Control En&.Practioe, Vol. 2, No. 1, pp. 113-121,1994 0967-0661/94 $6.00 + 0.

00
Printed in Great Britain. All fights re*erred. ~ 1994 Pergamonl ~ s s Ltd

NEURAL NETWORKS AND FUZZY LOGIC IN POWER ELECTRONICS

Bor-Ren Lin* and R.G. Holt**
*Department of Electrical Engineering, National Yunlin Institute of Technology, Touliu, Yunlin, Taiwan
**Department of Electrical and Computer En&insering, University of Missouri, Columbia, Missouri, USA

Abstract Power electronics has developed rapidly during recent years. It involves the technologies of power
semiconductor devices, convener circuits, electrical machines and VLSI circuits, as well as advanced control
theory. This paper reviews neural network and fuzzy logic approaches in buck converters, inverters, induction
motor drives and DC motor drives. Neural network and fuzzy logic applications are presently one of the most
active r~arch areas in power electronics.

Key Words. Neuralnets; Fuzzycontrol;Powerelectronics;Powersystemcontrol;Conveners;Electricmotor

1. INTRODUCTION parallel computation. There has been some research

on using neural networks for the control of dynamic
Power electronics integration promises to reduce the systems. In particular, Psaltis et a l . (1987) recently
cost and increase the reliability of future motion studied a number of interesting techniques for using
control systems. The PIC (power integrated circuit) the backpropagation method to control a nonlinear
makes it possible to combine substantial power plant.
handing and sophisticated control functions on the
same silicon chip. There are several PIC chips Fuzzy control has been one of the most active and
available for applications in step, DC brushless, DC fruitful areas for research in the applications of fuzzy
and AC motor drives. Sophisticated controls are also set theory during the past several years, especially in
required in a number of other power electronic the area of the industrial processes which lack
applications, such as switching mode power quantitative data regarding the input-output relations.
supplies, induction heating inverters, electronic Fuzzy logic is a logical system which is much closer
welding, active power line filtering and static var to human thinking and natural language than
compensation. These applications involve DC-DC, traditional logical systems. The fuzzy logic
DC-AC, AC-DC and AC-AC power conversion mad controller provides a means of converting a
control with power switching devices. linguistic control strategy based on expert
knowledge into an automatic control strategy.
Artificial neural networks is an area of research Basically it provides an effective means of capturing
which in recent years has aroused much interest in the approximate, inexact nature of the real world.
fields as disparate and varied as cognitive science, The essential part of the fuzzy logic controller is a
physiology, psychology, computer science and set of linguistic control rules related by the dual
engineering. In contrast to other machine learning concepts of fuzzy implication and the compositional
techniques, neural networks can modify their rule of inference. The methodology of the fuzzy
behavior in response to the environment, have the logic controller appears very useful when the
flexibility of easily handing different problem sizes processes are too complex for analysis by
and have the potential for hardware implementation. conventional quantitative techniques, or when the
Neural networks model a large number of neurons available sources of information are interpreted
that are dynamically coupled. Neural network qualitatively, inexactly or uncertainly. Thus, fuzzy
architectures have the capability to represent logic control may be viewed as a step toward a
nonlinear system models. The)' can handle large method between conventional precise mathematical
amounts of sensory information, perform collective control and human-like decision making.
processing and learning and provide powerful
processing capabilities with great potential for highly This paper briefly reviews the recent neural network

113
114 Bor-Ren Lin and R.G. Hoft

and fuzzy logic controller advances in power Each connection has associated with it a weighting
electronics for buck converters, inverters, DC motor which acts to modify the signal strength. There are
drives and induction motor drives techniques. various techniques for optimizing criterion functions
to train the neural network. One important
2. NEURAL NETWORK CONCEPTS characteristic of neural network classifiers is that
T h e r e are many artificial neural network their training usually requires iterative techniques.
architectures that have been proposed. One The backpropagation classifier is the most popular
architecture has been predominant; the feedforward technique trained by using the gradient descent
neural network (FNN). The standard neuron method. The advantage of this method is that it is
structure illustrated in Fig. 1 is adopted which is simple and easy to understand, but the disadvantage
comprised of a summer and a logistic function f( net is. that convergent speed is slow and the
p,i ) that can be either a sigmoid or a linear function. backpropagation classifier is not so robust.

The backpropagation algorithm is now introduced.

In the training phase, the weights are updated using
Op-1,I ~.~, I .

the standard backward error propagation algorithm,

netp,i also known as the generalized delta rule. The system
first uses the input vector to produce its own output
Op_t,3 P' I Op,i
vector and then compares this with the desired
output, or target vector. If there is no difference, no
learning takes place. Otherwise the weights are
changed to reduce the difference. Now the problem
//Tr p,i,n ep,i is to minimize the total mean-squared error E given
,. Op-l,n by

n
Fig. 1. Structure of an elementary neuron
E = 1/2" ~ (tL, k -OL, k )2 (3)
The equation for the i-th neuron of the p-th layer k-1
structure is
where t t.,k is the k-th target output at the output
layer, O I_,k is the k-th actual output at the output
net p,i = ~ Wp,i,k Op-l,k + 0p,i (1)
layer. This error can be minimized by taking partial
k=l
derivatives of E with respect to each weight. In
Rumelhart (1986), the generalized delta rule used is
Op,i = f( net p,i ) (2)
Wp,j, i = Wp,j, i + A Wp,j, i (4)
where Op,i is the output
Op_l, k is the k-th output at the (p- 1)-th layer
~ , j = ( tp,j - Op,j ) * f'( neb,,i )
Wp,i,k is the weight from the k-th input of (p-
,for output layer (p=L)
1)-th layer to the i-th output of the p-th layer n
Op,i is the bias.
or ~.~ ( t~p÷l,k Wp÷l,k, j ) * f'( netp,j ) (5)
k-1
These neurons are organized in layers as shown in ,for iudden layer (p-L)
Fig. 2. Scaled data enters the network at neurons of
the input layer and is propagated to the output
through intermediate layers. where A Wp,j, i iS the weight change between the
j-th neuron at the p-th layer and the i-th neuron at
the (p-l)-th layer. To get the correct generalization
LS~ler I 2 L- I L of the delta ride, A Wp,j, i is proportional to -iiE/i~
Wp,.j,i . In order to implement gradien! descent Ior E
, weight changes sholdd be made according to
0L,2
A Wp,.j,i = 'rI 6p,j Op_l, i (6)

where r I is the learning rate. For practical purposes,

r I is chosen as large as possible without leading to
oscillation. One way to increase the learuing rate
Fig. 2. Feedforward neural network architecture without leading to oscillation is to modify the
 Neural Networks and Fuzzy Logic 115

generalized delta rule to include a momentum term. Let A and B be two fuzzy sets in U with membership
This can be accomplished by the following rule function PA and/~B respectively where U is the
universe of discourse. The universe of discourse is a
A Wp,,j,i (k + 1) = r I 6p,j Op_l,i + ct A Wp,j,i (k) (7) collection of fuzzy variable possible values {x}.

where r I is the learning rate and ct is a constant Definition 1: Union"

which determines the effect of past weight changes The membership function Y(ALIB) of the union
on the current direction of movement in the weight ALJB is defined for all x ~ U by
space. If f(.) is a logistic function,
/~(AUB)(X) = max{/~A(X),/~B(X) } (9)
i.e., f(netp,j) = 1/( 1+ exp( -netp,j ) ),
Definition 2: Intersection
then f'( netp,j ) = f( netp,j ) * ( 1 - f( netp,$ ) ) The membership function P ( A N B ) o f the
= Op,j * ( 1- Op,j ) (8)
intersection Af')B is defined for all x ~ U by
The above generalized delta rule is used to update ]~(Af-)B)(X)= min{ PA(X),/,~B(X) } (I0)
the neural network weights until the total output
error is less than a defined small value.
Definition 3: Complement
The membership function/~ of the complement
3. FUZZY CONTROL CONCEPTS of a fuzzy set A is defined for all x ~ U by

Fig. 3 shows the basic closed-loop fuzzy control /ai~(x) = 1- yR(x) (11)
configuration. Since the converter produces
nonfuzzy measurements, these have to be fuzzied by Fig. 4 shows these three basic fuzzy set operations.
fuzzification. Similarly, since the converter cannot
respond directly to fuzzy controls, the fuzzy control
sets generated by the fuzzy algorithm have to be
defuzzified by defuzzification. The fuzzy control
p.(x)
algorithm consists of a set of fuzzy control rules
which are related by the concepts of fuzzy
implication and the compositional rule of inference.

Fuzzylosiccontroller
Reference
signal ¢ I II
.'-'~-$r'~ "- ti H Rule-basedfuzzy
' ~ Fuzz]nea on~ ¢ontrvlalgorithm Defuzzifieation[
"~:J. i
v~X
|

GR i p RuB(X)

Plant IA u
I-

Fig. 3. Closed-loop control configuration Lf~ IJ-AnI~(x)

These fuzzy control rules are combined by using the
sentence connectives and and also. A fuzzy variable
is expressed by natural language. For example, the
error of the output voltage can be defined by
linguistic variables "zero (Z)", "positive small (PS)",
\//
"positive medium (PM)", "positive big (PB)",
"negative small (NS)", "negative medium (NM)",
"negative big (NB)", etc. The basic fuzzy set Fig. 4. Basic fuzzy operations • Union, Intersection
operations are union, intersection and complement. and Complement
116 Bor-Ran Lin and R.G. Holt

Normally, the fuzzy rule has the if .. then., structure, fuzzy rules defined as
such as
RI: ifx is L1 and y is M1 then z is N1
If x is L and y is M then z is N R2: if x is L2 and y is M2 then z is N2
R3: if x is L3 and y is M3 then z is N3
where x and y are input fuzzy variables, z is the
output fuzzy variable; L, M and N are fuzzy subsets
in the universe of discourses X, Y and Z Rn: if x is Ln and y is Mn then z is Nn
respectively. A fuzzy subset can be NB, NM, NS, Z,
PS, PM and PB. Fig. 5 shows the basic fuzzy These individual rules are combined by using the
partition of membership functions for error, change union operator.
of error and change of control action. TaMe 1 shows
the fuzzy model based on fuzzy rules. There are n R = R I U R2UR3 U ........ U R n

A fuzzy controller should compute the actual output

control signal for a specific input signal from a given
set of fuzzy rules. The compositional operator is
applied to infer the output from the given process
states x, y and the fuzzy relation R.

(a) err z=yo(xoR) (12)

Y err
where o is sup-star composition. Since the control
action z cannot directly control the converter, the
defuzzification operation can be performed by the
(b) err center of gravity method in Fig. 6,

Yerr
~ P B h.~

(c) err

Fig. 5. Basic fuzzy membership functions (a)

Triangular membership function (b) Trapezoidal
membership function (c) Bell-shaped membership
function
Fig. 6. Basic fuzzy reasoning methods

The centroid defuzzification method is uniquely

determined from the fuzzy response set R as follows
Table 1. Fuzzy model
(Table entries are change in control output)
error z, - f z , t (z)dz
(continuous case) (13)
•~error
NB NB
NB N M
NB
NS Z P5 P M
NB NB NM
PB
NS Z
f. (z)dz
NM NB NB NB NM NS Z PS
NS NB NB NM NS Z PS PM
Z NB NM NS Z P$ PM PB or z.= iXt~c,(z,) (discrete case) (14)
PS NM N$ Z PS PM PB PB I

PM NS Z PS PM PB PB PB
PB Z PS PN PB PB PB PB Then the output signal zo is applied to control the
converter.

The principal fuzzy control algorithm is described as

NB: Negative Big , NM: Negative Medium, NS: follows :
Negative Small, Z: Zero, PB: Positive Big, PM: I. Sample the output signal of the plant
Positive Medium, PS: Positive Small 2. Calculate the error and change of error
 NeursINetwm'kssad Fuzzy Logic 117

3. Determine the fuzzy subset and membership will keep two output waveforms of PS from
function for error and change of error overlapping. If the two output waveforms from PS
4. Determine the change of control action overlap, the upper and lower transistors in one
according to the individual fuzzy rule inverter leg will conduct at the same time, which will
5. Calculate the actual change of control action damage the transistors. Finally, as expected, the
by def-Tzification operation three current output waveforms of the VSI will
6. Send the change of control action to control the follow the three sinusoidal reference currents.
converter
7. Goto step I

4. NEURAL NETWORKS IN POWER

I I_T
B .ECrRONICS

Two applications of neural networks in power

+
electronics are described in Lin (1993a)and Induc*~n
Weerasooriya (1991). These authors apply neural
networks to current regulation of inverter drives
(Lin, 1993a)and DC motor drives (Weemsooriya~
1991). i¢

4.1. Current Ret.~nlationof Inverter Drives (Lin,

1993)

Fig. 7 shows the neural network inverter control. The IV- l,a . r, f
input signal to the inverter is an analog signal such
as current error or voltage error. The output signal of O u t p u t L ~ e r ' ~ InputLa~er G :Gain
P$ : Pulse Separation
the inverter is a binary signal. Therefore, the HiddenL~er
controller has the properties of a nonlinear function
to map the analog input into binary output. Neural Fig. 7. Neural network inverter control
network architectures have the capacity to learn
nonlinear system models. They can perform A simulation and comparison of fixed band
collective processing and learning and provide hysteresis, sinusoidal band hysteresis and neural
powerful processing capabilities with great potential network controls for Fig. 7 is included in Lin (1993).
for highly parallel computation. In the neural The total harmonic distortion (THD) in the motor
network method, training is required to learn current for these three control methods is shown in
something about the plant behavior. The inputs to Fig. 8. The neural network control method has lower
the neural network are three-phase current errors and THD in all cases except when the hysteresis band is
the outputs are the voltage vectors. Current errors very small. The neural network
can be randomly generated at the neural network
input and the backpropagation method is used to
update the weights so as to decrease the current 5 THD
errors. The three sinusoidal reference currents are
i a , r , f , ib,r~, f , i t , r e f. Reference c u r r e n t s ib,ref and 4
it,re f are phase shifted 1200 and 2400 respectively • Fixed-band
from i=,re f. First, the three-phase currents from the i• Smwoided-band
3
VSI are measured and compared with the three
reference currents. The error signals are multiplied • NeulalNetwork
by a given coefficient, G. and then input to the
2
neural network. The neural network is trained to
have minimum output error. The training rule is I

i f I ira,re f - im I < c =¢' k e e p O m at t h e s a m e state o

if ~,r~-~ > t =* let Om --- I 0.01 0.04 0.07 0.1 0.13 0.16 0.19 0.22 0.25
if im,rt,f - i m <-~ =~letOm=0 (15) hysteresis band (A)

where m = a, b or c, and Om is the output of the Fig. 8. Comparisons of the total harmonic distortion
neural network. PS is the pulse separation circuit that
118 Bor-Ren Lin and R.G. Hoft

also has fast processing speed and fault tolerance. system to reduce harmonic voltage distortion in an
The fault tolerance property will allow the neural electrical power system was proposed by Akira
network to work well even if one of three current Kumamoto (1992). Fig. 10 shows the learning
sensors failed. Thus, neural networks appear to have control algorithm for the active filter control
significant advantages in power electronic incorporating neural networks so as to automatically
applications. adjust the resonant network to minimize the
harmonic current drawn from the power source.

4.2. DC Motor Control (Weerasooriya ,1991)

Weerasooriya (1991) proposed an artificial neural- Line Impedance __

network-based high-performance speed control
system for a d c motor. The unknown nonlinear
dynamics of the motor and load are captured by an
artificial neural network. The neural network
identifier is trained with a desired reference model to
achieve trajectory control of speed. The objective of
a speed controller is to manipulate the terminal
voltage in such a manner as to make the rotor speed
follow a specified trajectory with minimum
deviation. One of the main difficulties with
conventional tracking controllers for electric drives
Fig. 10. System configuration for harmonic voltage
is their inability to capture the unknown load
distortion reducing
characteristics over a wide range of operating points.
The ability of neural networks to learn a large class
Neural network techniques have also been applied to
of nonlinear functions is well known. The neural
forecast the short-term load for a large power system
network can be trained to emulate the unknown
(Lee, 1992). The main objective of short-term load
nonlinear plant dynamics by presenting a suitable set
forecasting is to predict the hourly loads, one day or
of input/output patterns generated by the plant. Fig. 9
even one week beforehand, which are necessary for
shows the basic concept of identification and control
the power system. The electric load demand is a
of the dc motor using a neural network. The scheme
function of weather variables and customer
is very similar to indirect model reference adaptive
requirements. The neural network approach is used
control. First, the motor is identified from its input
to forecast the short-term load on the Taiwanese
and output variables using neural networks. The
power system (Ho, i992). T o determine the
weights from the trained neural network identifier
c o n n e c t i o n weights b e t w e e n neurons, the
are then used in the controller to calculate the
conventional back propagation learning technique
terminal voltage which will asymptotically drive the
with different learning rates is employed in the
motor shaft speed wp(k) towards the reference model
training process. Inputs to the artificial neural
output wm(k ).
network (ANN) are past loads and the output of the
ANN is the load forecast for a given day.

Fault detection and diagnosis of power systems

using artificial neural networks is presented by
Swarup (1991). The model power system employed
for fault diagnosis is shown in Fig. 11. The power
system model consists of three buses, two
-, ........rFlIde~,fl,~uonMo~] ~- I"£%a) transformers and two transmission lines. The
J + protective elements of the power system consists of
circuit breakers and relays. The input to the neural
network consists of the status of the relays and
circuit breakers of the power system. The neural
network is designed so that it captures the behavior
Fig. 9. Neural-network-based identification and of the power system represented in the form of
control system signals from relays and circuit breakers. An input
pattern corresponding to a particular fault condition
is fed to the neural network and the desired output
5. NEURAL NETWORKS IN POWER SYSTEMS pattern corresponding to the fault condition is
impressed at the output. The neural network is
Recently a neural network active AC harmonic filler trained until the error in the weights between
 NeuralNetwor~ and Fuzzy Logic 119

successive iterations is less than a specified converters in Lin (1993b),The basic system
minimum value. When an unknown output pattern configuration is shown in Fig. 12. The performance
corresponding to the fault condition is given at the of fuzzy logic control is much better than that using
output layer of the neural network, the diagnostic conventional methods (Lin, 1993b).
system classifies the typ% duration and severity of
the fault.
$v,Z iL ,,_

A1
A2
BUS A L_ r0
RI rR2
%° I l 1 iRL ~'e
BI
BUS B
B2
fuzzy logic E
--F--- L1 --1-"/.2 controll, r
C1 I
TM gr¢ f
BUS C
C2
Fig. 12. System configuration of buck converter
based on fuzzy logic control
Fig. 11. Model power system for fault diagnosis.
6.2. Current Remdation of Inverter Drives

6. FUZZY LOGIC CONTROL IN POWER The current regulation of inverter drives by fuzzy
ELECTRONICS logic control is shown in Fig. 13. Fuzzy control rules
are designed to control the three-phase sinusoidal
6.1. Buck Converter current from the inverter output. Fig. 14 shows the
three-phase sinusoidal currents delivered by the
The modeling of networks which contain switches inverter using the fuzzy logic method.
has drawn much attention because of the unusual
properties of switches in comparison with other
circuit elements. The difficulties in the modeling of
switched networks are mainly due to the nonlinear
and time varying nature of switches. Most modeling
in power electronics is intended to convert the I 8
nonlinear and time-varying model to an ideal or non-
ideal switch model (Verghese, 1986; Eggers, 1986).
eb
Then, the state space method is used to solve the
state equation for the system. Recently, sliding-
mode controlled DC-DC converters have been
introduced (Cardoso, 1992). The main advantage of
uj(nT)~tb(nT)
~ IM(n~T)
the sliding-mode control is insensitivity to plant
parameter variations and external disturbances,
which leads to invariant dynamic and steady-state Z.+--
response in the ideal case.

Fuzzy set theory is a theory about vagueness or

uncertainty and enables one to use nonprecise, ill- Fig. 13. System configuration for current regulation
defined concepts. The fuzzy control does not need of inverter drives by fuzzy logic method
an accurate mathematical model of a plant.
Therefore, it is applicable to a process where the 6.3. Induction Motor Control
plant model is unknown or ill-defined. The fuzzy
control is also nonlinear and adaptive in nature, Recently fuzzy logic control of the induction motor
which gives it robust performance under parameter has been studied (Miki, 1991; Mir,1992), where
variations and load disturbances. Fuzzy algorithms fuzzy control is used to achieve robusmess against
based on linguistic rules describing the operator parameter variations and load torque disturbance
control strategy are applied to control DC-DC effects. Both coarse and fine control with the help of
120 Bor-Ren Lin and R.G. Hoft

shaft. The fuzzy rule for the estimation of the

6(A) absolute rotor position is implemented with the help
4 of the look-up tables. The basis fuzzy logic controller
for the speed control of a permanent magnet
2 synchronous machine (Azevedo, 1993) is the
0 analysis of the permanent magnet synchronous
machine (PMSM) transient response and fuzzy logic.
-2
The fuzzy controller generates the varations of the
reference current vector of the PMSM speed control
based on the speed error and its change.
-6 , , ; ; , ,

50 I00 150 200 250

T i m e : * 1/19800 (sec) 6.5. Unmterruptible Power Supply Control

A new control method for real-time digital control of

Fig. 14. Three phase sinusoidal current by fuzzy
the single-phase PWM inverter for UPS applications
logic method
is presented by Lin (1993_c).An approach for
combining the deadbeat control and the fuzzy logic
look-up rule tables are used to improve transient
compensator is used. Using the state variables
response. The fuzzy logic controller chooses the
measured at every sampling time, the controller
switching states based on a set of fuzzy variables.
calculates the inverter pulse width for the next
Error in flux magnitude and error in torque are used
sampling time. A fuzzy logic controller is used to
as fuzzy state variables. Fuzzy rules are determined
compensate for the voltage drop due to nonlinear
by observing the vector diagram of flux and currents.
loads. The control scheme is verified for nonlinear
A basic system configuration for induction machine
load conditions, sudden load changes, and system
control based on the fuzzy logic method is shown in
parameter variations. Fig. 16 shows the single phase
Fig. 15.
output voltage control scheme with fuzzy logic
compensator. For the same switching frequency, the
fuzzy logic control scheme works better than two-
Vdc [ Inverter level deadbeat control scheme.
I
Flux reference ~ ~
it..._~
flux_.~
Stato~
flux 7. FUZZY LOGIC CONTROL IN POWER
Stator flux ant SYSTEMS
Torque reference torque estimatoj
Fuzzy Expert system techniques are already used in power
controller Torque ] systems control. Recently, fuzzy logic controllers
applied to power flow control (Miranda, 1992),
power system voltage control (Hsu, 1992), power
Flux angle system fault diagnosis (McDonald, 1992), stability
analysis (Hua, 1991) and stability of electric power
Fig. 15. Fuzzy logic control for induction motor systems (Hassan, 1991) have been proposed. Power
system control based on fuzzy logic has very good
Recently variable-structured robust controller whose performance(Hassan, 1991). New research is
structure is continuously changed by fuzzy logic, anticipated using fuzzy logic applied to power
such that the controller makes the system respond systems for reactive power compensation, harmonic
quickly~ was proposed by Suyitno (1993). The distortion compensation and power factor correction.
algorithm is applied practically to speed control for
an induction sevormotor.
8. CONCLUSION
6.4..Permanent Magnet Motor Control

Control of permanent magnet servo motors and Neural network and fuzzy logic controls have been
synchronous machines have been studied by Kim successfully applied to power electronics aqd power
(1992) and Azevedo (1993). A fuzzy control systems. Both methods can be used for digital
algorithm based on the min-max compositional rule current regulation in motor drives, converters and
of inference is used for the estimation of the absolute power systems, and these controls can be
rotor position which is essential to electical implemented with hybrid analog/digital circuitry.
commutation of the permanent magnet AC servo New research is anticipated using both methods on
motor with an incremental encoder coupled to the power electronic active filter design, harmonic
 Neural Networks and Fuzzy Logic 121

elimination, power factor correction and resonant Lee, K. Y. (1992). Short-Term Load Forecasting
converter control. Using an Artificial Neural Network. IEEE
Transactions on Power Systems, pp. 124-131, Vol.
7. No. 1.
Lin, B.- IL and Richard G. Hoft (1993a). Power
Electronics Inverter Control with Neural Networks.
Proceeding of APEC, San Diego, March 7-11.
Lin, B. - R. (1993b). Analysis of Fuzzy Control
Method. Applied to DC - DC Converter Control.
E , Proceeding of APEC, San Diego, March 7-11.
Lin, B . - R. and Richard G. Hoft (1993c) Real-Time
Fuzzyl,~e Digital Control of PWM Inverter With Fuzzy
corr~enm for
Logic Compensator for Nonlinear Loads.
Fig. 16 Single phase inverter with fuzzy logic To appear: Proceeding of lAS.
McDonald, J. R. (1992). Alarm Processing and Fault
compensator.
Diagnosis Using Knowledge Based Systems For
Transmission and Distribution Network Control.
9. REFERENCES
IEEE Transactions on Power Systems, pp. 1292-
Azevedo, H. R. and K. P. Wong (1993). A Fuzzy 1298, Vol. 7, No. 3.
Logic Controller for Pemmnent Magnet Mild, Ishiro (1991), Vector Control of Induction
Synchronous Machine. Proceeding of APEC, pp. Motor with Fuzzy PI Controller. pp. 341-346,
122-127. Proceeding of 1AS.
Cardoso, B. J. (1992), Analysis of switching MAr, Sayeed A. (1992). Fuzzy Controller For
frequency reduction methods applied to sliding Inverter Fed Induction Machines. pp. 464-471,
mode controlled DC-DC converters. Proceeding Proceeding of lAS.
of APEC, pp. 403-410. Miranda, V. (1992). Fuzzy Modelling of Power
Eggers, G. (1986). Fast switches in linear networks. System Optimal Load Flow. IEEE Transactions
IEEE Trans. Power Electronics, Vol. 1, No. 3, pp. on Power Systems, pp. 843-849, Vol. 7, No. 2.
129-140. Psaltis, D., A. Sideris, and A. Yamamura (1987).
Hassan, M. A. M. (1991). A Fuzzy Logic Based Neural Controllers. First International Conference
Stabilizer For a Synchronous Machine. IEEE on Neural Networks, Vol. 4, pp. 551-558, San
Transactions on Energy Conversion, pp. 407-413, Diego.
Vol. 6, No. 3. Rumelhart, David E. (1986). Parallel Distributed
Ho, Kun - Long (1992). Short - Tem~ Load Processing. MIT Press.
Forecasting Using a Muitilayer Neural Network Suyitno, A., J. Fujikawa, H. Kobayashi, and
with an Adaptive Learning Algorithm. pp. 141- Yasuhiko Dote (1993). Variable-Structures Robust
149, IEEE Transactions on Power Systems, Vol. 7, Controller by Fuzzy Logic for Servomotors. IEEE
No. 1. Transactions on Industrial Electronics, pp. 80-88,
Hsu, Yuan - Yih (1992). Voltage Control Using a Vol. 40, No. 1.
Combined Integer Linear Programming and Rule- Swarup, K. S. and H. S. Chandrasekharaiah (1991).
Based Approach. IEEE Transactions on Power Fault Detection and Diagnosis of Power System
Systems, pp. 744-752, Vol. 7, No. 2. Using Artificial Neural Networks. IEEE First
Hua, Yuan - yih (1991). A Rule - Based Expert International Forum on Applications of Neural
System for Steady-State Stability Analysis. IEEE Networks to Power Systems, pp. 97-101
Transactions on Power Systems, pp. 771-777, Vol. Verghese, G. C., M. E. Elbduk and J. G. Kassakian,
6, No. 2. (1986). A general approach to sampled data
Kim, Dong - II (1992). Control of Permanent modeling for power electronic circuits. IEEE
Magnet AC Servo Motors via Fuzzy Reasoning. Trans. Power Electronics, Vol. 1 No. 2, pp. 76-89.
pp. 482-489, Proceedim? of lAS. Weerasooriya, Siri (1991). Identification and
Kumamoto, Akira (1992). Suppression of Harmonic Control of a DC Motor Using Back - Propagation
Voltage Distortion by Neural Network Controlled Neural Networks. IEEE Transactions on Energy
Active Filter. Proceeding oflAS, pp. 754-761. Conversion, pp. 663-669, Vol. 6, No. 4.