MoC2.

Proceedings of 2011 8th Asian Control Conference (ASCC)

Kaohsiung, Taiwan, May 15-18, 2011

An Improved Fuzzy Neural Networks Approach

for Short-Term Electrical Load Forecasting
Gwo-Ching Liao IEEE Member
AbstractTo solve the Short-Term Load Forecasting (STLF)
tasks, this paper proposes to use a new method, namely,
Quantum Genetic Algorithm (QGA) merged with Fuzzy
Neural Networks (FNNs), here after called the QGA-FNN
method. With the QGA method, we encode all the networks
weights and biases into several Artificial Neural Network
(ANN) system particle swarms, and then we train the network
parameter values using the QGA method proposed in this paper to locate the networks optimal parameter solution. Next,
we resolve the optimal STLF with the FNNs derived. The results by the proposed method are compared with that by other
commonly-used load forecasting methods, such as the Artificial Neural Network, the Evolutionary Programming combined with ANN (EP-ANN) and the Genetic Algorithm combined with ANN (GA-ANN). The comparisons indicate that the
proposed method renders smaller load forecasting discrepancies, with significant improvement rates ranging from 6.2% to
43.4%, signifying the proposed methods advantage in load
forecasting.

Index Terms Fuzzy Neural Network, Quantum Genetic

Algorithm, Load forecasting

I. INTRODUCTION

TLF is vital to an electric power systems thermal

power generation planning, thermal-hydro power generation coordination, power generation units economic
dispatch and power exchange planning. The traditional
electric power forecasting often uses vital linear factors
which will influence the load; such forecastings advantage
lies at its simplicity with the forecasting model. However,
since the system load is multi-dimensional and there exists
nonlinearity among the influential factors, the traditional
method hardly can forecast the load series with enough accuracy.
In recent years, there have been a few methods adopted
in load forecasting, such as the Time Series Method [2], the
Gray Theory [3], the Least Square Method [4] and the
Neural Network (NN) [5]-[9] etc. Due to its capabilities in
responding nonlinearity and its general approximating, the
neural network method is extremely suitable for load forecasting. The NN method has been adopted by several
methods [10]-[12] for STLF. However, this method still has
its limitation; for example, it is not quite suitable for load
forecasting future weekends and some special days, and it
requires huge data amount, such as weather data, load
variations, load cycles, and so forth. The traditional neural

Gwo-Ching Liao is with Department of Electrical Engineering, Fortune

Institute of Technology, Kaohsiung Taiwan.
(E-mail:liaogwo@ms68.hinet.net)

network method still has some shortcomings to be resolved,

such as low training speed, weaker searching capability for
the overall optimal solution, and so on.
To improve the shortcomings of various kinds of STLF
methods and to correct the drawbacks of the neural network
method, this paper introduces the Quantum Genetic Algorithm (QGA) method [13]-[19] for STLF. The QGA method
can enable the ANN to locate the optimal solution for the
networks weights, biases, etc. to make the true optimal solution achievable for load forecasting. Meanwhile, to arrive
more flexibility in treating the influences of various factors
(the results can get from section five of this paper), such as
temperature, humidity, rainfall, holidays, and so forth, we
adopt seven different FNNs, each of which represents the
load forecasting model for each day of the week, seven days
a cycle, along with the consideration of other special factors.
We next train the network parameter values using the
Quantum Genetic Algorithm method proposed in this paper
to locate the networks optimal parameter solution. Finally,
the optimal STLF was resolved with the FNNs derived.

II. THE ARCHITECTURE OF THE FUZZY

HYPER-RECTANGULAR COMPOSITE NEURAL NETWORKs
(FHCNNs) [9]

2.1 Fuzzy Hyper-rectangular Composite Neural Networks

(FHCNNs)

We use a fuzzy logic system merging with neural networks to proceed with the calculation of ANNs parameters,
including weights, biases, etc., the basic steps of which are
shown in Fig. 1. Such method can be expressed as
Fuzzy rule R j :
IF ( x1 is A1j ) and ( x 2 is A 2j ) and and ( x is A pj )
p

THEN y is B j

(1)

where, x represents a P-dimensional input vector, y the

output, A 1j the membership function between the input
variable x i and Rule j , and B j the relationship between the output variable y and Rule j .
A fuzzy logic system comprises four key factors, namely
1). Fuzzification structure, 2). Fuzzy rule base, 3). Inference
Engine, and 4). Defuzzification structure. The w j shown
in Fig. 2 represents the weight from the hidden layer to the
output layer.

- 596 -

Fuzzy Rule R

2.3

The FHRCNNs Fitness Function

w1

x1

xp

Fuzzy Rule R

wj

The relation between the FHCNNs inputs and outputs

can be written as:

d
(6)
y k (x t) = f (C id (x t) ) t =1, 2, , n d
Where

Fig. 1 The Fuzzy Neural Networks structure

2.2 FHCNNs mathematical structure

The FHCNNs as shown in Fig. 2 can be expressed as

y k (x t) = w jk m j (x t) T jk

(x t) =

(x t) are the input numbers, and

Ci

(7)

C1 (x t), C 2 (x t), , C p (x t)

(8)

Ci

(2)

nout is the number of output nodes,

m j (x t) = exp  S j[ net j (x t)  net j ]

d
y k (x t) is the expected output value mapped from the

(3)

(M ijmij)

Ci

(4)

net j (x t)=  max(Mij mijxit mijMij xit)

(x t) input.

The object function ( O v ) can be expressed with the

mean absolute percentage error (MAPE) as


(9)
O v 
1 e MAPE 

i 1

(5)

i 1

x t = ( x 1 t ,  x 2 t ,    x pt )
are the training data of the
input networks.
p is the input variables number of dimensions.
x it  ( i 12   p ) is its input variable, M ij and m ij
)

j 1

net j =

y k (x t) = y1 (x t), y 2 (x t), , y nout (x t)

in which

are the upper-bound value and lower-bound value between

the i -th input and the j -th hidden layer nodes. j =1,
2, , n h , n h denotes the number of the hidden layer

 y k  ( x t )  y k ( x t )
d

 T

MAPE  

n out 

n out

k 1

n d 

 

nd

(10)

y k ( x t )
d

t 1

where n d is the number of input-output data pairs.

The fitness function, O v , is mainly used to judge if we

sensitivity factor of the membership function. w jk is the

can make the networks MAPE arrive its minimum by using

the QGA method to adjust the value of the network parameters [  w jk T jk  S j M ,m ij ] .

weight from the j-th hidden layer node to the k -th output
node. k = 1, 2, , n out T jk is the bias from the j-th

III. THE BASIC ARCHITECTURE OF THE PROPOSED QGA-FNN

METHOD [17]

nodes. m j is the membership function value. S

hidden layer node to the

is the

-th output node.

In p u t

x1t

x pt

x2t

H id d e n
la y e r
m

( x )

y1 t y 2 t

ij

( x  )

o u tp u t

( x  )

3.1 The integrate of QGA and FNN

The proposed QGA-FNN method, the flow chart of

which is shown in Fig. 3, applies the Quantum Genetic Algorithm method proposed in this paper to search for each
parameters optimal value for the fuzzy neural networks.
Said parameter values are then inputted to the fuzzy neural
networks to perform the load forecasting.

yk t

Fig. 2 The architecture of the FHCNNs

- 597 -

QGA is a probabilistic algorithm similar to EA. The algorithm

flowchart
is
shown
in
Figure
1.
t
t
t
t
,
(
h
=1,
2,
,
l
)
where
h
is
H ( t ) Q , Q , ..., Q ..., Q

^ q , q , ..., q ..., q ` ,

the size of the population, Q ( t )

l

where n is the number of generator unit, t is the evolution

generation, q t is the binary coding of the generation volj

Fig. 3

The flow chart of the QGA-FNN method

3.2 The concept of quantum genetic algorithm

Quantum Evolutionary Algorithm (QEA) developed

from the basic quantum information science is an evolutionary algorithm based on the concept of quantum computing. It incorporates concepts such as superposition state
etc. in quantum computing and adopts unique coding format
to achieve better experimental results.
!
3.3 Steps to implement the quantum genetic algorithm for
economic dispatch!
Step 1: Initialize population

ume of the jth generator unit. Its chromosome is defined as

follows:
!
t
t
t

.
.
.
t
D
D
D
m
1
2
!
(13)

q j =
t
t
t

.
.
.
E m
E 1 E 2
!
)k!=1, 2, , n*)m is the length of the quantum chromosome* During the initialization of

The smallest information unit in QGA is a quantum bit.

The state of a quantum bit can be either 0 or 1, which can be
represented as:
!!!!!!!!!!
! < = D 0 + E 1 !!!!!!!!!!!!!!!!!!!

2). Quantum chromosome

Frequently used coding methods in EA are binary,
decimal, and symbolic coding. In QGA, a new coding
method based on quantum bit is adopted, i.e. using a pair of
complex numbers to define a quantum bit. A system with m
quantum bits can be described as
!

D
E

In the equation,

.....

.....

D i

D
E
2

m
m

E i

(i=1, 2, , m) in

Q l (t )

and all the q t are

j

initialized, it means that all the possible linear superposition

states will occur with the same probability. During the step

S (t ) from H (t ) , a common solution

set S (t ) is generated by observing the state of H (t ) ,
of generating

the
tth
generation,
t
t
t
t
t
t
t
S (t ) P1, P 2, ..., P h..., P l , Pl (t ) = x1 , x2 ,..., x j..., xtn , every
t
( j =1, 2, , n) is a series, ( x , x , ..., x ..., x ), of

in

length m, which are obtained from the amplitude of quan2

2
or
( i =1, 2, , m). The corretum bit
t
t
D i
E i

(12)

= 1 (i = 1, 2, ,

m). This representation method can be used to express any

linear superposition of states.

0, otherwise.
Step 3: Perform individual measurement to each object in

S (t ) .

ability for the quantum bit to be in the 0 and 1 state respectively.

H (t ) , if

sponding process in the binary situation is to generate a [0,1]

number randomly. take 1 if it is larger than
; take
t

(11)

In the equation, D E are two complex numbers representing the probability of occurrence for the corresponding
state: ( D 2  E 2 = 1 ) D 2 , E 2 represent the prob-

D
E

where

1).!Quantum bit

Use a fitness evaluation function to evaluate each individual object in S (t ) and keep the best object in the generation. If a satisfactory solution is obtained, stop the algorithm; otherwise, continue to the 4th step.
Step 4: Use a proper quantum rotation gate U ( t ) to update

S (t )
The traditional genetic algorithm uses mating, and mutation etc. operations to maintain the diversity of the population. Quantum genetic algorithm applies logic gate to the
probability amplitude of quantum state to maintain the diversity of the population. Therefore, here, a quantum rotation gate of quantum logic gate is adopted for the new
quantum genetic algorithm.

Step 2: Perform coding and individual measurement for the

generating units in the population

- 598 -

cos T
sin T

- sin T
cos T

(14)

Where, T is the quantum gate rotation angle. Its value

is taken as
!!!! T

k f (D

i ,E i

) !!!!!!!!!!!!!!!!!!!

(15)

We define k as a variable related to the generation of the

evolution so that it will adjust the size of the grid
t
self-adaptively. k S exp ( 
) . Where, t is the

iter m ax

evolution generation, S is an angle, iter m a x is a constant depending on the complexity of the optimization
problem. The purpose of the function f ( D i , E ) is to
i

make the algorithm search along the optimal direction. The

search strategies listed in Table 1 are adopted here. Its principle is to make the current search solution approach the
optimal solution gradually and, thereby, determine the direction of the quantum rotation gate. In Table 1, D 1 and
E 1 are the probability amplitude for the optimal solution

found, d
E

u E

,
1

ta n

1

) ,

and

is the probability amplitude for the current solution,

u E

ta n

1

. When d

and

are all larger than 0 at the same time, It means that the

current solution and the optimal solution found are both in

the first or the third quadrant. When [
, it
! [
1

shows that the current solution should be rotated counterclockwise, which has a value of +1; otherwise, its value
should be -1.
Table 1. Look up table for

!!!! d

boe!d

d 1 ! 0 !!! d

! 0 ![

f (D i, E i )
!

!
2

IV. EMPLOYING THE QGA-FNN METHOD AS A MODEL FOR

LOAD FORECASTING

Each neural network is equipped with one output node,

representing the next hours load forecasting values, the
network architecture of which is shown in Fig. 4. In forecasting, there are four vital input data, namely 1). The
loads historical data, 2). The temperature data, 3). The collected rainfall data, and 4). Holiday data collected. In this
network for STLF, we use 83 input nodes--the first 24
nodes representing the 24 hours load values prior to the
forecast day, Noses 25 to 48 representing the 25th to 48th
hours' load amount prior to the forecast day, Noses 49 to 52
representing the load value of the same date last week,
Noses 53 to 76 representing the load values of the same
date two weeks ago, Nodes 77 to 81 standing for the rainfall
forecast index, and Nodes 82 to 83 denoting the holiday index values. We mainly use off-line network training and
only proceed with on-line load forecasting after the off-line
network training is completed to save the actual load forecasting time. The actual Quantum Genetic Algorithms
neuron architecture is shown in Fig. 4.

function

V. SIMULATION AND RESULTS

f (D i , E i ) !
1

In order to solve the problem of QGA being prone to

be trapped in local extreme value better, we perturb the
population. It is found by using the QGA analysis that
when the best individual of the current generation is a local
extreme value, it is very hard for the algorithm to extricate
itself. Therefore, when the best individual does not change
in successive generations, the algorithm is trapped in the
local extreme. At this point of time, a perturbation should
be applied to the population to extricate itself out of the local optimal and start a new search.

!!!!True

!!!True!

+1!

- 1!

!!!!True

False!

+1!

+1!

!!!!False

True!

- 1!

- 1!

!!!!False!

False!

- 1!

+1!

The training data used in this study covers the period

from August 5, 2007 to August 4, 2008. The number of
trainings adopted for the network training is 500. The actual
load forecasting period starts from September 9, 2008 till
September 8, 2009. The related equations adopted are:

yk (x t ) = w jk m j (x t ) + T jk

In this way, the procedure to apply quantum rotation gate

to all the probability amplitude of individual object in the
population, i.e. using quantum rotation gate U (t ) to update
as:

where
the parameters are defined in (2) to (5),
J =16, 22, 28, 34, 40, or 46,
k =1, 2, , 24.

S (t ) , in quantum genetic algorithm can be expressed

!

S (t  1) U (t ) u S (t ) !!!!!!!!!!!!!!!!!!!!!(16)
In the above equation: t is the evolution generation,
U ( t ) is the tth generation quantum rotation gate, S (t ) is
the tth generation probability amplitude of a certain object,
S ( t  1 ) is the t+1th generation probability amplitude of
the corresponding object.

(14)

j 1

The MAPE value is calculated by

 y k  ( xt ) y k ( xt )

d

MAPE  



  
24 k 1 n d  t 1
24

nd

(15)

y k ( xt )
d

We utilize QGA to calculate [w jkT jkS jM ,mij] .

ij
After the FNNs accept the solution [w jkT jkS jM ,mij]
ij

Step 5: Perturbation

delivered upon the completion of the QGA trainings, the

FNNs will proceed with the load forecasting. The QGAs

- 599 -

main task is to train and adjust the parameter values,

[w jkT jkS jM ij ,mij] , in the FNNs. In this study, we
adopt Matlab language to write our programs needed; the
software codes are executed by a regular PC for STLF
simulation. As mentioned, we set up seven different types
ANNs, each representing a day in the week. Each FNN is
equipped with one output node, representing the next hour
forecast output in each day. The plottings for five actual
days load forecasting are shown in Figs. 5 through 6, which
include one regular day, one holiday. In each plotting, the
solid line denotes the actual load series, while the dotted
line stands for the forecasted one.

loads(M W )

30000
25000
20000
15000
10000
1

Next One Hour Load

Forecasting Output

9 11 13 15 17 19 21 23
hours

Fig. 5 The forecast results for a regular work day (September 17,
2008, Wed.)

Output

predicted
actual

predicted
actual

25000

24



 



Past 24
hr loads

25

48



 



Past 48 hr
loads

49

28

52 53

76 77

81

82

Hidden
Layer

loads(MW)

20000

15000

10000

83 Input




 


 

 

 


 

 


 
Temp.max,
Present
24
hr
Rainfall
Holiday
Temp.min
Temp. Forecast Index
Index
(d-1)

11

13 15

17 19

hours
Fig. 6 The forecast results for a national holiday in Taiwan (October 10,
2008, Fri.)

Fig. 4 The Quantum Genetic Algorithm networks neuron architecture

There exist improvement rates, on the errors, of 43.4,

38.8 27.9 and 6.2, respectively, over the ANN, the
EP-ANN the GA-ANN and the QGA-ANN methods for the
proposed QGA-FNN method. The STLFs errors by the
proposed method lie approximately at 1.237(minimum) to
1.392(maximum), well meeting the required criteria of this
method.The STLFs MAPE values by the proposed
QGA-FNN method for typical weeks and atypical weeks
for the four seasons in a full year, starting from Sept. 2008
till Sept. 2009, are tabulated in Table 3, where a typical
week stands for no rainy days, typhoon days, national holidays, and the like, in that week.

The STLFs MAPE and Max.MAPE values obtained

from different methods, including the ANN, the EP-ANN,
the GA-ANN and the proposed QGA-FNN methods are
tabulated for comparison in Table 2. Among them, the ANN
method is a basic NN trained by Back-Propagation (BP).
The EP-ANN method adopts a regular ANN, as its base,
trained by EP for its network parameters. The GA-ANN
method adopts a regular ANN, as its base, trained by the
GA for its network parameters. The QGA-FNN method is
the key one newly proposed in this paper. The period for
Table 2 starts from Sept. 12 till Oct. 31, 2008, where the
MAPEs, respectively, are: ANN=1.85, EP-ANN=1.79,
GA-ANN=1.65, QGA-ANN=1.37 and QGA-FNN=1.29.

Table 2 The MAPE and Max. MAPE obtained by different NN methods for various day nature.
ANN

Day
Best

Nature
Work day
Non-work-day
Natl holiday
Rainy day
Avg.

Max.

EP-ANN
Best

Max.

GA-ANN
Best

Max.

QGA-ANN
Best

MAPE MAPE MAPE MAPE MAPE MAPE MAPE

1.74
1.86
1.87
1.93
1.85

2.29
2.23
2.17
2.39
2.27

1.67
1.74
1.79
1.93
1.79

2.05
2.17
2.09
2.25
2.14

21 23

1.58
1.55
1.72
1.73
1.65

- 600 -

1.94
2.05
2.17
2.13
2.08

1.27
1.45
1.37
1.38
1.37

QGA-FNN

Max.

Best

Max.

MAPE

MAPE

MAPE

1.98
2.17
2.08
2.02
2.06

1.27
1.29
1.25
1.37
1.29

1.97
1.85
1.93
1.99
1.94

Table 3 The STLFs MAPE values by the proposed QGA-FNN method for typical weeks and atypical weeks for the
four seasons in a full year (Fall- Sept. till Nov. 2008; winter- Dec. 2008 till Feb. 2009; spring- Mar. till May 2009;
summer- June till Aug. 2009)

Date

Spring
Typical Atypical
Week
Week

Mon.
Tue.
Wed.
Thur.
Fri.
Sat.
Sun.

1.325
1.278
1.368
1.338
1.239
1.337
1.346

1.337
1.355
1.392
1.362
1.333
1.359
1.358

Summer
Typical
Atypical
Week
Week
1.341
1.276
1.336
1.321
1.256
1.321
1.339

1.352
1.345
1.342
1.325
1.336
1.352
1.353

VI. CONCLUSION
An accurate load forecasting can lead to accurate power
generation amount, save energy and simplify energy management work, making the entire power system operation
more efficient. For a better STLF, the particle swarm optimization method is first introduced in this paper; this method
encodes all the networks weights and biases into several
FNN system particle swarms. Then, we train the network
parameter values using the QGA method proposed in this
paper to locate the networks optimal parameter solution and
finally, resolve the optimal STLF with such networks derived.
The strength of the proposed QGA method lies at its simplicity and briefness, resulting in easy feasibility without the
need of many parameter adjustments, and gradient infortion
is not required.
This study resolves 24-hour solutions, as an example,
using the current load of Taipower in Taiwan for optimal
load forecasting and compares our forecast results with that
of other often-used methods. The comparisons indicate that
the proposed method significantly improves the error rates
over other often-used methods, ranging from 6.2% to
43.4%-- a sign demonstrating the proposed methods advantage in load forecasting.

VII. REFERENCES

[1] T. Haida and S. Muto, Regression bsed peak load forecasting using a
transformation technique, IEEE Trans. On Power System, Vol. 9,
pp.1788-1794, Nov. 1994.
[2] Charytoniuk W, Chen M. S, Van Olinda P. Nonpparametric regression
based Short-Term Load Forecasting, IEEE Trans. On Power System,
Vol. 13, pp-725-730, Mar. 1998.
[3] Akagi H. New trends in active filter for power conditioning, IEEE
Trans. On Power Electronics. Vol. 32, pp.1312-1322, Jun. 1996.
[4] Mastorocostas P. A, Theochairs J. B, Bakirtzis A.G, Fuzzy modeling
for short-term load forecasting using the orthogonal least squares
method, IEEE Trans. On Power System, Vol. 14, pp.29-36, Jan. 1999.
[5] A. Khotanzed, R. C. Hwang, A. Abaye and D. Maratulam, An Adaptive Moddular Artificial Neural Network Hourly Load Forecaster and
its Implementation at Electric Utilities , IEEE Trans. on Power Systems, Vol. 10, No. 3, pp1716-1722, August, 1995.
[6] H. Yoo, R. L. Pimmel, Short-term Load Forecasting Using a
Self-Super Vised Adaptive Neural Network , IEEE Trans. on Power
Systems, Vol. 14, No. 2, pp.779-784, May 1999.
[7] J.C. Sousa, L. P. Neves and H. M Jorge, Load Forecasting Based on
Neural Networks and Load Profiling , IEEE PowerTech, pp.1-8, June
28 2009-July 2 2009.
[8] V. H. Ferreira and S. A. P. Alves da Silva, Toward Estimating
Autonomous Neural Network-Based Electric Load Forecasters , IEEE

Fall
Typical
Atypical
Week
Week
1.331
1.237
1.287
1.313
1.312
1.341
1.338

1.352
1.289
1.345
1.331
1.323
1.357
1.338

Winter
Typical
Atypical
Week
Week
1.272
1.285
1.273
1.332
1.266
1.322
1.325

1.321
1.327
1.305
1.343
1.311
1.355
1.345

Trans. on Power Systems, Vol. 22, Issue 4, pp.1554-1562, Nov. 2007.

[9] G. C. Liao and T. P. Tsao, Application of a Fuzzy Neural Network
Combined with a Chaos Genetic Algorithm and Simulated Annealing
to Short-Term Load Forecasting, IEEE Transactions on Evolutionary
Computation, Vol. 10, No. 3, pp. 330-340, June 12 2006.
[10] A. Khotanzed, R. C. Hwang, A. Abaye and D. Maratulam, An Adaptive Moddular Artificial Neural Network Hourly Load Forecaster and
its Implementation at Electric Utilities, IEEE Trans. on Power Systems, Vol. 10, No. 3, pp1716-1722, August, 1995.
[11] H. Yoo, R. L. Pimmel, Short-term Load Forecasting Using a
Self-Super Vised Adaptive Neural Network, IEEE Trans. on Power
Systems, Vol. 14, No. 2, pp.779-784, May 1999.
[12] C. N. Elias and N. D. Hatziargyriou, An Annual Midterm Energy
Forecasting Model Using Fuzzy Logic, IEEE Trans. on Power Systems, Vol. 24, Issue 1, pp.469-478, Fed. 2009.
[13] J. Gaul, E. Handschin and C. Lehmkoster, Establishing a Rule Base
for a Hybrid ES/XPS Approach to load Management, IEEE Trans. On
Power Systems, Vol. 13, No. 1, pp.86-93, 1998.
[14] S. Sailesh Babu, D.B. Das and C. Patvardhan, Real Parameter Quantum Evolutionary Algorithm for Economic Load Dispatch, IET
Proc.-Gener. Transm. Distrib., Vol. 2,Issuel 1, 2008, pp. 22-31.
[15] J. Tylavsky, G.T. Heydt, Quantum Computing Power System Simulation, IEEE Power Engineering Society General Meeting 2003, Vol. 2,
2003, pp. 13-17.
[16] Z. Zhisheng, Quantum-behaved particle swarm optimization algorithm for economic load dispatch of power system, Expert Systems
with Applications, Vol. 37, Issue 2, pp. 1800-1803, Mar. 2010.
[17] J. Sun, W. Fang, D. Wang and W. Xu, Solving the economic dispatch
problem with a modified quantum-behaved particle swarm optimization method, Energy Conversion and Management, Vol. 50, Issue 12,
pp. 2967-2975, Dec. 2009.
[18] S. Zhao, G. Xu, T. Tao and L. Liang, Real-coded chaotic quantum-inspired genetic algorithm for training of fuzzy neural networks,
Computers & Mathematics with Applications, Vol. 57, Issue 11-12, pp.
2009-2015, Jun. 2009.
[19] L. D. S. Coelho and V. C. Mariani, Reactive power and voltage control based on general quantum genetic algorithms, Expert Systems
with Applications, Vol.36, Issue 3, Part 2, pp. 6118-6126, Apr. 2009.

- 601 -

VIII. BIOGRAPHIES
Gwo-Ching Liao (M04) He received his
MS-EE from the National Cheng Kung
University, Tainan, Taiwan in 1991 and the
Ph.D degree from National Sun Yat-Sen
University, Kaohsiung. He works on
Department of Electrical Engineering, Fortune
Institute of Technology, Kaohsiung County,
Taiwan. His interests are deregulation, power
system operations and AI application in power
system.