IEEE Transactions on Power System. Vol. 10. No. 4.

November 1995

1919

LARGE-SCALE ECONOMIC DISPATCH BY GENETIC ALGORITHM

Po-Hung Chen and Hong-Chan Chang, Member, IEEE
Department of Electrical Engineering, National Taiwan Institute of Technology Taipei, Taiwan, ROC
Abstract- This paper presents a new genetic approach for solving the economic dispatch problem in large-scale systems. A new encoding technique is developed. The chromosome contains only an encoding of the normalized system incremental cost in this encoding technique. Therefore, the total number of bits of chromosome is entirely independent of the number of units. The salient feature makes the proposed genetic approach attractive in large and complex systems which other methodologies may fail to achieve. Moreover, the approach can take network losses, ramp rate limits, and prohibited zone avoidance into account because of genetic algorithm's flexibility, Numerical results on an actual utility system of up to 40 units show that the proposed approach is faster and more robust than the well-known lambda-iteration method in large-scale systems.

Keywords: Economic dispatch, Genetic algorithm, Network losses, Ramp rate, Prohibited zone, Lambda-iteration.
1. INTRODUCTION

The operation planning of a power system is characterized by having to maintain a high degree of economy and reliability. Among the options that are available to the engineer in choosing how to operate the system, the most significant choice is in economic dispatch (ED) [1,2,3]. Previous efforts at ED have applied various mathematical programming methods and optimization techniques. These include the lambda-iteration method [4,5], the base point and participation factors method [4,5], the gradient method [4,6], and a unit-based genetic algorithm method [7,8] has also been reported recently. Among these methods, the lambda-iteration method has been applied in many software packages and used by power utilities for ED due to its ease of implementation. However, the characteristics and feasibility of the lambda-iteration method have not yet been extensively investigated. Since the lambda-iteration method requires a continuous problem formulation, it cannot be directly applied to the ED problem with discontinuous prohibited zones. Moreover, the experimental results in this paper show that the lambdaWS 95 WM 168-5 P R A paper recommended and approved by the IEEE Power System Engineering Committee of the IEEE Power Engineering Society for presentation a t the 1995 IEEE/PES Winter Meeting, January 29, t o February 2, 1995, Nw Pork, Np. Manuscript submitted e December 30, 1993; made available for printing
November 23, 1994.

iteration method has oscillatory problems in large-scale, mixed-generating unit systems which result in a slower solution time. Recently, a global optimization technique known as genetic algorithm (GA) has become a candidate for many optimization applications due to its flexibility and efficiency. GA is a stochastic searching algorithm. It combines an artificial survival of the fittest with genetic operators abstracted from nature to form a surprisingly robust mechanism that is suitable for a variety of optimization problems. One of the advantages of GA is using stochastic operators instead of deterministic rules to search a solution. GA hops randomly from point to point, thus allowing it to escape from local optimum in which other algorithms might land. Therefore, the global optimum of the problem can be approached with high probability [9,10]. Another attractive property of GA is that it searches for many optimum points in parallel. GA has been successfidly applied in various areas such as feeder reconQuration [11 , hydrogenerator governor tuning 1 [12], load flow problems [13], and so on. Sheble, et al. [7,8] first presented papers where GA solves the ED problem. In the papers, an example with three units was explored. Even though the example shows that GA is an efficient method for solving the ED problem in a small-size system, the proposed unit-based encoding technique restricts the applicability of CA to large-scale systems. In unit-based encoding, the number of bits of chromosome increase with the number of units. For example, suppose we use 11 bits to represent the generation output of one unit [7,8], we will then have 440 bits of chromosome in total for a 40-unit system. It is Hicult to obtain a solution with such a large bit dimension. This paper develops a lambda-based GA approach for solving the ED problem. A salient feature of the proposed approach is that the solution time grows approximately linearly with problem size other than geometrically. This feature is attractive in large-scale problems. Comparative studies on the Taipower system, a realistic system in Taiwan, show that the proposed approach is numerically more robust and efficient than the lambda-iteration method in large-scale systems.
2. PROBLEM DESCRIPTION

The list of symbols used in this section is as follows: F: total generation cost of the system : power generation of unit i

0885-8950/95/$04.00 0 1995 IEEE

1920

: minimum generation of unit i

_ . _ .

2.3 Ramp rate limits

: maximum generation of unit i n : total number of units ( 4 ) : generation cost for 4 : power generation of unit i at previous hour
pD: system load demand qoss : system transmission network losses

B,, : loss coefficients

U : ramp rate limit of unit i as generation increases . OR,: ramp rate limit of unit i as generation decreases
pp;, PG : bounds of a prohibited zone
2.1 Basic economic dispatchformulation

In ED research, a number of studies have focused upon the economical aspects of the problem under the assumption that unit generation output can be adjusted instantaneously. Even though this assumption simplifies the problem, it does not reflect the actual operating processes of the generating unit. The operating range of all on-line units is restricted by their ramp rate limits [2,15]. Fig. 1 shows three possible situations when a unit is on-line from hour t-1 to hour t. Fig l(a) shows that the unit is in a steady operating status. Fig l@) shows that the unit is in an increasing power generation status. Fig. l(c) shows that the unit is in a decreasing power generation status.

The objective of ED is to minimize the total generation cost of a power system over some appropriate period (one hour typically) while satisfying various constraints. In equation form, this becomes a constrained optimization problem:
n

p,
t-1
(a)

t-1

(b)

t-1
(C)

Fig. 1. Three possible situations of an on-line unit.

Problem Minimize F =
1=1

(4)

Then, the inequality constraints due to ramp rate limits are given: 1) if generation increases

q -5"SUR,
2) if generation decreases

(4)

In (l), the generation cost function expressed as a quadratic polynomial:

A ( 4 ) is

q O - 4 lDR, (5) The ramp rate limits UR, and OR, are both MW/h unit
usually Practically, OR, is greater than U&. Combining ( ) (5), and 4, (l),the constrained optimization problem is modified as:
n

where a,, bi ,and ci are constants.

2.2 Network losses

Problem Minimize F =
i=I

( 4)
(6)

PBI: Subject to:

Since the power stations are usually spread out geographically, the transmission network losses must be taken into account to achieve true economic dispatch. Network loss is a function of unit generation. To calculate network losses, two methods are in general use [2,141. One is the penalty factors method and the other is the B coefficients method. The latter is commonly used by the power utility industry. In the B coefficients method, network losses are expressed as a quadratic function:

Mi(e,e0-DR,)-<C M i n ( z , P," + U . ) 2.4 Prohibited operating zone

1 4 = PD +40ss 1=I

4s =zzmyP, 0
1=1j=1

(3)

where B,, are constants called B coefficients or loss coefficients. The B coefficients method is used for this study.

Fig. 2 shows the input-output performance curve for a typical thermal unit. The prohibited operating zones in the curve are due to steam valve operating or vibration in a shaft bearing [2,18]. Several studies in the literature [2,8,16 discuss the effects of the prohibited zone in the ED problem For example, Walters & Sheble model the effects of the prohibited zone as a recurring rectified sinusoid function [SI However, in practice, the shape of the input-output curve in the neighborhood of the prohibited zone is difficult to determine by actual performance testing or operating records In actual operation, the best economy is achieved by avoiding

1921

operation in these areas [16,17,18]. As such, a heuristic algorithm is developed in this paper to adjust the generation output of a unit in order to avoid unit operation in the prohibited zones.

of overcoming this shortcoming. It is an occasional (with small probability) random alternation of a chromosome position as shown in Fig. 4. This provides background variation and occasionally introduces beneficial materials into the population.
ChildA: NewChildA:

1 1

1 1 0 1 0

5
1 0
1 0

1 0

Fig. 4. The binaty mutation.

I
GZ-

%+

outpd(Mw)

4. GENETIC ALGORITHM SOLUTION METHODOLOGY

Fig. 2. Typical input-output curve of a thermal unit.

3. OVERVIEW OF THE GENETIC ALGORITHM

GA is a search algorithm based on the mechanics of natural genetics and natural selection [9]. It combines the adaptive nature of the natural genetics or the evolution procedures of organs with functional optimizations. By simulating "the survival of the fittest" of Darwinian evolution among chromosome structures, the optimal chromosome (solution) is searched by randomized information exchange. In every generation, a new set of artificial chromosomes is created using bits and pieces of the fittest of the old ones. While randomized, GA is not a simple random walk. It efficiently exploits historical information to speculate on new search points with expected improved performance [9,10]. GA is essentially derived from a simple model of population genetics. The three prime operators associated with the GA are reproduction, crossover, and mutation. Reproduction is simply an operation whereby an old chromosome is copied into a "mating pool" according to its fitness value. More highly fitted chromosomes (i.e., with better values of the objective function) receive a higher number of copies in the next generation. Copying chromosomes according to their fitness values means that chromosomes with a higher value have a higher probability of contributing one or more offspring in the next generation. Crossover is an extremely important component of the GA. It is a structured recombination operation. This operation is similar to two scientists exchanging information. This study applies a new crossover technique known as "uniform crossover'' as shown in Fig. 3. Syswerda [191 shows that convergence speed of the uniform crossover is faster than the popular one-point crossover and two-point crossover.
ParentA: ParentB: A random mask : Child A : ChildB:

The detailed solution methodology includes: the encoding and decoding techniques, constrained generation output calculation, the fitness function, parent selection, and parameter selection. These are described in more detail below.
4.1 Encoding and Decoding

Implementation of a problem in a GA starts from the parameter encoding (i.e., the representation of the problem). The encoding must be carefully designed to utilize the GA's ability to efficiently transfer information between chromosome strings and objective function of problem. The proposed approach uses the equal system A (equal system incremental cost) criterion as its basis. The only encoded parameter is the normalized system incremental cost, Anm, where 0 5 An"' II. The advantage of using system A instead of units' output as the encoded parameter is that the number of bits of chromosome will be entirely independent of the number of units. This is particularly attractive in large-scale systems. The resolution of the solution depends upon how many bits are used to represent P. In other words, the more encoding bits there are, the higher the resolution. However, on the other hand, the more encoding bits there are, the slower the convergence. In this paper, we use 10 bits to represent Anm. Fig. 5 shows the encoding scheme of Anm.
x
2-1

x
2-2

x
2-3

x
24

x
2-5

x
2-6
I

x
2-1

x
2-8

x
2-9

x
2-10

where d, E { O , I }

=1,2,....., 10

Fig. 5. The encoding scheme of km

1 0 0 1
0 0

1 0 0 1 0 1 1 0 0 1
1

0 0 0 1

Evaluation of a chromosome is accomplished by decoding the encoded chromosome string and computing the chromosome'sfitness value using the decoded parameter. The decoding of Anm can be expressed as:

Fig. 3. The uniform crossover.

Anm= z ( d i ~
i =Z

10

(7)

Although reproduction and crossover effectively search and recombine existing chromosomes, they do not create any new genetic material in the population. Mutation is capable

1922

When the fitness of each chromosome is calculated, the The relationship between the actual system incremental cost, A?t, and the normalized system incremental cost, Xm, "roulette wheel parent selection" technique [lo] is used to select the "best" parents according to their fitness. It consists is: of the following steps:

where A, and Aqs are the maximum and minimum values

aact -+anm q =A , ( -iVs -

of system incremental cost.

4.2 Generation output with ramp rate and prohibited zone constraints

1. Sum the fitness of all chromosomes in the population; call it the FITSUM. 2. Generate a random number, p, between 0 and FITSUM. 3. Return the fust chromosome whose fitness, added to the fitness of preceding chromosomes, is greater than or eaual to D.
Fig. 6. The roulette wheel parent selection algorithm.

Applying the methods of the LaGrange function and Kuhn-Tucker conditions [4] to the constrained optimization problem PBI, the ED problem can be reformulated as:

4.4 Parameter selection

subject to the power balance constraint. For a given A&, the generation output of each unit can be determined from (9). As mentioned in Section 2.4, in actual operation, the best economy is achieved by avoiding unit operation in the prohibited zone. If the generation output calculated ?om (9) is located in a prohibited zone, it needs heuristic adjustment to leave this area. If the dispatching hour in a load increasing period (the forecasted load of the next hour is greater than

2aZ4+bl =Aact for M a x ( 4 , e o -D&)Ie I Min(z,

eo+w)

2a,4 +bl SAact for 2a14+bl >Aact for

4 = M i n ( z , c o +U&)

(9)

4 = Max(4,4" -DR,) -

Like other stochastic methods, the GA has a number of parameters that must be selected. These include: size of population, probability of crossover, and probability of mutation. Usually, a relatively small size of population, high crossover probability, and low mutation probability is recommended. According to the authors' experiments, the foUowing values of parameters are appropriate far the ED problem:

o Size of population = 16 o Probability of crossover = 0.8 o Probability of mutation = 0.1

Fig. 7 shows the general flow chart of the proposed genetic approach for ED. The detailed evaluation process is shown in Fig. 8.

this hour), the adjusting point is the upper bound, P:, to follow the load fluctuation. On the contrary, when in a load
falling period, the adjusting point is the lower bound, P;.
4.3 Thefitnessfunction and parent selection

Initialize a population of chromosomes

Implementation of a problem in a genetic algorithm is realized within the fitness function. Since the proposed approach uses the equal incremental cost criterion as its basis, the constraint equation can be rewritten as:

c
Evaluate each chromosome (see Fig. 8 )

i
Rank chromosomes accordingto their Fitness

I I
I

I
Then, the converging rule is when E decreases to within a specified tolerance. In order to emphasize the "best" chromosomes and speed up convergence of the iteration procedure, fitness is normalized into the range between 0 and 1. The fitness function adopted is: 1 FIT = (1 1) & I +k(-)
PD

Select

-1

parents for reproduction

Evaluate new chromosomes and insert best into population displacing weaker chromosomes

.. .

output Fig. 7. General flow chart of the proposed approach.

where k is a scaling constant (k=200 in this study).

1923

I Decode

the chromosome to obtain

= A !]

o The loss formula coefficients are:

Translate

1-

a
i= 1
YS C

P to 2 9 using(8) '

Bu = 0.00001 75
0.000184

0.000136

0.0000175 0.000154 0.000283

0.000184 0.000283 0.00161

o hd = 300 MW a
Calculate. Pi using(9)
o Converging criterion : &

< 0.3MW (0.001 * h d a)

Succeeding iterations are shown in the following. Note that only the three best chromosomesin ranking are listed.
Rank
1 2 3 Chromosome 1000111010 1001010101 1000101011
An initial population AE(MW)

Fitness

I
L N o -

i=last unit 7
VS C

0.556641 0.583008 0.541992

3.4625 6.5155 9.7654

0.8125 0.6972 0.6057

Rank
1 2 3

The population after 1st iteration Chromosome am E(MW) 1001000000 1001001000 1000111010 0.562500 0.570313 0.556641 1.2371 1.7308 3.4625

Fitness 0.9241 0.8966 0.8125

Rank
2 3 Fig. 8. Detailed evaluation process.

The populationafter 2nd iteration (converging) E(MW) Fitness Chromosome A"" 1001000010 1001000000 0.564453 0.562500 0.4898 1.2317 0.9684 0.9241

The convergence is obtained in the second iteration. A

5. NUMERICAL RESULTS

summary of the results and the computation time are

presented below: Four example cases are studied in this section to illustrate o PI = 194.265 MW the performance of the proposed approach in practical Pz= 50.0 MW application. The software was written in Salford Fortran P3 = 79.627 MW language and executed on a 486-33 personal computer. (Pz does not locate in prohibited zone) Taking network losses, ramp rate limits, and prohibited zone avoidance into account only increases the solution time by o Pro, = 24.011 MW 5% because of GA's flexibility in handling constraints. o System incremental cost (aacf)=10.7028 $/MWhr Throughout the study, the lambda-iterationmethod is used as o CPU time : 0.016 s. the main benchmark of comparison for the proposed Example 2 approach.
Example I

In this example, a simple system with three thermal units [4] is used to demonstrate how the proposed approach works. The unit characteristics are given in Tables l(a) and l o ) .

Ir

Table 1@). Generating units ramp rate limits and prohibited zones. unit I I URiVh) ID R , ( M W ~I ) Prohibitednones(MW)
-*

1 2 3

215 72.0 98.0

55.0 55.0 45.0

95.0 78.0 64.0

[105,117][165,177] [50,601p2,1021 [25,32][60,67]

This example investigates the applicability of the proposed approach in mixed-generating cases. Three test runs were conducted in this example. Table 2 shows the test data (the complete 40 units data of the Taipower system are listed in the Appendix), and Table 3 shows the convergence behavior with reference to the proposed approach and the lambda-iteration method randomly executed 100 times, respectively. In Test 1, three oil units with similar coefficients of U,, b,, and c, are used for testing. Since the lambda-iteration method uses deterministic transition rules and the incremental cost rate of each unit is similar, it converges very rapidly in Test 1. However, the converging speed of the lambda-iteration method obviously slows down in the mixed-generating cases of Test 2 and Test 3.

1924

On the other hand, since the GA uses probabilistic transition rules to search for the optimum solution, the convergence speeds in these three tests are almost the same.

Example 4

Table 3. The convergencebehavior of randomly executing 100 times for each method. Proposed Approach I Lambda-Iteration Average I Maximum I Average I Maximum Iterations Iteraticins iterations Iterations 3.72 7 5 Test2 I 3.85 I Test3 I 4.12 I olambda-iterationmethod does not take prohibited zone into account.

I I

This example compares the solution cost, number of iterations, and solution time in different scale systems. Test data are the same as in Example 3. To obtain solutions under the same constraint conditions, the prohibited zones are released in this example. After randomly executing each method 100 times, the results are given in Fig. 10. It is obvious that the proposed approach can produce almost the same cost solution with the lambda-iteration method, and the solution time is less than the lambda-iteration method when the number of u i s is greater than 35. More significantly, nt since the number of iterations of the proposed approach increases negligibly with the number of units (see Fig. lo@)), the solution 'time of the proposed approach grows approximately linearly with problem size other than geometrically (see Fig. lO(c)). The salient feature makes the proposed GA approach attractive particularly in large-scale systems.

Example 3

_____
-AP

This example investigatesthe convergencebehavior of the proposed approach on the Taipower system. Five test runs were conducted for the test system. Table 4 shows the unit combination under test and the load level. For comparison, the test results obtained by a newly introduced unit-based GA method [SI, termed Method [8], are also included. A summary of test results are shown in Table 5. Since the proposed approach uses the same number of bits of chromosome to encode the normalized system incremental cost, irrespective of the system size, the number of iterations of the proposed approach increases negligibly with the number of units. By contrast, the number of iterations needed in method [8] increases dramatically with the increase in system size. This example shows the potential of the proposed approach in the use of GA for solving large-scale problems.
No. ofunits I 5 I
Table 4. Test data of Example 3 . Unit Combination 9.15.24.313 8

LambdaIteratian ropoicd Approach

10

20

30

40

N u m b e r o f units

(a). Average fuel cost vs. number of units

Iteration -*-Proposed Approach

io

.
0

I /

/-

io

20
N u m b e r of units

30

40

(b). Average number of iterations vs. number of units

Approach

I Load(Mw)
I
1100
4
0.04
0.02 0

10

20

30

40

N u m b e r o f units

(c). Average CPU time vs. number of units

Table 5 . The convergence behavior of randomly executing 100
Fig. 10. Solution cost, number of iterations, and solution time i n

comparison with the well-known lambda-iteration method.

6 CONCLUSION .

This paper suggests a different approach for solving the ED problem. In comparison with other ED methods, four differences--direct use of a coding, search of many optimum points in parallel, blindness to auxiliary information, and using probabilistic rules (rather than deterministic rules)--

1925

impart to the proposed approach a robust and global optimization algorithm. A salient feature of the proposed approach is that the solution time in solving ED problem increases approximately linearly with the number of units. This feature is attractive in large-scale problems. T i hs approach can also take network losses, ramp rate limits, and prohibited zone avoidance into account to make the dispatch more practical. Evaluation results based on the Taipower system show that the approach is faster than the well-known lambda-iteration method in large-scale systems.
ACKNOWLEDGMENT

[18] Lee, F.N., and A.M. Breipohl, "Reserve condrained economic dispatch with prohibited operating zones," IEEE Trans. on Power Systems,Vol. 8, No. 1, pp. 246-254, February 1993. [19] Syswerda, G., "Uniform crossover in genetic algorithms," Proceeding of the 3rd International Conference on Genetic Algorithms, pp. 2-9, June 1989.

BIOGRAPHIES

Po-Hung Chen was born on April 21, 1963, in

Taipei, Taiwan. He received his BSEE and MSEE degrees from the National Taiwan Institute of Technology (NTIT) and N t o a Tsinghua ainl University in 1986 and 1989, respectively. He has been with the S. John and S . Mary Institute of t t Technology since 1989, where he is an instructor in the EE Dept. He is currently a Ph.D. student in the EE Dept. of WIT. His research interests include the application of artificial intelligence to power systems operation and planning. Hong-Chan Chang was born in Taipei, Taiwan on March 5, 1959. He received his B. S , M. S , . . and Ph. D. degrees in electrical engineering from National Cheng Kung University in 1981, 1983, and 1987, respectively. In August of 1987, he joined National Taiwan Institute of Technology as a faculty member where he is presently an associate professor in the Electrical Engineering Department. His major areas of research include power system stability and neural network applications to power systems. APPENDIX. Taipower 40 units daia. unit
1

T i research was supported by the National Science hs Council of the Republic of China, under Grant No. NSC830404-EO11-028.
REFERENCES
Chowdhury, B.H. and S. Rahman, "A review of recent advances in economic dispatch," IEEE Trans. on Power Systems, Vol. 5, No. 4, pp. 1248-1259, NOV.1990. IEEE Committee Report, "Present practices in the economic operation of power systems," IEEE Trans. on Power Apparatus and Systems, Vol. PAS-90, 1768-1775, JUly/AUg. 1971. pp. Happ, H.H., "Optimal power dispatch a comprehensive survey," IEEE Trans. on Power Apparatus and Systems, Vol. PAS-96, pp. 841-854, May/June, 1971. Wood, AJ. and B.F. Wollenberg, Power Generahon, Operation and Control, John Wiley & Sons, New York,1984. Chen, C.L. and S.C. Wang, "Branch-and bound scheduling for thermal generating units," IEEE Trans. on Energy Conversion, Vol. 8, No. 2, pp. 184-189, June 1993. Lee, K.Y., al., "Fuel cost minimization for both real- and reactiveet power dispatches," IEE b C, Gener. Trsns. & Distr., 131, (3). pp. 85. 93,1984. Sheble, G.B. and K. Brittig, "Refined genetic algorithm-economic dispatch example," IEEE Paper 94 W M 199-0 PWRS, presented at the IEEE/PES 1994 Winter Meeting. Walters, D.C. and G.B. Sheble, "Genetic algorithm solution of economic dispatch with valve point loadiig," IEEE Trans. on Power Systems, Vol. 8, NO.3, p ~ 1325-1332, August 1993. . G l b r , D.E., Genehc Algorithms in Search, Ophmizahon, and odeg Machine Learning, Addison-Wesley, 1989. Davis, L., Handbook of Genehc Algorithms, Van Nostrand Reinhold, 1991. Nara, K ,et al., "Implementation of genetic algorithm for distribution . systems loss minimum re-configuration," IEEE Trans. on Power Systems, Vol. 7, No. 3, pp. 1044-1050, August 1992. Lansbeny, J.E. and L. Wozniak, "Adaptive hydrogenerator governor tuning with a genetic algorithm," IEEE Paper 93 WM 136-2 EC, presented at the IEEERES 1993 Winter Meeting. Y n X and N. Getmay, "Investigationson solving the load flow problem i, by genetic algorithm," EPSR, Vol. 22, pp. 151-163,1991. Gross, C.A., Power System Analysis, 2nd edition, John Wiley & Sons, New Y r , o k 1986. Wang C. and S.M. Shahidehpour, "Effects of ramp-rate limits on unit commitment and economic dispatch," IEEE Trans. on Power Systems, Vol. 8, NO. 3, pp. 1341-1350, August 1993. Shoults, R.R., et al., "A dynamic programming based method for developing dispatch curves when incremental heat rate curves are nonmonotonically increasing," IEEE Trans. on Power Systems,Vol. PWRS1, pp. 10-16, February 1986. Happ, H.H., et al., "Economic system operation considering valve throttling l s e , Part I & 11, NEE PAS, No. 64, pp. 609622,1963. oss"

pi - p i40 60 80 24 26 68
110

a,
003073 002028 000942 008482 009693 001142 000357 000492 000573
000605

bi
8336 70706 81817 69467 65595 80543 80323 6999 6602 12908 12986 12796 12501 8 8412 91575 91575 91575 79691 7955 79691 79691 66313 66313 66611 66611 71032 71032 33353 33353 33353 13052 2 887 1 10244 83707 26258 96956 71633
16339

ci
17044 30954 36903 13548 13519 22233 28771 391 98 45576 72282 6352 65469 9134 17604 17283 17283 17283 64785 64969 64783 64781 78596 785% 79453 79453 801 32 801 32 1055 1 10551 1055 1 12078 81079 12477 12192 64143 11128 10444
83224

U4
35 40 50 42 42 40 65 65 65 65 55 55 80 80 80 80 80 70 70 70 70 70 70 70 70 70 70

4
60 70 90 42 42 75 100 100
100

2 3 4 5 6 7 8 9
10

80 120
190

42 42
140

11 12 1 3 14 I5 1 6 17 1 8 1 9 20 2 1 22 23 24 25 26 27 28 29 30 3 1 32 33 34 35 36 37
38

135 135 130 94 94 125 125 125 125 125 220 220 242 242 254 254 254 254 254 254
10 10

10

20 20 20 20 18 1 8 20
2.3

300 300 300 300 375 375 500 500 500 500 500 500 500 550 550 550 550 550 550 550 550 150 150 150 70 70 70 70 60 60 60
60

000515 000569 000421 000752 000708 000708 000708 000313 000313 000313 000313 000298 000298 000284 000284 000277 000277 052124 052124 052124 025098 016766 02635 030575 018362 032563 033722
023915

100 95 95 120 120 120 120 120

110 110 110 110 110 110 110 110 110 110

90
90

90
70 70 70 70 60 60 60
60

150 150 150 70 70 70 70 60 60 60

60

39 40

25 25

60 60

023915 023915

16339 16339

83424 10352

60 60

60 60

o Subjectto l i t e d space, the 40 units' l s coefficients are not listed. os

1926

Discussion
J.T. MA and L.L. LAI, (Energy Systems Group, City University, London EClV OHB, England, UK) : The authors are to be congratulated for an interesting paper. It would be appreciated if the following points could be clarified:
1. The paper develops a lambda-based GA approach. However, the incremental cost lambda has to be differentiated from the input-output cost curve. The paper presents a prohibited operating zone to avoid the nondifferentiable area. Could this be always feasible? If so, the conventional optimization method can solve the ED problem more efficiently with the equal system lambda criterion. GA has the ability to solve nonsmooth, non-continuous and non-differentiable problem. It seems that the capability of GA has not been fully made use of.

What is the relationship of the fitness function in Equation (11) with the cost function in Equation (l)?
2.
3 . In the paper, the population size is fixed at 16. Is this selection appropriate for any system size?

4. In the last paragraph of Section 4.2, the paper talks about the heuristic adjustment to avoid the prohibited zone. It seems to us that changing the generator output is not based on the result during optimization procedure but on the load fluctuation. Does this mean that the procedure is an hourlychanging, dynamic procedure? Even in the dynamic procedure, the generator output is still determined by the optimization result in ED problem, even though it indirectly reflects the hourly load change.

Manuscript received February 21, 1995.

P. H. Chcn and H. C. Chang: We would like to thank the discussers for their insightful questions rose in the discussion. We will reply to their questions in order: 1.The input-output curve is espressed as a quadratic polynomial with prohibited zones in this paper. Since the shape of the curve in the non-differentiable prohibited area is dificult to determine by actual performance testing or operating records, the prohibited zones problem is simplified by avoiding operation in these zones. This is feasible in actual operation. Indeed, the conventional optimization method can also solve the ED problem. However, for a large-scale, mixed-generating

system such as the Taipower system where coal-fired, oilfired, gas-fired, gas turbine, diesel, combined cycle, and nuclear units are all present, the conventional method has oscillatory problem which results in a slower solution time. GA is a powerful technique with many advantages. The proposed GA-based approach is faster than the conventional method in the study system when the number of committed units is greater than 35. The salient feature of the method is that the solution time grows approximately linearly with problem size other than esponentially. This feature is attractive in large-scale problems. The authors agree with the discussers' opinion that the capability of GA has not been fully made use of in this paper, particularly in dealing with the prohibited zone problems. Further researches along this direction are undergoing. 2. Using the equal system lambda criterion, the minimal solution cost of the ED problem, as shown in equation (l), can be obtained, when the fitness value of equation (1 1) is equal to 1. This relationship could be easily checked from Exairiple I . 3. The population size must be determined espcri~nentally, depending significantly on the problem being solvcd. There is no single value suitable for different systems. Usually, a relatively small populatioiz size is recommended [ 1,2]. The authors tried the population size from 10 to 20 and found that "16" has the fastest convergence speed for the ED problem on the Taipower system. 1. The proposed approach is dewloped for solving the ED problem in thermal units scheduling environment. The heuristic adjustment following the load fluctuation is for the purpose of restraining the units from moving through the prohibited zones too frequently. If we adjust the operation point of a unit based only on the optimization procedure, it may oscillate between the upper bound and the lower bound of a prohibited zone. This is unwished in the actual operation. Again, the authors would like to thank Dr. Ma and Dr. Lai for their genuine interest in this paper. References
D.E., Genetic Algoritlinrs i i i Search, Optimization, and Macliine Learniiig, Addison-Wesley, 1989. [2]Davis. L , Haidbook of Genetic A I ~ ~ I . I ~ I ~ I Van Ns,
[ 11 Goldberg,

Nostrand Reinhold, 1931.

Manuscript received April 17, 1995.