You are on page 1of 5

Improved Fast Evolutionary Program for Economic Load

Dispatch with Non-smooth Cost Curves

N Sinha, Member
Prof (Dr) R Chakrabarti, Fellow
Prof P K Chattopadhyay, Fellow

This paper presents an improved fast evolutionary programming (IFEP) technique for solution of economic load
dispatch problem with generating units having non-smooth fuel cost curves. The superiority of the method over classical
evolutionary programming (CEP) method has been demonstrated on two test cases. Both the algorithms are
implemented in MATLAB 5.2 command line. It has been observed that this improved fast evolutionary program
outperforms the classical evolutionary program in solving power system economic load dispatch problems of smaller as
well as larger system in terms of execution time, convergence rate, quality and success rate.

Keywords: Classical evolutionary programming; Fast evolutionary programming; Self-adaption; Economic load dispatch;
Gaussian mutation; Cauchy Mutation

INTRODUCTION In this respect stochastic search algorithms like genetic algo-


The fundamental requirement of power system economic load rithm (GA)2,3,4 evolutionary strategy (ES)5,6 evolutionary
dispatch is to generate, at the possible lowest cost adequate programming (EP)5-14 and simulated annealing (SA)15 may
quantity of electricity to meet the demand. To meet the prove to be very efficient in solving highly nonlinear ELD
stringent quality requirements accurate tools based realistic problems without any restrictions on the shape of the cost
models with faster solution speed and a high degree of reliabili- curves. Although heuristic methods to not always guarantee
ty is required. To achieve higher reliability, improved the globally optimal solution, they provide a reasonable
security, less environmental impact the utilities are solution (suboptimal near globally optimal) in a short period
implementing tighter control on operation of their facilities. of time. Recent research endeavors, therefore, have been
These have brought about the necessity of greater directed towards application of these techniques. These
sophistication in power system planning, operation and evolutionary algorithms (EAs) are search algorithms based on
control. the simulated evolutionary process of natural selection and
Traditional classical dispatch algorithms, (eg, the Lagrangian genetics. EAs are more flexible and robust than conventional
multiplier method) require the incremental cost curves to be methods. Recently, a lot of interests has been shown towards
monotonically increasing or piece-wise linear. The input- developments and applications of evolutionary programming
output characteristics of modern units are inherently highly techniques to a variety of power system problems, like: (i) unit
nonlinear (with valve point loadings, rate limits etc) and commitment10; (ii) generation scheduling11; (iii) reactive
having multiple local minimum points in the cost function. power planning12; and (iv) optimal power flow13 etc.
However, their characteristics are approximated to meet the Since its first application to prediction tusks more than three
requirements of classical dispatch algorithms leading to decades ago, there have been a number of developments in
suboptimal solutions and therefore, resulting in huge revenue evolutionary programming techniques in view of speed,
loss over the time. Consideration of highly nonlinear robustness and quality. In EP, mutation is the only operator
characteristics of the units demands for highly robust used to generate new offspring. The mutation is often
algorithms to avoid getting stuck at local optima. The classical implemented by adding a random number from a certain
calculus based techniques fail in solving these types problems. distribution viz Gaussian distribution in the case of classical
Unlike the traditional algorithms, dynamic programming EP (CEP) and Cauchy distribution in case of fast EP (FEP) to
(DP)1 imposes no restrictions on the nature of the cost curves the parent. Recently, Xin Yau, et al 7 have proposed an
and hence solve inherently nonlinear and discontinuous ELD improved FEP (IFEP) that uses both the Gaussian and Cauchy
problems. However, this method suffers from the ‘curse of mutations for creation of offsprings from the same parent and
dimensionality’ or local optimality. better ones are chosen as offsprings for the next generation.
N Sinha, Prof (Dr) R Chakrabarti and Prof P K Chattopadhyay are with They have demonstrated the superiority of this method over
Department of Electrical Engineer, Jadavpur University, Kolkata 700 032. other EP methods on a number of benchmark functions. They
This paper was received on October 29, 2001. Written discussion on this paper key factor behind the success of IFEP appears to be that it uses
will be received until November 30, 2004. both smaller (Gaussian mutation) as well as larger (Cauchy

110 IE (I) Journal—EL


mutation) step sizes towards the creation of new offsprings. In where a j , b j , c j are the fuel cost coefficients of the jth unit and
highly non-linear systems, it is difficult to constrain the search
e j and f j are the fuel cost coefficients of the jth unit with valve
space to a sufficiently small region when there is little
knowledge about the global optimum. As no unique step size point effects.
is likely to give better results for all types of functions. IFEP The generating units with multi-valve steam turbines exhibit a
appears to be more promising simply because it uses both greater variation in the fuel cost functions. The valve-point
smaller as well as larger step sizes. effects introduce ripples in the heat rate curves.
Though there are some reports9 on the performances of CEP
using scaled cost based adaptation of strategy parameters in EP BASED ELD WITH ADAPTATION OF
solving power system ELD, the performance IFEP with the STRATEGY PARAMETER USING SCALED COST
same type adaptation is yet to be tested. Let pi = [ P1 , P2 ,....., Pn ] be a trial vector denoting the
In view of the above, the main objectives of the present work individual of a population to be evolved. The elements of pi are
are: the real power outputs of the committed generating units
(i) To develop and study the performance of a CEP subject to their respective capacity constraints in equation (3).
program with Gaussian mutation as described by9 for To meet exactly the load demand in equation (2), a dependent
the solution of ELD problem; unit is arbitrarily selected among the committed n units. The
(ii) To develop and study the performance of an IFEP power output of the dependent unit, Pd is calculated by
program with better of the Gaussian and Cauchy n
mutations as proposed by Xin Yao, et al 7 with Pd = D + PL − ∑ Pj
adaptation based on scaled cost for the solution of j =1 (6)
≠d
ELD; and
(iii) To investigate into the comparative performances of In this work the power loss is not considered. However, it
CEP and IFEP in solving ELD in view of speed, may be calculated by an iterative algorithm or by using
convergence rate, success rate and quality. directly B-loss matrix of the power system.
ELD PROBLEM FORMULATION Initialization
The economic load dispatch problem can be described as an
The initial parent trial vector pi , i = 1, 2,..., N , is determined
optimization (minimization) process with the following
objective function by setting the jth component P j ~U ( Pj min , Pj max ) , where
n j = 1, 2, . . . . . , n. U ( Pj min , P j max ) denotes a uniform random
Min ∑ FC j ( Pj ) (1)
j =1 variable ranging over [ Pj min , Pj max ] .
where FC j ( Pj ) is the fuel cost function of the jth unit and Pj is Creation of Offspring
the power generated by the jth unit. (i) By Gaussian mutation (in CEP) as in Yang, et al 9.
Subject to power balance constraints
An offspring vector pi′ is created from each parent by adding
n to each component of pi , a Gaussian random variable with a
D = ∑ P j − PL (2)
j =1 zero mean and a standard deviation proportional to the scaled
cost values of the parent trial solution, ie,
where D is the system load demand and PL is the transmission
loss, and generating capacity constrains pi′ = [ P1′, P2′, ....., Pn′ ] (7)

P j min ≤ P j ≤ Pj max for j = 1, 2, . . . . . , n (3) p ′j = Pj + σ j N ( 0, 1) for j = 1, 2, . . . . . , n (8)


where P j min and P j max are the minimum and maximum where N (0, 1) represents a Gaussian random variable with
power outputs of the jth unit. mean 0 and standard deviation 1.
The fuel cost function without valve point loadings of the (ii) By choosing better one from two offsprings generated from
generating units are given by each parent, one by Gaussian mutation and the other by
Cauchy mutation (in IFEP).
FC j ( Pj ) = a j + b j Pj + c j Pj2 (4)
Let P1′j and P2′ j are the two offsprings generated from the
And the fuel cost function considering valve point loadings of
parent Pj .
the generating units are given as
P1′j = Pj + σ j N j ( 0, 1) (Gaussian mutation) (9)
FC j ( P j ) = a j + b j P j + c j P j2 e j × sin( f j × ( P j min − P j max ))
(5) P2′ j = Pj + σ j C j ( 0, 1) (Cauchy mutation) (10)

Vol 85, September 2004 111


where Cj is a Cauchy random variable with scale parameter, Table 1 Units data for test case – I (3 units case) with valve point loadings
t = 1 and centred at zero that is generated a new for each value
Gene- Pmin, Pmax, a b c e f
of j. The values of the objective function of both the offsprings
rator MW MW
are evaluated, compared and better individuals are chosen as
the offsprings for next generation. In each method mentioned 1 100 600 0.001562 7.92 561 300 0.0315

above, the standard deviation, σ j is given by the expression 2 50 200 0.004820 7.97 78 150 0.0630

σ j = β f i / f min ( Pj max − P j min ) (11) 3 100 400 0.001940 7.85 310 200 0.0420

where f min is the minimum cost value among the N trial


8450
solutions, β is a scaling factor, f i ≅ f ( pi ) is the value of the
objective function associated with the trial vector pi . Here, σ j 8400

is adapted based on its scaled values of the parent trial solution.


8350 CEP
If f i is relatively low, the step size will be small and the

Cost, $
offspring trial solution is created near the current solution and
8300
if f i is relatively high, the step size will be larger and next trial
solution will be searched within a wider range thereby
8250
providing an adaptive effect based on the fitness (cost) value.
Also, as f min keeps on updating towards globally optimum 8200
point, the adaptation will also be towards the global point. 0 10 20 30 40 50
Number of Generation
COMPETITION AND SELECTION Figure 1 Comparative convergence nature of CEP and IFEP in
Each individual in the combined population of N parent trial test case-I (3 units’ case)
vectors and their corresponding N offsprings has to compete
with R randomly selected other individuals to have a chance The value of Scale factor β has been taken as 0.01 after verifi-
to survive to the next generation. A weight value wi is assigned cation. Penalty function method has been used in satisfying
to the individual according to the competition as follows the constraints of power balance and a penalty figure of 1000
R
was found satisfactory. The value of population size (N ) was
wi = ∑ w r (12) taken as 20 after verification with population size 10, 20 and 30
r =1 Figure 1 shows convergence nature of both CEP and IFEP. It
can be observed that IFEP converges faster than CEP towards
fi the global optimum.
w r = 1, if u1 >
f r + fi
Table 2 shows that IFEP takes less cpu time than CEP. Table 3
= 0, otherwise shows the convergence rates of both the EPs in solving the test
where R is the number of competitors; f r is the fitness value case–I. Investigation reveals that IFEP converges at a
of rth randomly selected competitor from 2 N trial solutions significantly faster rate than CEP towards the global
optimum.
based on r = [2 N u 2 + 1] ; [ x ] denotes the greatest integer less
than or equal to x ; fi is the fitness value of pi ; u1 and u 2 are
Table 2 The results of the test case-I (3-units system with valve point
uniform random numbers ranging over [0, 1]. When all
loadings) with load demand 850 MW
individuals obtain their competition scores, they will be
ranked in decending order of their corresponding score, wr . Evolution Mean Time, s Best Time, s Minimum Cost, $
The first N individuals are selected and transcribed along with Method
their corresponding fitness values fi to be the parents in the CEP 20.46 18.35 8234.07
next generation.
IFEP 6.78 6.11 8234.07
SIMULATION RESULTS AND EVALUATION
All the proposed EP algorithms have been implemented in Table 3 Convergence rates in test case-I
command line in MATLAB 5.2 for solution of two test cases
of economic load dispatch. Evolu- After 10 After 15 After 20 After 25 After 30 After 40
tion Genera- Genera- Genera- Genera- Genera- Genera-
TEST CASE – I Method tions tions tions tions tions tions
This case adapted from Goldberg3 comprises of three CEP 5.82 7.32 7.94 7.27 6.72 5.12
generating units with quadratic cost functions together with
the effects of valve point loadings as given in Table 1. IFEP 20.28 13.65 — — — —

112 IE (I) Journal—EL


8250 Table 4 Results of the test case-II (40 units syetem) with value point
8248 loadings for load demand 1050 MW

8246 Evolution Mean Time, s Best Time, s Minimum Cost, $


8244 Method
Cost, $

8242 CEP 1956.93 1955.20 123488.29


8240 IFEP 1167.35 1165.70 122624.35
8238
8236 Table 5 Statistical test results of 50 runs with different initial trial
8234 solutions in test case-II

0 10 20 30 40 50 60 70 80 90 100 Evolution Mean Cost, $ Maximum Cost, $ Minimum Cost, $


Number of Run Method
Figure 2 Cost distribution of the final solution with 100 different initial
CEP 124793.48 126902.89 123488.29
solutions by CEP in test case-I
IFEP 123382.00 125740.63 122624.35
8250
8248
Table 6 Convergence rate of EPs in test case-II
8246
Evolu- After 50 After 100 After 200 After 300 After 400
8244
tion Genera- Genera- Genera- Genera- Genera-
Cost, $

8242 Method tions tions tions tions tions


8240
CEP 137.08 89.82 50.91 37.99 29.58
8238
IFEP 198.83 103.03 62.85 47.27 36.85
8236
8234 CEP. Also, the minimum cost achieved is far better than that
0 20 40 60 80 100 by CEP. Table 5 depicts the statistical results of 50 runs of
Number of Run both algorithms with 50 different initial trial solutions. IFEP
Figure 3 Cost distribution of the final solution with 100 different initial achieves far better average cost than CEP. Both the average
solutions by IFEP in test case-I cost as well as minimum cost achieved by CEP are inferior to
a great extent compare to IFEP. In this test case also the
To investigate into effects of initial trial solution on the final performance of IFEP is significantly better in terms of speed,
results and also success rate, the problem is solved with 100 convergence rate, average cost minimum cost.
different initial trial solution and results are reported in Table 4 shows the convergence rate of both the algorithms.
Figures 2 and 3. IFEP has evidently faster convergence rate than CEP.
Here, the convergence rate at a particular generation is calcu- Figures 4, 5 and 6 show the convergence nature of the
lated as the difference of the costs at the initial generation and algorithms. It may be observed that IFEP converges faster
the particular generation divided by the number of genera- than CEP towards the global optimum.
tions and the success rate is the number of times out of 100 the
algorithm has succeeded in converging near the global ×10 5
optimum point. The success rates of CEP and IFEP in test 1.38
case-I are 76% and 100%, respectively. 1.36

It is evident from Figure 1 through Figure 3 and Tables 2 and 3 1.34


that IFEP outperforms CEP in terms of speed, convergence 1.32
Cost, $

rate and success rate.


1.30

TEST CASE – II 1.28


CEP
Test case-II consisting of 40 generating units has been adapted 1.26
from4 with modifications to incorporate the valve point load-
1.24 IFEP
ings. A population size of 60 and a scale factor of 0.5 have been
used after verification. A penalty of 100 is used in this test case. 1.22
0 200 400 600 800 1000
Simulation results are given in Tables 4-6.
Number of Generation
Table 4 shows the best solution time and the minimum cost Figure 4 Comparative convergence nature of CEP and IFEP in
achieved by EP algorithms. IFEP is significantly faster than test case-II

Vol 85, September 2004 113


×10 5 comments during the works for this paper. The authors are
1.270
also thankful to Jadavpur University for providing necessary
1.265 facilities for carrying out the research.
1.260
1.255
REFERENCES
Cost, $

1.250 1. Zi-Xiong Liang and J D Glover. ‘A Zoom Feature for a Dynamic


Programming Solution to Economic Dispatch Including Transmission Losses.’
1.245
IEEE Transaction of Power Systems, vol 7, no 2, May 1992, pp 544-549.
1.240
2. D C Walter and G B Sheble. ‘Genetic Algorithm Solution of Economic
1.235
Dispatch with Valve Point Loading.’ IEEE Transactions of Power Systems, vol 8,
1.230
no 3, August 1993, pp 1325-1332.

0 10 20 30 40 50 3. D E Goldberg. ‘Genetic Algorithms in Search, Optimization and Machine


Number of Run Learning.’ Addison Wesley, MA, 1989.
Figure 5 Cost distribution of the final solution with 50 different initial
4. P -H Chen and H -C Chang. ‘Large-scale Economic Dispatch by Genetic
solutions by CEP in test case-II
Algorithm.’ IEEE Transaction of Power Systems, vol 10, no 4, November 1995,
×10 5 pp 1919-1926.
1.270
1.265 5. Th Back and H -P Schwefel. ‘An Overview of Evolutionary Algorithms for
1.260 Parameter Optimization.’ Evolutionary Computation, vol 1, no 1, 1993, pp 1-23.

1.255 6. D B Fogel. ‘Evolutionary Computation: Toward a New Philosophy of


Cost, $

1.250 Machine Intelligence.’ IEE Press, Piscataway, NJ, 1995.


1.245 7. X Yao, Y Liu and G Lin. ‘Evolutionary Programming Made Faster.’ IEEE
1.240 Transactions on Evolutionary Computation, vol 3, no 2, July 1999, pp 82-102.
1.235
8. D B Fogel. ‘A Comparison of Evolutionary Programming and Genetic
1.230 Algorithms on Selected Constrained Optimization Problems.’ Simulation,
June 1995, pp 397-404.
0 10 20 30 40 50
Number of Run 9. H T Yang, P C Yang and C L Huang. ‘Evolutionary Programming Based
Figure 6 Cost distribution of the final solution with 50 different initial Economic Dispatch for Units with Non-smooth Fuel Cost Functions.’ IEEE
solutions by IFEP in test case-II Transactions on Power Systems, vol 11, no 1, February 1996, pp 112-118.

10. K A Juste, H Kita, E Tanaka and J Hasegawa. ‘An Evolutionary


Hence, IFEP proves to be far better than CEP for solution of
Programming Solution to the Unit Commitment Problem.’ IEEE Transactions
power system ELD problems of bigger size more
on Power Systems, vol 14, no 4, 1999, pp 1452-1459.
nonlinearities.
CONCLUSION 11. P -C Yang, H -T Yang and C -L Huang. ‘Scheduling Short-term
Hydrothermal Generation Using Evolutionary Programming Techniques.’
Evolutionary programs, CEP and IFEP, with adaptations IEE Proceedings of Gener Transm Distrib, vol 143, no 4, July 1996, pp 371-376.
based on scaled costs have been developed and their
performances are examined on two test cases of power system 12. J T Ma and L L Lai. ‘Evolutionary Programming Approach to Reactive
ELD. Investigations reveal that IFEP performs much better Power Planning.’ IEE Proceedings, Part-C, vol 143, no 4, July 1996, pp 365-370.
than CEP with Gaussian mutation in terms of convergence
13. J Yuryevich and K P Wong. ‘Evolutionary Programming Based Optimal
rate, solution speed, quality of solution and success rate in
Power Flow.’ IEEE Transactions on Power Systems, vol 14, no 4, November
smaller as well as larger systems involving non-linearities like
1999, pp 1245-1250.
valve-point loadings. Investigations also reveal that the
superiority of IFEP over its counterpart becomes more 14. K P Wong and J Yuryevich. ‘Evolutionary Programming Based Algorithm
prominent in case of larger systems. for Environmentally-constrained Economic Dispatch.’ IEEE Transactions on
ACKNOWLEDGEMENT Power Systems, vol 13, no 2, May 1998, pp 301-306.

The authors would like to thank Prof D B Fogel, Prof Ernst 15. K P Wong and C C Fung. ‘Simulated Annealing Based Economic Dispatch
Stadlober and Ko-Hsin Liang for their helpful suggestions and Algorithm.’ IEE Proceedings Part-C, vol 140, no 6, 1992, pp 544-550.

114 IE (I) Journal—EL

You might also like