You are on page 1of 4

A Modified Particle Swarm Optimization for Solving

the Non-Convex Economic Dispatch


Aniruddha Bhattacharya Pranab Kumar Chattopadhyay
Jadavpur University Jadavpur University
Department of Electrical Engineering Department of Electrical Engineering
Kolkata –32, India Kolkata-32, India

Abstract- This Paper presents a novel Particle Swarm bees and fishes [12]. In PSO each potential solution is a "bird"
Optimizer combined with Roulette selection operator to solve the in the search space, which is called "particle". Although PSO
economic load dispatch (ELD) problem of thermal generators of a
power system. Several factors such as quadratic cost functions locates the region of the optimum position faster than many
with valve point loading, transmission Loss, generator ramp rate other soft computing techniques, it suffers from the risk of
limits and prohibited operating zone are considered in the premature convergence. For single-modal problem, PSO can
computation models. This new approach provides a new find out the global optimal solution accurately and rapidly. But
mechanism to restrain the predominating of super (global best) for complex multi-modal optimization problems, PSO often
particles in early stage and can effectively avoid the premature
convergence problem and speed up the convergence property. gets trapped in a local optimum or time required reaching
The proposed method has been applied to a 3 & 10 generator global optimum point increases.
power system whose cost functions are non-convex in nature. Since its introduction, numerous variations of the basic PSO
Results obtained by this method have been compared with those algorithm have been developed to speed up the convergence
obtained from Particle Swarm Optimization (PSO) method. The process and to avoid premature solution. In this paper, the
experimental results show that the proposed modified PSO
method is indeed capable of obtaining solution in less time and in classical PSO algorithm has been modified with a roulette
fewer numbers of iterations. selection operator [4] inspired from genetic algorithms [9]. In
this paper, to solve ELD problems, this modified PSO (MPSO)
Keywords— Economic Load Dispatch, Genetic Algorithm,
Particle Swarm Optimization, Prohibited-operating zone, Valve- technique has been proposed and explained. The feasibility of
point loading. the proposed method has been demonstrated for a 3 generator
[1], [3] and a 10-generator system and compared with
I. INTRODUCTION classical PSO technique in terms of solution quality and
computation efficiency.
Economic Load Dispatch (ELD) seeks "the best" generation
Section II of this paper provides a brief introduction to ELD
schedule for the generating plants to supply the required
problem. The original PSO technique is described in Section
demand plus transmission losses with the minimum production
III. The modified PSO algorithm has been demonstrated in
cost. Previously a number of conventional approaches such as
Section IV. The parameter settings for the test system to
gradient method, linear programming algorithm, lambda
evaluate the performance of MPSO and the simulation studies
iteration method, quadratic programming, non-linear
are discussed in Section V.
programming algorithm, Lagrange relaxation algorithm etc
have been applied for solving the ELD problems [1]. These II. PROBLEM DESCRIPTION
traditional classical dispatch algorithm require incremental cost The ELD may be formulated as a nonlinear programming
curves that are monotonically increasing/ piece-wise linear in problem. Basically the problem consists in minimizing an
nature. But normally the input output characteristics of objective function subject to a no. of non-linear constraints.
modern generating units are highly non-linear in nature due to A. The Objective Function
valve-point effects, ramp-rate limits etc having multiple local In ELD problem with ‘Valve point loading’ effect, the
minimum points in the cost function. Due to such generation cost Fi(Pi) is represented by a more complex
approximation the solution is subnormal and hence a huge formula
amount of revenue loss occurs over the time. Dynamic
Fi ( Pi ) = ai + biPi + ciPi 2 + ei × Sin{ fi × ( Pi min − Pi )}
Programming [1] imposes no restrictions on the nature of the
cost curves but this method suffers from the curse of And objective function F is given by
dimensionality in the solution procedure. Several artificial
m m
intelligence methods, such as genetic algorithm [5], [6], F = ∑ Fi ( Pi ) = ∑ ai + biPi + ciPi 2 + ei × Sin{ fi × ( Pi min − Pi )}
artificial neural networks [2], simulated annealing and tabu i =1 i =1
search, evolutionary programming [10], [11] have been
(1)
developed and applied successfully to solve ELD problems. In
the mid 1990s, Kennedy and Eberhart invented Particle Swarm where P i is the power output of the i-th generator; Fi(Pi) is
Optimization (PSO). Particle Swarm Optimization (PSO) was the cost function of i-th generator expressed in terms of
originated as a simulation of simplified social model of birds, coefficients ai, bi and ci of a quadratic polynomial and

978-1-4244-3388-9/09/$25.00 ©2009 IEEE


coefficients ei and fi; to take care of valve point effect; m is the particle among all the particles in the group is represented by
number of committed generators. the gbest. The modified velocity and position of each particle
B. The Constraints can be calculated as per following formulas:
k+1 k k k
The above objective function is to be minimized subject to vid = w . vid + c ×rand × (pbest – x ) + c ×rand × (gbest
the following constraints k
1 1 id id 2 2
i) Power balance constraints: – xid ) (9)
m k+1 k k+1
(10)
∑P −P
i =1
i D− PL = 0 (2) x
id
=x +v
id id
k
The inertia weighting function (w ) in equation (9) is usually
Where, PD is the load demand for the power system. The total calculated using following equation
transmission losses PL is a function of unit power outputs that
w max − w min
can be expressed using B-coefficients as wk = w max − × iter (11)
N N N itermax
PL = ∑∑ PiBijPj + ∑ B 0iPi + B 00 (3) Where, i=1, 2, 3, …….n; d=1, 2, 3, ………….N;
i =1 j =1 i =1
‘n’ is the number of particles in a group; ‘N’ is the number
ii) Generator Capacity Constraints:
of members in a particle. Velocity of particle at any iteration
Pi min ≤ Pi ≤ Pi max (4) should lie in between its upper limit vdmax and lower limit vdmin .
Where, Pi min and Pi max are the minimum and maximum M. Clerc has derived a constriction coefficient K [12].
power outputs of the i-th unit. 2 , φ=c1+c2, φ>4
K=
iii) Generator Operation Constraints: 2φ − φ 2 − 4φ
The Ramp-Up and Ramp-Down rate limits of i-th generator are
given by The new velocity and position updating equation are
k+1 k k k
As generation increases
vid = K x [w . vid + c ×rand × (pbest – x ) + c ×rand ×
Pi − Pi 0 ≤ URi (5) 1 1 id id 2 2
k
As generation decreases (gbest – xid )] (12)
Pi − Pi 0 ≤ DRi (6) k+1 k k+1
(13)
and x =x +v
id id id
max( Pi min , Pi 0 − DRi ) ≤ Pi ≤ min( Pi max , Pi 0 + URi ) (7)
IV. THE PROPOSED MODIFIED PSO ALGORITHM
Where Pi is the current output power and Pi0 is the output
power in the previous interval of the i-th generator unit. URi is In this paper, a new Modified PSO (MPSO) has been
the up-ramp rate limit of the i-th generator and DRi is the proposed where the basic PSO algorithm (with constriction
down-ramp rate limit of the i-th generator. factor approach) is combined with a roulette wheel selection
iv) Prohibited operating zone operator [7] inspired from genetic algorithms [8]. Here a
Mathematically the feasible operating zones of unit can be roulette wheel selection operator has been introduced into the
described as follows: basic PSO algorithm. Here, the major work focuses on the
Pi min ≤ Pi ≤ P l i , 1 selection of ‘gbest’. To ensure the selection probability of a
particle is in inverse proportion to its original fitness, and the
P u i , j − 1 ≤ Pi ≤ Pl i , j; j = 2,3,.....ni (8) scaled fitness is non-negative, the following fitness scaling
P u i , ni ≤ Pi ≤ Pi max function has been used:
a
Where, j represents the number of prohibited operating FS ( f ( x ) ) = (14)
zones of unit i. Pi,j-1 is the (j-1)-th prohibited operating zone of a + f ( x ) − GM
i-th unit. Pi,j is the j-th prohibited operating zone of i-th unit. Where GM is the estimated extreme of the objective function,
Total number of prohibited operating zone of i-th unit is ni. ‘a’ is a positive constant denotes the scaling degree, and ‘f(x)’
is the original fitness of a particle. With the scaled fitness,
III. OVERVIEW OF PARTICLE SWARM OPTIMIZATION
Roulette wheel selection method has been adopted to randomly
Kennedy and Eberhart invented Particle Swarm choose a particle. The selection of particle ‘i’ is computed by:
Optimization (PSO) in 1995 [12]. The PSO can be best FS ( f ( x [i ])) (15)
understood through an analogy of a swarm of birds in a field. q [i ] = n
Without any prior knowledge of the field, the birds move in ∑ FS ( f ( x [i ]))
i =1
random locations with random velocities looking for foods.
In PSO, particles change their positions (states) with time. After calculating the value ‘q[i]’ for each particle as per
Let ‘x’ and ‘v’ denote a particle coordinates (position) and its equation (15), the mean value of ‘q[i]’ has been calculated as
corresponding flight speed (velocity) in a search space per following equation
respectively. The best previous position of the ith particle is
recorded and represented as pbesti. The index of the best
n
B. Test case 2
∑ q [i ] (16) 10-generator Systems: The load demand is 2000 MW. The unit
Mean( q ) = i =1

n characteristics data are given in Tables VI and VII. The system


Where ‘n’ is the size of population. i=1,2, ……n. B loss coefficients are given in Table-VIII. The best solutions
Index ‘i’ is calculated as per following equation obtained from proposed MPSO, PSO methods are shown in
q [i ] Table IX. Convergence characteristics of the PSO and MPSO
i= (17) method for the system are shown in Fig.2.
Mean(q)
TABLE IV
That set of population, which gives maximum value of ‘i’ BEST OF 3-UNIT SYSTEM (BEST INDIVIDUAL)
among total population size, is nothing but index . The position Unit Power Output MPSO PSO
of particle ‘i’ is used to replace ‘gbest’ in the equation (12). P1 (MW) 199.5997 199.5997
P2 (MW) 233.7630 233.7630
V. NUMERICAL EXAMPLES AND RESULTS
P3 (MW) 598.6655 598.6655
Total Power Output (MW) 1032.0282 1032.0282
The proposed MPSO technique has been applied to two Ploss (MW) 182.0282 182.0282
different test systems e.g. 3 Generators [3] and 10 Generators Total Generation Cost ($/h) 10097.5752268421 10097.5752268421
Power Systems. The software programs were written in TABLE V
COMPARISON BETWEEN THE METHODS (50 TRIALS), 3 GENERATOR SYSTEMS
MATLAB-6.5 language and executed on a 1.7 GHz Pentium Method: - MPSO PSO
IV personal computer with 512-MB RAM. The following Generation Cost Min 10097.5752268421 10097.5752268421
MPSO and PSO parameters have been used after a no. of Max 10104.1372631893 10123.5361910523
($/hour)
Average 10098.2314304768 10100.8275177926
careful experimentation. CPU Time Min 1.7575 1.8670
i) PSO Method:- Population Size = 20; Iterations = 1000; wmax required for Max 2.1036 2.2512
= 0.95 and wmin = 0.2; vdmax = 0.5 Pmax and vdmin = - 0.5 Pmax ; Convergence(Sec) Average 2.0300 2.0585
C =2.05 and C =2.05; Constriction factor = 0.729. Iterations required Min 370 370
1 2
for Convergence Max 430 420
ii) MPSO Method:- Population Size = 20; Iterations = 1000; Average 390 400
wmax = 0.95 and wmin = 0.2; vdmax = 0.5 Pmax and vdmin = - 0.5 No. of hits to Global
45
40
Minimum (out of 50)
Pmax ; C =2.8 and C =1.3; Constriction factor = 0.729; The
1 2 Convergence Characterestics of 3 Generator Systems
4
value of ‘GM’ = 0; and the constant ‘a’ =100. 1.0106
x 10
PSO
A. Test Case 1 1.0105 MPSO
3-generator Systems: The unit characteristics data are given in 1.0104
Table I and II. The load demand is 850 MW. The B loss
Fuel Cost ($/hour)

1.0103
coefficients are given in Table-III. The best solutions of the
proposed MPSO, PSO methods are shown in Table IV. 1.0102
Convergence characteristics of the proposed MPSO and PSO 1.0101
method for 3 Generator systems are shown in Fig.1.
1.01
TABLE I
CAPACITY AND COST COEFFICIENTS OF 3 GENERATOR SYSTEM 1.0099
Quantities Unit-1 Unit-2 Unit-3
1.0098
ai($/hr) 78 310 562
bi($/MWh) 7.97 7.85 7.92 1.0097
0 100 200 300 400 500
ci($/(MW)2hr) 0.004820 0.001940 0.001562 Iteration No.
ei($/hr) 150 200 300
Figure 1. Convergence property of 10 Generator system using MPSO and PSO
fi(rad/MW) 0.063 0.042 0.0315
Pimin(MW) 50 100 100 Convergence Characterestics of 10 Generator System
5
Pimax(MW) 200 400 600 x 10
TABLE II 1.267 PSO
RAMP RATE LIMITS & PROHIBITED OPERATING ZONES OF 3 GENERATOR MPSO
SYSTEMS 1.266
Prohibited
Unit Pi0(MW) URi(MW/hour) DRi(MW/hour)
zones (MW)
Fuel Cost ($/hour)

1.265
[90 110]
1 170 50 90
[140 160] 1.264
[210 230]
2 350 80 120
[350 380] 1.263
[240 270]
3 440 80 120 1.262
[460 490]
TABLE III
B-COFFICIENTS FOR 3 GENERATOR SYSTEMS 1.261
⎡ 0.0006760 0.0000953 -0.0000507⎤
1.26
B= ⎢ 0.0000953 0.0005210 0.0000901⎥MW−1;B0 =[-0.07660 -0.00342 0.01890];B00 = [4.0357] MW 0 100 200 300 400 500
⎢ ⎥ Iteration No.
⎢⎣-0.0000507 0.0000901 0.0002940⎥⎦
Figure 2. Convergence property of 10 Generator system using MPSO and PSO
TABLE VI respective minimum generation cost. It is seen that average
CAPACITY AND COST COEFFICIENTS OF 10 GENERATOR UNITS
convergence time and number of iterations required for
Qua bi ci ei fi Pimin
ntiti ai($/hr) ($/M ($/(M ($/h (rad/ (MW
Pimax convergence in case of MPSO are lesser than Basic PSO. So it
2 (MW) may be concluded that overall performance of proposed MPSO
es Wh) W) hr) r) MW) )
Unit 786.79 38.53 0.1524 450 0.041 150 470 is comparatively less than that of basic PSO.
-1 88 79 TABLE IX
Unit 451.32 46.15 0.1058 600 0.036 135 470 BEST OF 10-UNIT SYSTEM (BEST INDIVIDUAL)
-2 51 91 Unit Power Output MPSO PSO
Unit 1049.9 40.39 0.0280 320 0.028 73 340
-3 977 65 P1 (MW) 226.6242 226.6242
Unit 1243.5 38.30 0.0354 P2 (MW) 232.7816 232.7816
260 0.052 60 300
-4 311 55 P3 (MW) 340.0000 340.0000
Unit 1658.5 36.32 0.0211 280 0.063 73 243 P4 (MW) 300.0000 300.0000
-5 696 78 P5 (MW) 243.0000 243.0000
Unit 1356.6 38.27 0.0179 310 0.048 57 160 P6 (MW) 160.0000 160.0000
-6 592 04
Unit 1450.7 36.51 0.0121 P7 (MW) 130.0000 130.0000
300 0.086 20 130 P8 (MW) 120.0000 120.0000
-7 045 04
Unit 1450.7 36.51 0.0124 340 0.082 47 120 P9 (MW) 80.0000 80.0000
-8 045 04 P10 (MW) 243.9484 243.9484
Unit 1455.6 39.58 0.1090 270 0.098 20 80 Total Power Output
-9 056 04 2076.3542 2076.3542
Unit 1469.4 40.54 0.1295 (MW)
380 0.094 10 455
-10 026 07 Ploss (MW) 76.3542 76.3542
Total Generation Cost
TABLE VII 126015.5072978053 126015.5072978053
RAMP RATE LIMITS & PROHIBITED OPERATING ZONES OF 10 UNIT SYSTEMS ($/h)
Prohibited TABLE X
Unit Pi0(MW) URi(MW/hour) DRi(MW/hour) Operating COMPARISON BETWEEN THE METHODS (50 TRIALS), 10 GENERATOR SYSTEMS
Zones (MW) Method: - MPSO PSO
[180 200] Generation Cost Min 126015.507297805 126015.507297805
1 90 80 120 [305 335] ($/hour) Max 129173.281139619 133397.183803809
Average 126494.824390074 127449.197446111
[390 420]
CPU Time Min 3.3944 3.4129
[210 230] required for Max 4.0190 4.0340
2 400 80 120 [365 395] Convergence(Sec) Average 3.9540 3.9780
[430 470] No. of Iterations Min 400 410
3 100 70 105 - required for Max 480 480
4 95 65 100 - Convergence Average 460 470
5 80 60 90 - No. of hits to Global
6 105 60 100 - 43 38
Minimum (out of 50)
7 100 130 130 -
8 90 100 120 - REFERENCES
9 40 80 80 [30 40][55 65]
[185 225] [1] A. J. Wood and B. F. Wollenberg, Power Generation, Operation, and
10 400 80 120 [305 355] Control, 2nd edition, Wiley, New York, 1996.
[420 450] [2] C.T. Su, G.J. Chiou, A fast-computation Hopfield method to economic
TABLE VIII dispatch of power systems, IEEE Transactions Power System 12 (4) (1997)
B-COFFICIENTS FOR 10 GENERATOR SYSTEMS 1759- 1764.
[3] C.T. Su, C.T. Lin, New approach with a Hopfield modeling framework to
⎡ 0.000049 0.000014 0.000015 0.000015 0.000016 0.000017 0.000017 0.000018 0.000019 0.000020 ⎤ economic dispatch, , IEEE Transactions Power System, 15 (2) (2000) 541-545.
⎢ 0.000014 0.000045 0.000016 0.000016 0.000017 0.000015 0.000015 0.000016 0.000018 0.000018 ⎥
⎢ ⎥ [4] Fang WANG Yuhui QIU, A Modified Particle Swarm Optimizer with
⎢ 0.000015 0.000016 0.000039 0.000010 0.000012 0.000012 0.000014 0.000014 0.000016 0.000016 ⎥ Roulette Selection Operator, Proceeding of NLP-KE'O5, IEEE, 2005.
⎢ ⎥
⎢ 0.000015 0.000016 0.000010 0.000040 0.000014 0.000010 0.000011 0.000012 0.000014 0.000015 ⎥ [5] H.K. Youssef, K.M. El-Naggar, Genetic based algorithm for security
⎢ 0.000016 0.000017 0.000012 0.000014 0.000035 0.000011 0.000013 0.000013 0.000015 0.000016 ⎥
B=⎢ ⎥ constrained power system economic dispatch, E.P.S.R. 53 (1) (2000) 47-51.
⎢ 0.000017 0.000015 0.000012 0.000010 0.000011 0.000036 0.000012 0.000012 0.000014 0.000015 ⎥ [6] J. Baker, "Adaptive selection methods for genetic algorithms," Proceedings
⎢ 0.000017 0.000015 0.000014 0.000011 0.000013 0.000012 0.000038 0.000016 0.000016 0.000018 ⎥
⎢ ⎥ of the 2nd International Conference on Genetic Algorithms, Lawrence
⎢ 0.000018 0.000016 0.000014 0.000012 0.000013 0.000012 0.000016 0.000040 0.000015 0.000016 ⎥ Erlbaum Associates, Hillsdale, NJ, 1985, pp. 100-1 11.
⎢ 0.000019 0.000018 0.000016 0.000014 0.000015 0.000014 0.000016 0.000015 0.000042 0.000019 ⎥
⎢ ⎥ [7] J. Holland, Adaptation in Natural and Artificial Systems, Cambridge: MIT
⎢⎣ 0.000020 0.000018 0.000016 0.000015 0.000016 0.000015 0.000018 0.000016 0.000019 0.000044 ⎥⎦
Press, MA, 1992.
MW-1 [8] J. Kennedy, R. C. Eberhart, "Particle swarm optimization," Proceedings of
B0 = [ 0.00000 0.00000 0.00000 0.00000 0.00000 0.00000 0.00000 0.00000 0.00000 0.00000] IEEE International Conference on Neural Networks. Piscataway, NJ: IEEE
Press Center, 1995, pp. 1942-1948.
B00 = [0.00000] MW
[9] J. Kennedy, R. C. Eberhart, Swarm Intelligence Morgan Kaufinann
C. Comparison of two Methods Publishers, 2001.
1) Solution Quality & Computation Efficiency: From the [10] J. Kennedy, R.C Eberhart, A discrete binary version of the particle swarm
results seen in Tables V and X, it is seen that the MPSO and optimization algorithm, Proceedings of the 1997 conference on Systems, Man,
and Cybernetics, pp.4104-4109, 1997.
PSO methods can obtain same minimum generation cost for [11] J.O. Kim, D.J. Shin, J.N. Park, C. Singh, Atavistic genetic algorithm for
the Systems. But for the average generation cost over 50 trial economic dispatch with valve point effect, E.P.S.R. 62 (3) (2002) 201-207.
runs, the MPSO method gives much better results than that of [12] M. Clerc; The swarm and the queen: towards a deterministic and adaptive
PSO method. It is seen that in comparison to PSO, MPSO out particle swarm optimization, Proceedings of IEEE congress on Evolutionary
Computation, Piscataway, NJ, IEEE Press Center, 1999, pp. 1951-1957.
of 50 trials 45 (3-Unit) and 43 trials (10-Unit) reaches their

You might also like