IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 16, NO.

2, MAY 2001 307

Nonconvex Economic Dispatch by Integrated

Artificial Intelligence
Whei-Min Lin, Member, IEEE, Fu-Sheng Cheng, and Ming-Tong Tsay, Member, IEEE

Abstract—This paper presents a new algorithm by integrating this problem [4], [5]. With parallel searching mechanism, EP
evolutionary programming (EP), tabu search (TS) and quadratic has a high probability to find optimal solutions [6]–[8]. For large
programming (QP) methods to solve the nonconvex economic scale complex problems, the solution might get trapped in a sub-
dispatch problem (NED). A hybrid EP and TS were used for
quality control, and the Fletcher’s quadratic programming optimal state, where the variation operators cannot produce any
technique for solving. EP and TS determines the segment of a offspring outperforming its parents. These applications involved
cost curve used, which is piecewise quadratic natured. Operation a large number of iterations and were susceptible to the related
constraints are modeled as linear equality or inequality equations, control parameters.
resulting in a typical QP problem. Fletcher’s QP was chosen to Fan and McDonald [9], and Lee and Breiphol [10] solved
enhance the performance. The fitness function is constructed
from priorities without penalty terms. Numerical results show EDPO by using a lambda iterative dispatch first. For infeasible
that the proposed method is more effective than other previously solutions, they redispatch the units using some heuristic rules to
developed evolutionary computation algorithm. probe the neighborhood for feasibility. These methods may not
Index Terms—Adaptive decay scale, distance, economic dis- be applicable if the problem contains too many nonlinear con-
patch with piecewise quadratic cost function (EDPQ), economic straints. Most artificial intelligent techniques have the ability to
dispatch with prohibited operating zones (EDPO), evolutionary conquer restrictions. Hopfield neural network models [11], the
programming (EP), genetic algorithm (GA), mutation scale, standard GA, and the deterministic crowding GA [12] were used
nonconvex economic dispatch problem (NED), quadratic pro-
gramming (QP), tabu search (TS). to solve this problem. However, performance becomes another
concern.
Tabu Search (TS) introduced by Glover and Hansen [13], [14]
I. INTRODUCTION has been successfully applied to a number of integer optimiza-
tion problems [15]–[17]. TS avoids cycling by storing informa-
T YPICAL economic load dispatch is to minimize the oper-
ation cost. Two major NED problems are
1) economic dispatch with piecewise quadratic cost function
tion of the past from the search. It easily gets trapped in the local
optimum since a single candidate is used to generate offsprings.
(EDPQ); With a piecewise quadratic cost function, many researchers
2) economic dispatch with prohibited operating zones have applied Beale’s [18], [19] and Fletcher’s [20], [21] QP
(EDPO). methods to solve the problem. Literature survey reveals that
Cost of generators are usually approximated by quadratic Fletcher’s QP method provides a better performance than
functions. However, it is more practical to consider EDPQ Beale’s QP method [21].
for valve point loading for fossil fired plants. Besides, the In this paper, we proposed a hybrid algorithm (ETQ) for
whole operating range may not be always available. Units solving EDPQ and EDPO. The algorithm integrates main
may have prohibited operation zones due to faults in machines features of EP, TS and QP. We solve the problem in two phases,
or associated auxiliaries, such as boilers, feed pumps, etc., the cost-curve-selection subproblem, and the typical ED
leading to instabilities in certain ranges. The prohibited region solving sub-problem. The first phase was solved with a hybrid
separates the decision space into disjoint subsets, constituting EP and TS, and the second phase by QP. The curve-selection
a nonconvex solution space, and EDPO becomes a nonconvex subproblem is further solved by using EP for mutation and TS
optimization problem. for selection as described later. Both EDPQ and EDPO are
Lin and Viviani presented the hierarchical structure approach similar problems when ETQ was employed.
to solve EDPQ [1]. Local optimality problem exists. Hopfield
neural network models were employed to study this problem II. PROBLEM STATEMENT
[2], [3]; these methods require tremendous time for training. A. The Objective Function
Recently, the three major types of evolutionary computations,
including the genetic algorithms (GA), evolutionary program- A piecewise quadratic cost function can be represented by
ming (EP) and evolution strategies (ES), were applied to solve
(1)
Manuscript received August 25, 2000.
W.-M. Lin and F.-S. Cheng are with the Department of Electrical Engineering, where
National Sun Yat-Sen University, Kaohsiung, Taiwan, R.O.C. : generated power output of unit .
M.-T Tsay is with the Department of Electrical Engineering, Cheng-Shiu In-
stitute of Technique, Kaohsiung, Taiwan, R.O.C. : fuel consumption cost coefficients of unit
Publisher Item Identifier S 0885-8950(01)03797-X. for fuel type .

0885–8950/01$10.00 © 2001 IEEE

308 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 16, NO. 2, MAY 2001

(a)

(b)
Fig. 2. (a) Coding example. (b) Equivalent decimal vector of one individual.
(a) (b)
the operation curve shown in Fig. 2. The population in ETQ is
Fig. 1. (a) Thermal unit with piecewise quadratic function. (b) With prohibited
operating zones. represented by an integer matrix of dimension , where
is the population size.
ED minimizes the total cost by B. Initialization

(2) Let be a trial vector denoting

the th individual of a population. The initial trial vector ,
where is the number of generating units. , is determined by setting its th components
by
B. The Constraints
round (7)
Unit with prohibited operating zones will have
(3) where
round : rounds to the nearest integers.
Also : a uniform random number between 0 and 1
: maximum number of segments of the operating
(4) curve of unit .

C. Adaptive Progressing Scheme

where
: minimum power generation An adaptive scheme is defined in this paper by
: maximum power generation (8)
: lower bound of an operating zone
(9)
: upper bound of an operating zone
: total load demand and
: power loss
if (10)
and spinning reserve (SR) can be expressed by
where
(5) :
maximum value of parameter .
:minimum value of parameter .
: the step size.
(6) : the regulating scale set to 1.25 in this paper.
: maximum allowable number of iterations.
The adaptive scheme was used by setting equal to various
III. DEVELOPMENT OF THE ALGORITHM parameters and in later sections. depends on the number
of generations and the complexity of the system.
In solving the EDPQ and EDPO, two variables needed to be
determined are i) the integer sequence number of segmented D. Mutation Based on EP
cost functions, and ii) the power output of each unit. The first is
a combinatorial optimization problem and the second is a non-
linear one. A hybrid EP and TS is proposed to solve the combi- round (11)
natorial optimization. The nonlinear optimization is then solved
by a quadratic programming routine. The algorithm that inte- where : a Gaussian random variable with mean and
grates EP, TS and QP is called ETQ in this paper. variance . If exceeds limits, it will be set to the limit. is
set to 0 in general, and can be expressed as
A. Solution Coding
(12)
Sequence numbers are assigned to the curves of EDPQ and
EDPO as shown in Fig. 1. The solution is a row vector called an where
“individual” with decimal numbers ( ), where : cost of the th individual in the old generation
is the number of units. Each element represents a segment of : maximum cost in the old generation
LIN et al.: NONCONVEX ECONOMIC DISPATCH BY INTEGRATED ARTIFICIAL INTELLIGENCE 309

: an adaptive mutation scale with ,

and , is set to 0.5 and is set
to 0.005 in this paper.

E. Repairing Strategy
With mutation operations, the reserve constraints might be
violated. A repairing strategy is used to restore the feasibility as
follows:
1) Pick at random one element in the individual with
violation.
2) If violating (5), add one to the selected element; if vio-
lating (6), minus one from the selected element.
3) Check the reserve constraint for violations, if not, go to
the next step, otherwise, go to 1).
Fig. 3. The concept of distances.

F. Fletcher’s QP Method
Evaluate each chromosome by solving the ED using QP.
Fletcher’s QP method [20] can be used to minimize a quadratic
function presented in Section II. The formulation becomes
minimizing

(13)

subject to
(14)
(15)
(16)
where is the equality constraints and is the in-
equality constraints evaluated at .

G. Fitness Function with “Distance” Term

Assign a rank (i.e., the number place , etc.) to
the calculated cost of each new individual, , .
A combined population with individuals is formed. Un-
like previously developed statistic methods, the concept of “dis-
tance” was added to the fitness function to prevent from being Fig. 4. The flow chart of the proposed algorithm.
trapped in a local minimum. The fitness score of the th indi-
vidual is modified by new population does not include the current best solution, the
(17) best solution must replace the last individual in the new popu-
lation (Elitism).
where is an adaptive decay scale with , and The current best solution (point in Fig. 3) may not reach the
, is set to 0.6 and is set to 0.005 in this global optimum. If a more attractive solution exists, it could be
paper. relatively far away (point ). Generally, solutions with slightly
is the rank of assigned to the th individual, where better cost ( and ) prevail, so solution prematures. To prevent
is the distance from the individual to the current best solution prematurity, point with slightly worse cost than , needs a
vector, and is given by higher rank to be selected. That is, a higher is awarded to
(18) a longer .
Stopping Criteria: Stopping criteria is given in the order of
where
1) number of iterations reached without improving the cur-
: vector of the th individual in the combined
rent best solution.
population
2) maximum allowable number of iterations reached.
: the current best solution vector.
Fig. 4 shows the flow chart of the proposed algorithm.
H. Selection Based on TS
Individuals will be ranked in ascending according to their fit- IV. CASE STUDY
ness scores by a sorting algorithm. The first PS individuals are The ETQ-based ED algorithm was implemented with C lan-
transcribed along with their cost for the next generation. If the guage on a 600-MHz Pentium III computer with 128 MB RAM.
310 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 16, NO. 2, MAY 2001

TABLE I
COMPARISON OF VARIOUS METHODS WITH 100 TRAIL TESTS FOR
2400 MW LOAD

Note:
PS: Population size.
MAG: Maximum allowable number of iterations (generations).
NWI: The number of iterations without improving the current best solution.
MGC: Mean value of number of generations to converge.
MCT: Mean value of computation time. Fig. 5. Convergence comparison.

TABLE II
In order to show the effectiveness of the proposed algorithm, THE PROHIBITED OPERATING ZONES
test results of two typical cases [2], [12] are shown for example.
Case 1 includes 10 generating units with piecewise quadratic
cost functions (EDPQ). Case 2 contains 15 units with prohib-
ited operating zones for some units (EDPO).

A. Case 1 [2]
A sample system with 2400 MW [2] was used for tests. Three
hybrid methods were developed in this paper for tests. TABLE III
1) ETQ: the algorithm previously developed; EP for muta- COMPARISON OF THREE METHODS FOR 2650 MW LOADS
tion, TS for selection, and QP for solving.
2) EQ: EP used for both mutation and selection, QP for
solving. is set the same as in ETQ.
3) TQ: Executes TS producer [16], QP for solving. The tabu
list size is set to 7 in this paper. A choice of tabu list size in
a range centred around 7 seems to be very effective [16],
[22].
Different runs were tested with results also shown from
other evolutionary computation algorithms in [5], for which a
required population size of 100 is known to generate satisfac-
tory results. For fair comparison, 30 individuals were randomly
selected for the proposed algorithms from the 100 initial
population used by other evolutionary computation algorithms.
100 test runs were conducted for each method. Elitism was
applied to all methods. Sample results are given in Table I and
Fig. 5. Every method could get the optimal solution. It can be
seen that ETQ is a better performer in terms of solution quality,
number of generations to converge, and the performance.

B. Case 2 [12]
A 15-unit system is used for tests. 4 units are with prohib-
an active repairing strategy to probe for the new solution, while
ited operating zones, and the other 11 units will contribute reg-
GA uses penalty function to passively test feasible and infea-
ulating reserves. This system supplies a load of 2650 MW with
sible solutions.
200 MW spinning reserve. Parameters of the 15 units are in [12].
The prohibited zones embedded in unit 2, 5, 6, and 12, are given
V. CONCLUSION
in Table II. Table III shows the comparison chart of the proposed
algorithm against other GA’s in [12]. Again, ETQ is a better per- The implementation of ETQ differs from EQ and TQ in three
former in both quality and performance. Note that ETQ adopts aspects. Firstly, its fitness function considered the distance to
LIN et al.: NONCONVEX ECONOMIC DISPATCH BY INTEGRATED ARTIFICIAL INTELLIGENCE 311

avoid converging to a local optimum. The selection in ETQ is [11] C. T. Su and G. J. Chiou, “A Hopfield network approach to economic
more deterministic than the one in EQ. Secondly, it also forms dispatch with prohibited operating zones,” in Proceedings of the 1995
Energy Management and Power Delivery International Conference,
a parallel structure and can weaken the dependence of the con- 95TH8130, pp. 382–387.
vergence rate on the initial solutions such as TQ. Thirdly, for [12] S. O. Orero and M. R. Irving, “Economic dispatch of generators
ETQ, vectors with fitness score good enough would be used as with prohibited operating zones: A genetic algorithm approach,” IEE
Proc.—Gener. Transm. Distrib., vol. 143, no. 6, pp. 529–533, Nov.
candidates to create new solutions regardless of tabu restrictions 1996.
such as TS. The aspiration criteria in ETQ are flexible than the [13] F. Glover, “Tabu search—Part I,” ORSA Journal on Computing, vol. 1,
one in TQ. no. 3, pp. 190–206, Summer 1989.
[14] , “Tabu search—Part II,” ORSA Journal on Computing, vol. 2, no.
The implementation of the ETQ differs from other evolu- 1, pp. 4–32, Winter 1990.
tionary computation (GA, EP, and hybrid with ES) in four [15] Y. C. Huang, H. T. Yang, and C. L. Huang, “Solving the capacitor
aspects. Firstly, it chooses the combination of all units’ curves placement problem in a radial distribution system using tabu search ap-
proach,” IEEE Trans. on Power Systems, vol. 11, no. 4, pp. 1868–1873,
from finite sets rather than infinite points as in other evolu- Nov. 1996.
tionary computation, and needs few generations to converge. [16] A. H. Mantawy, Y. L. Abdel-Magid, and S. Z. Selim, “Unit commitment
Secondly, it has the advantage of QP to conquer various by tabu search,” IEE Proc.—Gener. Transm. Distrib., vol. 145, no. 1, pp.
56–64, Jan. 1998.
constraints without using the fitness function with penalties, [17] F. S. Wen and C. S. Chang, “Tabu search approach to alarm processing
and it can perform better. Thirdly, the solution is coded with in power system,” IEE Proc.—Gener. Transm. Distrib., vol. 144, no. 1,
a decimal representation, saving computer memory as well as pp. 31–38, Jan. 1997.
[18] E. M. L. Beale, “On quadratic programming,” Nav. Res. Logist. Q., vol.
computation time outperforming the GA search. Fourthly, the 6, pp. 227–243, 1959.
gradually decaying parameters ( and ) can satisfy a succes- [19] J. Y. Fan and L. Zhang, “Real time economic dispatch with line flow
sive statistic searching process by first using the diversification and emission constraints using quadratic programming,” IEEE Trans.
on Power Systems, vol. 13, no. 2, pp. 320–325, May 1998.
(bigger parameters) to reserve the larger attractive region, and [20] R. Fletcher, “A general quadratic programming algorithm,” J. Inst.
then the intensification (smaller parameters) to search the small Math. and Appi., vol. 7, pp. 76–91, 1971.
neighborhood of an elite solution. [21] J. Nanda, D. P. Kothari, and S. C. Srivastava, “New optimal power-dis-
patch algorithm using Fletcher’s quadratic programming method,” IEE
Proceedings, pt. C, vol. 136, no. 3, pp. 153–163, May 1989.
REFERENCES [22] F. Glover, “Artificial intelligence, heuristic frameworks and tabu
search,” Managerial Decis. Econ., vol. 11, pp. 365–373, 1990.
[1] C. E. Lin and G. L. Viviani, “Hierarchical economic dispatch for piece-
wise quadratic cost functions,” IEEE Trans. on Power Apparatus and
Systems, vol. PAS-103, no. 6, pp. 1170–1175, June 1984.
[2] J. H. Park, Y. S. Kim, and K. Y. Lee, “Economic load dispatch for
piecewise quadratic cost function using Hopfield neural network,” IEEE
Trans. on Power Systems, vol. 8, no. 3, pp. 1030–1040, Aug. 1993. Whei-Min Lin was born on October 3, 1954. He received the B.S.-E.E. degree
[3] K. Y. Lee, A. S. Yome, and J. H. Park, “Adaptive Hopfield networks for from the National Chao-Tung University, the M.S.-E.E. degree from the Uni-
economic load dispatch,” IEEE Trans. on Power Systems, vol. 13, no. 2, versity of Connecticut, and the Ph.D. E.E. degree from the University of Texas
pp. 519–526, May 1998. in 1985. He worked at Chung-Hwa Institute for Economic Research, Taiwan,
[4] J. H. Park, S. O. Yang, H. S. Lee, and Y. M. Park, “Economic load dis- as a Visiting Researcher after his graduation. He joined Control Data Crop. in
patch using evolutionary algorithms,” in Proceedings of the 1996 Intel- 1986 and worked with Control Data Asia in 1989. He has been with National
ligent Systems Applications to Power System International Conference, sun Yat-Sen University, Taiwan, since 1989. Dr. Lin’s interests are GIS, distri-
96TH8152, pp. 441–445. bution system, SCADA and automatic control system. Dr. Lin is a Member of
[5] J. H. Park, S. O. Yang, and K. J. Mun, “An application of evolutionary IEEE and Tau Beta Pi.
computations to economic dispatch with piecewise quadratic cost
functions,” in Evolutionary Computation Proceedings of the 1998 IEEE
World Congress on Computional Intelligence, 98TH8360, pp. 289–294.
[6] H. T. Yang, P. C. Yang, and C. L. Huang, “Evolutionary programming
based economic dispatch for units with nonsmooth fuel cost functions,”
Fu-Sheng Cheng was born on January 4, 1963. He received the M.S.-E.E. de-
IEEE Trans. on Power System, vol. 11, no. 1, pp. 301–306, Feb. 1996.
gree from the Cheng-Kung University in 1990. Currently, he is pursuing the
[7] K. P. Wong and J. Yuryevich, “Evolutionary-programming-based algo-
Ph.D. degree in the National Sun Yat-Sen University. His interests are energy
rithm for environmentally-constrained economic dispatch,” IEEE Trans.
management systems and distribution automatic system.
on Power Systems, vol. 13, no. 2, pp. 301–306, May 1998.
[8] L. L. Lai and J. T. Ma, “Application of evolutionary programming to
reactive power planning-comparison with nonlinear programming ap-
proach,” IEEE Trans. on Power Systems, vol. 12, no. 1, pp. 198–206,
Feb. 1997.
[9] J. Y. Fan and J. D. McDonald, “A practical approach to real time eco- Ming-Tong Tsay was born on October 10, 1964. He received the B.S.-E.E. de-
nomic dispatch considering unit’s prohibited operating zones,” IEEE gree from Feng-Chia University in 1988. M.S.-E.E. and Ph.D. degrees from the
Trans. on Power Systems, vol. 9, no. 4, pp. 1737–1743, 1994. National Sun Yat-Sen University in 1990 and 1994. He has been with the De-
[10] F. N. Lee and A. M. Breipohl, “Reserve constrained economic dispatch partment of Electrical Engineering of Cheng-Shiu Institute of Technology &
with prohibited operating zones,” IEEE Trans. on Power Systems, vol. Commerce, Taiwan, since 1994. He is interested in GIS applications on distri-
8, no. 1, pp. 246–254, 1993. bution planning and power system operation.