic dispatch with nonsmooth cost functions,”
ACKNOWLEDGMENT IEEE Trans. Power Syst., vol. 20, no. 1, pp. 34–42, Feb. 2005.
[24] Z. L. Gaing, “Constrained optimal power flow by mixed-integer par-
The authors would like to thank the anonymous reviewers for ticle swarm optimization,” in Proc. IEEE Power Eng. Soc. General
Meeting, San Francisco, CA, 2005, pp. 243–250.
their valuable comments. [25] Z. L. Gaing, “Particle swarm optimization to solving the economic dis-
patch considering the generator constraints,” IEEE Trans. Power Syst.,
vol. 18, no. 3, pp. 1187–1195, Aug. 2003.
REFERENCES [26] R. D. Zimmerman, C. E. Murillo-Sanchez, and D. Gan, MATPOWER:
A Matlab Power System Simulation Package 2006 [Online]. Available:
[1] J. Carpentier, “Contribution e létude do Dispatching Economique,” http://www.pserc.cornell.edu/matpower
Bull. Soc. Franc. Elect., pp. 431–447, 1962. [27] R. Yokoyama, S. H. Bae, T. Morita, and H. Sasaki, “Multiobjective op-
[2] J. D. Weber and T. J. Overbye, “An individual welfare maximization timal generation dispatch based on probability security criteria,” IEEE
algorithm for electricity markets,” IEEE Trans. Power Syst., vol. 17, Trans. Power Syst., vol. 3, no. 1, pp. 317–324, Feb. 1988.
no. 3, pp. 590–596, Aug. 2002. [28] “Present Practices in the Economic Operation of Power Systems,”
[3] Y. He, Y. H. Song, and X. F. Wang, “Bidding strategies based on bid IEEE Trans. Power App. Syst., vol. PAS-90, no. 4, pp. 1768–1775,
sensitivities in generation auction markets,” Proc. Inst. Elect. Eng., vol. Nov. 1971.
149, no. 1, pp. 21–26, 2002. [29] T. Kulworawanichpong and S. Sujitjorn, “Optimal power flow using
[4] J. A. Momoh, R. Adapa, and M. E. El-Hawary, “A review of selected tabu search,” IEEE Power Eng. Rev., vol. 22, no. 6, pp. 37–55, Jun.
optimal power flow literature to 1993—I: Nonlinear and quadratic pro- 2002.
gramming approaches,” IEEE Trans. Power Syst., vol. 14, no. 1, pp. [30] L. Coelho and V. C. Mariani, “Combining of chaotic differential evo-
96–104, Feb. 1999. lution and quadratic programming for economic dispatch optimization
[5] J. A. Momoh, M. E. El-Hawary, and R. Adapa, “A review of selected with valve-point effect,” IEEE Trans. Power Syst., vol. 21, no. 2, pp.
optimal power flow literature to 1993—II: Newton, linear program- 989–996, May 2006.
ming and interior point methods,” IEEE Trans. Power Syst., vol. 14,
no. 1, pp. 105–111, Feb. 1999. M. R. AlRashidi (S’00) received the B.S. degree
[6] M. Avriel and B. Golany, Mathematical Programming for Industrial from the University of Portland, Portland, OR, and
Engineers. New York: Marcel Dekker, 1996. the M.S. degree from Youngstown State University,
[7] C. Wang, H. Yuan, Z. Huang, J. Zhang, and C. Sun, “A modified par- Youngstown, OH, both in electrical engineering,
ticle swarm optimization algorithm and its application in optimal power in 1997 and 2000, respectively. He is currently
flow problem,” in Proc. Int. Conf. Machine Learning and Cybernetics, pursuing the Ph.D. degree in electrical engineering
Guangzhou, China, 2005, vol. 5, pp. 2885–2889. at Dalhousie University, Halifax, NS, Canada.
[8] R. Ma, P. Wang, H. Yang, and G. Hu, “Environmental/economic His research interests include power system oper-
transaction planning using multiobjective particle swarm optimization ation, optimization techniques, and artificial intelli-
and non-stationary multi-stage assignment penalty function,” in Proc. gence.
IEEE/PES Transmission and Distribution Conf. Exhibition: Asia and
Pacific, Dalian, China, 2005, pp. 1–6.
[9] B. Zhao, C. X. Guo, and Y. J. Cao, “Improved particle swarm optimiza- M. E. El-Hawary (F’90) received the B.S. degree
tion algorithm for OPF problems,” in Proc. IEEE/PES Power Systems in electrical engineering with distinction from the
Conf. Exposition, New York, 2004, pp. 233–238.
University of Alexandria, Egypt, in 1965, and
[10] P. N. Biskas, N. P. Ziogos, A. Tellidou, C. E. Zoumas, A. G. Bakirtzis,
and V. Petridis, “Comparison of two metaheuristics with mathematical the Ph.D. degree from the University of Alberta,
programming methods for the solution of OPF,” Proc. Inst. Elect. Eng., Edmonton, AB, Canada, in 1972, where he was a
vol. 153, no. 1, pp. 16–24, 2006. Killam Memorial Fellow.
[11] S. He, J. Y. Wen, E. Prempain, Q. H. Wu, J. Fitch, and S. Mann, “An He is an Associate Dean of Engineering and
improved particle swarm optimization for optimal power flow,” in Proc. has been a Professor of electrical and computer
Int. Conf. Power System Technology, The Pan Pacific, Singapore, 2004, engineering at Dalhousie University, Halifax, NS,
vol. 2, pp. 1633–1637. Canada, since 1981. His research interests are
[12] M. A. Abido, “Optimal power flow using particle swarm optimization,” in power system economics and computational
Int. J. Elect. Power Energy Syst., vol. 24, no. 7, pp. 563–571, 2002. methods in power systems engineering.