IEEE Transactions on Power Delivery, Vol. 9, No. ... POWER USING NEURAL NETWORKS ... Comparing to the conventional methods, this neural network.
IEEE Transactions on Power Delivery, Vol. 9, No.3, July 1994
1253
MEASUREMENT AND EVALUATION OF INSTANTANEOUS REACTIVE POWER USING NEURAL NETWORKS
T.W.S.Chow and Y.F.Yam
City Polytechnic of tiong Kong, Hong Kong
ABSTRACT
-
accurately.
Overall success depends on accurate, and f a s t
The e r r a t i c disturbance caused by a n electric arc f u r n a c e
convergence
requires a f a s t and a c c u r a t e V A r evaluation algorithm f o r
algorithm 161.
compensation.
This paper describes t h e development of a
novel
using
method
Networks
(ANN)
the to
approach
evaluate
of
the
Artificial
Neural
VAr.
instantaneous
Because
in
the
t h e level of
based algorithm i s capable of operating at a much lower
methods.
sampling rate and delivering a n a c c u r a t e and f a s t response
based
output.
By h a r d w a r e implementation of t h i s algorithm using
instantaneous
neuron
chips,
erratic
the
VAr
fluctuation
can
be
arc
characteristic,
Comparing t o t h e conventional methods, t h i s neural network
the
furnace
there
VAr
instantaneous
has
a
has
been
an
evaluation
highly
interest
fluctuating in
predicting
a r c f u r n a c e disturbances by using s t a t i s t i c a l
This paper describes a novel method, on
artificial
converging
neural
VAr.
a
With
real-time
networks, more
to
accurate
evaluate and
VAr
instantaneous
which i s
fast
evaluation
algorithm, t h e amount of disturbance generated by t h e arc
accurately estimated f o r compensation.
f u r n a c e can b e b e t t e r controlled and suppressed. INTRODUCTION The f i r s t p a r t of t h e paper p r e s e n t s a mathematical model The electric arc f u r n a c e hitherto s t i l l provides
t h e most
that
will be
used
in l a t e r
sections f o r t h e
analysis of
efficient way of producing alloy s t e e l s f r o m s c r a p metals.
conventional
In t h e United States alone, between 1975 and 1981, some 15
instantaneous VAr evaluation algorithm.
million
t h i s paper briefly discusses, and compares t h e limitations
tons
of
newly
installed
arc
furnace
capacity
brought t h e arc f u r n a c e s s h a r e s up t o one t h i r d of total
production
highly
arc
The
Ill.
non-linear
plant;
it
VAr caused by t h e rapid,
furnace,
generates
however,
a very
the is
a
fluctuating
large and e r r a t i c variations in
f u r n a c e c u r r e n t which i s always considered as a disturbance t o t h e electrical-supply an
unacceptable
consequence,
level
causes
network.
fluctuating
of
voltage
severe
light
reactive
fluctuation flicker
plant
s t a t i c VAr compensation system.
and
conventional methods
VAr evaluation.
real-time
and,
to
other
as
a
used
Generator
a
loads
be
capable
real-time
instantaneous
Source Inductance
responsive
I
..................................................
high speed s t a t i c V A r compensators t3-51, a n advancement in
must
novel
Q ".
Induct lve
v PCC
Arc Furnace
compensators alone cannot provide a complete solution f o r
compensator
this
The second p a r t of
approach.
Although t h e innovation in
a n e l e c t r i c arc f u r n a c e compensation.
for
of
ANN and describes t h e formation of a
power electronics technology h a s enabled t h e development of
the
development
instantaneous VAr evaluation method based on t h i s
To o p e r a t e such a
requires
the
The t h i r d p a r t of t h i s paper introduces
t h e Back-propagation
This disturbance causes
connected t o t h e same supply g r i d [21. rapidly
of
methods
A s t h e controller of of
evaluating
the
instantaneous V A r f l o w s into t h e uncompensated supply g r i d
94 WM 255-0 PWRD A paper recommended and approved by the IEEE Power System Instrumentation t Measurements Committee ofthe IEEE Power Engineering Society for presentation at the IEEE/PES 1994 Winter Meeting, New York, New York, January 30 - February 3, 1994. Manuscript submitted October 2, 1992; made available for printing December 29, 1993.
Figure I
Single phase equivalent c i r c u i t f o r a n arc furnace
ARC FURNACE MODEL An a r c f u r n a c e for
the
mathematical
investigation
0885-8977/94/$04.00 Q 1994 IEEE
of
model
i s f i r s t l y established
conventional
VAr
evaluation
.
1254 methods
and
the
later
instantaneous
VAr
evaluation
single
inductive
phase
development
load
of
an
a1gori:hm with
varying
based
ANK
In
Fig.
resistance
400
I
a
1.
fed
through a source impedance of negligible r e s i s t a n c e is used t o represent the
supply
a
typical
grid
electric The
[6].
a r c f u r n a c e connected erratic
fluctuation
of
to 0
the
0 3
0.2
0.1
f u r n a c e resistance caused by t h e melting of metal s c r a p s is Figure 2a = V
*
0.7
0.6
0.8
The system p a r a m e t e r s a r e defined as:
modelled by R L .
v
0.5
0.4
Time ( m s )
*
Variation in a r c f u r n a c e c u r r e n t with t i m e
sin(wt)
By solving t h e following differential equation diL
iLRL
V
- + -= dt
LL
LL
sin(wt)
0
0 2
0 1
03
0 5
0 4
0 6
08
07
Time ( m s )
[
"s
iL =
sin(wt-#2)
+ s i n ( $ -wt 2
e-' 1
1
Figure 2 b
Variation in
with time
v PCC
V +
2 sin(wt I - 6 1
e-A
STATEMENT OF PROBLEMS
ZL I
R L Z (t - t i )
There a r e f o u r conventional methods commonly used f o r t h e
A=-
evaluation
LL
of
VAr
The
161.
instantaneous
can
VAr
be
evaluated by A ) c u r r e n t sampling at voltage c r o s s over, B) v
pcc
= v 5
- iLXs
volt-ampere product, C) p a r a m e t e r cross product .and D) r a t e of
where v
is t h e voltage a t t h e point of common coupling, PCC
t
is t h e time a t which t h e load impedance varies f r o m Z L1
change
performance
of
parameters.
of
each
VAr
..Using evaluation
Eqtns.
(1-31, t h e
algorithms
will
be
discussed and compared.
t o Z L 2 because of t h e change in t h e a r c f u r n a c e r e s i s t a n c e from R
L1
to R
L2
.
Where Z L 1 , ZL2. 9, and #
a r e defined a s
f 01lows:
A)
Current sampling a t voltage Cross Over
The
z
output
of
corresponding
this
to
algorithm
the
inductive
a
quantised of
level
the
load
c u r r e n t given by
L1
I x = I Lm s i n ( x n - 6X ZL2=/
is
component
RL:+
(4)
+ d c component
XL2
where I
zero
i s t h e load c u r r e n t measured a t t h e x t h v PCC
c r o s s over, and I Lm is t h e peak value of t h e load c u r r e n t . To minimise t h e e f f e c t of c u r r e n t harmonics and t r a n s i e n t s , The load VAr over M cycles is
t h e average value is used. then proportional t o
Based on t h e above equations, Fig. 2 a r e t h e simulation of and load c u r r e n t
t h e variation of t h e v
i
due t o R
PCC
In
Fig.
3c.
it
varying a t pseudo-random in t h e range between 50Q t o 2502.
algorithm has a
and t h e frequency of variation ranged f r o m 2 H z t o 100 H z .
seems
to
Apparently, In
l a t e r sections of
of
d i f f e r e n t types of
this
paper,
real-time
analysis and V.4r
evaluation
will all be based on this a r c f u r n a c e model.
development algorithms
cannot
shows
that
noticeable
be
fast,
the
overall
cope
satisfactorily.
with This
it
an
the
output
delay. has
step
a
significant
response
abrupt
algorithm
response
of
this
Although t h e response
change is
not
of in
load
suitable
application of a r c f u r n a c e V A r compensation.
overshoot.
this
algorithm current for
the
1255 B) V A r evaluation by volt-ampere product The integral over one period of t h e product of voltage and
a q u a r t e r of cycle delayed c u r r e n t can be used t o evaluate t h e steady state V A r flow.
-
Q =
-
vp c c (ut) i L [ u ( t - G ) ] d ( u t )
O
(6)
-T
where T i s t h e period of supply.
This Volt-Ampere Product
algorithm can be implemented by t h e voltage and c u r r e n t samples as given by
pr
- .
QL,Pn-
where
i
'pcc,x
L,x-n/4
x=P+1
Q L , P n corresponds t o
volt-ampere
(7)
n
the
running
average of
product and n i s t h e number of
the
samples per
cycle of supply voltage.
The integral properties of t h i s algorithm h a s a n e f f e c t of reducing
t h e bandwidth
o r d e r harmonics. Fig.
3d
is
approximation
V A r flowing
fundamental
in attenuating
am
aac
ai
am
ao6
aiz
ai4
ais
ai8
Qz
higher
The output of t h i s algorithm shown in closer
a
and r e s u l t s
f
in t h e
to
the
network
steady
because of
state the
reduced bandwidth.
The output response of t h i s algorithm
i s slightly sluggish.
This may not be acceptable t o a f a s t
iI
response compensation system.
C) VAr evaluation by p a r a m e t e r c r o s s product The reactive power flow in t h e system may a l s o be evaluated by
obtaining
the
differential
cross
product
of
the
alternately delayed voltage and c u r r e n t p a r a m e t e r s as given by Q
L
1 =-(P 2
2
- P )
(8)
1
where (9)
P
1
= v
pcc
( u t ) iL[u(t
-
-I_)]
" nc=
(8)
(10)
Figure 3a & 3b 3c 3d 3e 3f
This can be represented by d i s c r e t e p a r a m e t e r samples and implemented as follows, V
pcc,x-n/4
=
1
L,x 2
- v
pcc,x
i
L.x-n/4
(11)
To implement t h i s technique as a running average and hence reduce t h e distortion
due t o isolated voltage and c u r r e n t
spikes, we have P+N
where P r e p r e s e n t s t h e running average summation point and
N V
pcc,x-n/4
i
L,x n
-
v
pcc.x
i
- Camputer simulation of V ~ C Cand 1, The output of algorithm A - The output of algorithm B (N=25) - The output of algorithm C (N=25) - The output of algorithm D (N=25)
-
is the
number
of
samples
representing t h e summation
period. L.x-n/4
(12) Fig. 3e shows t h a t t h e response of t h i s algorithm converges much f a s t e r t h a n t h e previous two, but i t h a s a significant overshoot and undershoot.
1256 D)
V A r evaluation as a function of t h e r a t e of change of
sampling point, N , a very high sampling frequency, l / ~ , i s
parameters.
required.
The s t e a d y s t a t e value of VAr-flow in t h e electrical supply
evaluation
network’can be evaluated by t h e function shown below
networks,
In l a t e r section, t h e development of a new V A r algorithm, will
be
based
on
thoroughly
back-propagation
described.
neural
a special
With
t r a i n i n g procedure, t h e output of t h i s neural network based
(13) where
k
is
the
desired
integration
interval
and
is
Q,
(14) c a n be simplified t o
By using
delivering
a
higher
accuracy
without
the
BACK-PROPAGATION N E U R A L NETWORKS
A
neural
2
mLm
sin6
t h e voltage
(15)
and
current
samples,
Eqtn.
(14) i s
known
is
network
processing
1 Q=-VI L
of
requirement of operating a t a higher sampling frequency.
defined by
and Q
algorithm can be very close t o t h e instantaneous V A r and i s capable
a
structure
parallel,
consisting
neurons
as
unidirectional
signal
distributed
interconnected channels
information
processing
ot
called
elements
together
with
connections.
Each
neuron h a s a single output connection t h a t branches into as
given as
many
collateral
connection
connections
carries
the
desired.
as
same
Each
signal.
The
collateral
strength
of
connection between neurons is represented by a value called a weight. (16)
Each neuron can have its local memory, which
r e p r e s e n t s t h e s t a t e of neuron.
The neuron output depends
only upon t h e c u r r e n t values of t h e input signals arriving where T i s t h e sampling period.
a t t h e neuron
via
impinging connections
s t o r e d in t h e neuron’s local memory. aver t h e integration
i s t h e running average of Q
QLpN
L.X
period k.
characteristics
of
each
neuron
and upon values
The input t o output
are
described
by
the
activation function and local memory of each neuron. P+N
(17)
Neural net models a r e specified by t h e net topology, node characteristics,
k = Nr
w a s used. where P r e p r e s e n t s t h e running average summation point and
N
is t h e
number
and
or
training
learning
rules.
of
samples
representing
the
summation
arranged
point.
This network
i s made up of
in t h r e e or more layers.
s e t s of
shows t h a t
t h e response
of
this
algorithm
is
capable of providing a very good balance in response time and
damping.
Because of
i t s differential -property,
h a s a n e f f e c t on increasing
r e s u l t s in higher o r d e r harmonics amplification. t h i s algorithm amplification,
h a s t h e drawback it
can
generally
this
t h e bandwidth
and
Although
of
high frequency noise
be
compensated
by
neurons
These a r e a n input
network,
the
output
of
neurons
in
a
t r a n s m i t t e d t o neurons in t h e upper layers.
algorithm
our
layer, an output layer and one or more hidden layers. this
Figure 3f
In
application, a neural network of a f e e d f o r w a r d a r c h i t e c t u r e
the
feedforward
network
does
‘ n e u r o n s in t h e same layer.
not
are
A neuron of
interact
Except
In
layer with
f o r t h e input
other layer
neurons, t h e net input t o each neuron i s t h e sum of t h e weighted
outputs
of
the
neurons
in
the
prior
layer.
Therefore, t h e net input t o neuron j , i s described by
(18)
the
system low-pass f i l t e r . w . . i s t h e weight of connection f r o m neuron i t o neuron j , J I
In
accordance
with
the
above
analysis,
selected f o r t h e development of From Eqtn. (17), t h e evaluated Q instantaneous
VAr
by
algorithm
D
0,
is
a n ANN based algorithm. L.PN
can be assumed as a n
reducing t h e integration
interval
k.
very
small
integration
interval,
k.
In o r d e r t o achieve a and
a
very
large
is j
The output of
neuron j i s then obtained by t h e operation of a nonlinear function on t h e net input. o
I t is also noticed t h a t t h e evaluation accuracy i s a f f e c t e d
by t h e number of sampling point N.
i s t h e output of neuron i in t h e previous layer, (3
t h e local memory (threshold) of neuron j .
The
J
function
=
This can be described a s
f(net.1
(19)
J
f(net )
is t h e
nonlinear
activation
J
of t h e neurons, given by t h e sigmoidal function:
function
1257 1
f(x1 =
(20)
OUTPUT b Y E R
1 + exp(-x)
neural networks is mainly due t o
The learning ability of its
capability
training modify
to
adjust
algorithm the
network.
the
devised
weights
weights.
by
of
Back-propagation
multilayer
..........
I71 i s used t o
Rumelhart
feedforward
0
HIDDEN LAYER
neural
In our application, i t i s appropriate t o consider
back-propagation
neural
networks
as
a
method
to
solve
The output of a
nonlinear function approximation problems.
INPUT IAYER
f e e d f o r w a r d neural network can be considered a s a function f ( x , W ) of input vector x and weight matrix W. feedforward function
is
network
used
c R"+
f(x) : A
Rm.
(x y 1, ( x 2 , y 2 ) ,
examples
1'
to
1
Assume a
approximate In
the
a
...,( xk .y k 1, ...
of
the
yk = f ( x ) a r e presented t o t h e neural network.
can approximate
network weights
W
the
function
bounded
training
by
Back-propagation neural network f o r VAr
Figure 4
stage,
evaluation
mapping
The neural
adjusting
its
so t h a t t h e error E can be reduced, where E i s
defined as
The learning capability
of
is exploited t o develop a
ANN
special training procedure such t h a t the modified algorithm is capable
of
delivering
a
better
performance
than
the
conventional approaches. E =
1 (y 1 -
2
&x i I)',
(21)
X ~ E A
The network The weights are updated
in t h e gradient descent direction
of E, i.e.
where Aw
=
JI
-
r)
aE
F,
where w JI
E Ji
w
(221
and t h e positive constant n determines t h e learning wJ,, r a t e of t h e neural network. The larger t h e constant q is, t h e f a s t e r t h e change in t h e weights is. may
lead
feedforward function,
to
oscillation
neural
in
weight
However, a large space.
For
a
network with sigmoidal activation aE has been determined 171. aw
t h e evaluation of
Ji
In
order
to
speed
process without
up
the
convergence
of
the
training
leading t o oscillation, Rumelhart suggested
t o modify t h e training algorithm in Eqtn. (22) t o include momentum t e r m , such t h a t
are
generated
using
where n labels t h e iteration in t h e learning process and a
the
is explained
procedure pairs of v
reduces
weight
This momentum t e r m provides a damping e f f e c t and the
amount
of
oscillations
neural
during
the
training
process.
INSTANTANEOUS V A r EVALUATION VIA ANN
5.
special
L ,PN'
training
In this. example,
101
A-sampling frequency of 5000 Hz is
evaluated Q,, P I 0 0
The V A r , Q,,,,,,
is
the algorithm D.
from these samples by
This
shown in Fig. 5 is used a s a t a r g e t value f o r the
network training.
Instead of
using t h e 101 p a i r s of v PCC
and
samples a s inputs t o t h e neural
i
shows t h a t only 5 pairs of v
and i
network,
Fig.
5
a r e sub-sampled from
PCC
the
same sample t r a i n
a s the
101 s e t s
for
t h e network
this arrangement,
the output of
the trained network is equal t o Q L.Pl00
+ c,
where c is the
network
training
error
always
inherited
this
been
With
from
the
investigation,
kept
below
network error
in
The
required f o r a 50 Hz line frequency.
time will be much longer.
space.
the
PCC
and c u r r e n t variation.
i s a constant which determines t h e e f f e c t of past weight movement
model.
in Fig.
changes on t h e
of
Before
and i L a r e sampled f r o m one cycle of voltage
Throughout
direction
The network
4.
network can %e applied f o r V A r evaluation, a large number
inputs (i.e. m=5).
(231
current
in Fig.
of training patterns, which include t h e v p c c , i L and Q
i s t h e amount of change f o r t h e weight component
JI
is shown
instantaneous V A r .
the 'corresponding Aw
n
topology
m pairs of v p c c ( x ) and i L ( x ) , and t h e output is
inputs a r e
this
2%.
is preferred
network
Apparently,
but
a
procedure. error very
has small
its corresponding training
This method has t w o advantages. because
the
202
to
10.
Secondly, t h e sampling frequency can be much lower.
In
this
is
Firstly,
the
number
of
network input
neural
network
only 250 Hz. in
Fig.5,
size
neuron
data
is is
approach,
much
smaller
reduced
the
from
sampling
frequency
Although the d a t a a r e sampled f r o m one cycle can
be
collected
from
a
fraction
of
the
that
the
The back-propagation neural network i s applied t o model t h e
cycle
Eqtns (16) & (17) f o r t h e evaluation of instantaneous VAr.
evaluated result can be closer t o the instantaneous VAr.
to
reduce
the
integration
period
k,
so
1258
pCc(1)
-
Conventional calculation using 101 sets of inputs per cycle 5 0' vpe, and IL aamplea M used to (raln
--- -- --
Conventional calculation using 5 sets of inputs per cycle
network'
the neur.l
......................
For nalwork
T
Output using neural network model and 5 sets of inputs per cycle
XI06
I .6
101 w(. of vpec and IL rampler targets c%looc M uaed to wduat. V*r for UaInIng the
neurd network
1
It
Figure 5 Neural network t r a i n i n g mechanism RESULTS Figure
4
a
shows
three-layer
employed f o r t h i s application. neurons f o r 5 p a i r s of v
back-propagation
network
The input layer contains 10 and iL samples.
The output
0.2' 0
50
100
150
200
PCC
250
300
350
400
450
I 500
Tmie (s)
layer h a s one neuron, which gives t h e instantaneous VAr. The number of neurons in t h e hidden layer w a s chosen t o be 50. The learning r a t e
1)
and momentum t e r m a w e r e chosen t o
Figure 6 Output comparison based on samples collected i n one cycle
be 0.2. I t w a s found t h a t t h e network worked well with In t h i s t h i s number of hidden neurons and parameters. investigation,
all
simulations
were
carried
out
using
a
33MHz 80386 PC AT with 80387 coprocessor. Each t r a i n i n g set contains numerous patterns. consists of a QL,p,oo
Each p a t t e r n
and 5 p a i r s of corresponding v PCC
and iL samples. The t r a i n i n g s e t w a s generated so t h a t t h e values of Q L , P 1 O Ow e r e evenly distributed in t h e r a n g e of
V A r variation.
After having been trained, t h e network w a s
applied t o evaluate t h e V A r using inputs samples d i f f e r e n t
8-
f r o m t h a t f o r training.
7-
very
capable of
i t w a s found t h a t t h e network i s
generating Q
L.PlO0
f r o m any 5 sets of
inputs.
.
9.
P
3 ,-
1
g58 4-
Very promising r e s u l t s have been obtained and are shown in Fig. 6.
In t h e s e simulations, t h e v
and iL samples f o r PCC
each the
w e r e collected in one cycle as discussed in
QL.Pl00
last
section.
Using
the
algorithm
D, t h e
mean
percentage e r r o r between t h e evaluated V A r based on 101
sets and 5 sets d a t a sub-sampled f r o m t h e s a m e 101 sets
3-
:i
1
1
'0
.
300
100
400
of
500
600
700
900
800
Time (s)
c u r r e n t and voltage samples w a s found t o be about 20 %. This
illustrates
insufficient training
to
iterations,
evaluate VAr. output
that represent
of
sub-sampled
the
five one
samples cycle.
neural
network
per
cycle
After is
is
applied
neural
network
using
5 sets of
data
f r o m t h e same sample t r a i n as t h e 101 sets.
The mean percentage e r r o r between t h e output of
Output comparison based on samples collected in
114 cycle
to
The dotted line in Figure 6 r e p r e s e n t s t h e
trained
Figure 7
100,000
neural
In Fig. 7, t h e voltage and c u r r e n t samples f o r each Q L,P100 in one-quarter of a cycle and t h e neural
w e r e collected network iterations.
was trained
with
the
same number
of
training
A sampling frequency of 20,000 Hz i s required.
network and t h e evaluated V A r based on 101 sets of c u r r e n t
With
and voltage samples i s below 2%.
response t o t h e changing V A r is f a s t e r .
a f a s t e r sampling frequency and
t h e same
N, t h e
Fig. 7 shows t h a t
1259 t h e t r a i n e d neural nktwork can deliver a n a c c u r a t e VAr with sampling frequency of 1000 Hz.
Conventional calculation using 101 sets of input p e r cycle
The mean percentage e r r o r
between t h e evaluated VAr using 101 sets and 5 sub-sampled
sets of d a t a i s about 2 %.
_____________Output using neural network model a n d 5 sets of inputs p e r cycle
The mean percentage e r r o r
XI08
between
the
output
of
neural
network
model
and
the
evaluated VAr using 101 s e t s of samples i s less t h a n 2%. Figure 7 a l s o shows t h a t using 5 samples t o r e p r e s e n t 1/4 cycle i s more acceptable in conventional approach.
The e f f e c t of t r a i n i n g i t e r a t i o n s on network performance i s illustrated
in
8.
Fig.
When
the
number
of
training
i t e r a t i o n s is 5,000, Fig. S a shows t h a t t h e VAr generated from
t h e output
neuron
cannot
converge t o t h e i r t a r g e t
In addition, a f t e r a change in f u r n a c e resistance,
values.
zt
t h e mean values of VAr do not converge t o t h e t a r g e t value.
0
Figure. 8 b shows t h e comparison between VAr output by t h e
200
100
300
400
neural network a f t e r 100.000 i t e r a t i o n s and t h a t evaluated by conventional approach using 101 sets of samples. error
and
fluctuation
between
t h e VAr
output
relationship
between
error
i t e r a t i o n s is shown in Fig. 9. training
iterations,
it
and
number
of
500
600
BOO
io0
900
1000
Time (ms)
The
from the
neural network and the t a r g e t values i s negligible.
i
F
Output comparison after t h e network h a s been
Figure 8 b
t r a i n e d .for 100,000 i t e r a t i o n s
The
training
By increasing t h e number of
is possible
t o further
reduce
the
e r r o r in t h e network output.
DISCUSSIONS
Eqtn. (17) can only evaluate t h e VAr a t a c e r t a i n i n s t a n t a f t e r t h e voltage and c u r r e n t samples are captured. evaluated VAr lags t h e sampling process.
The
By reducing t h e 0
'
.
.
I
0
.
4
3
' 6
5
*
I
*
7
0
8
10 1104
Training Iterations
Conventional calculation using 101 sets of i n p u t s p e r cycle Figure 9 Output using n e u r a l network model a n d 5 sets of inputs p e r cycle
.-_---I------
'
*
2
Variation in error w i t h d i f f e r e n t number of iterations
portion of each cycle taken f o r samples, t h e evaluated VAr will approach t o t h e t r u e VAr flowing in t h e power supply network.
In t h e
present
investigation,
although
w e r e only taken in one-quarter of a cycle. i s possible 1/400 the
t o reduce t h e portion of
cycle. same
In addition, accuracy
t h e cycle down t o
t h e new algorithm can deliver
as
conventional
methods
substantially less voltage and c u r r e n t samples.
4 0
Y 100
200
300
400
500
800
with
Thus, t h e
evaluation system can be operated at a lower sampling rate.
700
800
900
1000
Time (ms) Figure 8a
samples
In p r a c t i c e i t
The ,long computational
time required by s e r i a l computers
can
hardware
be
eliminated
by
chips
implementation
using
Output comparison a f t e r t h e network h a s been
technology
allows 8
t r a i n e d for 5,000 i t e r a t i o n s
number
synapses can be expanded
of
neuron
.
algorithm
neurons
with
Recently,
of the
8 synapses each
this VLSI
(the
by adding e x t e r n a l
1260 h a r d w a r e s ) t o be f a b r i c a t e d in a single MD1220 Neural Bit
S.E.Haque
Slice.
capacitor-thyristor
This neuron chip h a s an equivalent processing power
of 55 MIPS [SI.
and
N.H.Malik,
"Operation
controlled
reactor
a
of
fixed
(FC-TCR) power
f a c t o r compensator," IEEE Trans. on Power Apparatus
The compensation system constructed by
these neuron chips can evaluate t h e instantaneous V A r with
and Systems, Vol. PAS-104, No. 6, pp. 1385-1390,
negligible processing time a f t e r t h e weights obtained f r o m
1985.
the
learning
chips.
process
a r e download
This performance
can
t o the
never
be
of
RAM
achieved
these by
S . Etminan and R .
the
compensators." this
study,
has
it
Sotudeh, "Fast converging reactive
power measurement techniques f o r high speed s t a t i c V A r
conventional methods.
In
June
also
been
demonstrated
that
the
Proc. of 22th
UPEC 87,
pp. 7081-
7084, Sunderland, U.K. 1987.
neural network model using only 5 s e t s of input can deliver a n equivalent that
the
enhanced
by
training. decreased training
accuracy of
performance
101 input
of
this
selecting
sets.
This suggests
algorithm
appropriate
can
sets
be
of
D.
further
input
Rumelhart,
"Learning
for
E.
G.
internal
propagat;on,"
Hinton.
and
R.
J.
representations
Parallel
Distributed
Williams,
by
error
Processing,
D.
The e r r o r of t h e network output is found t o be
Rumelhart and J . McClelland (Eds.), vol. 1, M I T P r e s s ,
by
1986.
increasing
set.
In
the
number
addition,
Fig.
of
9
patterns
illustrates
in
the
the error
can substantially be decreased by increasing t h e number of
Neural Network Devices,
Micro Devices, 1990.
iterations. Tommy W
CONCLUSIONS
S
Chow obtained
his B.Sc
( f i r s t c l a s s honors)
d e g r e e and Ph.D f r o m t h e University of Sunderland, England. By
introducing
approximating t o estimate
back-propagation
t h e VAr the
neural
evaluation
network
function,
instantaneous V A r
for
is possible
it
as t h a t
a s accurately
obtained by conventional methods operating at a very high sampling r a t e .
The methodology developed in t h i s paper is
successful and r e s u l t s
a r e promising.
He
had
been
working
on
current
Supercomputing Homopolar Machines.
collection
system
L e c t u r e r in t h e Electronic Engineering Department Polytechnic of Hong Kong.
of
He i s now a Senior of City
His c u r r e n t r e s e a r c h activities
are Artificial Neural Networks and i t application.
The neural network
model and training mechanism developed in t h i s paper a r e
Y
very flexible and versatile t o provide any specified output
Polytechnic of Hong Kong.
requirement.
r e s e a r c h a s s i s t a n t in t h e Electronic Engineering Department
accuracy
Because of
properties
e r r a t i c fluctuation
of of
the f a s t this
VAr
ANN
convergence based
generated
by
G.J.McManus,
121
"Electric
limitations
electric
and
networks, system identification and adaptive control.
IEEE
"Flicker Trans.
on
No.9, pp.
2627-2631, September 1985.
P. T. Ho and S . K. Tso, " S o f t w a r e control of s t a t i c reactive
current
compensation,"
IEE
Proc.
vol.
135
pt.c, NO. 6 , pp. 518-527, NOV 1988.
141
R.Sotudeh and
J.Holmes,
compensation,"
Proc.
Dundee t i . K , April 1984.
"High of
19th
speed UPEC.
from
of t h e s a m e college.
electric
Power Apparatus and Systems. Vol. PAS-104,
131
(Hons) degree
reactive pp.
power
329-332,
City
He i s now a M.Phil s t u d e n t and
arc
an
D.J.Ward,
utilities."
B.Eng
the
f u r n a c e succeeds i n technology
J.F.Buch of
his
high
Iron Age, Mp.7 t o hlp.18, Feb 1980.
R.C.Seebald,
obtained
and
REFERENCES
and profit,"
Yam
algorithm,
furnace can be accurately estimated f o r compensation.
111
F
His research i n t e r e s t s a r e in neural