42

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 22, NO. 1, FEBRUARY 2007

A New Particle Swarm Optimization Solution

to Nonconvex Economic Dispatch Problems
A. Immanuel Selvakumar, Member, IEEE, and K. Thanushkodi

AbstractThis paper proposes a new version of the classical particle swarm optimization (PSO), namely, new PSO (NPSO), to solve
nonconvex economic dispatch problems. In the classical PSO, the
movement of a particle is governed by three behaviors, namely, inertial, cognitive, and social. The cognitive behavior helps the particle to remember its previously visited best position. This paper
proposes a split-up in the cognitive behavior. That is, the particle is
made to remember its worst position also. This modification helps
to explore the search space very effectively. In order to well exploit
the promising solution region, a simple local random search (LRS)
procedure is integrated with NPSO. The resultant NPSO-LRS algorithm is very effective in solving the nonconvex economic dispatch problems. To validate the proposed NPSO-LRS method, it
is applied to three test systems having nonconvex solution spaces,
and better results are obtained when compared with previous approaches.
Index TermsEconomic dispatch (ED), local search, nonconvex
solution space, particle swarm optimization (PSO).

I. INTRODUCTION

CONOMIC dispatch (ED) is one of the important optimization problems in power systems that has the objective of dividing the power demand among the online generators
economically while satisfying various constraints [1]. Since the
cost of the power generation is exorbitant, an optimum dispatch
saves a considerable amount of money. Traditional algorithms
like lambda iteration, base point participation factor, gradient
method, and Newton method can solve the ED problems effectively if and only if the fuel-cost curves of the generating units
are piece-wise linear and monotonically increasing [2].
The basic ED considers the power balance constraint apart
from the generating capacity limits. However, a practical ED
must take ramp rate limits, prohibited operating zones, valvepoint effects, and multifuel options into consideration to provide the completeness for the ED formulation. The resulting ED
is a nonconvex optimization problem, which is a challenging
one and cannot be solved by the traditional methods. Dynamic
programming (DP) [3] can solve such type of problems, but it
suffers from the curse of dimensionality.
This paper considers three types of nonconvex ED problems,
namely, ED with prohibited operating zones (EDPO), ED with
Manuscript received February 28, 2006; revised August 30, 2006. Paper no.
TPWRS-00115-2006.
A. Immanuel Selvakumar is with the Department of Electrical Sciences,
Karunya Deemed University, Coimbatore 641 114, Tamil Nadu, India (e-mail:
[email protected]).
K. Thanushkodi is with the Department of Electrical Engineering, Government College of Technology, Coimbatore 641 114, Tamil Nadu, India (e-mail:
[email protected]).
Digital Object Identifier 10.1109/TPWRS.2006.889132

valve-point loading effects (EDVL), and ED with combined

valve-point loading effects and multifuel options (EDVLMF).
A considerable amount of work has been contributed to solve
the EDPO problem. The DP approach [3], decomposition technique [4], advantageous decision spaces approach [5], genetic
algorithm (GA) [6], deterministic crowding genetic algorithm
(DCGA)[7], artificial intelligence (AI) technique[8], and PSO
approach [9] are the important contributions to the solution of
EDPO.
The GA[10], GA combined with simulated annealing (SA)
[11], evolutionary programming (EP)[12], improved Tabu
search (ITS)[13], improved fast EP (IFEP) [14], modified PSO
(MPSO) with a dynamic search-space reduction strategy [15],
and evolutionary strategy optimization (ESO) [16] are the
modern heuristic techniques that have been used to solve the
EDVL.
The solution methodology for ED with multifuel options
(EDMF) has been developed by hierarchical method [17],
neural networks [18], [19], and EP [20]. Recently, both EDVL
and EDMF are combined and solved by an improved GA with
multiplier updating (IGA_MU) [21].
This paper introduces a new PSO (NPSO) and its solution
to the above-mentioned nonconvex ED problems. PSO is one
of the modern heuristic algorithms, which can be used to solve
nonlinear and noncontinuous optimization problems [22]. It has
been used for many power system problems such as optimal
design of power system stabilizers [23], distribution state estimation [24], and optimal reactive power dispatch [25], [26]
apart from ED. After the introduction of PSO, many variations
have been proposed for the basic PSO by various researchers
[27][32].
In the classical PSO, three aspects, namely, inertial, cognitive,
and social, govern the movement of a particle. The cognitive
behavior helps the particle to remember its previously visited
best position. This paper proposes a split-up in the cognitive
behavior. That is, the particle is made to remember its worst
position also. This modification helps in exploring the search
space very effectively to identify the promising solution region.
Moreover, to exploit the promising region well, a simple local
random search (LRS) procedure, which is a modification of a direct search procedure [33], is integrated with NPSO. The resultant NPSO-LRS algorithm is very effective in solving the nonconvex ED problems.
To validate the proposed NPSO-LRS method, it is tested on
three test systems having nonconvex solution spaces. The results of the proposed NPSO-LRS and those of the previous approaches are compared. The outcome of the comparisons shows
the effectiveness of the proposed NPSO-LRS method in terms

0885-8950/$25.00 2007 IEEE

SELVAKUMAR AND THANUSHKODI: NEW PARTICLE SWARM OPTIMIZATION SOLUTION

of solution quality and consistency. For all the test systems considered, NPSO-LRS achieves better results compared to the existing results.
II. NONCONVEX ECONOMIC DISPATCH PROBLEMS
The basic ED becomes a nonconvex optimization problem if
the practical operating conditions are included. In this paper,
three different formulations of the ED problems, which reflect
the real-time operating conditions, are used.
A. EDPO

43

B. EDVL
The valve opening process of multivalve steam turbines produces a ripple-like effect in the heat rate curve of the generators,
and it is taken into consideration in the ED problem by superimposing the basic quadratic fuel-cost characteristics with the
rectified sinusoidal component as follows:

(6)
, and are the fuel-cost coefficients of genwhere
erator . The objective of EDVL is to minimize
with the
constraints (2)(4).

The objective is
(1)
is the total generation cost ($/hr),
is the fuel-cost
where
is the number of generators,
function of generator ($/hr),
is the real power output of generator (MW), and
,
are the fuel-cost coefficients of generator . The basic
and
constraints are the real power balance and the real power operating limits

C. EDVLMF
For a power plant with
generators and
fuel options for
each unit, the cost function of the generator with valve-point
loading is expressed as

if
(2)
(3)
where
is the total load in the system (MW), and
is
the network loss (MW) that can be calculated by matrix loss
and
are the minimum and maximum
formula.
power generation limits of generator . The other important constraints are as follows.
Generator Ramp Rate Limits: If the generator ramp rate
limits are considered, the effective real power operating limits
are modified as follows:

fuel option

(7)

and
are the minimum and maximum power
where
generation limits of generator with fuel option , respectively;
, and
are the fuel-cost coefficients of genand
suberator for fuel option . The objective is to minimize
ject to the constraints (2)(4).
III. PROPOSED NEW PARTICLE SWARM OPTIMIZATION (NPSO)

(5)

PSO is a population-based, self-adaptive, stochastic optimization technique [22]. The basic idea of the PSO is the
mathematical modeling and simulation of the food searching
activities of a swarm of birds (particles). In the multidimensional space where the optimal solution is sought, each particle
in the swarm is moved toward the optimal point by adding a
velocity with its position. The velocity of a particle is influenced
by three components, namely, inertial, cognitive, and social.
The inertial component simulates the inertial behavior of the
bird to fly in the previous direction. The cognitive component
models the memory of the bird about its previous best position,
and the social component models the memory of the bird
about the best position among the particles (interaction inside
the swarm). The particles move around the multidimensional
search space until they find the food (optimal solution). Based
on the above discussion, the mathematical model for PSO is as
follows.
Velocity update equation is given by

where
and
are the lower and upper boundaries
of prohibited operating zone of generator in (MW), respecis the number of prohibited operating zones of gentively;
erator ; and
is the number of generators with prohibited
operating zones.

(8)

(4)
where
is the previous operating point of generator
,
are the down and up ramp limits of the generator .
and
Prohibited Operating Zones: A generator with prohibited regions (zones) has discontinuous fuel-cost characteristics. The
concept of prohibited operating zones is included as the following constraint in the ED:

44

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 22, NO. 1, FEBRUARY 2007

Position update equation is given by

(9)

The positions are updated using (9). The inclusion of the

worst experience component in the behavior of the particle gives
additional exploration capacity to the swarm. By using the bad
experience component, the bird (particle) can bypass its previous worst position and always try to occupy a better position.

where
IV. LOCAL RANDOM SEARCH (LRS)

iteration count;
dimension of the velocity of particle at iteration
;
dimension of the position of particle
iteration ;

at

inertia weight;
acceleration coefficients;
dimension of the own best position of particle
until iteration ;
dimension of the best particle in the swarm at
iteration ;
dimension of the optimization problem (number
of decision variables);
number of particles in the swarm;
two separately generated uniformly distributed
random numbers in the range [0, 1].

The metaheuristic algorithms like GA, EP, SA, and PSO are
performing well for small dimensional and less complicated
problems. However, they fail to locate global minima for the
complex multiminima functions. Although they locate the
promising area, they fail to exploit the promising area to get
good quality solutions [6], [9], [11], [14], [15]. With a single algorithm, it is difficult to control and to strike a balance between
exploration of whole search space to locate the promising area
and exploitation of the promising area to get global minima.
Several hybrid methods have been proposed by combining the
metaheuristics methods with simple local search algorithms.
This paper uses a simple LRS procedure, which is a modification of a direct search technique proposed in [33]. The LRS
procedure is outlined below. The initial search point is taken as
, and the objective function value at
is
.
Step 1) The initial local search range is selected around
as follows:
(11)

This paper proposes a new variation in the classical PSO by

splitting the cognitive component of the classical PSO into two
different components. The first component can be called good
experience component. That is, the bird has a memory about its
previously visited best position. This component is exactly the
same as the cognitive component of the basic PSO. The second
component is given the name bad experience component. The
bad experience component helps the particle to remember its
previously visited worst position. To calculate the new velocity,
the bad experience of the particle is also taken into consideration. This gives the new model of the PSO as below.
The new velocity update equation is given by

(12)
(13)
and
are the lower and upper
where
boundaries of the local search region; is the local
and
are the vectors
area parameter;
is the initial local
of power generation limits; and
search range. The iteration count is set to 1.
(best search point at the beginning of LRS) and
(optimum search point) are set to
.
local search points are randomly generated
Step 2) The
as follows:

(14)
(10)
where
acceleration coefficient, which accelerates the
particle toward its best position;
acceleration coefficient, which accelerates the
particle away from its worst position;
dimension of the own worst position of particle
until iteration ;
three separately generated uniformly distributed
random numbers in the range [0, 1].

is a random number vector of length

where
, whose elements are randomly generated beand 1. If any local search point violates
tween
the limits, it is forced within the boundaries.
Step 3) For each local search point, the objective function
values are calculated. Then the minimum objective
, and the correfunction among all is taken as
sponding is taken as
. The optimum values
are updated as follows:
If
Otherwise

then

and
and

SELVAKUMAR AND THANUSHKODI: NEW PARTICLE SWARM OPTIMIZATION SOLUTION

Step 4) The search range is reduced as

(15)

45

without ramp rate limits and based on (4) for generators with
ramp rate limits. The velocities of the particles are initialized as
follows:

where is the range reduction parameter.

Step 5) If maximum iteration for local search
is
not reached, the iteration count is incremented by
one and the above procedure is repeated from step
and
are taken as the op2). Otherwise,
timum results found by the LRS algorithm.
V. IMPLEMENTATION OF NPSO-LRS FOR EDPO,
EDVL, AND EDVLMF PROBLEMS
A. Hybridization of NPSO With LRS
In a stochastic optimization algorithm like PSO, the solution
process moves through a random trajectory to locate the optimum point. The trajectory may not be the same for each run.
A robust algorithm should locate the optimum solution irrespective of the starting point. Moreover, the stochastic optimization
algorithms are good enough to locate the promising areas. If a
promising area where the global optimum is residing is identified at the end of the optimization process, the algorithm is
able to exploit the promising area to get global optimum. On the
other hand, if the promising area with global optimum is identified at the earlier stages of the optimization, there is a possibility of missing that area without exploitation. Considering the
above fact, the LRS is employed whenever a promising area is
obtained by NPSO. The resultant NPSO-LRS hybrid algorithm
is robust in finding the global optimum for large dimensional
nonconvex functions.
B. Solution of EDPO, EDVL, and EDVLMF Problems
With NPSO-LRS
The main objective of EDPO, EDVL, and EDVLMF is to
obtain the amount of real power to be generated by each of
the committed generators while achieving minimum generation cost within the constraints. This section provides the solution methodology for the three types of ED problems with
NPSO-LRS.
Representation of the Swarm: Since the decision variables
for the ED problems are real power generations, they are used
to form the swarm. The real power output
of all generators
is represented as the positions of the particles in the swarm.
generators, the particle position is represented
If there are
. Again if there are
particles in
as a vector of length
the swarm, the complete swarm is represented as a matrix as
follows:
(16)
where
is the position vector of the particle . It represents
one of the possible solutions for the optimization problem. The
of
is the th position component of particle ,
element
and it represents the real power generation of generator of the
possible solution .
Initialization of the Swarm: Each element of the swarm matrix is initialized randomly within the effective real power operating limits. The initialization is based on (3) for generators

(17)
where is a small positive number. This velocity initialization
scheme always guarantees to produce new particles satisfying
real power operating limit constraints [15].
Penalized Fuel Cost Function: The nonconvex ED problems
involve many constraints. Out of these constraints, limiting the
movement of the particles imposes the effective real power operating limits. The real power balance and prohibited operating
zone constraints are handled by including penalty terms to the
original objective function as follows:

(18)
is the penalized objective function,
is the
where
is the
penalty factor for real power balance constraint,
penalty factor for prohibited operating zone constraint, and
is an indicator of falling into the prohibited operating zone. The
and
are used to penalize the fuel cost
penalty factors
proportional to the amount of constraint violations. If there are
is set to zero.
no prohibited zones,
Initialization of the Best and Worst Positions: In the strategy
and global best
of PSO, the particles best position
are the key factors. The best position of a
position
, and the
particle is the position, which gives the minimum
is taken as
. In this
best position out of all the
is introduced. At
paper, the particles worst position
and
the beginning of the iteration process, the
for all the particles are taken as the same as the initial positions.
at
is taken as
.
The
Moving the Particles: The particles in the swarm are moved
to new positions with the help of new velocities. The new velocities are calculated using (10) and the position of the particles are
is taken as
. If any
violates
updated using (9) where
the effective real power operating limit constraints, its value is
taken as the limiting value.
Updating the Best and Worst Positions: The particles are
. Then
and
evaluated in the new positions by
of particle are updated as follows:
if
if
if
if

(19)

46

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 22, NO. 1, FEBRUARY 2007

VI. TEST RESULTS AND ANALYSIS

A. Description of the Test Systems
In order to validate the proposed NPSO-LRS, it is tested with
three test systems having nonconvex solution spaces. The first
test system consists of six generators with prohibited operating
zones and has a total load of 1263 MW [9]. All the generators
are having ramp rate limits. The network losses are calculated
matrix loss formula. The best generation cost reported
by
until now is $15 450/h [9]. The previous best generation cost
was $15 459/h [9].
The second test system consists of 40 generators with valvepoint loading effects and has a total load of 10 500 MW [14].
The system has many local minima, and the global minimum
is very difficult to determine. The best generation cost reported
until now is $122 122.16 [16].
The third system consists of ten generators with multifuel options and valve-point effects [21]. The first generator is having
two fuel options, and the remaining generators are having three
fuel options. The best generation cost reported so for is $624.
5178 [21].
B. Determination of Parameters for NPSO-LRS

Fig. 1. Flowchart for NPSO-LRS.

where
is the penalized objective function value of particle
at iteration . The best position out of all the new
is
, and
at
is taken as
.
taken as
is better than
,
Employing LRS Procedure: If
and
for the LRS
the LRS subroutine is invoked. The
and
, respectively. If
obtained
are taken as
and
are refrom LRS is better than
and
, respectively.
placed with
Stopping Criterion: There are different criteria available to
stop a stochastic optimization algorithm. Tolerance, number of
function evaluations, and maximum number of iterations are
some examples. In this paper, in order to compare with previous
results, maximum number of iterations is taken as the stopping
criterion. The overall NPSO-LRS optimization process is shown
in Fig. 1.

To successfully implement the NPSO, the values of the parameters

, and
have to be determined. The inis linearly decreased from 0.9 to 0.4 over the
ertia weight
is taken as 2, since
iterations, and the acceleration coefficient
these settings are suitable for many power system problems [9],
[24]. The number of particles is selected as 20. To find suitable
and
, the following procedure is used.
values for
1) The
, and
are fixed at 0.9, 0.4, and 2, respectively.
is increased from 1.0 to 1.9 in steps 0.05, and
is
2)
decreased from 1.0 to 0.1 in steps 0.05. For each combiand
, 100 independent trials have been
nation of
made with 1000 iterations per trial.
The EDVL formulation of the 40-generator system is used for
testing. The minimum generation cost for this system reported
so far are $122 624.35 [14], $122 252.265 [15], and $122 122.16
[16]. Hence, for each trial, the minimum generation cost is tested
whether it lies in the region below $122 500 or in the range between $122 500 to $123 000. Based on the outcome of the exand
are chosen as 1.6 and
periments given in Table I,
0.4 (case 13), respectively. They give the minimum generation
cost more consistently, and the obtained generation cost is also
less among the remaining cases.
To implement the LRS, the number of iterations for LRS
, the number of local search points
, local area
, and the range reduction parameter
are to be
parameter
determined. Since has been taken as 0.05 for most of the problems [33], it is fixed at 0.05. If the tolerance for range reduction
can be calculated as 90. To deteris taken as 1%, the
and , the following experiment is permine the parameters
formed on the 40-generator system. With the above-determined
is varied from 10 to 20 in steps of 5.
NPSO parameters,
, five different values (0.1 to 0.5 in steps 0.1) are
For each
and combination, 100 independent trials
tested. For each
have been made with 1000 iterations per trial. As before, for

SELVAKUMAR AND THANUSHKODI: NEW PARTICLE SWARM OPTIMIZATION SOLUTION

47

TABLE I
INFLUENCE OF PARAMETERS ON NPSO PERFORMANCE

Fig. 2. Comparative convergence behaviors of the three PSO strategies for sixgenerator system.

each trial,
and
are calculated. Based on the results,
and are chosen as 10 and 0.4, respectively.
C. Testing Strategies
Since the proposed NPSO-LRS is the hybridization of NPSO
and LRS, it is necessary to find the relative strength of each
constituent. So, three different testing strategies are applied on
the EDPO, EDVL, and EDVLMF problems with a swarm of 20
particles.
1) PSO-LRS: The classical PSO with standard parameters
, and
) is in(
tegrated with LRS. This strategy is selected to analyze the
performance of LRS in PSO environment.
2) NPSO: The proposed NPSO is applied without LRS.
3) NPSO-LRS: The proposed NPSO is integrated with LRS.
The coding is written with MATLAB 6.5 programming language and executed in the Pentium IV, 1.5-GHz, 128-MB RAM
processor. In order to find the effectiveness and superiority of
the NPSO-LRS algorithm, the test results are compared with the
results obtained by other algorithms available in the literature.
D. Convergence Test
The convergence test is carried out to determine the quickness
of the three PSO strategies in terms of the number of main PSO
iterations. The three PSO strategies are tested with the first test
system, and the result is shown in Fig. 2.
The NPSO and NPSO-LRS are almost similar in convergence and show their superiority over the PSO-LRS algorithm.
The NPSO algorithm performs well due to the extra diversification provided by the worst experience component. However,
NPSO-LRS is slightly better than NPSO due to the local
searching ability.
The results of convergence test on the 40-generator system
are shown in Fig. 3. For this system also, the NPSO-LRS is
the best performer. It is very fast when compared to the other
two strategies in terms of main PSO iterations. The PSO-LRS is

Fig. 3. Comparative convergence behaviors of the three PSO strategies for

40-generator system.

slow in convergence when compared to NPSO and NPSO-LRS,

which indicates that the combined strength of PSO and LRS is
inferior to those of NPSO and NPSO-LRS.
Recently, the multiple fuel option and valve-point loading formulations of ED have been combined and solved by IGA_MU
[21]. The IGA_MU could achieve quality solutions with considerable amount of speed compared with conventional genetic
algorithm with multiplier updating (CGA_MU) [21]. The same
EDVLMF problem is solved using the proposed PSO, NPSO,
and NPSO-LRS strategies with 20 particles, and the convergence behaviors of the three PSO strategies are shown in Fig. 4.
Here also, NPSO-LRS is faster in convergence in terms of
main PSO iterations. The behavior of NPSO falls between
those of PSO-LRS and NPSO-LRS. For all the three test
cases, NPSO-LRS stands first in the performance ladder.
The proposed worst experience component in the velocity

48

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 22, NO. 1, FEBRUARY 2007

TABLE III
BEST POWER OUTPUT FOR 40-GENERATOR SYSTEM

Fig. 4. Comparative convergence behaviors of the three PSO strategies for tengenerator system.

TABLE II
BEST POWER OUTPUT FOR SIX-GENERATOR SYSTEM

update equation and the local search procedure strengthen the

NPSO-LRS.
E. Comparisons of the Best Solutions
The best power output of the six-generator system obtained
by the three PSO strategies are compared with those of GA
[9] and PSO [9] in Table II. Except GA, all the other algorithms give the same minimum generation cost. However, the
losses obtained by PSO_LRS, NPSO, and NPSO-LRS algorithms are less when compared to the remaining methods. Moreover, NPSO-LRS achieves best generation schedule with minimum network loss in addition to minimum generation cost.
For the 40-generator system, the best power output that
results from the three proposed PSO strategies are listed in
Table III. The number of iteration is taken as 1000 to match
with the previous analysis, and 20 particles are used. The comparison of generation cost obtained by IFEP [14], MPSO [15],

ESO [16], and the three PSO strategies is given in Table IV.
All the 40 generators are having valve-point effects, and the
solution space has multiple minima. The optimal generation
cost is difficult to achieve, and the minimum generation cost
reported so far is $122 122.16 [16]. However, the three PSO
strategies have the ability to obtain lower generation cost when
compared to $122 122.16. Among the three PSO strategies, the
NPSO-LRS algorithm provides minimum generation cost.
The best power output and the fuel options for the ten-generator EDVLMF problem, obtained by different methods, are
given in Table V. The three PSO strategies are able to obtain
better results compared to IGA_MU[21]. However, NPSO-LRS

SELVAKUMAR AND THANUSHKODI: NEW PARTICLE SWARM OPTIMIZATION SOLUTION

TABLE IV
MINIMUM GENERATION COST OBTAINED BY DIFFERENT METHODS

49

TABLE VII
FREQUENCY OF CONVERGENCE FOR 40-GENERATOR SYSTEM

TABLE V
BEST SOLUTION FOR TEN-GENERATOR SYSTEM

TABLE VIII
COMPARISON AMONG DIFFERENT METHODS AFTER
100 TRIALS (40-GENERATOR SYSTEM)

TABLE VI
COMPARISON AMONG DIFFERENT METHODS AFTER
50 TRIALS (SIX-GENERATOR SYSTEM)

proves its superiority among its competitors by providing minimum generation cost.
F. Robustness Test
Owing to the randomness of the heuristic algorithms, their
performance cannot be judged by the result of a single run.
Many trials with different initializations should be made to acquire a useful conclusion about the performance of the algorithm. An algorithm is robust, if it gives consistent result during
all the trials.
The comparison of results after 50 independent trials with the
first test system is shown in Table VI. From the results, the superiority of the PSO_LRS, NPSO, and NPSO-LRS strategies over
GA [9] and PSO [9] can be noticed. Moreover, the maximum
and average values obtained by NPSO-LRS are very close to
the minimum value, which proves that NPSO-LRS is more robust.
For the second test system, 100 independent trials have been
made. In order to compare the results in a statistical manner, the

TABLE IX
FREQUENCY OF CONVERGENCE FOR TEN-GENERATOR SYSTEM

frequencies of attainment of a cost within the specific ranges are

presented in Table VII. The minimum, maximum, and average
costs of 100 independent trails are presented in Table VIII.
Tables VII and VIII reveal the consistency of NPSO-LRS in
achieving minimum generation cost. NPSO also exhibits similar
characteristics, but its average cost is slightly greater than the average cost of NPSO-LRS. Table VII discloses that NPSO-LRS
has the higher probability of attaining quality solution. NPSO
stands next in performance.
Tables IX and X provide the result of robustness test after 100
trials for the EDVLMF problem. Here also, the performances of
the three PSO strategies are superior. They provide good quality
solutions when compared to IGA_MU.
G. Computational Efficiency
In comparison to the basic PSO, the NPSO-LRS has two additional components, i.e., the bad experience component and the
LRS procedure. These extra burdens necessitate the analysis of

50

IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 22, NO. 1, FEBRUARY 2007

TABLE X
COMPARISON AMONG DIFFERENT METHODS AFTER
100 TRIALS (TEN-GENERATOR SYSTEM)

TABLE XI
CPU TIME COMPARISON

the computational efficiency of the NPSO-LRS. The mean CPU

and
time taken to complete the fixed number of iterations
the mean CPU time taken to converge into the lower solution
for 100 trials are shown in Table XI. The
range
has been calculated only for succeeded trials.
For the 40-generator system, the basic PSO takes an average
CPU time of 4.63 s to complete 1000 iterations. The
increases with the addition of worst experience component and
the LRS. The NPSO-LRS takes 12.1 s more than the NPSO to
complete 1000 iterations. Nevertheless, NPSO-LRS takes only
0.41 s more than the NPSO to converge into the lower solution
range ($120 000$122 500). Similarly, the NPSO-LRS takes
only 0.17 s more than the NPSO to converge into the lower
solution range ($623.5$624.5) of the ten-generator EDVLMF
problem. For both the test cases, the basic PSO is not able to
converge into the lower solution range within the specified fixed
iterations, and the PSO-LRS converges into the lower solution
range at the margin of the fixed iterations. It is observed that,
even though the NPSO-LRS contains the burden of LRS, it has
a mean CPU time of converging into the lower solution range
closer to NPSO and has obtained quality solutions
with more robustness than NPSO. If the solution quality and
the robustness of NPSO-LRS are considered, a slight increase
due to the bad experience component and the LRS
in
procedure can be tolerated.
From all the findings, it is concluded that the three PSO strategies perform well for the entire test systems selected. The performance of NPSO is better than PSO-LRS due to the extra
diversification capability provided by the proposed bad experience component. Since NPSO-LRS has the strength of both
NPSO and LRS, it performs well among the three PSO strategies and outperforms the previous achievements. Hence, the
NPSO-LRS is suggested as a powerful optimization tool for
nonconvex ED problems.
VII. CONCLUSION
A NPSO approach is developed and integrated with an LRS
procedure to form a powerful optimization tool called NPSO-

LRS. The suitable parameters for NPSO-LRS are determined

by appropriate experiments. To prove the ability of the proposed
NPSO-LRS in solving nonconvex optimization problems, ED
problems with nonconvex solution spaces are considered and
solved with three different PSO strategies (PSO-LRS, NPSO,
and NPSO-LRS). The three strategies are tested for convergence
and robustness to find the relative strength of NPSO and LRS.
With the aid of comparisons of the results obtained by the three
PSO strategies and the results of earlier methods available in the
literature, it is proved that the proposed NPSO-LRS method is
very effective in giving quality solutions consistently for nonconvex ED problems.
REFERENCES
[1] B. H. Chowdhury and S. Rahman, A review of recent advances in economic dispatch, IEEE Trans. Power Syst., vol. 5, no. 4, pp. 12481259,
Nov. 1990.
[2] A. J. Wood and B. F. Wollenberg, Power Generation, Operation, and
Control, 2nd ed. New York: Wiley, 1996.
[3] Z.-X. Liang and J. D. Glover, A zoom feature for a dynamic programming solution to economic dispatch including transmission losses,
IEEE Trans. Power Syst., vol. 7, no. 2, pp. 544550, May 1992.
[4] F. N. Lee and A. M. Breipohl, Reserve constrained economic dispatch
with prohibited operating zones, IEEE Trans. Power Syst., vol. 8, no.
1, pp. 246254, Feb. 1993.
[5] J. Y. Fan and J. D. McDonald, A practical approach to real time economic dispatch considering units prohibited operating zones, IEEE
Trans. Power Syst., vol. 9, no. 4, pp. 17371743, Nov. 1994.
[6] P.-H. Chen and H.-C. Chang, Large-scale economic dispatch by
genetic algorithm, IEEE Trans. Power Syst., vol. 10, no. 4, pp.
19191926, Nov. 1995.
[7] S. O. Orero and M. R. Irving, Economic dispatch of generators with
prohibited operating zones: A genetic algorithm approach, Proc. Inst.
Elect. Eng., Gen., Transm., Distrib., vol. 143, no. 6, pp. 529534, Nov.
1996.
[8] W.-M. Lin, F.-S. Cheng, and M.-T. Tsay, Nonconvex economic dispatch by integrated artificial intelligence, IEEE Trans. Power Syst.,
vol. 16, no. 2, pp. 307311, May 2001.
[9] Z.-L. Gaing, Particle swarm optimization to solving the economic dispatch considering the generator constraints, IEEE Trans. Power Syst.,
vol. 18, no. 3, pp. 11871195, Aug. 2003.
[10] D. C. Walters and G. B. Sheble, Genetic algorithm solution of economic dispatch with valve point loading, IEEE Trans. Power Syst.,
vol. 8, no. 3, pp. 13251331, Aug. 1993.
[11] K. P. Wong and Y. W. Wong, Genetic and genetic/simulated -annealing approaches to economic dispatch, Proc. Inst. Elect. Eng.,
Gen., Transm., Distrib., vol. 141, no. 5, pp. 507513, Sep. 1994.
[12] H. T. Yang, P. C. Yang, and C. L. Huang, Evolutionary programming
based economic dispatch for units with non-smooth fuel cost functions, IEEE Trans. Power Syst., vol. 11, no. 1, pp. 112118, Feb. 1996.
[13] W.-M. Lin, F.-S. Cheng, and M.-T. Tsay, An improved tabu search for
economic dispatch with multiple minima, IEEE Trans. Power Syst.,
vol. 17, no. 1, pp. 108112, Feb. 2002.
[14] N. Sinha, R. Chakrabarti, and P. K. Chattopadhyay, Evolutionary programming techniques for economic load dispatch, IEEE Trans. Evol.
Comput., vol. 7, no. 1, pp. 8394, Feb. 2003.
[15] J.-B. Park, K.-S. Lee, J.-R. Shin, and K. Y. Lee, A particle swarm
optimization for economic dispatch with nonsmooth cost functions,
IEEE Trans. Power Syst., vol. 20, no. 1, pp. 3442, Feb. 2005.
[16] A. Pereira-Neto, C. Unsihuay, and O. R. Saavedra, Efficient evolutionary strategy optimisation procedure to solve the nonconvex economic dispatch problem with generator constraints, Proc. Inst. Elect.
Eng., Gen., Transm., Distrib., vol. 152, no. 5, pp. 653660, Sep. 2005.
[17] C. E. Lin and G. L. Viviani, Hierarchical economic dispatch for piecewise quadratic cost functions, IEEE Trans. Power App. Syst., vol. PAS103, pp. 11701175, Jun. 1984.
[18] J. H. Park, Y. S. Kim, I. K. Eom, and K. Y. Lee, Economic load dispatch for piecewise quadratic cost function using Hopfield neural network, IEEE Trans. Power Syst., vol. 8, no. 3, pp. 10301038, Aug.
1993.
[19] K. Y. Lee, A. Sode-Yome, and J. H. Park, Adaptive Hopfield neural
network for economic load dispatch, IEEE Trans. Power Syst., vol. 13,
no. 2, pp. 519526, May 1998.

SELVAKUMAR AND THANUSHKODI: NEW PARTICLE SWARM OPTIMIZATION SOLUTION

[20] T. Jayabarathi and G. Sadasivam, Evolutionary programming-based

economic dispatch for units with multiple fuel options, Eur. Trans.
Elect. Power, vol. 10, no. 3, pp. 167170, 2000.
[21] C.-L. Chiang, Improved genetic algorithm for power economic dispatch of units with valve-point effects and multiple fuels, IEEE Trans.
Power Syst., vol. 20, no. 4, pp. 16901699, Nov. 2005.
[22] J. Kennedy and R. Eberhart, Particle swarm optimization, in Proc.
IEEE Int. Conf. Neural Networks, 1995, pp. 19421948.
[23] M. A. Abido, Optimal design of power-system stabilizers using particle swarm optimization, IEEE Trans. Energy Convers., vol. 17, no.
3, pp. 406413, Sep. 2002.
[24] S. Naka, T. Genji, T. Yura, and Y. Fukuyama, A hybrid particle
swarm optimization for distribution state estimation, IEEE Trans.
Power Syst., vol. 18, no. 1, pp. 6068, Feb. 2003.
[25] A. A. A. Esmin, G. Lambert-Torres, and A. C. Zambroni de Souza, A
hybrid particle swarm optimization applied to loss power minimization, IEEE Trans. Power Syst., vol. 20, no. 2, pp. 859866, May 2005.
[26] B. Zhao, C. X. Guo, and Y. J. Cao, A multiagent-based particle
swarm optimization approach for optimal reactive power dispatch,
IEEE Trans. Power Syst., vol. 20, no. 2, pp. 10701078, May 2005.
[27] Y. Shi and R. C. Eberhart, A modified particle swarm optimizer, in
Proc. IEEE Int. Conf. Evolutionary Computation, 1998, pp. 6973.
[28] Y. Shi and R. C. Eberhart, Empirical study of particle swarm optimization, in Proc. IEEE Int. Congr. Evolutionary Computation, 1999,
vol. 3, pp. 101106.
[29] R. C. Eberhart and Y. Shi, Tracking and optimizing dynamic systems
with particle swarms, in Proc. IEEE Congr. Evolutionary Computation, 2001, pp. 9497.
[30] R. Mendes, J. Kennedy, and J. Neves, The fully informed particle
swarm: Simpler, maybe better, IEEE Trans. Evol. Comput., vol. 8, no.
3, pp. 204210, Jun. 2004.
[31] F. van den Bergh and A. P. Engelbrecht, A cooperative approach to
particle swarm optimization, IEEE Trans. Evol. Comput., vol. 8, no.
3, pp. 225239, Jun. 2004.

51

[32] A. Ratnaweera, S. K. Halgamuge, and H. C. Watson, Self-organizing

hierarchical particle swarm optimizer with time-varying acceleration
coefficients, IEEE Trans. Evol. Comput., vol. 8, no. 3, pp. 240255,
Jun. 2004.
[33] R. Luss and T. H. I. Jaakola, Optimization by direct search and systematic reduction of the size of the search region, AIChE J., vol. 19,
no. 4, pp. 760766, 1973.

A. Immanuel Selvakumar (M03) was born in

Tamil Nadu, India, on March 19, 1974. He received
the B.E. and M.E. degrees in electrical engineering
form Thiagarajar College of Engineering, Madurai,
India. Currently, he is pursuing the Ph.D. degree at
Anna University (GCT Campus), Coimbatore, India.
His research topics include power system operation and control.

K. Thanushkodi was born in Theni District, Tamil

Nadu, India, in 1948. He received the B.E. degree in
electrical and electronics engineering and the M.Sc.
(Engg) degree from Madras University, Chennai,
India, in 1972 and 1976, respectively, and the Ph.D.
degree in electrical and electronics engineering from
Bharathiar University, Coimbatore, India, in 1991.
He is currently a Professor of electrical Eengineering at Anna University, Coimbatore. His research
interests include computer modeling and simulation,
computer networking, and power systems.