You are on page 1of 7


JEEE Transactions on Power Systems, Vol. 11, No. 1, February 1996

Hong-Tzer Yang Member, IEEE Pai-Chum Yang Ching-Lien Huang Member, IEEE

Department o f Electrical Engineering National Cheng Kung University Tainan, 701 TAIWAN
Abstract - This paper develops an efficient, general economic dispatch (ED) algorithm for units with non-smooth fuel cost functions. Based on the evolutionary programming (EP) technique, the new algorithm is capable of determining the global or near global optimal dispatch solutions in the cases where the classical Lagrange based algorithms cease to be applicable. Effectiveness of the new algorithm is demonstrated on two example systems and compared to that of the dynamic programming, simulated annealing, and genetic algorithms. Practical application of the developed algorithm is additionally verified on the Taiwan power (Taipower) system. Numerical results show that the proposed EP based ED algorithm can provide accurate dispatch solutions within reasonable time for any type of fuel cost functions. Keywords: Economic Dispatch, Evolutionary Programming, Combinatorial Optimization

Economic dispatch (ED) is an important daily optimization task in the power system operation. To solve effectively the ED problem, most calculus based industrial algorithms, e.g., the Lagrangian multiplier method, require the incremental cost curves to be of monotonically increasing nature or of piece-wise linearity. For the generating units virtually having non-monotonically increasing or non-linear incremental cost curves, the conventional procedure either ignores or flattens out the portions of the incremental cost curve that are not continuous or monotonically increasing. Inaccurate dispatch results are thus induced. In these cases, to obtain accurate dispatch results, needed are more general approaches without restrictions on the shape of fuel cost functions. Dynamic programming (DP) solution is one of the approaches to solving the inherently non-linear and discontinuous ED problem [1,2]. However, the problems of "curse of

dimensionality" [ l ] or local optimality [2] with the DP method are caused in the solution procedure. With the advent of stochastic search algorithms, the simulated annealing (SA) [3] and the genetic algorithms (GA) [4] were devoted to solving the highly non-linear ED problem without restrictions to the shape of fuel cost functions. Nevertheless, the SA based algorithm is difficult to tune the related control parameters of the annealing schedule and may be too slow when applied to a practical power system. On the other hand, the operations of reproduction and crossover in the GA, which are based on competitions within a single generation of population and based on probabilistic exchange of segments of encoded binary strings, may be detrimental to actually achieving an optimal solution [5]. Besides, the GA requires an encoding scheme on the decision variables to conduct its genetic operations. Different encoding schemes have significant influences on convergence of the GA. Encoding and decoding for each solution searched and the operations of crossover and mutation on binarycoded variables also waste much computer time. These problems degrade the effectiveness of the GA in solving the ED problem. In this paper, a novel evolutionary programming (EP) algorithm [5,6] is proposed to solve the ED problem for units with non-smooth fuel cost units. The stochastic mechanics, which combine offspring creation based on the performance of current trial solutions and competition & selection based on successive generations, form a considerably robust scheme for large-scale real-valued combinatorial optimization. The weaknesses of the algorithms mentioned above are circumvented. The proposed EP approach is capable of not only solving the ED with any type of fuel cost functions, analytical or empirical curves, but also obtaining the global or
near global minimum solution considering transmission

95 WM 167-7 PWRS A paper recommended and approved by t h e IEEE Power System Engineering Committee of the IEEE Power Engineering Society f o r presentation a t t h e 1995 IEEE/PES Winter Meeting, January 29, t o February 2 , 1995, New York, NY. Manuscript submitted J u l y 29, 1994; made a v a i l a b l e f o r p r i n t i n g November 23, 1994.

losses or not within reasonable execution time. Encoding and decoding schemes essential in the GA approach are not needed; considerable computation time can thus be saved. To validate our proposed approach, the EP algorithm is tested on three test systems including one practical power system. Numerical results obtained from the proposed EP, the DP [2], the SA [3], and the GA [4] metliods are compared. The practical system is used to verify that the proposed EP approach is applicable to a physical-sized power system ED problem.

0885-8950/96/$05.00 0 1995 IEEE

113 Nomenclature fuel cost function of the j t h unit. score of p. power generated by the combined-cycle unit. . fuel cost coefficients with valve-point effects.7]. a power generation system usually uses combined-cycle unit as peak load serving unit. due to wire drawing as each steam admission valve starting to open.) = aj + bjPgJ + cjPjJ + The generating units with multi-valve steam turbines exhibit a greater variation in the fuel cost functions. Example input-output curve with five valve points Figure 3 . C. scaling factor. thejth unit minimum power output. Typically. the ith individual of a population.. capacity of heat-recovery steam generator. 3. ECONOMIC DISPATCH FORMULATION The ED problem studied in this paper can be described as an optimization process with the following objective function and constraints: N Min Pq j=1 FC.(Pgj) s. The ED with combinedcycle units becomes a mixed continuous-discrete optimization problem. Before describing the solution procedure using the EP. a unified approach using the EP technique is proposed to solve the above ED problem without any restriction to the fie1 cost characteristics. Example net heat rate curve of a combined cycle unit with three gas turbines 11. the valve-point effects.. number of combustion turbines in the combined cycle <PgJ5 P g J m f o r j = l . PgJ. number of individuals in a population. thej t h unit maximum power output. 1. system load demand. transmission 1os s. B. The general scheme of the EP algorithm for optimization is summarized as follows [5. E Operating points of admission valves Figure 1 . maximum number of generation. capacity of one combustion turbine. the operation of the combined-cycle unit is performed on a zero-one discrete basis. Since the combustion unit is most economic when loaded with full capacity once it is placed on line. power balance constraint: N D=C. Schematic diagram of the evolutionary programming algorithm . produce ripple like heat rate curve [4. power generated by the jth unit. In this paper. EVOLUTIONARY PROGRAMMING ALGORITHM The schematbc diagram of the EP algorithm is depicted in Fig..t. . Vector Representation of the Decision Vanables Competing pool output (Mw) A. the general schemes of the EP are briefed in the subsequent section. N (31 The fuel cost functions of the generating units are generally characterized by third-order polynomials: (4) FCj(P. number of committed thermal generating units. illustrated in Fig. The actual unit curve data therefore can not be represented by the polynomial function as in (4). for economic operation the combustion turbine should be either filly loaded or completely turned down [8] in the ED. Figure 2 shows the net heat-rate curve of a combined-cycle unit with three gas turbines. output W) Figure 2:. fuel cost coefficients of the jth unit. Besides.PL F1 (2) and generating capability constraints: Pg. 111. which comprises a series of single-cycle gas turbines in conjunction with some heat-recovery steam generators (HRSGs). Hence. after competition. power generated by the dependent unit. objective function value ofp.6].

3 ) Competition & Selection: Each'individual in the competing pool must stochastically strive against other members The Np of the pool based on the functions @ i ) and&:). Cmation ofofffpring An offspring vector p: is created from each parent p . a Gaussian random variable with zero mean and a standard deviation proportional to the scaled cost values of the parent trial solution. unit. generated by adding a Gaussian random variable with zero mean and preselected standard deviation to each component of p.. . the offspring trial solution is created near the current from among the committed N units except for the combined solutionp.iffpr is relativein (2).P.> PLNI P. ...As gven in (l). c A.. . i = 1. 4) Stopping Rule: The process of generating new trials and selecting those with best function values are continued until the function values are not obviously improved or the given count of total generations is reached... j is not a combined-cycle unit where N(p. The standard deviation 5j indicates the range the offspring is created around the parent trial solution. PgJ.114 0 ) Vector Representation: The real-valued variabks to be determined in the EP algoritlp are represented as a trial n-dimensioned vectorp that is associated with the extremum of an objective functional f @ ) : R"-+ R. i =1.fp..= Pg. by adding to each component of p. In the example system of case 1 demonstrated NP wp.. N.i = 1.") (8) f o r j =1. The score for ized to calculate the power loss from the outputs of generateach trial vector after a stochastic competition is given by ing units [9]. the B-loss matrix is used.belonging to (1. [XI denotes the greatest integer less than or equal to x.. The individuals in the survivor set are new parents for the next generation. . then Pgd is calculated by : For the combined cycle units... is straints in (3).t 1 and N. and uI.... .. . an iterative algorithm or The Np parent trial vectors p. EVOLUTIONARY PROGRAMMING BASED ECONOMIC DISPATCH utions. l}. if subsections.. i =1.... to meet exactly the load demand created according to the relative value offpi . . P. cally given a discrete number n..Np ... Initialization The initial parent trial vector p l .0.. The power generation of the combined-cycle unit P.).-. PgN] be a trial vector desigN nating the ith individual of a population to be evolved. . the next trial solution will cycle units. 11. i. is selected randomly from a feasible range in each dimension. I ) Initialization: An initial population of parent individuals p?.O...P..p ~ . is determined by setting its jth components PgJ-U(PgJ. = [Pil> PG2> .. 0. The offspring p . is the capacity of HRSG..= n. x P . . The elements of p . individuals including parents and offspring exist in the competing pool (Fig. inclusive to avoid the combined cycle unit off-line.. 2) Creation o f Offspring: Equal number of offspring p: . (7) P. are uniform random number ranging over [0. corresponding offspring p . is the capacity of one combustion turbine. U. o1is given according to the following equation: (9) where frmn is the minimum cost value among the Np trial solI V .. Competition & Selection ent unit Pgd and the power loss PL .. Let Pgd be the power output of the dependent be searched within a wider range. . +Ps (6) ' where P. to determine the output of dependC.. N p and their loss penalty factors from power flow program are required. 3).") represents a Gaussian random variable with mean p and standard deviation ol. =0.Np contend for survive More directly. Therefore. In this paper.. is to be minimized.The combined-cycle unit is stochastiover fPgJm. Typically. B-loss matrix of the power system can be util. Since the power loss is a function of the generation outputs and system topology. i = 1. N . which are subjected to the capacity con. and j is not of a combined cycle unit.with each other within the competing the value of the objective function associated with the trial vectorp.e.> PiJ> . . a dependent generating unit is arbitrarily selected ly low.i. . V. N.... fp.Pgz. are the real power outputs of the committed N generating units. N.} that indicates thc number of combustion turbines fully loaded in the combined-cycle unit.) denotes a uniform random variable ranging PgJm]. the distribution of initial trials is uniform. but is limited to between Il. .. P is a scaling factor. Let p I = [Pgl. otherwise where the competitor p. B. The EP method implemt=l ented to solve the ED problem is stated in the following wt= 1. N.2.. . U( then obtained by Pg. (5) f o d y distributed among ( selected at random from among the 2Nptrial solutions based on r = [2Npuz + 11.= wi in Sec.e. Each vectorp is an individual of the population to be evolvqd. +N(O.... where j=1. the number of combustion N turbines fully loaded ricin (6) is added a random number uniPgd = D 4 . individuals with the best function values (minimum for the minimization problem) are selected to form a survivor set according to a decision rule... .iffpl is relatively high. 2..

The search process is stopped as the count of generations reaches N. The following section illustrates the results of the proposed EP approach to solving the ED problem. one of which is the global minimum solution.657 100.448 62. Since the cost function is directly employed as the performance index of the solution rather than the incremental costs (derivative of the cost functions). the SA [3]. Further tuning the scaling factor of the cost ($) 6644 T 6642 1 1 0 + ++ + +I I I I I I 6640 - 6638 I I I I I 20 40 60 80 100 Number of run Figure 4. generation number ( N .546 376. to be set before starting the evolutionary process. the practical Taiwan power (Taipower) thermal system in year 1987 had 40 generating units with 3 combined-cycle units. the ED problem is solved by using the EP process without concerning the form of the objective function. The fuel cost funcvalve-point effects described in Sec. iterations) N. 1 tions of the generating units were expressed as FC.4 Unit 3 Total Gen.273 1. Any shape of fuel cost functions of units can be dealt with in the EP search process. Obviously.434 43. Each combined-cycle unit consisted of 3 combustion turbines and 1 HRSG unit. where coefficients of fuel cost function in (1 1) can be found. The associated incremental costs are non-monotonically increasing.769 62. are ranked in descending order of the score obtained in (IO). A maximum number of generations (is. the test system is composed of three generating units whereas the fuel cost functions take into account the 1 . the 2Nptrial solutions. the solution space of the problem has two local optima. APPLICATION EXAMPLE A. SA. including the parents and the offspring.1 30 50 0. TABLE 1 BESTPARAMETER SETTINGOF THE EP Parameter Case 1 Case2 Case3 Scaling factor (p) Population size (Np) 0.152 1. The EP algorithm has 3 parameters.639. Fuel cost curves of the units were represented by thirdorder polynomial functions. ) TABLE 2 DISPATCH b S U L T X FROM EP.8 1. AND DP IN CASE 1 FOR LOAD OF 1400 M W Methods Unit 1 Unit 2 (Mw) DP SA' SAz (Mw) 406.01 50 1.000 scaling factory = 80.734 677.504 997.434 6. is given. + e.462. = FC. Total capacities from 6345 MW to 9300 MW were committed to meet the 24-hour load demands from 5320 MW to 8708 MW.2 359. as described in Sec.1 30 30 0. control parameters of the EP can be easily tuned to obtain satisfactory results. Fuel cost functions of the generation units and the associated B-matrix for transmission loss calculation can be found in [2]. unlike the annealing schedule of the SA algorithm [3]. The test system in case 1 comprises three generating units. only the local optimal solution of $6642. IV. and the DP approach with a zoom feature [2] are summarized in Table 2. As described above.921 1..4 6.443. The fuel cost curves of the units were approximated by second-order polynomial functions such that the classical Lagrangian multiplier algorithm (h-iteration method) [9] can be used to verify the developed EP program.639.052 986.trial solutions survive and are transcribed along with their objective functions fp. In Case 2. the h-iteration method was moldified to enumerate all combinations of the combustion turbines in the combined cycle unit. Testing Resulis Case 1 Dispatch results obtained from the proposed EP.462. Description of the Test Systems The proposed EP approach was numerically tested on three cases of ED problem.01 scaling factory = IO. the differentiability of cost function and the monotonically increasing property of the incremental cost curves are not required any more as in classical Lagrange based algorithms.459 406. Results of the EP were compared with those from the DP with a zoom feature [2] and the SA approach [3] reported recently. After testing and evaduating different parameter combinations.000 Max. All the programs were developed using Turbo C language and run on PC-4186 computer.115 After competing. into the survivor set as the basis of the next generation.459 can be found by the DP and the SA approaches (as shown in Table 2).848 100. I I B.181 ' EP 364. Multiple local minimum points therefore exist in the objective function of the system. 1 V. As noted in Table 1. Results of these three testing cases are given below.448 6.4 43.642. The first N.123 676.769 6. Considering the nature of combined-cycle unit in practical operation. In Case 3. However..642. x sin(5 x (pglrnm -p g l ) ) (11) The same problem has been solved by the GA approach in [4]. parameters of the EP used in the three cases are listed in Table 1 for clarity. Cost distribution of final solutions with 100 different initial trials in Case 1 . Losses Total cost (MW) (Mw) ( M Y ($1 360.443.

234.01 to 0. . the same test showed that only 14% (when y= 0.000172 Oh00120 * Combined-cycle units were committed.0 400.02 8.002937 21 22 23 Err % Hour h-iteration 0.795 150. Different random initial trial solutions were further used to explore the converging characteristics of the EP.0 850.00020 16' 6162956 6162966 0.000197 0. 11* 12' h-iteration EP 7151778 7151778 0. in the GA.1 149.22 8234. 4.222.4 399. exact balance of power generation and load demand can be satisfied by the EP.8 (MW) 849.). I).4 400.005316 0.234. different initial trial solutions were given to the EP approach.234.4 850. Table 3 shows the final dispatch results from the proposed EP and the GA approaches. compared to 0.000495 0.001744 0.000426 0.000345 15' 6020843 6020855 0.09 sec.0 850. about 95% of the solutions from the EP are located at the global optimum.1 300. The CPU time used by the EP program for this case averaged 0.4 299.0 400. Cost distribution of final solutions with 100 different initial trials in Case 2 'different runs of GA in [4]. To investigate effects of the initial trial solutions on the final results.002848 0.07 8. Case 2\ In this case.001918 0.00 (MW) GA' GA' GA' GA' 498.2 850. Apparently. Figure 5 exhibits the final results when 100 random initial solutions were given.041 150. used by the SA approach. In contrast. As noted in Table 3.234.0 400.1) of the solutions found by the SA approach were the optimal with the rest of the solutions entrapped in the local minimum solution.11 8.07 8. the EP did not encode and decode the decision variables and did not resort to crossover and mutation operations on the binary-coded variables to perform the evolutionary computing.000042 0. In the meantime.0 399.00130 14' 6372545 6372567 0.000157 0.239.225. one can obtain the global minimum solution of $6639.205 299.12 sec.0 EP' EP' EPZ EP2 *3 20 40 60 80 Number of run Figure 5.000184 0. 3optimal solution used for verification in [4].133 150.234.001197 8911330 8911344 9585475 9585479 9387110 9387150 8719088 8719103 8327709 8327719 8202102 8202102 0. Although multiple local minimum solutions exkt in this case. the results obtained revealed that the power generation can not meet the load demand.736 149. Total cost 100.867 300.75 1 + 0 + +I 100 50. the proposed EP approach has more powerful ability to achieve the global optimum solution.29 8.01) to 53% (when y= 0.116 TABLE 3 DISPATCH RESULTS FROM EP AND GA IN CASE 2 FOR LOAD O f 850 Methods Unit 1 Unit 2 cost ($) 8235. SA to enlarge the search neighbourhood of the current solution (changing y from 0. the encoding and decoding schemes were required for the power generations to make feasible the genetic operators.30 8.1 398. Since the power balance constraints of the ED problem were handled in conjunction with the fuel cost as the objective function to be minimized in the GA approach [4]. Videlicet.0 Unit 3 (MW) Total Gen.92 8. where results from the better encoding scheme of the GA in [4] were used for comparison.24 8.9 848.000046 0.0 (MW 250.264 300. Also. 'different runs of EP.7 149.3 848. However.0 20 8711369 10706244 11066826 10860027 8711373 10706297 11067019 10860346 0.) is far less than that by the GA (10.6 sec. much less execution time (about one seventh) was needed for the proposed EP approach.959 299. A lot of computer memory and computing time can thus be saved.0 ($1 8.7 392. the proposed EP approach can easily achieve the global optimal solution using the parameter setting given in Table 1.389.000162 17' 5994532 5994543 6045002 6045009 6674603 6674603 0.003670 0. The CPU time used by the EP (0.6 307.0001 16 18 13 19 24 9141785 9141803 11269335 11586870 10998487 10572727 9357091 11269656 11587486 10998698 10573115 9357203 0.0 850. the fuel cost characteristics considering the valve-point effects were employed to test the performance of the proposed EP approach when compared to the GA approach. the EP can still find out the optimal or TABLE 4 OPnMAL COSTS (NT$$)FROM EP AND h-ITERATION h"0DS M CASE 3 FOR 24-HOUR &ONOMIC DISPATCH 7 8 9 Hour 1 2 3 4 5 6 10.0 6616323 6616409 0. As shown in Fig. Figure 4 shows the dispatch results of the EP when run 100 times with different random initial solutions.0 850.5 150.0 sec.504 with higher probability.0 400. while different encoding schemes resulted in different performance of GA.0 EP E r r % 0.

Jan. 1984. pp. of CPU time for the 40-unit. 1992. 1991. Vol. University. Tran. in 1973. PWRS-8. VIII. Though the practical application of the EP to solve the ED problem is verified in a power system with monotonically increasing incremental cost curves. PAS-90. When actual fuel cost models of units without any approximation are considered. the EP approach totally consumed about 30 min. No. PWRS-7. Taiwan. 509-515. Hsin-Chu. respectively. and power system planning. 93 WM 157-8 PWRS. the optimal ED solutions from the EP approach can be used as the benchmark to off-line evaluate any approximated ED approach. Kim. "A zoom feature for a dynamic programming solution to economic dispatch including transmission losses. Japan. VI. D. ' An evolutionary programming approach has been proposed and demonstrated to be a powerful optimization process in solving the highly non-linear and discontinuous economic dispatch problem where the classic Lagrange based algorithms can not be directly applied.B. Tainan. Also.J. 24-hour ED problem with an average of 75. National Cheng-Kung University. The proposed EP approach was compared to the related methods intended to serve this purpose. degree in electncal engineering from National Cheng-Kung. VII.Q. the solutions yielded lby the EP method would be inevitable references.D. 2060-2088. This is one of the significant contributions possessed by the proposed EP approach to solving the ED problem. Lianf and J." IEEE Trans.No.3 sec. John Willey. are all near the optimal solutions with a maximum error of 0. power system switching surge and protection. University. Z. for one-hour ED. respectively.S. where he is now a professor. degree from National Taiwan Institute of Technology.D. OH. the proposed EP approach can be applied to the difficult dispatch problem with any shape of fuel cost functions to which the classical Lagrange based method is not amenable. July/Alugust 1971. 2. Ross and S. on Power Apparatus and Systems.B. Vol. He received his Ph. D. 1980. Case 3 This case demonstrates the proposed EP approach is applicable to solve the practical ED problem. Taiwan. Vol.S.S.W. degrees in electrical engineering from National Cheng-Kung University. intellectual schemes of encoding and decoding entailed by the GA approach are not needed in the proposed EP approach. Taiwan in 1982 and 1984. No. New York. Fogel." IEEE Trans. pp. in 1957.X. degree from Osaka University. the SA. "Simulated annealing based economic dispatch algorithm. on Power Systems. on Power Apparatus and Systems. Dispatch results obtained from the EP and the h-iteration methods are shown in Table 4 for 24-hour load demands. Glover. Sheble. in 1986 and 1988. 1352-1331." IEEE Trans. PAS-99. pp. "A ctombinatorid optimization based economic constrained dispatch with raise and lower control reserve constraint and combinedcycle unit discrete optimum loading.P.S." IEEE Trans. Ching-Lien Huang received B. Wong and C. Walters and G. most of which are even smaller than the 'optimal' solution of $8234. Yang is a member of Phi Tau Phi.07 and $8234. 1993. 5. on Power Systems. Wood and B.F. IEEE Committee Report.117 near-optimal solutions that distribute between $8234. A. ACKNOWLEDGEMENTS Financial support from the National Science Council for the work reported in this paper is gratefully acknowledged. disadvantages of huge memory size required by the DP and long execution time required by the SA method are ameliorated. In the practical ED problem. 6. Since 1964. He is now a technical superintendentof Chung Shan Institute of Science and Technology. and M. Moreover. Operation and Control.S. can be dealt with by the proposed EP . K. Power Generation.S. Pai-Chuan Yang received the B. the solutions obtained by the hiteration method should be optimal solutions.E. Tainan. 1993. such as the DP with a zoom feature. His research interests are on the production simulation and unit commitment of power systems.D. His major research interests are on high voltage engineenng. 140. He is currently working toward his Ph. 3.pp. Needham.0053 16% when compared to those of the Literation method. 1994.73. Vol. on NarralNetworks. Osaka. 1768-1775. 1993. and the GA approaches.S. Sysi'em Identification through Simulated Evolution: A Machine Learning Approach to Modelling. any form of fuel cost functions. degree in electrical engineering from National Tsing-€ha University. U. degree from National TsingHua University.B. D. and the M. "Genetic algorithm solution of economic dispatch with valve point loading. Hong-Tzer Yang received the B. smooth or not. No. pp. "Dynamic economic dispatch of generation.E. Since the fuel cost functions were simplified and represented by second-order polynomials on purpose.E. Feb. degree at National Cheng-Kung University. The solutions achieved by the EP. Tainan.A. Hsin-Chu.E. Wollenberg." presented at the IEEEPES 1993 Winter Meeting. The problem of power unbalance previously existing in the solution of GA is circumvented as well in this paper. 544-550. MA. In view of evaluating the approximated ED approach which simplifies the unit fuel cost models.C. 3-14. Dr. His present research interests are neural networks and expert system applications in power systems. as shown in Table 4. Fung. and M. Taipei. Taiwan in 1989. Ginn Press. the EP approach avoids entrapping in local optimal solutions as in the DP with a zoom feature and the SA methods. In this case. 1. Paper No. Vol." IEEE Trans. pp. 6. He is also an associate professor of Electrical Engineering Department at National Cheng-Kung. "Present practices in the economic operation of power system. "An introduction to simulated evolutionary optimization.C. By means of stochastically searching multiple points at one time and considering trial solutions of successive generations.22 obtained by a verification program using dynamic programming [4]. and the IEEE PES and CSS. D.E. No. Fogel.E. either in the representation of unit fuel cost models or in the heuristic optimization process. CONCLUSIONS approach. Vol. he has been with the Department of Electrical Engineering. REFERENCES D." IEE Proceedings-C.

generation Number of encoded binary bits GA 30 50 100 100 11 0 95 0. P. Do the authors use the same size of population for both EP and GA in the comparison to get the conclusion that EP is better than GA in CPU execution time? Could the fitness computation based on B-loss-matrix need less computer time than that for the encoding and decoding? population size. London E C l V OHB. EP is better than GA in searching f o r a global optimum but not in CPU execution time. such as SA. The one-hour ED execution time of EP in Case 3 is only 75. GA and lambda iteration. Tahle D1 COMPARISON OF EP AND GA APPROACHZS IN TEE CASE 2 OF OUR PAPER 2. due to lack of e e Scaling fictor 01 10 Execution time (sec) 0. The fuel cost functions were simplified and represented by second-order polynomials on purpose that the h-iteration method can be used to achieve the optimal solution for comparison to that of the proposed EP approach. while GA computes fitnesses only for one population size at each generation.. City University. 1995. England. Manuscript received April 17. the optimal solution of the EP approach can be used as the benchmark to off-line evaluate any approximated ED approach. .118 Discussion J. as mentioned in our paper. based on competitions within a single generation of population and based on probabilistic exchange of segments of encoded binary strings. However. the time is much shortel.12 encoding and decoding the decision variables. 3 . because EP has to compute fitnesses for individuals of twice the population size. 1. Obviously. (Energy Systems Group. Besides. and other control parameter settings of both EP and GA approaches in the Case 2. the problem to which the classical Lagrange based algorithms (e.g. The averaged time for one-hour ED execution time when using the h-iteration method is 0. the proposed EP approach is focused on the ED problem with non-smooth fuel cost functions. The authors present a comparison between results using EP and other algorithms. Comparison of execution time. 1995. and C. Yang.01 Manuscript received February 21. However. Case 3 of our paper aimed to demonstrate that the proposed EP approach is capable of solvlng the practical ED problem.C.than that needed by the EP approach. Yang. According to our experience. - Crossoverprobability Mutation probability N. As pointed out in our paper. Huang: We are grateful to the discussersfor their interest in ow paper. the h-iteration method) cease to be applicable.T. if the power losses are to be taken into account. number of maxiiiium generation.T. The questions of the discussers are clarified as follows.4 sec. the operations of reproduction and crossover in the GA niay be detrimental to actually achieving an optimal solution. either in the representation of unit fuel cost models or in the heuristic optimnization process. It is noted that the power loss calculations (B-loss matrix) were not considered in Case 2 where the results of EP approach were compared to those of the GA. Nevertheless. the EP approach indeed has to compute fitness functions for twice the population size. LAI.MA and L.3 seconds? What is the time when using the lambda iteration method? EP Population size Number of max.L. the EP approach needs less total execution time than does the GA method. 2. UK): The authors are to be congratulated for an interesting paper. the computer time for the fitness computation based on B-loss matrix or other schemes is additionally needed for both EP and GA approaches. In terms of the number of individuals evaluated per generation. It would be appreciated if the following points could be clarified: 1. is further summarized in Table D1.L. and crossover and mutation operations on encoded binary bits according to a uniform-distributed random number generator.