Available online at www.sciencedirect.

com

Electric Power Systems Research 78 (2008) 626–633

Hybrid genetic algorithm for economic dispatch with valve-point effect

He Da-kuo ∗ , Wang Fu-li, Mao Zhi-zhong
Key Laboratory of Process Industry Automation, Ministry of Education, Northeastern University, Shenyang 110004, China
Received 19 January 2006; received in revised form 18 September 2006; accepted 14 May 2007
Available online 2 July 2007

Abstract
This paper presents an efficient method for solving the economic dispatch problem (EDP) through combination of genetic algorithm (GA), the
sequential quadratic programming (SQP) technique, uniform design technique, the maximum entropy principle, simplex crossover and non-uniform
mutation. The proposed hybrid technique uses GA as the main optimizer, the SQP to fine tune in the solution of the GA run. Based on the maximum
entropy principle, the cost function of EDP is approximated by using a smooth and differentiable function to improve the performance of the SQP.
An initial population obtained by using uniform design exerts optimal performance of the proposed hybrid algorithm. The effectiveness of the
proposed method is validated by carrying out extensive tests on two different EDP with incremental fuel-cost function taking into account the
valve-point loadings effects. The result shows that the proposed hybrid genetic algorithm improves the solution accuracy and reliability compared
to other techniques for EDP considering valve-point effects.
© 2007 Elsevier B.V. All rights reserved.

Keywords: Hybrid genetic algorithm (HGA); Sequential quadratic programming; Uniform design; Maximum entropy principle; Economic dispatch problem

1. Introduction nique methods in conventional approaches to the problems are

calculus-based techniques, and require a smooth and convex cost
Economic dispatch is one of the most important problems function and strict continuity of the search space. Dynamic pro-
to be solved in the operation of a power system. Improvements gramming (DP) [2] imposes no restrictions on the nature of the
in scheduling the unit outputs can lead to significant cost sav- cost curves and therefore it can solve EDP with inherently non-
ings. The primary objective of the economic dispatch problem linear and discontinuous cost curves. This method, however,
(EDP) of electric power generation is to schedule the committed suffers from the “curse of dimensionality” or local optimality
generating unit outputs so as to meet the required load demand [1].
at minimum operating cost while satisfying all unit and system GA is a stochastic optimization technique, which is based
equality and inequality constraints [1]. This makes the EDP a on the principle of natural selection and genetics [3,4]. It com-
large-scale highly non-linear constrained optimization problem. bines solution evaluation with randomized, structured exchanges
The input–output characteristics of large units are inherently of genetic information between solutions to obtain optimality.
highly non-linear because of valve-point loadings, generating Also it searches multiple solutions simultaneously in contrast to
unit ramp rate limits, etc. Furthermore they may generate mul- conventional optimal algorithms. Therefore, the possibility of
tiple local minimum points in the cost function. In light of the finding global optimal solution is increased. The main advan-
non-linear characteristics of the units, there is a demand for tech- tage of GA is that it finds near optimal solution in relatively
niques that do not have restrictions on the shape of the fuel-cost short time compared with other random searching methods.
curves. To obtain accurate dispatch results, approaches with- In recent years, the interest in these algorithms has been rising
out restriction on the shape of incremental fuel-cost functions fast, for that they provide robust and powerful adaptive search
are needed. Whereas both lambda-iterative and gradient tech- mechanisms [5]. GA has an immense potential for applications
in the field of power systems and it has been successfully applied
to solve various problems in electric power systems such as eco-
∗ Corresponding author. nomic dispatch [6,7], unit commitment, reactive power control,
E-mail address: [email protected] (D.-k. He). hydrothermal scheduling, and distribution system planning, etc.

0378-7796/$ – see front matter © 2007 Elsevier B.V. All rights reserved.
doi:10.1016/j.epsr.2007.05.008
 D.-k. He et al. / Electric Power Systems Research 78 (2008) 626–633 627

When compared with the foregoing conventional techniques, where Pi min and Pi max are the minimum and maximum power
GA is well appreciated for their global optimality in complex outputs of the ith unit.
search space (multiple local optima, multi-objective, non-linear, The inclusion of valve-point loading effects makes the model-
discontinuous and highly constrained space). Despite the afore- ing of the incremental fuel-cost function of the generators more
mentioned success, GA is only capable of identifying the high practical. This increases the non-linearity as well as number of
performance region at an affordable time and displays inherent local optima in the solution space. Also the solution procedure
difficulties in performing local search for numerical applications can easily trap in the local optima in the vicinity of optimal value.
[8]. The incremental fuel-cost function of the generating units with
To overcome premature convergence and speed up the search valve-point loadings is represented as follows [9,11]:
process, a hybrid method that integrates the GA with a gradient
search algorithm called SQP [9] is proposed to take advantage Fi (Pi ) = ai Pi2 + bi Pi + ci + |ei sin(fi (Pi min − Pi )| (4)
of both GA and the local search techniques. GA is capable of
where, ai , bi and ci , are the fuel-cost coefficients of the unit, ei
exploring a large space, yet is slow in fine tuning local search. In
and fi are the fuel-cost coefficients of the unit with valve-point
contrast, SQP techniques can climb hills rapidly; however, they
effects.
are blind to the potential hills in the neighbourhood area and
The economic dispatch of generation of real power of the
sensitive to the initial starting points. The hybrid GA uses a GA
generating units is to be done to the required load demand by
to identify the potential hill within a reasonably short period of
satisfying the above constraints. The incremental fuel-cost func-
time, while SQP technique subsequently takes over and rapidly
tion can be modeled in a more practical fashion by including
climbs the remaining hill. Therefore, this algorithm increases
the valve-point effects [12]. The generating units with multi-
the possibility of finding global optimal point and improves the
valve steam turbines exhibit a greater variation in the fuel-cost
convergence speed. The proposed hybrid technique uses GA as
functions. The valve-point effects introduce ripples in the heat-
a base level search towards the optimal region and SQP method
rate curves, thereby the number of local optima is increased.
as optimization to do the fine tuning.
Hence, a technique that overcomes these complexities has to be
In general, the hybrid method offers an exact solution only
evolved.
when the function is smooth and its gradient information are
known [8]. The method is proposed to approximate EDP by
using a smooth and differentiable function based on the max- 3. Entropic smoothing approximation function
imum entropy principle. In this way, the performance of the
hybrid method is improved. At the same time, to improve Let Ψ i : Rn → R, i = 1, 2, . . ., m, be differentiable and define
rationality of the distribution of initial population, the hybrid a max-type function Ψ : Rn → R, i = 1, 2, . . ., m by
technique integrating the uniform design with the genetic algo-
Ψ (x) = max Ψi (x) (5)
rithm (UHGA) is proposed. i
In order to validate the performance of the proposed UHGA,
two economic dispatch problems with incremental fuel-cost Given any μ > 0, consider the following entropy-type func-
functions taking into account the valve-point loading effects tion as a smoothing approximation function of Ψ ,
were tested and the results obtained were compared with those 
m  
Ψi (x)
reported in literatures [1,10]. Ψ (x, μ) = μ ln exp (6)
μ
i=1
2. EDP formulation Note that, for μ > 0,

m  
The classic EDP minimizes the following incremental fuel- Ψ (x) − Ψi (x)
cost function associated to dispatchable units: Ψ (x, μ) = Ψ (x) + μ ln exp (7)
μ
N  i=1
P
min F = min Fi (Pi ) (1) Moreover,
i=1
Ψ (x) ≤ Ψ (x, μ) ≤ Ψ (x) + μ ln m, ∀x ∈ Rn and μ > 0 (8)
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 Therefore, Ψ (x, μ) → Ψ (x) as μ → 0. This fact allows us
constraints: to solve the problem without facing the non-differentiability
problem of Ψ (x). Since the function (6) can be derived from

NP
the dual problem of an entropy optimization problem, we call
Pi = PD + Ploss (2) function (6) an entropic smoothing approximation function. Li
i=1
[13] discovered a few properties of this function and named
where PD is the system load demand and Ploss is the transmission it as the aggregate function. Related work can be found in
loss, and generating capacity constraints: [14–17].
To EDP, the incremental fuel-cost function of the generat-
Pi min ≤ Pi ≤ Pi max , for i = 1, 2, . . . , NP (3) ing units with valve-point loadings (1) are represented as the
628 D.-k. He et al. / Electric Power Systems Research 78 (2008) 626–633

following function: hi = n + 1 − hj (i = j) (13)

Fi (Pi ) = ai Pi2 + bi Pi + ci + |ei sin(fi (Pi min − Pi )| where ϕ(n + 1) is the Euler function of n + 1; hi and hj are the
different generating elements in generating vector.
= ai Pi2 + bi Pi + ci + max{ei sin(fi (Pi min − Pi ),
Theorem 2. If (12) and (13) are satisfied, the optimal gen-
− ei sin(fi Pi min − Pi )} (9)
erating vector of uniform design with n dimensionalities and
Based on the maximum entropy principle the incremental (1/2)ϕ(n + 1) − 1 factors is composed of natural numbers which
fuel-cost function can be approximated by entropic smoothing are prime numbers with n + 1 in [1, n + 1/2) with any one taken
approximation function: out.

Ψ (Pi , μ) = ai Pi2 + bi Pi + ci + μ ln[eei sin(fi (Pi min −Pi )/μ Based on Theorems 1 and 2, for the uniform designs with
(1/2)ϕ(n + 1) or (1/2)ϕ(n + 1) − 1 factors, the usable tables can
+ e−ei sin(fi (Pi min −Pi )/μ ] (10) be obtained immediately without any measurement for its uni-
The function Ψ (Pi , μ) provides a good approximation to the formity. The good lattice point method is employed to obtain
function Fi (Pi ) in the sense that the usable table with optimal generating vector. For the uniform
designs with certain factors, n is calculated by formula of Euler
0 ≤ Ψ (Pi , μ) − Fi (Pi ) ≤ μ ln 2 and μ > 0 (11) function.
It is clear that this function uniformly approximates the
5. Hybrid genetic algorithm
maximum function Fi (Pi ) when a parameter μ tends to infinites-
imal and is greater than Fi (Pi ) with an error bound less
5.1. Selection operator
than μ ln 2.
Because of approximation, the result obtained by SQP
After the evaluation of the initial population, the GA begins
method is approximation of exact result. To handle this prob-
the creation of the new generation. The selection used in this
lem, the hybrid method proposed in this paper only uses SQP
paper depends on individual fitness. The best individuals of the
method to find approximation solution not optimum cost. In
present population are kept for the next population. The fitness
other word, entropic smoothing approximation function is only
value in this paper contains the penalty function and does not
used in SQP method not in whole HGA. Thus, approximation
represent the true objective function.
does not influence accuracy of HGA.

4. Uniform design 5.2. Simplex crossover operator

The distribution of the individuals of initial population The primary genetic operator is the crossover operator. The
directly influences the globe convergence and searching effi- purpose of crossover operator is to produce new chromosomes
ciency. Therefore, the reasonable setting of initial population that are distinctly different from their parents, yet remain some
is an important problem in the application of GA to per- of their parents characteristics. It has important function in the
form optimization calculation. Because of introducing SQP capability of searching the optimal solution.
method, if the distribution of the individuals of initial pop- Simplex algorithm is a method which finds optima through
ulation is not reasonable, the HGA sometimes converges to improving inferior point by searching from inferior point
local optima and cannot reach the global optimal point. There- towards the pivot of n + 1 initial points of the Simplex for the n-
fore, initial population is more important to the HGA proposed dimensional problem. Simplex algorithm is similar to crossover
in this paper. In order to exert optimal performance of GA, operator of GA in utilizing information of multi-points. Based
the initial population of GA must reflect the information of on aforementioned analysis, simplex crossover operator [19] is
solution space scientifically. However, it is difficult to eval- employed in this paper to improve convergence speed and lead
uate the distribution of the individuals of initial population. the population to the global optimum because of the character-
The uniform design is an effective method to solve this istic of improving inferior point.
problem. It is employed to obtain initial population in this
paper. 5.3. Non-uniform mutation operator [20]
Because the dimensionality of EDP is large, it is difficult to
obtain optimal generating vector of the usable tables of uniform Non-uniform mutation operator is employed in this paper.
design. So the following method is used in this paper. This operator is described below.
If P t = [P1t , P2t , . . . , Pnt ] is a power output chromosome vec-
Theorem 1. If (12) and (13) are satisfied, the optimal gen- tor and Pit is ith unit’s output that is chosen to be mutated, the
erating vector of n dimensionalities and (1/2)ϕ(n + 1) factors new output Pimut will be after mutation
uniform design is composed of natural numbers which are prime
numbers with n + 1 [18]. Pit + (t, Pi max − Pit ), if r = 0
Pimut = (14)
hi = hj (i = j) (12) Pit − (t, Pit − Pi min ), if r = 1
 D.-k. He et al. / Electric Power Systems Research 78 (2008) 626–633 629

where r is a random bit, and function (t, y) returns a value in For each iteration of the QP sub-problem the direction dk is
the range [0, y] such that the probability of the value returned is calculated using the above Eq. (16). The solution obtained is
close to 0 increases with t used to form a new iterate given by:
b
(t, y) = y × (1 − ξ (1−t/gen) ) (15) Pk+1 = Pk + αk dk (23)

where ξ is a random floating-point number in the interval [0, 1]; t The step length value αk is determined to produce a consid-
the current generation; gen the maximum number of generations; erable reduction in an augmented Lagrangian merit function:
b is a parameter that determines the degree of dependence on ρ
the number of generations. In this way, the operator makes a LA (P, λ, ρ) = FT (P) − λT (P) + c(P)T c(P) (24)
2
uniform search at the beginning of the evolution and in later
where, λ is the vector of Lagrangian multiplier and ρ is a non-
stages narrows the search around the local area of the parameter
negative scalar. The procedure will be repeated until the value
resembling a hill-climbing operator. For our experiments, b was
of sk has reached some tolerance value.
chosen equal to 2.
6. Solution methodology
5.4. Sequential quadratic programming operator [9,21]
The proposed HGA for EDP with valve-point effects can be
SQP method seems to be the best non-linear programming summarized as follows:
methods for constrained optimization. It outperforms every
other non-linear programming method in terms of efficiency, Step 1: Get the data for the system.
accuracy, and percentage of successful solutions over a large Step 2: Initialize parameter of algorithm and count t.
number of test problems. The method resembles closely to New- Step 3: Generate initial population by the usable tables of uni-
ton’s method for constrained optimization just as is done for form design.
unconstrained optimization. At each iteration an approxima- Step 4: Evaluate the objective function and update count t.
tion is made by the Hessian of the Lagrangian function using Step 5: Identify the Fitbest (t) of the current run t.
a Broyden–Foldfarb–Shanno (BFGS) quasi-Newton updating Step 6: If Fitbest (t) < Fitbest (t − 1) replace Fitbest (t − 1) with
method. This is then used to generate a quadratic programming Fitbest (t), otherwise go to Step 7.
(QP) sub-problem whose solution is used to form a search direc- Step 7: Generate selection offspring using selection operation.
tion for a line search procedure. In this paper, SQP is used as Step 8: Generate crossover offspring using simplex crossover
a local optimizer to fine-tune the region explored by GA in its operation.
run. Step 9: Generate mutation offspring using non-uniform muta-
First let us formulate the QP sub-problem for the problem as tion operation.
stated by (1) subject to (2) and (3). Step 10: Take the individuals selected randomly as the initial
1 starting point for the SQP and generate SQP offspring
min ∇FT (Pk )T dk + dkT Hk dk (16) by the final solution obtained using the SQP.
2
Step 11: Generate whole offspring with selection offspring,
subject to crossover offspring, mutation offspring and SQP off-
c(Pk )+∇c(Pk )T dk = 0 (17) spring.
Step 12: While (termination criterion not met).
Pmin ≤ Pk + dk ≤ Pmax (18)
The termination is done when a specified number of iterations
where, Hk is the Hessian matrix of the Lagrangian function at the met.
kth iteration, dk the search direction at the kth iteration, Pk the
real power vector at the kth iteration, and c(Pk ) is the constraint 7. Simulation results
given by Eq. (2)

L(P, λ) = FT (P) + c(P)T λ (19) The proposed UHGA approach was tested with two test cases
of EDP with valve-point effects. The software was written in
and is constructed from a quasi-Newton update formula given MATLAB 7.0 and executed on a Pentium-IV 2.99 GHz personal
by: computer. Hereinafter, the results represent the average of 30
runs of the proposed method for both the two test cases.
Hk sk (sk )T Hk qk (qk )T Since ϕ(28 + 1) = ϕ(29) = 28 = ϕ(13 + 1) and ϕ(82 + 1) =
Hk+1 = Hk − + (20)
(sk )T Hk sk (qk )T sk ϕ(83) = 82 = 2(40 + 1), the optimal generating vector of uniform
design with 28 dimensionalities and 13 factors and the optimal
where
generating vector of uniform design with 82 dimensionalities
sk = Pk+1 − Pk (21) and 40 factors are composed of natural numbers which are
prime numbers with 29 in (1, 14.5) with any one taken out and
qk = ∇L(Pk+1 , λK+1 ) − ∇L(Pk , λK+1 ) (22) 83 in [1, 41.5) with any one taken out, respectively, by Theorem
630 D.-k. He et al. / Electric Power Systems Research 78 (2008) 626–633

2. In this paper, the optimal generating vector of uniform design Table 1

with 28 dimensionalities and 13 factors are all natural numbers Comparison of fuel costs for PD = 1800 MW for case 1
in [1, 13]. The optimal generating vector of uniform design with Method Mean time (s) Best cost (US$ h−1 ) Best cost (US$ h−1 )
82 dimensionalities and 40 factors are all natural numbers in [1, EP [1] 157.43 17,994.07 18,127.06
40]. Then the good lattice point method is employed to obtain EP-SQP [9] 121.93 17,991.03 18,106.93
the usable table with optimal generating vector. Hence the PSO [9] 77.37 18,030.72 18,205.78
population size was chosen as 28 for test case 1 and 82 for test PSO-SQP [9] 33.97 17,969.93 18,029.99
case 2. UHGA 15.33 17,964.81 17,992.92
The selection ‘crossover’ SQP and mutation operation pro-
portions were chosen as 5:8:8:7 for test case 1 and 10:20:20:32
Table 2
for test case 2. The simulation parameters of the UHGA for Best result obtained for PD = 1800 MW for case 1 using UHGA
the two test systems are fixed as follows, b = 2, μ = 1, and
itermax = 30. To validate the comparison of results obtained using Power Generation Power Generation
the proposed technique with the results obtained using the EP, 1 628.21 8 60.00
EP-SQP, PSO and PSO-SQP techniques, solution procedure of 2 223.94 9 109.71
the UHGA technique is terminated when the maximum number 3 149.30 10 40.00
4 109.71 11 40.00
of iterations is reached. 5 109.71 12 55.00
The fitness function Fit(Pi ) is given as 6 109.71 13 55.00
 7 109.71
F, if Pi is feasible
Fit(Pi ) = (25)
F ± F × gmax , otherwise

where F is the value of objective function, symbol ‘±’ is used points as there are more possibilities for any method to stick on
to keep penalty, gmax is given as any one of the local optimum points.
The final fuel costs obtained using the EP, EP-SQP, PSO,
gmax = max{0, gi (P), |hj (P)|, i = 1, 2, . . . , m1 , j PSO-SQP and the proposed method for power demand of
= 1, 2, . . . , m2 } (26) 1800 MW were summarized in Table 1. The best results obtained
for solution vector, with UHGA with minimum cost of US$
where gi (P) are the inequality constraints, hi (P) the equality con- 17,964.81 h−1 is given in Table 2. Table 3 reports the dispatch
straints, m1 and m2 are the number of the inequality constraints results of the various methods [23], EP-SQP, PSO-SQP and the
and the equality constraints, respectively. proposed method for a load demand of 2520 MW. The prob-
lem is solved for two different power demands in order to show
7.1. Case 1 the effectiveness of the proposed method in producing quality
solutions.
This test case comprises of thirteen generating units, the com- It is clear from the Tables 1 and 3, minimum cost and the mean
plexity and non-linearity to the solution procedure is increased. cost value obtained by the proposed method is comparatively less
The expected power demands to be met by the all thirteen gener- compared to all the other methods. To show the effectiveness
ating units is 1800 [1] and 2520 MW [22]. The system data can of UHGA, the test problems were also experimented using the
be found from [1]. The problem has a number of local optimum HGA without initial population obtained by the uniform design.

Table 3
Dispatch results for a PD = 2520 MW for case 1
Generator Unit generation (MW)
GA-SA [9] EP-SQP [9] PSO-SQP [9] UHGA

1 628.23 628.3136 628.3205 628.2330

2 299.22 299.1715 299.0524 299.0288
3 299.17 299.0474 298.9681 299.0288
4 159.12 159.6399 159.4680 159.6077
5 159.95 159.6560 159.1429 159.6077
6 158.85 158.4831 159.2724 159.6077
7 157.26 159.6749 159.5371 159.6077
8 159.93 159.7265 158.8522 159.6077
9 159.86 159.6653 159.7845 159.6077
10 110.78 114.0334 110.9618 77.1613
11 75.00 75.0000 75.0000 77.1613
12 60.00 60.0000 60.0000 89.5992
13 92.62 87.5884 91.6401 92.1414
Total cost (US$ h−1 ) 24,275.71 24,266.44 24,261.05 24,172.25
 D.-k. He et al. / Electric Power Systems Research 78 (2008) 626–633 631

Table 4
Comparison of fuel costs for case 2
Method Mean time Best cost Mean cost
(s) (US$ h−1 ) (US$ h−1 )

EP [1] 1167.35 122,624.35 123,382.00

EP-SQP [9] 997.73 122,323.97 122,379.63
MPSO [24] – 122,252.27 –
PSO [9] 933.39 123,930.45 124,154.49
PSO-SQP [9] 733.97 122,094.67 122,245.25
DEC(2)-SQP(1) [25] 14.26 121,741.98 122,295.13
UHGA 333.68 121,424.48 121,602.81

Table 5
Best result obtained for case 2 using UHGA
Power Generation Power Generation

1 110.8056 21 523.2789
2 110.8000 22 523.2799
Fig. 1. Convergence characteristics of the UHGA and HGA for case 1 for a 3 97.4052 23 523.2799
PD = 1800 MW. 4 179.7314 24 523.2832
5 87.8939 25 523.2823
6 140.0000 26 523.2884
Fig. 1 shows the convergence characteristics of the UHGA 7 259.6016 27 10.0000
and HGA method for power demand of 1800 MW. Fig. 2 shows 8 284.6084 28 10.0000
the convergence characteristics of the UHGA and HGA method 9 284.6046 29 10.0000
10 130.0000 30 97.0000
for power demand of 2520 MW. The average number of iter-
11 94.0000 31 190.0000
ations to reach the optimum solution of the UHGA for power 12 168.8002 32 190.0000
demand of 1800 MW is 8–12 and 2520 MW is 7–10 in all the 13 214.7600 33 190.0000
30 trial runs. 14 304.5239 34 164.8113
15 394.2796 35 200.0000
16 394.2790 36 200.0000
7.2. Case 2 17 489.2820 37 110.0000
18 489.2799 38 110.0000
This test case comprises of 40 generating units. This is a 19 511.2804 39 110.0000
larger system. The number of local optima, complexity and non- 20 511.2803 40 511.2803
linearity to the solution procedure is enormously increased. The
required power demand to be met by all the forty generating
SQP, MPSO, DEC(2)-SQP(1) and the proposed method were
units is 10,500 MW. The system data can be found from [1].
summarized in Table 4. It is clear from Table 4, UHGA has the
The final fuel costs obtained using the EP, EP-SQP, PSO, PSO-
best probability of the mean cost value and the minimum cost
amongst all the methods in this test case. It is noticeable that the
mean cost value obtained using UHGA is less than the minimum
cost obtained using other methods. Though the solution time
of UHGA in case 2 is higher than DEC(2)-SQP(1), it offered
better solution quality as compared to DEC(2)-SQP(1). The best
results obtained for solution vector, with UHGA with minimum
cost of US$ 121,424.48 h−1 is given in Table 5. Fig. 3 shows
that the UHGA performed better than HGA in the convergence
characteristics. The average number of iterations to reach the
optimum solution of the UHGA is 16–18 in all the 30 trial runs.
Hence, for power system ELD problems of greater size with
more non-linearities, the proposed method is proved to be the
best algorithm amongst all the methods.

8. Discussion and conclusion

Traditionally, to solve the EDP effectively, conventional tech-

niques require convex cost function and strict continuity of the
Fig. 2. Convergence characteristics of the UHGA and HGA for case 1 for a search space. But practically the incremental fuel-cost curves
PD = 2520 MW. of the generating units are inherently highly non-linear and
632 D.-k. He et al. / Electric Power Systems Research 78 (2008) 626–633

Appendix A. List of symbols

b parameter that determines the degree of dependence on

the number of generations
F total production cost (US$)
Fi (Pi ) incremental fuel-cost function (US$ h−1 )
gen the maximum number of generations
NP number of generating units
PD the system load demand (MW)
Pi real power output of the ith unit (MW)
Pi min /Pi max minimum/maximum limit of the real power of the
ith unit (MW)
Pimut ith unit’s new output after mutation (MW)
Pit ith unit’s output that is chosen to be mutated (MW)
Ploss transmission loss (MW)
r a random bit
t the current generations
Fig. 3. Convergence characteristics of the UHGA and HGA for case 2 for a
P = 10,500 MW. Greek letters
ϕ(n + 1) Euler function of n + 1
ξ a random floating-point number in the interval [0, 1]
non-continuous. GA is an important tool for solving complex
optimization problems, being applied to solve various problems References
in various diverse fields. It was also effectively used in solving
complex problems in the power system field such as EDP. [1] N. Sinha, R. Chakrabarti, P.K. Chattopadhyay, Evolutionary programming
GA is faster in finding the high performance region but techniques for economic load dispatch, IEEE Trans. Evol. Comput. 7 (1)
(2003) 83–94.
displays difficulties in performing local search for complex func-
[2] Z.-X. Liang, J.D. Glover, A zoom feature for a dynamic programming
tions. It leads to premature convergence and also has a poor fine solution to economic dispatch including transmission losses, IEEE Trans.
tuning of the final solution. To overcome these drawbacks, GA Power Syst. 7 (5) (1992) 544–549.
was integrated with SQP. This technique is used to solve the EDP [3] J.B. Park, Y.M. Park, J.R. Won, K.Y. Lee, An improved genetic algorithm
with incremental fuel-cost functions taking valve-point effects for generation expansion planning, IEEE Trans Power Syst. 15 (3) (1992)
916–922.
into account. SQP proves itself as a best non-linear programming
[4] J.-R. Wona, Y.-M. Park, Economic dispatch solutions with piecewise
method to solve the constrained optimization problem. The SQP quadratic cost functions using improved genetic algorithm, Electr. Power
can explore the search space quickly with a gradient direction Energy Syst. 25 (5) (2003) 355–361.
and guarantee a local optimum solution. But the method is sensi- [5] S.O. Orero, M.R. Irving, A genetic algorithm modeling framework and
tive to the initial point. The hybrid approach for solving the EDP solution technique for short term optimal hydrothermal scheduling, IEEE
Trans. Power Syst. 13 (2) (1998) 501–518.
of units with value-point effects produces quality solutions as
[6] G.B. Sheble, T.T. Maifeld, K. Brittig, G. Fahd, S. Fukurozaki-Coppinger,
compared to the one produced by these techniques when applied Unit commitment by genetic algorithm with penalty methods and a com-
separately. parison of Lagrangian search and genetic algorithm-economic dispatch
It is advantaged to improve the performance of the SQP that example, Electr. Power Energy Syst. 18 (6) (1996) 339–346.
the cost function of EDP is approximated by using a smooth [7] Gerald B. Sheble, K. Brittig, Refined genetic algorithm-economic dispatch
example, IEEE Trans. Power Syst. 10 (1) (1995) 117–124.
and differentiable function based on the maximum entropy prin-
[8] S. Baskar, P. Subbaraj, M.V.C. Rao, Hybrid real coded genetic algorithm
ciple. An initial population obtained by using uniform design solution to economic dispatch problem, Comput. Electr. Eng. 29 (3) (2003)
reflects the information of solution space scientifically and exerts 407–419.
optimal performance of the proposed hybrid algorithm. [9] T. Aruldoss Albert Victoire, A. Ebenezer Jeyakumar, Hybrid PSO-SQP for
The performance of the hybrid method was tested for two economic dispatch with valve-point effect, Electric Power Syst. Res. 71 (1)
(2004) 51–59.
EDP test cases with valve-point effects included and compared
[10] D. Bhagwan Das, C. Patvardhan, Solution of economic load dispatch using
with the results obtained using the methods reported in recent real coded hybrid stochastic search, Int. J. Electr. Power Energy Syst. 21
literature. The results show that the proposed UHGA performed (3) (1999) 165–170.
much better than other methods compared in terms of conver- [11] D.C. Walters, G.B. Sheble, Genetic algorithm solution of economic dis-
gence performance, minimum cost and probability of attaining patch with valve-point loadings, IEEE Trans. Power Syst. 8 (3) (1993)
1325–1331.
better solutions.
[12] B.H. Chowdhury, S. Rahman, A review of recent advances in economic
dispatch, IEEE Trans. Power Syst. 5 (4) (1990) 1248–1259.
Acknowledgement [13] X.-S. Li, An aggregate function method for nonlinear programming, Sci.
China (Ser. A) 34 (12) (1991) 1467–1473.
[14] X.-S. Li, S.-C. Fang, On the entropic regularization method for solving
This work is supported by the National Nature Science Foun- min-max problems with applications, Math. Methods Oper. Res. 46 (1)
dation of China under Grant 60374063. (1997) 119–130.
 D.-k. He et al. / Electric Power Systems Research 78 (2008) 626–633 633

[15] H.-D. Qi, L.-Z. Liao, Z. Lin, Regularized smoothing approximations to [21] T. Aruldoss Albert Victoire, A. Ebenezer Jeyakumar, Deterministically
vertical nonlinear complementarity problems, J. Math. Anal. Appl. 230 (1) guided PSO for dynamic dispatch considering valve-point effect, Electric
(1999) 261–276. Power Syst. Res. 73 (3) (2005) 313–322.
[16] P. Tseng, D.P. Bertsekas, On the convergence of the exponential multiplier [22] K.-P. Wong, Y.-W. Wong, Genetic and genetic/simulated-annealing
method for convex programming, Math. Program. 60 (1) (1993) 1–19. approaches to economic dispatch, IEEE Proc. Gener. Trans. Distrib. 141
[17] S.-C. Fang, J.-Y. Han, Z.-H. Huang, et al., On the finite termination of an (5) (1994) 507–513.
entropy function based non-interior continuation method for vertical linear [23] P. Attaviriyanupap, H. Kita, E. Tanaka, J. Hasegawa, A hybrid EP and
complementarity problems, J. Global Optim. 33 (3) (2005) 369–391. SQP for dynamic economic dispatch with nonsmooth incremental fuel cost
[18] X.-F. Sun, Y.-Z. Fan, W.-R. Ning, On the uniformity of U* uniform designs, function, IEEE Trans. Power Syst. 17 (2) (2002) 411–416.
Chin. J. Appl. Prob. Stat. 17 (4) (2001) 341–345. [24] J.-B. Park, K.-S. Lee, J.-R. Shin, K.Y. Lee, A particle swarm optimization
[19] D.-K. He, Y.-Q. Li, F.-L. Wang, Hybrid genetic algorithm based on the for economic dispatch with nonsmooth cost unction, IEEE Trans. Power
operator of pattern search, Inf. Control 30 (3) (2001) 276–278. Syst. 20 (1) (2005) 34–42.
[20] I.G. Damousis, A.G. Bakirtzis, P.S. Dokopoulos, Network-constrained eco- [25] L. dos, S. Coelho, V. Cocco Mariani, Combining of chaotic differential
nomic dispatch using real-coded genetic algorithm, IEEE Trans. Power evolution and quadratic programming for economic dispatch optimization
Syst. 18 (1) (2002) 198–205. with valve-point effect, IEEE Trans. Power Syst. 21 (2) (2006) 989–996.