Available online at www.sciencedirect.

com

Electrical Power and Energy Systems 30 (2008) 3138

www.elsevier.com/locate/ijepes

A hybrid genetic algorithm approach based on dierential evolution

for economic dispatch with valve-point eect
Dakuo He *, Fuli Wang, Zhizhong Mao
Key Laboratory of Process Industry Automation, Ministry of Education, Northeastern University, Shenyang 110004, China
Received 22 August 2006; received in revised form 14 June 2007; accepted 17 June 2007

Abstract
An ecient hybrid genetic algorithm (HGA) approach for solving the economic dispatch problem (EDP) with valve-point eect is
presented in this paper. The proposed method combines the GA algorithm with the dierential evolution (DE) and sequential quadratic
programming (SQP) technique to improve the performance of the algorithm. GA is the main optimizer, while the DE and SQP are used
to ne tune in the solution of the GA run. To improve the performance of the SQP, the cost function of EDP is approximated by using a
smooth and dierentiable function based on the maximum entropy principle. An initial population obtained by using uniform design
exerts optimal performance of the proposed hybrid algorithm. The combined algorithm is validated for two test systems consisting of
13 and 40 thermal units whose incremental fuel cost function takes into account the valve-point loading eects. The proposed combined
method outperforms other algorithms reported in literatures (EP, EPSQP, PSO, PSOSQP) for EDP considering valve-point eects.
2007 Elsevier Ltd. All rights reserved.
Keywords: Hybrid genetic algorithm (HGA); Dierential evolution (DE); Maximum entropy principle; Uniform design; Economic dispatch problem

1. Introduction
Scarcity of energy resources, increasing power generation cost and ever-growing demand for electric energy
necessitates optimal economic dispatch in todays power
systems [1]. Economic dispatch is one of the most important problems to be solved in the operation of a power system. The primary objective of the economic dispatch
problem (EDP) of electric power generation is to schedule
the committed generating unit outputs so as to meet the
required load demand at minimum operating cost while
satisfying all unit and system equality and inequality constraints [2].
Improvements in scheduling the unit outputs can lead to
signicant cost savings. The input output characteristics of
large units are highly nonlinear because of valve-point
loadings, generating unit ramp rate limits, etc. Furthermore, they may generate multiple local minimum points
*

Corresponding author.
E-mail address: hedakuo@mail.neu.edu.cn (D. He).

0142-0615/$ - see front matter 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ijepes.2007.06.023

in the cost function. In light of the nonlinear characteristics

of the units, there is a demand for the techniques that do
not have restrictions on the shape of the fuel-cost curves.
To obtain accurate dispatch results, approaches without
restriction on the shape of incremental fuel cost functions
are needed. However, both lambda-iterative and gradient
technique methods in conventional approaches to the
problems are calculus-based techniques and require a
smooth and convex cost function and strict continuity of
the search space. Dynamic programming (DP) [3] imposes
no restrictions on the nature of the cost curves and therefore it can solve EDP with inherently nonlinear and discontinuous cost curves. This method, however, suers from the
curse of dimensionality or local optimality [3]. Recently,
as an alternative to the conventional mathematical
approaches, modern heuristic optimization techniques such
as simulated annealing (SA), genetic algorithm (GA), particle swarm optimization and taboo search, are proved to
be very eective in solving EDP.
GA is a stochastic optimization technique, which is
based on the principle of natural selection and genetics

32

D. He et al. / Electrical Power and Energy Systems 30 (2008) 3138

[4,5]. In recent years, the interest in these algorithms has

been rising fast, for that they provide robust and powerful
adaptive search mechanisms [6]. GA has an immense
potential for applications in the eld of power systems
and it has been successfully applied to solve various problems in electric power systems such as economic dispatch,
unit commitment, reactive power control, hydrothermal
scheduling, and distribution system planning etc. It
searches multiple solutions simultaneously in contrast to
conventional optimal algorithms. Therefore, the possibility
of nding global optimal solution is increased. The main
advantage of GA is that it nds near optimal solution in
relatively short time compared with other random searching methods. Despite the aforementioned success, GA is
only capable of identifying the high performance region
at an aordable time and displays inherent diculties in
performing local search for numerical applications [7].
Dierential evolution (DE) is a stochastic search algorithm that is originally motivated by the mechanisms of
natural selection. Like other evolutionary algorithms, DE
is very eective for solving optimization problems with
nonsmooth objective functions, since it does not require
derivative information. The DE algorithm was successfully
applied in the optimization of some well-known nonlinear,
nondierentiable and nonconvex functions by Storn and
Price [8]. The DE algorithm is an evolutionary algorithm
that uses a rather greedy and less stochastic approach to
problem solving than do classical evolutionary algorithms,
such as genetic algorithms, evolutionary programming and
evolution strategies [9]. The tness of an ospring is one-toone competed with that of the corresponding parent in DE.
This one-to-one competition will have a faster convergence
speed than other EAs. The potentialities of DE are its simple structure, easy use, local searching property and speediness. Nevertheless, this faster convergence yields in a
higher probability of searching toward a local optimum
or getting premature convergence. This drawback could
be overcome by employing a larger population. However,
by doing so, much more computation time is required to
estimate the tness function [10].
To overcome premature convergence and to speed up
the searching process, a hybrid genetic algorithm (HGA)
approach that integrates the GA and DE with a gradient
search algorithm called SQP [11] is proposed to take
advantage of them(GA, DE and the local search techniques). GA is capable of exploring a large space, yet is
slow in computation time and ne tuning local search. In
contrast, DEs important potentialities are its faster convergence speed and local searching property. GA and DE
have supplementary potentialities. SQP techniques can
climb hills rapidly while they are blind to the potential hills
in the neighbourhood area and sensitive to the initial starting points. Based on advantages of DE in computation
time and less stochastic approach and SQP in local searching property, DE and SQP are employed as operators of
GA to improve the search speed and exploitation. The
hybrid approach uses GA as a base level search to identify

the potential optimal regions within a reasonably short period of time, while DE and SQP technique subsequently
search these regions rapidly. Therefore, the proposed algorithms the possibility of nding global optimal point and
the convergence speed are improved.
In general, the hybrid method oers an exact solution
only when the function is smooth and its gradient information is known [7]. The method is proposed to approximate
EDP by using a smooth and dierentiable function based
on the maximum entropy principle in SQP operator. In this
way, the performance of the hybrid method is improved.
At the same time, to improve rationality of the distribution
of initial population, the hybrid technique integrating the
uniform design with the genetic algorithm is proposed.
Two ED problems with 13 and 40 thermal units with a
nonsmooth fuel cost function are employed to prove the
eectiveness of the proposed hybrid algorithm. The results
obtained through the combined approach are compared
with those reported in the literatures [12,13].
2. EDP formulation
The classic EDP minimizes the following incremental
fuel cost function associated to dispatchable units:
(
)
NP
X
min F min
F i P i ;
1
i1

where Fi(Pi) is the fuel cost function of the unit and ith is
the power generated by the ith unit, Pi subject to power
balance constraints:
NP
X

P i P D P loss ;

i1

where PD is the system load demand and Ploss is the transmission loss, and generating capacity constraints:
P imin 6 P i 6 P imax

for i 1; 2; . . . ; N P ;

where Pimin and Pimax are the minimum and maximum

power outputs of the ith unit.
The inclusion of valve-point loading eects makes the
modeling of the incremental fuel cost function of the generators more practical. This increases the nonlinearity as
well as number of local optima in the solution space. Also
the solution procedure can easily trap in the local optima in
the vicinity of optimal value. The incremental fuel cost
function of the generating units with valve-point loadings
is represented as follows [11,14]:
F i P i ai P 2i bi P i ci jei sinfi P imin  P i j;

where, ai, bi and ci, are the fuel-cost coecients of the unit,
ei and fi are the fuel cost-coecients of the unit with valvepoint eects.
The economic dispatch of generation of real power of
the generating units is to be done to the required load
demand by satisfying the above constraints. The incremental fuel cost function can be modeled in a more practical

D. He et al. / Electrical Power and Energy Systems 30 (2008) 3138

fashion by including the valve-point eects [15]. The generating units with multi-valve steam turbines exhibit a greater
variation in the fuel cost functions. The valve-point eects
introduce ripples in the heat-rate curves, thereby the number of local optima is increased. Hence, a technique that
overcomes these complexities has to be evolved.

33

Because of approximation, the result obtained by SQP

method is approximation of exact result. To handle this
problem, the hybrid method proposed in this paper only uses
SQP method to nd approximation solution not optimum
cost. In other word, entropic smoothing approximation
function is only used in SQP method not in whole HGA.
Thus approximation does not inuence the accuracy of HGA.

3. Entropic smoothing approximation function

4. Uniform design
Let Wi : Rn ! R, i = 1, 2, . . ., m, be dierentiable and
dene a max-type function W : Rn ! R, i = 1, 2, . . ., ,m by
Wx max Wi x:

Given any l > 0, consider the following entropy-type

function as a smoothing approximation function of W,
m
X
Wx; l l ln
expWi x=l
6
i1

Note that, for l > 0,

m
X
expWx  Wi x=l:
Wx; l Wx l ln

i1

Moreover,
Wx 6 Wx; l 6 Wx l ln m

8x 2 Rn

and l > 0:
8

Therefore, W(x, l) ! W(x) as l ! 0. This fact allows us

to solve the problem without facing the nondierentiability
problem of W(x). Since the function (6) can be derived from
the dual problem of an entropy optimization problem, we
call function (6) an entropic smoothing approximation
function. Li [16] discovered a few properties of this function and named it as the aggregate function. Related work
can be found in [1720].
To EDP, the incremental fuel cost function of the generating units with valve-point loadings (1) is represented as
the following function:
F i P i ai P 2i bi P i ci jei sinfi P imin  P i j
ai P 2i bi P i ci maxfei sinfi P imin  P i ;
 ei sinfi P imin  P i g

Based on the maximum entropy principle the incremental fuel cost function can be approximated by entropic
smoothing approximation function:
WP i ; l ai P 2i bi P i ci


l ln eei sinfi P imin P i =l eei sinfi P imin P i =l :
10
The function W(Pi, l) provides a good approximation to
the function Fi(Pi) in the sense that
0 6 WP i ; l  F i P i 6 l ln 2 and l > 0:
11
It is clear that this function uniformly approximates the
maximum function Fi(Pi) when a parameter l tends to
innitesimal and is greater than Fi(Pi) with an error bound
less than lln 2.

The distribution of the individuals of initial population

directly inuences the globe convergence and searching eciency. Therefore the reasonable setting of initial population is an important problem in the application of GA to
perform optimization calculation. On the other hand, the
main idea of the proposed approach is GA as main optimizer and DE and SQP as local optimizer. The initial population of DE is generated from the population of GA
randomly. So, the DE operators performance is depend
on distribution of the population of GA in initial phase
and evolutionary phase. The SQP is a gradient-based optimization method starting from a single point and using
gradient information to obtain a solution. If the distribution of the individuals of initial population is not reasonable, the HGA sometimes converges to local optima and
can not reach the global optimal point. The distribution
of the population of GA is important to the approach
for exerting the advantage of SQP. In order to exert optimal performance of the hybrid algorithm, the initial population of GA must reect the information of solution space
scientically. However it is dicult to evaluate the distribution of the individuals of initial population. The uniform
design is an eective method to solve this problem. It is
employed to obtain initial population in this paper.
Because the dimensionality of EDP is large, it is dicult to
obtain optimal generating vector of the usable tables of uniform design. So the following method is used in this paper.
Theorem 1. If (12) and (13) are satised, the optimal
generating vector of n dimensionalities and 12 un 1
factors uniform design is composed of natural numbers
which are prime numbers with n + 1 [21].
hi 6 hj

i 6 j;

hi 6 n 1  hj

12
i 6 j;

13

where u(n + 1) is Euler function of n + 1; hi and hj are different generating elements in generating vector.
Theorem 2. If (12) and (13) are satised, the optimal generating vector of uniform design with n dimensionalities
and 12 un 1  1 factors is composed of natural numbers
which are prime numbers with n + 1 in [1, n + 1/2) with any
one taken out.
Based on Theorem 1 and Theorem 2, for the uniform
designs with 12 un 1 or 12 un 1  1 factors, the
usable tables can be obtained immediately without any

34

D. He et al. / Electrical Power and Energy Systems 30 (2008) 3138

measurement for its uniformity. The good lattice point

method is employed to obtain the usable table with optimal generating vector. For the uniform designs with certain factors, n is calculated by formula of Euler function.

5.4. Sequential quadratic programming operator [11,2426]

5. Hybrid genetic algorithm

5.1. Selection operator
After the evaluation of the initial population, the GA
begins the creation of the new generation. The selection used
in this paper depends on individual tness. The best individuals of the present population are kept for the next population. The tness value in this paper contains the penalty
function and does not represent the true objective function.
5.2. Simplex crossover operator
The primary genetic operator is the crossover operator.
The purpose of crossover operator is to produce new
chromosomes that are distinctly dierent from their
parents, yet remain some of their parents characteristics.
It has important function in the capability of searching
the optimal solution.
Simplex algorithm is a method which nds optima
through improving inferior point by searching from inferior point towards the pivot of n + 1 initial points of the
Simplex for the n-dimensional problem. Simplex algorithm
is similar to crossover operator of GA in utilizing information of multi-points. Based on aforementioned analysis,
simplex crossover operator [22] is employed in this paper
to improve convergence speed and lead the population to
the global optimum because of the characteristic of
improving inferior point.
5.3. Nonuniform mutation operator [23]
Nonuniform mutation operator is employed in this
paper. This operator is described below.
If P t P t1 ; P t2 ; . . . ; P tn  is a power output chromosome
vector and P ti is ith units output that is chosen to be
mutated, the new output P mut
will be after mutation
i
 t
P i Dt; P imax  P ti if r 0;
P mut

14
i
P ti  Dt; P ti  P imin if r 1;
where r is a random bit, and function D(t, y) returns a value
in the range [0, y] such that the probability of the value returned is close to 0 increases with t
b

Dt; y y  1  n1t=gen ;

around the local area of the parameter resembling a hillclimbing operator. For our experiments, b was chosen
equal to 2.

15

SQP method seems to be the best nonlinear programming methods for constrained optimization. It outperforms every other nonlinear programming method in
terms of eciency, accuracy and percentage of successful
solutions over a large number of test problems. The
method resembles closely to Newtons method for constrained optimization just as is done for unconstrained
optimization. At each iteration an approximation is made
by the Hessian of the Lagrangian function using a BroydenFoldfarbShanno (BFGS) quasi-Newton updating
method. This is then used to generate a quadratic programming (QP) sub-problem whose solution is used to
form a search direction for a line search procedure. In this
combined method, SQP is used as a local optimizer to
ne-tune the region explored by GA in its run. The GA
is applied to the optimization problem, and the randomly
chosen solution obtained by GA is applied as the initial
starting point for the SQP and generate SQP ospring
by the nal solution obtained using the SQP.
5.5. Dierential evolution operator
The DE operator used in this paper can be described as
following steps and procedures.
Step 0. Set counter k = 1.
0 0
Step 1. Generate the initial individuals (points) x1 ; x2 ;
0
. . . ; xD randomly from the current GA population.
Step 2. While (stopping criterion is not met, namely
k < SSmax).
Generate D new individuals (points) as follows:
k
Step 2.1. Find the best individuals xb in the current
population;
k
k k
k
For each preset vector xi xi;1 ; xi;2 ; . . . ; xi;n ;
i 1; 2; . . . ; D.
k k
Step 2.2. (First selection). Chose two vectors xr1 ; xr2
randomly from the current population, where r1, r2 2
{1, 2, . . ., D}.
Step 2.3. (Mutation). Generate a vector v = (vi,1, vi,2,
. . ., vi,n) according to




k
k
k
k
k
fm xr1  xr2 :
16
v xi f m xb  xi
Step 2.4. (Crossover). Generate a new vector u = (ui,1,
k

where n is a random oating-point number in the interval

[0, 1]; t is the current generation; gen is the maximum number of generations; and b is a parameter that determines the
degree of dependence on the number of generations. In this
way, the operator makes a uniform search at the beginning
of the evolution and in later stages narrows the search

ui,2, . . ., ui,n) according to the preset vector xi xi;1 ;

k

xi;2 ; . . . ; xi;n and the vector v = (vi,1, vi,2, . . ., vi,n) generated in Step 2.3 as follows:
(
uij

vij if rand0;1 < CR or j rnbri;

xkij if rand0;1 > CR or j 6 rnbri;

j 1;2;...;n;

17

D. He et al. / Electrical Power and Energy Systems 30 (2008) 3138

where rand(0, 1) is a uniform random number in range

(0, 1), and rnbr(i) is a randomly chosen index in the set
{1, 2, . . ., n}, which insures the new vector to get at least
one parameter from the new generated vector v.
k
Step 2.5. (Second selection). If F u < F xi , set
k1
k1
k
xi
u and k = k + 1. Else set xi
xi .
End for
End while
Step 3. Output the DE population.

6. Solution methodology
The proposed hybrid approach for EDP with valvepoint eects can be summarized as follows:
Step 1: Get the data for the system.
Step 2: Initialize parameter of algorithm and count t.
Step 3: Generate initial population by the usable tables
of uniform design.
Step 4: Evaluate the objective function and update count t.
Step 5: Identify the Fitbest(t) of the current run t.
Step 6: If Fitbest(t) < Fitbest(t  1) replace Fitbest(t  1)
with Fitbest(t), otherwise goto Step 7.
Step 7: Generate selection ospring using selection
operation.
Step 8: Generate crossover ospring using simplex crossover operation.
Step 9: Generate mutation ospring using nonuniform
mutation operation.
Step 10: Generate DE ospring using DE operation.
Step 11: Generate SQP ospring using SQP operation.
Step 12: Generate whole ospring with selection ospring, crossover ospring, mutation ospring, DE ospring and SQP ospring.
Step 13: While (termination criterion not met).

35

generating vector of uniform design with 28 dimensionalities and 13 factors are all natural numbers in [1,13]. The
optimal generating vector of uniform design with 82
dimensionalities and 40 factors are all natural numbers in
[1].Then the good lattice point method is employed to
obtain the usable table with optimal generating vector.
Hence the population size was chosen as 28 for test case
1 and 82 for test case 2.
The selection, crossover, DE, SQP and mutation operation proportions were chosen as 3:3:7:1:14 for test case 1
and 6:15:20:1:40 for test case 2. The simulation parameters
of the proposed hybrid approach are xed as follows,
D = 7, SSmax = 30, F = 0.48, CR = 0.88, b = 2, l = 1,
itermax = 30 for test case 1 and D = 20, SSmax = 50,
F = 0.48, CR = 0.88, b = 2, l = 1, itermax = 60 for test case
2. To validate the comparison of results obtained using the
proposed technique with the results obtained using the EP,
EPSQP, PSO and PSOSQP techniques, solution procedure of the HGA technique is terminated when the maximum number of iterations is reached.
The tness function Fit(Pi) is given as

F if
P i is feasible;
FitP i
18
F  F  gmax otherwise;
where F is the value of objective function, is used to keep
penalty, gmax is given as
gmax maxf0; gi P ; jhj P j; i 1; 2; . . . ; m1 ;
j 1; 2; . . . ; m2 g;

19

where gi(P) are the inequality constraints, hi(P) are the

equality constraints, m1 and m2 are the number of the
inequality constraints and the equality constraints,
respectively.
7.1. Case 1

The termination is done when a specied number of iterations met.

7. Simulation results
In this paper, to assess the eciency of the proposed
hybrid approach, two case studies (13 and 40 thermal units
or generators) of ED problems were applied. The software
was written using MATLAB 7.0. Simulations were executed on a Pentium-IV 2.99 GHz personal computer. Hereinafter, in each case study, the results represent the average
of 30 runs of the proposed method.
Since u(28 + 1) = u(29) = 28 = 2(13 + 1) and u(82 +
1) = u(83) = 82 = 2(40 + 1), the optimal generating vector
of uniform design with 28 dimensionalities and 13 factors
and the optimal generating vector of uniform design with
82 dimensionalities and 40 factors are composed of natural
numbers which are prime numbers with 29 in [1, 14.5) with
any one taken out and 83 in [1, 41.5) with any one taken
out, respectively, by Theorem 2. In this paper, the optimal

This case study consisted of 13 thermal units of generation with the eects of valve-point loading, the complexity
and nonlinearity to the solution procedure is increased.
The expected power demands to be met by the all thirteen
generating units are 1800 MW [12] and 2520 MW [27]. The
system data can be found from [12]. The problem has a
number of local optimum points as there are more possibilities for any method to stick on any one of the local optimum points. The problem is solved for two dierent
power demands in order to show the eectiveness of the
proposed method in producing quality solutions.
The nal fuel costs obtained using the EP, EPSQP,
PSO, PSOSQP and the proposed method for power
demand of 1800 MW were summarized in Table 1, which
shows that the minimum cost and the mean cost value
obtained by the proposed method is comparatively less
compared to all the other methods. The best results
obtained for solution vector, with HGA with the minimum
cost of $17 963.83/h is given in Table 2. Table 3 reports the

36

D. He et al. / Electrical Power and Energy Systems 30 (2008) 3138

Table 1
Comparison of fuel costs for PD = 1800 MW for case 1

Table 4
Comparison of fuel costs for case 2

Method

Mean time (s)

Best cost ($/h)

Mean cost ($/h)

Method

Mean time (s)

Best cost ($/h)

Mean cost ($/h)

EP [12]
EPSQP [14]
PSO [14]
PSOSQP [14]
HGA

157.43
121.93
77.37
33.97
8.48

17 994.07
17 991.03
18 030.72
17 969.93
17 963.83

18 127.06
18 106.93
18 205.78
18 029.99
17 988.04

EP [12]
EPSQP [14]
PSO [14]
PSOSQP [14]
HGA

1167.35
997.73
933.39
733.97
84.21

122 624.35
122 323.97
123 930.45
122 094.67
121 418.27

123 382.00
122 379.63
124 154.49
122 245.25
121 784.04

Table 2
Best result obtained for PD = 1800 MW for case 1 using HGA
Power

Generation

Power

Generation

1
2
3
4
5
6
7

628.3185
222.7491
149.5996
109.8665
109.8665
109.8665
109.8665

8
9
10
11
12
13

60.0000
109.8665
40.0000
40.0000
55.0000
55.0000

Table 3
Dispatch results for a PD = 2520 MW for case 1
Generator

Unit generation (MW)

GASA [14]

EPSQP [14]

PSOSQP [14]

628.23
299.22
299.17
159.12
159.95
158.85
157.26
159.93
159.86
110.78
75.00
60.00
92.62

628.3136
299.1715
299.0474
159.6399
159.6560
158.4831
159.6749
159.7265
159.6653
114.0334
75.0000
60.0000
87.5884

628.3205
299.0524
298.9681
159.4680
159.1429
159.2724
159.5371
158.8522
159.7845
110.9618
75.0000
60.0000
91.6401

1
2
3
4
5
6
7
8
9
10
11
12
13
Total cost
($/h)

24 275.71

24 266.44

24 261.05

HGA
628.3184
299.1992
299.1988
159.7330
159.7329
159.7324
159.7330
159.7330
159.7331
77.3994
77.3996
87.6879
92.3992
24 169.92

dispatch results of the various methods [25], EPSQP,

PSOSQP and the proposed method for a load demand
of 2520 MW. It is clear from Table 3, the proposed method
succeeded in nding the best solution for the tested
methods.
7.2. Case 2
This case study involved 40 thermal units with quadratic cost functions together with the eects of valvepoint loading. This is a larger system. The number of
local optima, complexity and nonlinearity to the solution
procedure is enormously increased. The required power
demand to be met by all the forty generating units is
10 500 MW. The system data can be found from [12].
The nal fuel costs obtained using the EP, EPSQP,

Table 5
Best result obtained for case 2 using HGA
Power

Generation

Power

Generation

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20

111.3793
110.9278
97.4104
179.7331
89.2188
140.0000
259.6198
284.6570
284.6588
130.0000
168.8214
168.8496
214.7524
394.2848
304.5361
394.2987
489.2877
489.2869
511.2752
511.2857

21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40

523.2961
523.3202
523.2916
523.3014
523.2675
523.2787
10.0000
10.0000
10.0000
88.6376
190.0000
190.0000
190.0000
164.9795
165.9970
165.0464
110.0000
110.0000
110.0000
511.3005

PSO, PSOSQP and the proposed method were summarized in Table 4. The best result obtained for solution vector, with HGA with the minimum cost of $121 418.27/h is
given in Table 5. It is clear from Table 4, HGA has the
best probability of attaining better solutions, the mean
cost value, and the minimum cost amongst the ve methods in this test case. It is noticeable that the mean cost
value obtained using HGA is less than the minimum cost
obtained using other methods. Hence, for power system
EDP of greater size with more nonlinearities, the proposed method is proved to be the best algorithm amongst
the ve methods.
8. Discussion and conclusion
In the past, to solve economic dispatch problem
eectively, most algorithms require convex cost function
and the incremental cost curves to be of monotonically
smooth increasing nature and strict continuity of the search
space. For the generating units, the incremental fuel cost
curves are highly nonlinear and noncontinuous and
nonmonotonically.
GA is an ecient tool for solving complex optimization
problems. It is applied to solve various problems in various
diverse elds. It is also eectively used in solving complex

D. He et al. / Electrical Power and Energy Systems 30 (2008) 3138

problems in the power system eld such as EDP. GA is faster in nding the high performance region but displays difculties in performing local search for complex functions.
It leads to a poor ne tuning of the nal solution. The
DE is a population-based direct search algorithm. In general, the DE provided fast convergence. However, as it is
a greedy scheme, with an explicit bias towards the best ospring, DE sometimes runs into the problem of losing the
diversity of the population, which, in turn, leads to a premature convergence. It shows that the DEs main advantages are its simple structure, speed and local searching
property. The SQP as a best nonlinear programming
method can explore the search space quickly with a gradient direction but this local gradient search technique is sensitive to the initial starting points and always gets stuck at a
local optimum solution. The GA, DE and SQP have
advantages that complement one another.
To overcome these drawbacks and exert these advantages, this paper proposed new hybrid approach combining GA, DE and SQP to solve the EDP of electric energy
with the valve-point eect. Basically, the proposed algorithm consists of two parts. The rst part employs the
property of GA, which can provide a near global search
region. The second part employs DE to explore the
search region and SQP to ne tune solutions of the
GA run in order to obtain a local optimum solution.
The cost function of EDP is approximated by using a
smooth and dierentiable function based on the maximum entropy principle. This is advantaged to improve
the performance of the SQP. An initial population
obtained by using uniform design reects the information
of solution space scientically and exerts optimal performance of the proposed hybrid algorithm.
The eectiveness of the HGA is illustrated for two test
cases of EDP with valve-point eects and compared with
the results obtained using the EP, EPSQP, PSO and
PSOSQP methods. The numerical results reveal the superiority and feasibility of the proposed hybrid approach for
solving the EDP of units with value-point eects.
Acknowledgement
This work is supported by the National Nature Science
Foundation of China under Grant 60374063.
Appendix A
List of symbols
F
Fi(Pi)
Pi
NP
PD
Ploss
Pimin/Pimax

total production cost ($)

incremental fuel cost function ($/h)
real power output of the ith unit (MW)
number of generating units
the system load demand (MW)
transmission loss (MW)
minimum/maximum limit of the real power
of the ith unit (MW)

u(n + 1)
P ti
P mut
i
r
n
t
gen
b
k
SSmax
D
k
xb
k
xi
fm
CR
v
u
rand(0,1)
rnbr(i)

37

Euler function of n + 1
i th units output that is chosen to be
mutated (MW)
ith units new output after mutation (MW)
a random bit
a random oating-point number in the
interval [0, 1]
the current generations
the maximum number of generations of GA
a parameter that determines the degree of
dependence on the number of generations
current iteration number of DE operator
maximum iteration number of DE operator
population scale of DE operator
the best individuals in the k iteration
the i individuals in the k iteration
mutation parameter of DE operator
crossover parameter of DE operator
a vector after mutation in DE operator
a vector after crossover in DE operator
a uniform random number in range (0,1)
a randomly chosen index in the
set {1, 2, . . ., n}

References
[1] Jeyakumar DN, Jayabarathi T, Raghunathan T. Particle swarm
optimization for various types of economic dispatch problems.
Electric Power Energy Syst 2006;28 (1):3642.
[2] Abido MA. A novel multiobjective evolutionary algorithm for
environmental/economical power dispatch. Electric Power Syst Res
2003;65 (1):7181.
[3] Zixiong Liang, Glover JD. A zoom feature for a dynamic programming solution to economic dispatch including transmission losses.
IEEE Trans Power Syst 1992;7 (2):5449.
[4] Park JB, Park YM, Won JR, Lee KY. An improved genetic
algorithm for generation expansion planning. IEEE Trans Power
Syst 1992;15 (3):91622.
[5] Wona Jong-Ryul, Park Young-Moon. Economic dispatch solutions
with piecewise quadratic cost functions using improved genetic
algorithm. Electric Power Energy Syst 2003;25 (5):35561.
[6] Orero SO, Irving MR. A genetic algorithm modeling framework and
solution technique for short term optimal hydrothermal scheduling.
IEEE Trans Power Syst 1998;13 (2):50118.
[7] Baskar S, Subbaraj P, Rao MVC. Hybrid real coded genetic
algorithm solution to economic dispatch problem. Comput Electric
Eng 2003;29 (3):40719.
[8] Storn R, Price K. Dierential evolutionA simple and ecient
heuristic for global optimization over continuous spaces. J Global
Optim 1997;vol. 11 (4):34159.
[9] Coelho Leandro dos Santos, Mariani Viviana Cocco. Combining of
chaotic dierential evolution and quadratic programming for economic dispatch optimization with valve-point eect. IEEE Trans
Power Syst 2006;21 (2):98996.
[10] ChungFu C, JiJen W, JiPyng C, et al. Robust searching hybrid
dierential evolution method for optimal reactive power planning in
large-scale distribution systems. Electric Power Syst Res 2007;77 (56):4307.
[11] Aruldoss Albert Victoire T, Ebenezer Jeyakumar A. Hybrid PSO
SQP for economic dispatch with valve-point eect. Electric Power
Syst Res 2004;71 (1):519.

38

D. He et al. / Electrical Power and Energy Systems 30 (2008) 3138

[12] Sinha N, Chakrabarti R, Chattopadhyay PK. Evolutionary programming techniques for economic load dispatch. IEEE Trans Evol
Comput 2003;7 (1):8394.
[13] Bhagwan Das D, Patvardhan C. Solution of economic load dispatch
using real coded hybrid stochastic search. Int J Electr Power Energy
Syst 1999;21 (3):16570.
[14] Walters DC, Sheble GB. Genetic algorithm solution of economic
dispatch with valve-point loadings. IEEE Trans Power Syst 1993;8
(3):132531.
[15] Chowdhury BH, Rahman S. A review of recent advances in economic
dispatch. IEEE Trans Power Syst 1990;5 (4):124859.
[16] Li Xingsi. An aggregate function method for nonlinear programming.
Sci China (Ser A) 1991;34 (12):146773.
[17] Xingsi L, Shucherng F. On the entropic regularization method for
solving minmax problems with applications. Math Methods Operat
Res 1997;46 (1):11930.
[18] Houduo Q, Lizhi L, Zhenghua L. Regularized smoothing approximations to vertical nonlinear complementarity problems. J Math
Anal Applicat 1999;230 (1):26176.
[19] Tseng P, Bertsekas DP. On the convergence of the exponential
multiplier method for convex programming. Math Program 1993;60
(1):119.
[20] Shucherng F, Jiye H, Zhenghai H, et al. On the nite termination of
an entropy function based non-interior continuation method for

[21]
[22]

[23]

[24]

[25]

[26]

[27]

vertical linear complementarity problems. J Global Optim 2005;33

(3):36991.
Xianfang S, Yuezu F, Wenru N. On the uniformity of U* uniform
designs. Chin J Appl Probab Statist 2001;17 (4):3415.
Dakuo H, Yanqiang L, Fuli W. Hybrid Genetic Algorithm
Based on the Operator of Pattern Search. Inf Control 2001;30 (3):
2768.
Damousis Ioannis G, Bakirtzis Anastasios G, Dokopoulos
Petros S. Network-constrained economic dispatch using realcoded genetic algorithm. IEEE Trans Power Syst 2002;18 (1):
198205.
Aruldoss Albert Victoire T, Ebenezer Jeyakumar A. Deterministically
guided PSO for dynamic dispatch considering valve-point eect.
Electric Power Syst Res 2005;73 (3):31322.
Attaviriyanupap P, Kita H, Tanaka E, Hasegawa J. A hybrid
EP and SQP for dynamic economic dispatch with nonsmooth
incremental fuel cost function. IEEE Trans Power Syst 2002;17 (2):
4116.
Aruldoss Albert Victoire T, Ebenezer Jeyakumar A. A modied
hybrid EPSQP approach for dynamic dispatch with valve-point
eect. Electric Power Energy Syst 2005;27 (8):594601.
Wong KP, Wong YW. Genetic and genetic/simulated-annealing
approaches to economic dispatch. IEEE Proc Gener Trans Distrib
1994;14 (5):50713.