A Modified Particle Swarm Optimizer Applied to the Solution of the Economic

Dispatch Problem
S.F. Mekharner Yasser G. Moustafa Nehad El-Sherif and M. M. Mansour
Electrical Power and Machines Department, Arab Academyfor Science, Technology, Electrical Power and Machines Department,
Ain Shams Universitv, and Maritime Transportation Ain Sham University,
Abbassia, Cairo,EGpt Abbassia, Cairo,Egypt

Abstract- In this paper, a modiJed approach, based on the particle unlike all other EAs it doesn’t rely on the famous Darwinian
swarm optimizer (PSO) is presented and explained in detail. As a natural selection “survival of the fittest”, but mainly depends
case study, this PSO technique is applied to the solution of the on “constructive co-operation” among individuals (agents).
economic dispatch (ED) problem of thermal generating units. A Another important difference between PSO and EAs is the
piecewise quadratic function is used to represent the fuel cost of ability of PSO to keep track of the position, and the change in
each generating unit. The B-coefficient method is used to model the
transmission losses. A modification to the standard PSO algorithm
position (velocity) of each particle (agent), while EAs can
is proposed to allow dealing with the power balance equality only keep information regarding the position of the members
constraint. Unlike the traditional methods, this modification does of the population.
not depend on penalizing the infeasible solutions using a pre-defined This paper presents a modified PSO technique; the
penalty function. The application of the proposed method to several algorithm procedures are explained in detail. To show the
case studies shows that it is applicable to large networks with multi- superiority of this approach, the technique is to be used to
generating units including transmission losses with very promising solve the ED problem.
results.
11. PROBLEM FORMULATION
Index Terms- Economic dispatch, Quadratic cost function,
Evolutionary computation, Swarm intelligence, Particle swarm
optimization. A. The Standard ED Problem
In most practical cases a piecewise quadratic function
I. INTRODUCTION can be used to mathematically describe the operating cost of a
thermal plant in the form:
The main target of electric power utilities is to provide high
quality reliable supply to the customers at the lowest possible F, = a , + b,P, crPI2 +
cost while operating to meet the limits and constraints where
imposed on the generating units, and the network (load p, is the power output of unit i
demand). This formulates the well-known ED problem for a,, b, , c, are the fuel consumption cost coefficients of unit i
finding the optimal combination of the output power of all the
on-line generating units that minimizes the total fuel cost, Using the above piecewise linearization, the ED problem
while satisfying all constraints. Traditionally, the ED problem can be formulated as the minimization of the overall cost of
is solved using the mathematical optimization techniques generation that can be described by the objective function in
such as the lambda-iterative method, gradient methods like the form:
Newton-Raphson, and dynamic programming method [ 1,2].
During the last two decades, the interest in applying
artificial intelligence (AI) in optimization has grown rapidly.
The most AI tool that has been used intensively in the
C F, = C + biPi + ciPI2
U,
r=l r=l
literature is the artificial neural network (ANN), namely the
Hopfield model [3-51. The evolutionary algorithms (EAs) where
such as genetic algorithm [6], and evolutionary programming n is the number of on-line dispatchable units
[7], and integrated AI techniques [8] are also used. Recently,
A new evolutionary computation (EC) technique is proposed subjected to
by Kennedy and Eberhakt [9], which they call particle swarm
optimizer (PSO) that was reported as an optimization (i) Power balance constraint
technique as claimed by its co-founder. PSO is a new AI
technique that can be considered as a member of the wide
category of swam intelligence [IO]. It is used to solve a wide
2 P, = P, + P, (3)
variety of optimization problems such as ANN training [ 11-
i=l

131, and function minimization [14,15]. PSO is recently where

adopted by a lot of researchers due to its superiority to other Po is the total load demand
EAs regarding the memory, and computational time
PL is the transmission losses
requirements as it relies on very simple mathematical
operations, also requires very few lines of computer code to The method adopted in this work for including the effect
of transmission losses is the one developed by Kron and
implement [16]. A common feature between PSO and most
adopted by Kirchmayer, namely the loss coeflcient method
EAs is that it is initialized with a random population, but

0-7803-8575-6/04/$20.0002004 IEEE 725

or the B-coeficient method [1,2]. In the B-coeflcient method the best position visited by all the particles is memorized, i.e.
the total transmission losses are expressed as a quadratic the best position among all Pbesf positions referred to as
function of the generators power outputs in the form: “global best position” (&st). The particles of the swarm are
assumed to travel the problem search space in a discrete
n n rather than continuous time steps. At each time step
(4) (iteration) the velocity of each particle is modified using its
current velocity and its distance from Pbesr and according
to:
Using the matrix notation, the loss formula can be expressed
as: (pbest, - sIk)
Vlk+l = VIk + c , *rand()*
At
PL = PT[B]P +BOP+Boo
(gbest, - s,!)
where
P is a vector of the output powers of dimensions ( n * 1).
+ c2 *rand()* At
(7)
[B] is a square matrix of the same dimensions as P. where
Bo is a vector of the same length as P v,” is the ih velocity component at iteration k
Boois a constant term.
rand() is random number between 0 and 1
A detailed derivation of the loss formula can be found in [2]. s,” is the current position in the i a dimension
c I , c2 are the acceleration coefficients
(ii) Generators operating limits
pbest, is the personal best position in the i h dimension
gbest, is the global best position in the ia dimension
where At is the time step
Pi,,,,” is the minimum power output of unit i
Pimm is the maximum power output of unit I After that each particle is allowed to update its position
using its current velocity to explore the problem search space
The minimum load limitations are generally caused by for a better solution as follows:
the fuel combustion stability and inherent steam generator
design constraints, while the maximum limitations are sk+’= sIk+ vIk+’
* At (8)
imposed by the rated output power. Finally it worth
mentioning that the cost of transmission losses is not included
It is a common practice in PSO literatures to choose a
in the total cost function as the fuel cost of power losses is unity time step (At), accordingly (At) is set to one thought out
already included in the generation cost [7].
this work. Referring to equation (7), the velocity update
equation has three terms; the first term represents the
B. Particle Swarm Optimization particle’s memory of its current velocity (change in position)
PSO is a population-based optimization technique that
in the different dimensions of the search space, the second
was originally inspired by the sociological behavior
term is associated with “cognition” since it only takes into
associated with bird flocking and fish schooling [17,18]. One
account the particle’s own experience, while the third one
of the main advantages of PSO is that it needs no gradient
represents the ‘Lsocialinteraction” between the particles. A
information derived from the objective function or the error
better way to understand the mechanism of the stochastic
function. The aforementioned feature is a common property
search done by PSO is to think of each iteration not a process
of all EAs (including PSO) allowing them to be used on of replacing the previous population with a new one (death
functions where the gradient is either unavailable (due to the
and birth), but rather as a process of adaptation [17].
discontinuity of most of real functions), or computationally
Attempting to increase the rate of convergence of the
expensive to obtain. The main idea of the PSO algorithm is to
standard PSO algorithm to a global optimum the inertia
maintain a population of particles (agents), referred to as
weight is introduced in the velocity update equation [19,20].
“swarm” where each particle represents a potential solution to
The inertia weight is a scaling factor associated with the
the objective function under consideration. Each particle in
velocity during the previous time step. According to this
the swarm can memorize its current position that is
modification proposed by Shi and Eberhart, equation (7) is
determined by evaluation of the objective function, velocity,
modified to:
and the best position visited during its flying tour in the k
problem search space referred to as ‘>personal best position”
@best). Here we mean by the personal best position, the one
Vk+l = wv; + c1* rand()* (pbest, - si )
At
that yields the highest fitness value for that particle. For a
minimization task like our case, the position having a smaller
function value is regarded to as having a higher fitness. Also
+ c 2 * rand () * (gbest , - s: )
(9)
At

726
where In order to give the above process a- stochastic nature, the
w is the inertia weight reference generator is selected randomly every iteration.

The inertia weight governs how much of the previous (ii) ED with transmission losses included
velocity should be retained from the previous time step. In
this work a linearly decreasing inertia weight is used [17]. Using the loss formula given by equation (4), the power
The inertia weight is set to decrease linearly from 0.9 to 0.4 balance equation is reduced to:
during the course of a simulation. This setting allows the PSO
to explore a large area at the start of the simulation (when the
inertia weight is large), and to refine the search later by using
a small inertia weight. In brief the PSO algorithm for
unconstraint optimization can be summarized as follows [21]:
Obviously the problem now is more complicated than before.
1. Create an initial swarm, with a random distribution Assuming that the index of the reference generator is r, then
and random initial velocities.
2. Calculate a velocity vector for each particle, using
the particle’s memory and the knowledge gained by
the swarm. i#r
3. Update the position of each particle, using its
velocity vector and previous position. ifr
4. Go to step 2 and repeat until convergence, or the
stoppage criteria is met.
Solving equation (13) for P,
C. Handling the Equality Constraint of the ED Problem
The main problem that faces any of the EAs including
PSO during the process of solving any constraint
optimization problem is the existence of any equality
constraint. The stochastic nature of the search done by EAs
:. BrrC2+
L#*
C B l r e+Bo,-1)

+ ~ B o i ~ + B= Oo o(14)
leads to the generation of potential random solutions to the If?

optimization problem under consideration that must satisfy

all the constraints. In the case of inequality constraints it is It is clear that equation (14) is a second-order equation in P,
easier to find such random solutions while it is practically provided that the power outputs of the other (n-1) generators
impossible to generate such solutions that submit to a pre- are known, following the same methodology that is used
defined equality constraint. The method used in this work to during neglecting the transmission losses, i.e. PSO calculates
handle the power balance equality constraint of the ED the power outputs of the other (n-1) generators then these
problem is as follows: values are used to find the corresponding power output of the
reference generator that satisfies the power balance equation.
(i) ED with transmission losses neglected Solving (14) for Pr will give two values for the power output,
which means that we have three distinct cases:
In this case equation (3) is reduced to
Casel: both values of Pr lies within the limits of operation of
the generator

1=l In such case, the value chosen for P, is the one that gives
a lower overall cost of generation.
To overcome the previously mentioned problem, a generator
is selected randomly referred to as “reference generator” and Case2: both values of Pr violates the limits of operation of
its power output is calculated using the power balance the generator
constraint, while the power outputs of the other (n-1) In such case, another generator is chosen as the reference
generator can be chosen arbitrarily within limits by the PSO. generator and the calculations are repeated again for the new
Accordingly the power output of the reference generator (P,) generator.
is given by:
-
Prk = P D - c
i# r
Pik Case3: one of the two values of Pr lies within the limits of
operation of the generator
where
is the reference generator power output at @ iteration Such case is the most frequent one. By neglecting the
p,“ violating value, we get the combination of the power outputs
r is the index of the reference generator of all generators at that iteration.

727
 The common drawback of these methods is that each
Initialize the swarm with random positions and velocities
method works well for some problems, but doesn't work for
others. The methods based on penalty functions require many
additional parameters, the majors ones being the penalty
Evaluate fitness (objective function) coefficients themselves. An investigation of the performance
of the PSO for constrained optimization problems using the
.c penalty function can be reviewed in [22]. ,As for the methods
of the first category to maintain the feasibility of solution
I If fitness (x) > fitness (pkrl)then =x
I special operators have to be used such as arithmetic
crossover, non-uniform mutation, geometric mutation, sphere
crossover, etc.

I If fitness (x) > fitness (gbst)then g,, =x

I In this work, a modification to the standard PSO
algorithms in [14,21] is proposed that has the advantage that
no penalty function is used to handle the infeasible solutions.
The proposed technique of handling the constraint
Update the velocity of agents using eq. (9) optimization problems can be summarized as follows:

After position updating of each particle, a position check is

Update the position of agents using eq. (8) carried out to make sure that none of the particles flied out of
the search space bounds or violated the constraints. If a
4? violation is detected, a repair algorithm is used to force the
violated particle to return to the feasible region. The
Perform the position check
following pseudo code concisely summarizes the repair
.c algorithm:
Violation detected z
if violation is detectedfor thefirst time then
vy =c,*runq>*(pbesf-slk)+c**cranq)*
(gbesf-s:)
At At
elseif violation is detectedfor the second time then
(gbest, - sIk)
1 - + v,k+l= c* * rand()*

- Stoppage criteria met

No At
elseif violation is detectedfor the third time then
S:+' = pbest,

*I
endif
Yes
Stopping'Criteria: The iterative process is stopped when:
Output the optimal solution The change in the cost of generation is monitored until it is in
the desired accuracy, or the maximum allowable number of
Figure 1 iterations reached.
Flow chart of the proposed algorithm A flow chart summarizing the proposed algorithm is shown
in Fig. 1.
D. The Proposed PSO Algorithm
The ED problem is a constraint optimization problem. 111. IMPLEMENTATION AND RESULTS
The Lagrange multiplier method is the most technique used
frequently in solving the ED problem in the literature leading The Proposed PSO is implemented with MATLABTMon a
to the formulation of the incremental loss for each bus, from 1.3-GHz Pentium I11 computer with 128 MB RAM. In order
which the bus penalty factor can be calculated [1,2]. to demonstrate the effectiveness of the proposed algorithm,
According to [14], the methods of handling a constraint typical case studies used before by other researchers are
optimization problem can be classified into four categories: chosen so as to compare our results with theirs. The PSO
parameters used during the simulation are summarized in
1. Methods based on preserving feasibility of solution Table I. These parameters are selected from previous work by
2. Methods based on penalty function the authors and no attempt is made to fine-tune the algorithm
3. Methods that make a clear distinction between to the current example problems. The main emphasis of the
feasible and infeasible solutions present work is not to fine-tune the PSO algorithm to a
4. Other hybrid methods. specific problem, but rather to investigate the utility of using
the PSO algorithm in a constraint optimization environment,

728
applied to some real power system problems. To compare our
results with the results reported previously, the case studies -
are arranged in three sets, where each set is taken from a Case Method PI PZ P3 Total Total
certain reference. (MW) (MW) (MW) power cost$/h
- I (loss) I
TABLE I
Case Lambda- I 393.2 I 334.6 I 122.2 I 850 I 8194.36
1 iterative I
Parameter Value
PSOof 1 391.8 I 338.2 I 120 I 850 I 8194.4
[15] I
PSO I 392.83 I 334.88 I 122.29 I 850 I 8194.36
proposed
Case Lambda- 600 187.1 62.2 850 7252.83
2 iterative
PSOof 599.9 187.4 62.7 850 7252.85
[IS]
First set [15]: PSO 600 187.15 62.85 850 7252.83
-proposed I
Case Lambda- I 435.1 I 300 I 130.7 I 865.8 I 8344.36
This set contains 3 case studies; with 3 generators each. 3 iterative (15.8
In the first two cases, transmission losses are neglected, while
in the third, case losses are included. A full data of the 3 case (15.9
studies are found in [15]. The obtained simulation results
compared to the results of [ 151 and the lambda-iterative method - (15.83)
are summarized in Table 11. It is clear that the results of our
modified PSO shows high accuracy as it outperforms the
PSO of [I51 in all cases and the numerical methods in the
first 2 cases, while reaching a very close result in the third Method Newton-Raphson PSO proposed
case, while keeping the power balance constraint in all cases.
Also it is important to point out that it is reported in [151 that PI (MW) 23.5581 23.5367
a population of 20 agents has been used, while in our p2 (MW) 69.5593 69.5817
simulation only 5 particles (agents) are used, as shown in Fig. p3 (MW) 59.0368 59.0295
L. Total power (MW) 152. I542 152.1478
Power loss (MW) 2.1542 2.1478
Total cost ($k) 1596.9643 1596.9 194

TABLE IV
Simulation Results Of The 26-Bus Power System Network
Method I Newton-Rauhson I PSO urouosed I
PSO PSO of [I51
proposed
Method

Figure 2
Number of agents used for solving the ED problem

Second set [2]:

This set contains 2 case studies; the first one is a 5-bus Thjrdsetr5]:
power system network with 3 generators and a load demand
of 150 &, the second one is a 26-bus power system This set contains a 20-generator test system supplying a
network with 6 generating units and a total load of 1263 MW. total load demand of 2500 W .A full data of the system is
In both transmission losses are using the B- provided in [SI. Computational results from both the Hopfield
coefJiccient method. Input data including fuel cost functions ANN of 151 and the proposed pso are given in Table v.
coefficients and the loss expressions of the 2 systems are
given in the appendix. The simulation results of the first case
study obtained compared to the previous results of [2] are
summarized in Table 111, while the results of the second case
study are summarized in Table IV.

729
 TABLE V TABLE VI

Pi2 292.7647 294.435 Total 62456.634 62474.485 B,, = 0.000305 16

Pi3 119.1155 113.3914 Cost

outperforms the PSO of [ 151 in terms of computational speed TABLE VI1

and memory requirements as it required only 25 ‘KOof the The Cost Function Coefficients
agent used in [I51 (5 instead of 20) and still converging to a I Unit I a, I b, I C, I P,,,” I Plrnax I
better result. One of the advantages of the proposed algorithm i (Uh) I
(UMWh) ($/IvlWZh)I 1 (MW) I (MW)
is that no penalty function is used to handle the violating
..L
1 240
_ _ _ ~
7 I 1 0.007 I 100 1 500
solutions making it a problem independent, i.e. it’s applicable 200
~~ I 10 I 0.0095
~ I 50 I 200 I
to any other optimization problem regardless of the shape of 3 220 8.5 0.009 80 300
the feasible region or the constraints (linear or non-linear). 4 200 11 0.009 50 150
Also, for the proposed method, it is not required to compute 5 220 10.5 0.008 50 200
the incremental fuel cost in the solving process. Another 6 190 12 0.0075 50 120
important feature of the proposed technique is its
mathematical simplicity requiring no special operators to deal The B-matrices of the loss formula:
with constraints or any restrictions on the problem as
described before. This makes it easily implemented on a r 0.0017 0.0012 0.0007 -O.OOOI -0.0005 -O.OOO~
microcontroller, which is a feature that is not applicable to
most of the previous methods that based on ANN. This
0.00 12 0.00 14 0.0009 0.000 1 -0.0006 -0.000 1
makes our method more practical and could be used in real 0.0007 0.0009 0.0031 0 -0.001 -0.0006
world. We also believe that there is no great difficulty in B=
adding the adaptive feature to our proposed method. Also -0.0001 0.0001 0 0.0024 -0.0006 -0.0008
including more constraints to the standard ED problem such -0.0005 -0.0006 -0.001 -0.0006 0.0129 -0.0002
as security constraints, economic constraints, reserve
capacity, etc. This allows the proposed method to address the --0.0002 -0.0001 -0.0006 -0.0008 -0.0002 0.015 -

Where,
In this work, a modified PSO is presented as an All the B-matrices are given in per unit on a 100 MVA base.
optimization technique that can be used to solve highly
constrained engineering problems. The ED problem is used to VI. REFERENCES
examine the proposed PSO. The results of the simulation of
the case studies show that the proposed method had [I] .IWoods
. and B.F. Wollenberg, Power Generation, Operation and
performed well in most cases reaching the optimal solution Control, John Wiley & Sons, 1996.
with satisfying accuracy. [2] H. Sadaat, Power System Analysis, WCB/McGraw-Hill, 1999.
[3] T. Denise King, M. E. El-Hawary and Feria1 El-Hawary,
“Optimal Environmental Dispatch of Electric Power System via
V. APPENDIX an Improved Hopfield Neural Network Model”, IEEE Trans. on
PowerSystems, vol. 10, no. 3, pp. 1559-1565,1995.
Data of the 5-bus power system network:

730
C. T. Su and G. J. Chiou, “A Fast Computation Hopfield Method A. I. El-Gallad, M. E. El-Hawary and A. A. Sallam,“ Swarming
to Economic Dispatch of Power Systems”, IEEE Trans. on Power of Intelligent Particles for Solving The Nonlinear Constrained
Systems, vol. 12, no. 4, pp. 1759-1764, November 1997. Optimization Problem”, Engineering Intelligent Systems, vol. 9,
C. T. Su and G. J. Chiou, “New Approach with a Hopfield No. 3, pp. 155-163, September 2001.
Modeling Framework to Economic Dispatch, IEEE Trans. on A. I. El-Gallad, M. E. El-Hawary, A. A. Sallam and A. Kalas,“
Power Systems, vol. 15, no. 2, pp. 541-545, May 2000. Swarm Intelligent for Hybrid Cost Dispatch Problem”, IEEE
Bakirtzis A., Petridis V., Kazarlis S., “Genetic Algorithm Canadian Conf:, pp. 753-757,2001.
Solution to The Economic Dispatch Problem”, IEE Proc. Gener. J. Kennedy and R. Eberhart,” A New Optimizer using Particle
Transm. Disfrib., vol. 141, No4, July1994. Swarm Theory”, Proc. of The Sixth Int. Symposrzim on Micro
L. M. Proenp,, J. Luis Pinto and Manuel A. Matos, “Economic Machine and Human Science (Nagoya, Japan), IEEE Service
Dispatch in Isolated Networks with Renewables using center, Piscataway, NJ, pp. 39-43, 1995.
Evolutionary Programming”, IEEE Power Tech’99 Conference, F. Van den Bergh, “ An Analysis of Particle Swarm Optimizer”,
Budapest, Hungary, 1999. PhD thesis, University of Pretoria, Pretoria, South Africa, 2001.
httu:/~emnos.inescn.Dt/artipos/artI~o42.~df httu.//www cs.uo.ac.zalcs/fvdber~~uiiblications.ph~
Whei-Min Lin, Fu-Sheng Cheng and Ming-Tong Tsay, K. E. Parsopoulos and M.N. Vrahatis, “Recent Approaches to
“Nonconvex Economic Dispatch by Integrated Artificial Global Optimization Problems through Particle Swarm
Intelligence”, IEEE Trans. on Power Sysfems, vol. 16, no. 2, pp. Optimization”, Natural Computing 1: 2002, pp. 235-306. Kluwer
307-3 1 1, May 200 1. Academic Publishers. Printed in the Netherlands.
J. Kennedy and R. Eberhart,” Particle Swarm Optimization”, Y. Shi, R. Eberhart, “A Modified Particle Swarm Optimizer”,
Proc. IEEE Inf. Con$ on Neural Networks (Perth, Australia), Proc. IEEE Int. Con$ on Evolutionary Computation. pp. 69-73,
IEEE Service center, Piscataway, NJ, pp. IV: 1942 - 1948, 1995. 1998. ,
Peter Tarasewich and Patrick R. McMullen, “Swarm Intelligence: Y. Shi and R. Eberhart, ”Parameter Selection in Particle Swarm
Power in Numbers”, Appears in Communications of the ACM, pp. Optimization. In: Port0 VW, Saravanan N, Waagen D and Eiben
62-67, August 2002,45(8). AE (eds) Evolutionary Programming VII, pp. 61 1-616, 1998.
-
http://www.ccs.neu.edu/honie/tarase/TaraseMcMullSwarm.udf &://ww.ener iupui,edd-sh1/PSO/Paper/EP98/psof6/ep98 PSO.
R. Eberhart and Xiaohui Hu, “Human Tremor Analysis Using html
Particle Swarm optimization”, 1999. G. Venter and J. S . Sobieski, “Multidisciplinary Optimization of a
httu://web ics.purdue.edu/-huu/uauers/CEC 1999Huinan.udf Transport Aircraft Wing using Particle Swarm Optimization”,
http~//citeseer.ni.nec.com/eberhart99human html Published by the American Institute of Aeronautics and
A. I. El-Gallad, M. E. El-Hawary, A. A. Sallam and A. Kalas,“ Astronautics, Inc. with permission, 2002.
Swarm Intelligent Trained Neural Network for Power httu://wu~v,vrand.coin/uuh/vsmdo 2002.pdf
Transformer Protection”, IEEE Canadian Con$, pp. 265-269, K. E. Parsopoulos and M.N. Vrahatis, “Particle Swarm
2001. Optimization for Constraints Optimization Problems”, 2002.
F. Van de Bergh and A. P. Engelbrecht, “Training Product Unit httu.//~w.malh.upatras.ar/-kostasu/uapers/eisci.Qdf
Networks using Cooperative Particle Swarm Optimisers”, Proc. htt~://citeseer.ni.nec.com/530812.htinl
of The Int. Joint Conf: on Neural Nehvorks (IJCNN), pp. 126-132,
Washington DC, USA, July 200 1.

73 1