TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 12, No. 9, September
2014, pp. 67
3
2
~ 674
1
DOI: 10.115
9
1
/telkomni
ka.
v
12i9.464
2
6732
Re
cei
v
ed O
c
t
ober 7, 20
13;
Revi
se
d Apr
19, 2014; Accepted Ma
y 10
, 2014
Fault diagnosis of Electric Power Grid Based on
Improved RBF Neural Network
Luo Yi-Ping*
1,2
, Shen Lin
g
2
, Cao Yi-Jia
3
1
Huna
n Institute of Engi
n
eer
in
g, Xi
angta
n
, C
h
in
a
2
Institute of Electrical an
d Information En
gi
n
e
e
rin
g
, Huna
n U
n
iversit
y
, Ch
an
gsha, Ch
in
a
3
Institute of Informatio
n
Engi
n
eeri
ng, C
entra
l
South Univer
sit
y
, Sha
ngsh
a
,
China
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: l
y
p
8
6
88@s
o
hu.com, 447
01
355
1@q
q
.com,
y
j
ca
o@
hun
an.
edu.cn
A
b
st
r
a
ct
T
h
is pa
per i
n
troduc
es a n
o
vel
clusterin
g
a
l
go
rithm th
at co
mbin
es crisp
and
fu
zz
y
cl
usterin
g
. It no
t
only
has
the
hig
h
acc
u
racy
of fu
zz
y
cl
us
tering,
but a
l
s
o
red
u
ces
the
de
pen
de
ncy
on i
n
iti
a
li
z
a
t
i
o
n
.
Specific
ally, it
constitutes a fa
st learni
ng pr
o
c
ess and
th
ere
f
ore, the conv
e
r
genc
e rate a
n
d
the accur
a
cy
of
the RBF
N
N
ar
e gre
a
tly i
m
pro
v
ed. T
he s
i
mul
a
tion r
e
su
lts s
how
that this s
t
rategy is s
u
cc
essfully
ap
pli
e
d t
o
the fault dia
g
nosis of electr
ic pow
er grid.
T
he
training
spee
d and th
e fault-toler
a
n
c
e of informatio
n
aberr
ance, w
h
i
c
h co
mes fro
m
the
mal
ope
ration
of t
he
protectio
n
s a
n
d
bre
a
kers, ar
e sup
e
rior to
th
e
traditio
nal RBF
NN.
Ke
y
w
ords
:
fau
l
t diag
nosis, R
B
F
neural n
e
tw
ork, crisp clusterin
g
, fu
zz
y
clu
s
tering, el
ectric
pow
er grid
Copy
right
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
With a
su
bsta
ntial incre
a
se
of the type
s
and q
uantitie
s of the
g
r
id
electri
c
al
eq
ui
pments,
the increa
sin
g
compl
e
xity of operating
conditio
n
s
couple
d
with natural di
sa
st
ers a
nd mi
su
se
mak
e
the grid fault occ
u
rs frequently. If the partial
faul
t of the
po
we
r gri
d
can
not
received
time
ly
treatment, it
will lea
d
to
a
large
-
scale bl
ack
out,
whi
c
h
se
riou
sly e
n
dang
ers the
stable
ope
rati
on
of the po
we
r
system.
Ho
wever, in the
case
of t
he a
b
norm
a
l op
era
t
ion, multi-fau
l
t of prote
c
tive
relays a
nd ci
rcuit bre
a
kers,
fast and accurate faul
t di
agno
si
s is very difficult to
achi
eve for the
influx of massive
amou
n
t
s of informa
t
ion [1-3
]. In recent years, with the
developm
ent
of
comp
uter te
chnolo
g
y and
intelligent th
eory, a vari
e
t
y of artificial intelligent a
nd optimi
z
ati
on
method
s a
r
e
used in
po
wer sy
stem
fault diagn
osis, such a
s
fuzzy theo
ry, optimizati
on
techni
que
s, e
x
pert system
s, Petri
netwo
rks, data mini
ng [7, 8]. Artifi
cial Neural Networks with i
t
s
self-le
a
rning
ability, fault
tolera
nce, and
paralle
l info
rmation proce
ssi
ng capabili
ties, is more
and
more
used in
the study of
power
syste
m
fault diagn
osi
s
, espe
ci
al
ly the RBFNN that sho
w
s its
advantag
es i
n
practi
cal e
n
g
inee
ring
ap
plicatio
ns
[8].
RBF
NN
ha
s
a any fun
c
tio
n
app
roximati
on
ability in the
o
r
y, trainin
g
a
nd exe
c
utio
n
time
is le
ss than
other
co
mmonly u
s
e
d
network le
arning
algorith
m
s,
a
nd the
net
work h
a
s a
certain
deg
re
e of fault
tol
e
ran
c
e
for the n
on- trai
ning
detectio
n
sa
mples.
There are a v
a
riety of lea
r
n
i
ng alg
o
rithm
s
of RBF
N
N
durin
g the
RBFNN t
r
ainin
g
peri
od,
[9-13], a
clu
s
terin
g
al
gori
t
hm whi
c
h f
u
lly take
s int
o
acco
unt th
e data i
nhe
rent dist
ributi
on
relation
shi
p
i
s
p
r
opo
se
d in
this pa
per, t
he dia
gno
stic re
sult of this me
thod
i
s
co
mpared with
the
result simul
a
ted by traditio
nal fuzzy clu
s
terin
g
algo
rit
h
m (F
CM) [1
4, 15]. The si
mulation resu
lts
of the 4
-
b
u
s
test sy
stem
show
that
the
clu
s
terin
g
sp
eed and
the
accuracy of hybrid clu
s
te
ring
algorith
m
are
both better th
an FCM al
gorithm.
The rest of the pap
er i
s
orga
nized a
s
follows. In
section 2,
we
introdu
ce th
e
RNF
NN
about its structure. In secti
on 3, the
propo
se
d a
ppro
a
ch co
m
b
ined
with crisp a
nd fuzzy
clu
s
terin
g
alg
o
rithm
s
is prese
n
ted. In S
e
ction
4, it d
eals
with the
paramete
r
e
s
timation fo
r
the
training
of th
e
RBF
NN.
In
Section
5, th
e effect
iven
e
s
s
of su
ch
a
methodol
ogy is
inve
stigate
d
by
mean
s of sim
u
lation
s. Fina
lly, conc
lu
sio
n
s are drawn in Section 6.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fault diagn
osis of Electri
c
Powe
r Grid B
a
se
d
on Im
prove
d
RBF Ne
ural Netwo
r
k (Luo Yi-Pi
ng)
6733
2. RBF
NN S
t
ructur
e
RBFNN i
s
a
feedfo
r
ward
network
wit
h
three-
tie
r
structu
r
e; it
s t
opolo
g
y is
shown in
Figure 1. Inp
u
t layer nod
e
s
tran
sfer th
e
input si
g
nals to the hidden
layer, the hid
den laye
r nod
es
are
compo
s
e
d
of
radi
al a
c
tion fu
nctio
n
s
li
ke
Ga
ussi
an
ke
rnel
fu
nction,
and
t
he o
u
tput l
a
yer
node
s are usually simple l
i
near fun
c
tion
s. Whe
n
t
he input sig
nal is clo
s
e to th
e cente
r
of the
base function, the hidden layer nodes will produce
a larger output, which shows that this
netwo
rk ha
s
a ca
pa
city of local
app
rox
i
mation.
As t
he form
of the ba
si
s fun
c
tion, the m
o
st
comm
only used is Ga
ussia
n
function:
2
(
)
exp
i
1,2,...,
m
2
xc
i
i
Rx
i
(1)
Whe
r
e
x
is
n
–dimensi
onal i
n
put vecto
r
,
i
c
is the ce
nter
of the
i
th ba
si
s
function,
i
is
the width of Gau
ssi
an fun
c
tion,
m
is the n
u
mbe
r
of hid
den no
de
s. The Gau
s
sian
function a
bov
e
has the
cha
r
a
c
teri
stics of si
mple structu
r
e, good an
alyticity and any orde
r de
rivabl
e.
R
1
R
2
R
m
y
i
w
1
w
2
w
k
x
1
x
2
x
n
....
....
Figure 1. Structure of
RBF
N
N
For the
stru
ct
ure a
bove, th
e input la
ye
rs carry out the
nonline
a
r m
appin
g
of
()
i
x
Rx
,
while the o
u
tput layers
ca
rry out the linear map
p
ing of
()
i
k
R
xy
, that is
:
1
(
)
1
,
2
,
3
......
m
ii
k
i
i
yw
R
x
k
p
(2)
Whe
r
e
p
is the numbe
r of the output layer node
s.
3. H
y
brid Fu
zzy
Clusteri
ng Algorith
m
3.1. The Basi
c Theor
y
The co
re id
e
a
of the algorithm is tha
t
for those sample sets
whi
c
h ne
ede
d to be
clu
s
tere
d, all of the sample
s sh
ould b
e
d
i
vided into three cate
gori
e
s: one part of the sam
p
le
s a
r
e
only belon
g to one of the clu
s
ters, this
kind of sampl
e
s is
calle
d the crisp
clu
s
tering
sam
p
le
s;
anothe
r p
a
rt
of the
sampl
e
s,
whi
c
h a
r
e
call
ed
semi
- fuzzy
clu
s
te
ring
sa
mple
s [16], belon
g
to
s
o
me of the clus
ters
; the las
t
part of the
s
a
mp
l
e
s, whi
c
h are
called
full
fuzzy
sam
p
les, belon
g
t
o
all clu
s
te
r. F
r
om th
e exp
e
rime
ntal verification,
thi
s
idea
well
consi
ders th
e
inhe
rent
da
ta
distrib
u
tion re
lationship am
ong the
samp
les. Obje
ct
ive
function of th
e clu
s
teri
ng a
l
gorithm b
a
se
d
on the
above
idea i
s
give
n belo
w
. Th
e
obje
c
tive function
of crisp clu
s
te
ring [
17, 18], an
d
the
objec
tive func
tion
of fuzzy c
l
us
ter
i
ng
ar
e
used
to
do a s
i
mple affine ar
ithmetic
, wher
e the
para
m
eter
is a vari
able
to
co
ntrol th
e
cl
usteri
ng
sp
ee
d, clu
s
teri
ng
accuracy, a
n
d
de
pend
ency
on initiali
zatio
n
of the
alg
o
rithm. The
ma
thematical
ex
pre
ssi
on fo
rm of the
obj
e
c
tive fun
c
tion
is
sho
w
n in formula (3
):
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 9, September 20
14: 67
32 – 674
1
6734
2
11
nc
Hi
k
k
i
ki
Ju
x
v
2
2
11
(1
)
(
)
nc
ik
k
i
ki
ux
v
(3)
W
h
er
e
c
is
the numbe
r of
cl
u
s
ters,
n
is th
e
numbe
r
of sa
mples,
i
v
is
the rand
om clu
s
t
e
r
cente
r
s ch
osen
befo
r
e cl
usteri
ng,
[0
,
1
)
,
[0
,
1
]
ik
u
is
the memb
ership de
gre
e
o
f
the
k
-th
training
vecto
r
to the
i
-th cl
uster. If
0
, the objective function will
become F
C
M al
gorithm
with
2
m
; if
1
, then it
become
s
cri
s
p
clu
s
terin
g
algorith
m
. T
he con
s
traint
is sho
w
n in
the
following:
1
1,
c
ik
i
uk
(4)
Acco
rdi
ng to
the basi
c
p
r
i
n
cipl
e of the clu
s
terin
g
alg
o
rithm, the m
i
nimum value
of the
obje
c
tive fun
c
tion will
be
ob
tained
und
er
the con
s
traint
of the fo
rmul
a (3), if valu
e
s
of th
e d
e
g
r
ee
of membership
ik
u
and the
cl
uster cente
r
i
v
are
the
sta
gnation
point
of La
gra
nge
functio
n
(,
)
ik
k
Fu
which co
rrespond
s to
H
J
, so the followi
ng formul
a ca
n b
e
use
d
to sol
v
e the value:
2
11
(,
)
nc
ik
k
i
k
k
i
ki
Fu
u
x
v
2
2
11
(1
)
(
)
nc
ik
k
i
ki
ux
v
11
(1
)
nc
ki
k
ki
u
(5)
After partial differential, we get:
22
2
11
(,
)
2(
1
)
0
(,
)
(2
)
(
)
(
1
)
(
)
(2
)
(
)
0
ik
k
ki
i
k
k
i
k
ik
nn
ik
k
ik
k
i
ik
k
i
kk
i
Fu
xv
u
x
v
u
Fu
ux
v
u
x
v
v
(6)
ik
i
uv
,
and
k
can be solved by the above eq
uati
on:
2
1
2(
2
)
1
1
2(
1
)
2(
1
)
()
k
c
j
kj
c
xv
(7)
2
2(
2
)
1
2
(
1)
2
(
1)
()
j
ik
ki
vC
kj
C
u
xv
xv
(8)
2
1
2
1
(1
)
(
)
(1
)
(1
)
(
)
n
ik
ik
k
k
i
n
ik
ik
k
uu
x
vi
c
uu
(9)
For
0
ik
u
,
the following scali
ng i
nequ
ality can
be obtained,
this equation
can be u
s
ed
as a di
scrimin
ant to judge e
a
ch
sampl
e
b
e
long
s to whi
c
h cl
uste
r, its form is a
s
follows:
2
2
1
2(
2
)
1
1
()
ki
c
j
kj
c
xv
xv
(10)
Cha
nge Equ
a
t
ion (10
)
to Equation (11
)
:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fault diagn
osis of Electri
c
Powe
r Grid B
a
se
d
on Im
prove
d
RBF Ne
ural Netwo
r
k (Luo Yi-Pi
ng)
6735
2
2
2(
(
)
2
)
:
1
1
()
k
j
k
ikk
i
k
vT
ki
T
vT
x
v
T
xv
(11)
Obviou
sly, th
e Equation
(11) expresse
s that:
k
T
rep
r
ese
n
ts the set of the cluster
cente
r
s which co
ntain th
e
k
-th sampl
e
,
()
k
T
rep
r
e
s
ent
s th
e num
ber of the cl
uste
rs which th
e
k
-th sam
p
le is belon
ged t
o
. Appare
n
tly, when
()
1
k
T
,
1(
)
k
Tc
, and
()
0
k
T
, th
e
sampl
e
bel
on
gs to th
e crisp clu
s
te
ring
sampl
e
s,
se
mi fuzzy clu
s
tering
sam
p
l
e
s a
nd full fu
zzy
s
a
mples
,
respec
tively.
3.2. The Process o
f
H
y
brid Fuzzy
Algorithm
Based
on th
e fore
going
analysi
s
, the
gene
ral
pr
o
c
e
ss of
this algorith
m
is as
follo
ws.
Firstly, cla
s
si
fication, and
next, for those sam
p
le
s
which b
e
lon
g
to different
cl
usters, different
method
s a
r
e
adopte
d
to
ca
lculate
the
co
rre
sp
ondi
ng
d
egre
e
of
the membe
r
ship. Then, cal
c
ula
t
e
cluster centers,
and
check the
cl
uster center
to
see
whether it still
change
s. If it
changes, repeat
the above ste
p
s until the chang
e rea
c
h
e
s a certain
error threshol
d,
then stop the algo
rithm,
the
cluster center will be got. In
troduce iteration parameter
v
, the above
algorith
m
is rewritten a
s
a
n
iterative form.
(1
)
(1
)
2
(1
)
()
2
2(
(
)
2
)
:
1
1
()
v
k
j
v
v
k
ik
k
i
v
k
vT
ki
T
vT
x
v
T
xv
(12)
Equation
s
for calculating th
e degree of membe
r
ship
of different sa
mples a
r
e a
s
follows:
Whe
n
()
1
k
T
, the sa
mple
belon
gs to the
crisp
clusteri
ng
sa
m
p
les.
ik
u
i
s
cal
c
u
l
ated by
Equation (13).
2
2
1
1m
i
n
0
ki
k
j
jc
ik
if
x
v
x
v
u
othe
rwise
(13)
Whe
n
1(
)
k
Tc
,
t
he sa
mple b
e
lon
g
s
t
o
t
h
e
se
mi-f
uz
zy
cl
ust
e
ri
ng
sampl
e
s.
ik
u
is
cal
c
ul
ated
by Equation (14).
()
(1
)
2
2(
(
)
2
)
1
2(
1
)
()
v
k
j
v
k
ik
ki
vT
kj
T
u
xv
xv
2(
1
)
(14)
Whe
n
()
0
k
T
, the s
a
mple bel
on
gs to the full fuzzy clu
s
te
ring sample
s.
ik
u
is calculated b
y
Equation (15).
2
1
1
()
ij
c
ji
k
jk
u
xv
xv
(15)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 9, September 20
14: 67
32 – 674
1
6736
The deg
re
e o
f
membership
is norm
a
lize
d
by Equation
(16).
1
(1
,
)
ik
ik
c
jk
j
u
ui
j
c
u
(16)
The upd
ate formul
a of clu
s
ter
cente
r
is:
2
1
2
1
(1
)
(
)
(1
)
(1
)
(
)
n
ik
ik
k
k
i
n
ik
i
k
k
uu
x
vi
c
uu
(17)
The followi
ng
steps
sho
w
the pro
c
e
s
s of
the propo
se
d
hybrid fuzzy algorith
m
.
Select values for
c
and
. Randomly initialize
12
,
,
..
.,
c
vv
v
, s
e
t
it
e
r
0
.
(0
)
:(
)
k
kT
c
,
(0
)
12
,
,
...,
kc
Tv
v
v
Step 1
:
Set
ite
r
i
t
e
r
1
.
Step 2
:
Use Equation (12) to update the
sets
()
v
k
T
and their cardi
nalitie
s
()
()
(
1
)
v
k
Tk
n
.
Step 3
:
If
()
()
1
v
k
T
, use Equation (12)to calculat
e membe
r
shi
p
degrees
(1
;
1
)
ik
uk
n
i
c
;
if
()
0
k
T
, use Equ
a
t
ion (16
)
to calcul
ate mem
bership d
e
g
r
ees; el
se u
s
e Equation
(17
)
.
Step 4
:
If
0(
1
;
1
)
ik
uk
n
i
c
,
se
t
0
ik
u
.
Step 5
:
Then
use Equatio
n
(16) to initiali
ze mem
bersh
ip degree
s.
Step 6
:
Use Equation (17) to updat
e the
cluste
r cente
r
s.
Step 7
:
If there are no noti
c
ea
ble chan
g
e
s for the
clu
s
te
r
cente
r
s, then sto
p
, else turn to step
1.
4. Param
e
ter
Estim
a
tion of the
RB
FN
N
In this
clu
s
tering alg
o
rithm,
the nu
mbe
r
of the hid
den
nod
es
equ
al
s the
clu
s
te
rs
c
, while
the center of
the ra
dial
ba
sis
fun
c
tion i
s
the
cl
uste
ri
ng
cente
r
12
,
,
..
.,
c
vv
v
. For the
cal
c
ul
ation of
the width
of the radial
ba
si
s fu
n
c
tion,
two
que
stion
s
should be con
s
ide
r
ed. Fi
rst of all, the
wi
dth
value can not be too small, because the small
wi
dth will cause a small degree of overlap.
Ho
wever, th
e
deg
ree
of ov
erlap
can
not
be to
o la
rge,
be
cau
s
e
an
over e
s
timate
d be
havior wi
ll
be caused,
whi
c
h
will greatly redu
ce
the performance
of the ne
twork. So,
a new method to
cal
c
ulate the
width of the radial
ba
sis fu
nction, which
gives full con
s
ide
r
ation of the sp
ecific d
a
ta
distrib
u
tion
of
ea
ch
cl
ass, i
s
p
r
o
p
o
s
ed
in
this pa
per.
A thre
sh
old val
ue of
mem
b
e
r
shi
p
deg
ree
i
s
sele
cted, an
d
a credi
ble selectio
n is
0.00
1
[19]. Then the sample
s of each cla
s
s are re
-
scree
ned a
n
d
expresse
d u
s
ing Equ
a
tion
(18).
:(
0
,
1
)
ik
i
i
k
Gx
C
u
(18)
Next, the maximum distan
ce from the
cl
uster
cente
r
to the sampl
e
of the
i
G
cluste
r is
obtain
ed:
2
max
ma
x
ki
i
ki
xG
dx
v
(19)
Finally, the width
i
yields:
ma
x
2
(1
)
3
i
i
d
ic
(20)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fault diagn
osis of Electri
c
Powe
r Grid B
a
se
d
on Im
prove
d
RBF Ne
ural Netwo
r
k (Luo Yi-Pi
ng)
6737
Thus, the val
ue of RBFNN is obtaine
d b
y
the followin
g
equatio
n.
2
||
|
|
()
e
x
p
ki
ik
i
xv
gx
(21)
The value of
the radial b
a
si
s functio
n
is matrix
H
, which is
solved
by sub
s
tituting the
clu
s
ter
ce
nte
r
an
d the
wi
dth above
int
o
Equatio
n (21). App
a
ren
t
ly,
H
is
nc
, where
n
is the
total numbe
r of sample
s, a
nd c is the n
u
m
ber of cl
ust
e
rs.
For the
solvin
g of the weig
ht, assume th
e output of th
e trainin
g
sa
mples i
s
Y
, the actual
output of
the
network i
s
Y
, then
acco
rdin
g to th
e train
i
ng p
r
o
c
e
s
s
of the
wei
ght
, the follo
win
g
error fun
c
tion
can
obtain t
he minimu
m
only unde
r th
e pro
p
e
r
wei
ght
w
. An expression
of the
error func
tion:
2
()
E
WY
Y
(22)
Here, the l
e
a
s
t squa
re
s m
e
thod i
s
a
d
o
p
ted to
solve
the weight
v
a
lue whi
c
h m
a
ke
s
th
e
error fun
c
tion
to achi
eve the minim
u
m
value.
The fo
llowing
wei
g
h
t
calculation f
o
rmul
a can b
e
easily de
du
ce
d:
1
[]
TT
wH
H
H
Y
(23)
Whe
r
e,
1
[]
T
HH
represents the p
s
eudo
-inv
e
r
se
calculation
of the matri
x
. The well
trained
RBNNN i
s
as follo
ws.
1
()
(
)
m
kk
i
i
k
i
f
xy
w
g
x
(24)
5. Simulation and Analy
s
is on Fault
Diagno
sis
5.1. Fault Di
agnosis Simulation Ba
se
d on Impro
v
ed RBFNN
A four-bu
s
-ba
r
system
is u
s
ed
a
s
th
e ex
perim
ental
sy
stem, a
nd it
is
sho
w
n
in
Fi
gure
2.
The sy
stem is co
mpo
s
ed
of bus ba
rs B1~B4, a tr
an
sformer T
1
, an
d four tran
sm
issi
on line
s
L
1
~
L4. CB rep
r
e
s
ent
s the circuit brea
ke
r, MB represent
s the main p
r
ot
ection of th
e bus b
a
r, M
L
is
the mai
n
p
r
ot
ection
of the
t
r
an
smi
ssi
on li
ne, BL
i
s
th
e
backu
p p
r
ote
c
tion
of the
transmi
ssion
li
ne,
and MT i
s
the
main p
r
ote
c
tion of the tran
sform
e
r. Th
e values of
co
n
d
ition
attribut
es
a
r
e “0
”or “1”.
“1” in
dicates t
hat the clo
s
e
d
circuit brea
ker i
s
di
sco
n
necte
d or in
prote
c
tive sta
t
e, “0” re
pre
s
ents
that the circui
t breaker i
s
u
n
ch
ang
ed or
prote
c
tion is
non-ope
ratio
n
.
MB
1
B
1
CB
1
BT
MT
T
1
CB
3
MB
2
B
2
BL
4
ML
4
BL
5
BL
2
ML
5
ML
2
CB
4
CB
2
CB
5
BL
6
BL
7
BL
8
ML
6
ML
7
ML
8
CB
6
CB
7
CB
8
L
1
L
2
L
3
L
4
CB
10
CB
9
BL
10
BL
9
ML
10
ML
9
MB
4
B
3
MB
3
B
4
Figure 2. A Simple Power
Grid Stru
cture
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 9, September 20
14: 67
32 – 674
1
6738
39 sam
p
le
s a
r
e sel
e
cte
d
a
s
a trainin
g
samp
le set, so
the input and output of the neural
netwo
rk a
r
e
39
1
6
,
39
21
,
res
p
ec
tively. Set
0.
5
, randoml
y
initialize the number of cl
usters
and clu
s
ter
centers, such as
C3
0
, and th
e d
i
agno
stic re
sults a
r
e
sh
own in T
able
1.
For th
ese
39
1
6
-dimen
sion
al
training input
sample
s, we
have the followin
g descrip
tion: each of this 16 -
dimen
s
ion
a
l i
nput
sign
al
repre
s
e
n
ts th
e corre
s
po
nd
ing o
p
e
r
ation
of the
ci
rcuit
brea
ker a
n
d
the
prote
c
tion i
n
the ab
ove figu
re, the
o
r
d
e
r
of the protecti
on an
d
circuit
bre
a
ker i
n
th
e input
sign
al
is
as
follows
: (
12
4
5
6
7
,,
,
,
,
,
CB
CB
CB
CB
CB
CB
10
1
2
,,
,
,
CB
MB
MT
ML
78
9
4
7
,,
,
,
,
M
LM
L
M
LB
L
B
L
B
T
)
.Assum
e
on
e
of the in
put
vector i
s
(
1
0 0 0 0 0
0 0 0 0
1
0 0
0
0 0
), it rep
r
e
s
ent
s t
hat
1
CB
and
7
M
L
have actio
n
, resp
ectively. The dime
nsi
o
n
of the output training
sam
p
les is
39
21
, including
singl
e devi
c
e
fault and
du
al devices fa
ult. If ther
e i
s
a fa
ult of t
he devi
c
e, th
e corre
s
pon
d
i
ng
value ta
ke
s “1”, othe
rwise, the valu
e is
“0”.
Th
e
r
e
are 21
group
s
of input te
st
sampl
e
s,
so
the
input test matrix is
21
16
. Moreover, in ord
e
r to detect
the fault toleran
c
e of the
propo
se
d
algorith
m
, a
new
set of te
st sa
mple
s i
s
set by reversing all the
act
i
on value
s
of
1
M
B
. Simulation
results sho
w
that when the test sa
m
p
les a
r
e n
o
n
-
interfe
r
en
ce
sample
s, a
c
cura
cy of fault
diagn
osi
s
i
s
1
00%. When
the te
st
sampl
e
s
with
no
sy,
the output of
the neu
ral n
e
twork
is
sho
w
n
in Table 1, th
e co
rre
ct faul
t diagno
sis
result
s
have b
een ma
rked
with line
s
, on
ly the sample
2
and the sam
p
le 10 ca
n not correctly d
i
agno
se
d in
21 gro
up sa
mples, ap
parently, accu
ra
cy of
fault diagno
si
s is 9
0
%.Part
icula
r
ly attention t
hat sa
m
p
le 2 an
d sa
mple 10
are
compl
e
me
nta
r
y,
so
wh
en
dist
urbe
d, neith
e
r
sampl
e
s are
able
to id
enti
f
y by any dia
gno
stic m
e
th
ods [20]. If we do
not co
nsi
der these t
w
o
situations, di
a
gno
stic
a
c
curacy is
still 1
00%, much
highe
r than t
he
method in [20
]
.
Table 1. Simulation Results of Fault Dia
gno
sis Ba
sed
on Improved
RBFNN
Number
B
1
T
B
2
B
3
B
4
L
1
L
2
L
3
L
4
B
1
,T
B
2
,T
1
0.1991
-0.0245
-0.0126
-0.0163
-0.015
0.0284
0.0336
-0.0026
-0.0229
0.0249
-0.0171
2
0.0023
0.0226
0.0122
0.0163
0.014
-0.012
-0.0294
0.0208
-0.0016
0.0208
0.0264
3
-0.2193
-0.059
-0.0446
0.0113
-0.0274
-0.0602
-0.091
0.1153
0.3673
0.1782
0.0382
4
-0.0908
0.1005
0.1083
-0.0361
0.0161
0.0187
-0.0035
-0.0308
-0.007
0.0169
-0.0406
5
-0.1533
0.0373
-0.0283
-0.0267
-0.0185
0.0292
-0.1169
-0.1155
0.1816
-0.0394
-0.033
6
-0.3025
0.0626
-0.0382
0.0566
-0.0432
0.0294
-0.0674
0.0104
0.1043
-0.1172
-0.0305
7
-0.2617
-0.0652
-0.059
-0.0046
-0.0939
0.0458
0.0251
-0.0838
0.0968
-0.1278
-0.028
8
-0.2145
0.0719
0.0341
-0.013
0.0486
0.0352
-0.1075
-0.0204
0.1728
-0.1291
-0.0385
9
-0.155
0.0633
-0.0382
0.0265
-0.0218
0.0158
-0.129
-0.0553
0.1975
-0.0287
-0.0171
10
0.0073
-0.0067
-0.0195
-0.0092
-0.0301
0.0088
0.0098
-0.012
0.0056
-0.0016
-0.016
11
0.1239
0.0692
0.0432
0.02 0.0561
-0.0072
-0.1131
0.1115
0.1406
-0.032
1.0494
12
0.0629
0.0456
0.0839
-0.0024
0.0247
0.0204
0.0112
0.0415
0.1614
0.7472
0.0071
13
0.1119
0.166
0.0114
0.0611
0.0199
0.0614
-0.1401
0.055
1.1083
-0.0673
0.0234
14
0.0498
0.0224
0.0195
-0.044
0.0661
0.0681
-0.0804
1.0788
0.1798
-0.1002
0.0182
15
0.0964
0.4742
0.0483
0.0828
-0.0363
0.042
0.6092
0.1391
0.0176
0.1134
0.0228
16
0.0031
0.0654
-0.0024
0.0553
0.0697
1.1067
-0.087
0.0521
0.1154
-0.0077
0.0112
17 0.0737
0.1311
-0.0312
0.0344
0.9908
0.0801
-0.2165
0.0435
0.171
0.0299
0.0191
18 0.0329
0.0401
-0.0127
1.0705
0.0629
0.0736
-0.0553
-0.0394
0.1044
0.0304
0.0023
19 0.0774
0.0582
1.0167
0.0553
0.0229
0.0046
-0.0157
0.0093
0.0519
0.0751
0.0002
20 0.1165
0.5958
0.0873
-0.0187
0.0653
-0.0149
0.4101
-0.0057
0.1586
-0.0104
-0.0193
21 0.8054
-0.0106
0.0251
0.0278
0.0744
-0.0311
-0.0078
0.0384
0.0407
0.0518
-0.0104
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fault diagn
osis of Electri
c
Powe
r Grid B
a
se
d
on Im
prove
d
RBF Ne
ural Netwo
r
k (Luo Yi-Pi
ng)
6739
Table 1. Simulation Results of Fault Dia
gno
sis Ba
sed
on Improved
RBFNN (cont
inued
)
序号
B
2
,L
1
B
2
,L
2
B
2
,L
3
L
1
,L
2
L
2
,L
3
L
3
,L
4
L
2
,L
4
B
3
,L
4
B
3
,L
1
NO
1
-
0.1692
0.0365
-
0.1263
0.2161
-
0.3115
0.0948
0.0103
-
0.0287
0.2784
0.8245
2 0.6837
-
0.0409
0.0134
-
0.1186
0.0964
-
0.0172
0.012
-
0.0352
0.2427
0.0713
3
0.2631
0.0292
0.0379
-
0.1149
0.0819
0.022
0.1863
0.4587
-
0.3363
0.1633
4 0.3089
-
0.0748
0.0308
-
0.0548
0.0749
-
0.0086
0.7118
0.1199
-
0.3114
0.1516
5
0.3246
0.1079
0.2568
-
0.0863
0.0753
0.4628
0.0802
0.2216
-
0.3468
0.1876
6 0.3882
-
0.0003
0.1394
-
0.0125
0.8084
-
0.0377
0.109
0.1821
-
0.3954
0.1547
7
0.1716
0.128
0.0873
0.6047
0.2277
-
0.0835
0.1133
0.3049
-
0.2241
0.2263
8 0.2788
-
0.0781
0.7449
-
0.0686
0.0997
0.0395
0.0972
0.1456
-
0.2992
0.2005
9 0.302
0.6507
0.1022
-
0.1124
0.0427
-
0.1141
0.326
0.1045
-
0.3355
0.1758
10 0.2427
0.0611
-
0.0129
0.0819
-
0.0755
-
0.0038
0.0026
0.0133
0.8094
-
0.0551
11 0.4029
-
0.1669
-
0.0839
-
0.2815
-
0.0434
0.1132
0.1239
-
0.1066
-
0.5055
0.0861
12 0.33
-
0.1502
-
0.0333
-0.284
-
0.1999
0.0018
0.1295
0.2325
-
0.3593
0.1293
13 0.2788
-
0.1554
-
0.0524
-
0.3219
-
0.1005
0.0391
0.0658
0.0534
-
0.3151
0.0971
14 0.323
-
0.1007
-
0.2273
-
0.2907
-
0.0764
0.1147
0.191
0.0402
-
0.3554
0.1034
15 0.2905
-
0.1116
-
0.1179
-
0.2731
-
0.2478
0.1681
-
0.0558
-
0.0146
-
0.3241
0.0769
16 0.2657
-
0.1668
-
0.2331
-
0.3389
0.0323
0.1277
0.1727
-
0.0401
-
0.2934
0.0921
17 0.3287
-
0.2186
-
0.1654
-
0.2579
-
0.0731
-
0.0786
0.3863
0.0237
-
0.3692
0.098
18 0.316
-
0.2758
-
0.0419
-
0.3878
0.0537
0.12 0.1343
0.0195
-
0.3306
0.0828
19 0.335
-
0.2178
-0.006
-
0.2323
-
0.0013
0.0062
0.0267
-
0.0072
-
0.3552
0.0961
20 0.3517
-
0.1337
0.0109
-0.382
0.1828
0.0506
-
0.0197
-
0.1794
-
0.3413
0.0954
21 0.0023
-
0.1119
-
0.0635
0.1268
0.0292
0.2151
-0.024
-
0.1926
0.0073
0.0077
5.2. Compari
s
on bet
w
e
e
n
FCM and th
e Ne
w
M
e
tho
d
FCM i
s
ado
pted to t
r
ain
RBF
NN,
an
d test
re
sult
s a
r
e
compa
r
ed
with
the
re
sult
s
simulate
d by
the improved
method. Also
sele
ct t
he n
u
m
ber
of clu
s
t
e
rs C
= 3
0
, si
mulation
re
su
lts
are
sh
own in
Table
2. Fro
m
Table
2, th
e fault
d
i
ag
no
s
i
s
ac
cu
ra
c
y
o
f
R
B
FN
N ba
s
e
d
o
n
F
C
M is
85%, that is in 21 gro
u
p
s
o
f
sample
s,
on
ly 15 group
s
of diagno
si
s is co
rrect.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 9, September 20
14: 67
32 – 674
1
6740
Table 2. Simulation Results of Fault Dia
gno
sis of RB
FNN Ba
se
d o
n
FC
序号
B
1
T
B
2
B
3
B
4
L
1
L
2
L
3
L
4
B
1
,T
B
2
,T
1
0.0376
0.0105
0.0529
-0.0491
-0.0372
0.0269
0.025
0.0132
-0.0397
-0.0262
-0.0153
2
0.0374
0.0031
-0.0187
0.0318
0.0209
-0.0101
-0.032
-0.0456
-0.0534
-0.0282
0.1933
3
0.0114
-0.1069
-0.0763
0.0478
-0.0151
-0.0165
-0.0441
0.1828
0.1758
0.2316
0.0355
4
0.0713
0.014
0.2555
0.0089
-0.048
0.0703
0.0309
-0.0007
-0.0211
0.1185
-0.1321
5
0.0884
-0.0335
-0.09
-0.0441
0.0042
0.0419
-0.1448
-0.0807
0.1461
0.1327
-0.1111
6
0.0309
-0.1946
-0.0391
0.1521
-0.0417
-0.0411
-0.0461
0.0932
0.0041
0.1138
-0.0537
7
0.1123
-0.2232
0.0035
0.1154
-0.0491
-0.0011
-0.03
0.1339
-0.1001
0.1447
-0.1171
8
0.106
-0.0045
-0.0451
0.0086
0.1434
0.0765
-0.1052
0.1109
0.02
0.0142
-0.173
9
0.0257
-0.0098
-0.0711
0.0309
0.112
-0.0422
-0.0357
-0.0057
0.0066
0.0909
-0.0558
10
0.058
0.0302
0.0291
-0.0116
-0.0059
-0.0257
-0.0044
-0.084
-0.0645
-0.1007
0.2958
11
-0.1137
-0.0704
-0.0607
0.0272
0.0303
0.0287
0.0146
0.1522
0.125
0.2013
0.4311
12
-0.0942
-0.0367
-0.0138
-0.0589
0.0407
0.0314
0.1061
0.1699
0.005
0.523
0.0295
13
-0.0379
0.1285
-0.1082
0.1661
-0.0312
0.1556
0.0373
0.0406
0.4323
0.0657
0.01
14
-0.0716
-0.1051
-0.0655
0.0199
0.1636
0.1549
0.0327
0.4302
0.0123
0.2028
0.0155
15
-0.0739
0.2076
0.0096
0.1399
-0.0036
0.106
0.2601
0.0955
0.0895
0.2116
-0.0681
16 -0.0915
0.006
-0.0776
0.1205
0.1424
0.44
0.062
0.1767
0.1548
0.0908
-0.0899
17
-0.0732
-0.0311
0.0561
0.1256
0.3619
0.1309
-0.0554
0.1741
-0.051
0.084
-0.0938
18 -0.0112
-0.0025
0.1209
0.4426
0.1131
0.0883
0.0533
0.015
0.131
-0.0313
-0.1238
19 -0.0017
0.0001
0.3961
0.2223
0.1309
-0.013
0.0279
0.018
-0.0404
0.1095
-0.1203
20 -0.0208
0.3913
-0.0226
0.0804
0.0215
0.0475
0.2019
-0.043
0.1777
0.0668
-0.1521
21
0.1842
-0.0054
0.0156
0.0063
-0.0517
-0.0733
-0.0588
-0.0517
-0.0212
-0.0766
-0.0938
Table 2. Simulation Results of Fault Dia
gno
sis of RB
FNN Ba
se
d o
n
FCM (co
n
tinued
)
序号
B
2
,L
1
B
2
,L
2
B
2
,L
3
L
1
,L
2
L
2
,L
3
L
3
,L
4
L
2
,L
4
B
3
,L
4
B
3
,L
1
NO
1
0.0116
0.0482
-0.0107
0.0397
-0.0883
0.108
0.2865
-0.189
0.0469
0.7487
2
0.4919
-0.0476
-0.0193
0.0956
0.0047
-0.0088
-0.1424
0.0664
0.2808
0.1802
3
0.1699
0.0612
-0.2329
0.0833
0.3119
-0.0015
-0.3405
0.491
-0.2363
0.2679
4
0.1656
-0.2361
-0.1212
0.0522
0.0141
0.105
0.2026
0.236
-0.0706
0.2851
5
0.2135
0.2137
0.0387
0.0599
0.0073
0.33
-0.2058
0.3092
-0.1408
0.2649
6 0.1676
-0.098
0.0341
0.3773
0.4729
-0.1553
-0.3409
0.39
-0.1703
0.3451
7 0.2245
-0.0964
-0.2123
0.607
0.2325
0.0018
-0.1392
0.3011
-0.1369
0.2289
8 0.2632
-0.0806
0.4515
0.0627
0.1226
-0.0774
-0.2249
0.2211
-0.1224
0.2322
9 0.099
0.5487
-0.1736
0.1179
-0.0903
0.0483
-0.1565
0.3072
-0.0977
0.3511
10
0.2808
-0.0086
0.1375
0.0753
-0.0037
0.0649
0.1467
-0.1513
0.4321
-0.0899
11
0.4022
-0.0065
-0.2665
-0.1144
-0.0171
-0.1329
-0.2933
0.3178
0.1487
0.1964
12
0.1932
0.0361
-0.3406
0.0626
0.2426
-0.0028
-0.2676
0.3828
-0.2502
0.2417
13
0.1785
-0.004
-0.199
0.0216
0.0138
0.0029
-0.3902
0.4511
-0.2197
0.286
14
0.1642
-0.1786
0.1269
0.01
0.1911
-0.1967
-0.3107
0.3523
-0.2276
0.2791
15
0.2087
-0.2023
-0.3229
0.2416
0.2287
-0.2165
-0.2218
0.2282
-0.1639
0.2458
16
0.2138
-0.2171
0.1864
0.237
-0.4928
0.0171
-0.2303
0.246
-0.1778
0.2834
17
0.2242
0.4747
0.2519
-0.1618
-0.0953
-0.107
-0.4238
0.1315
-0.1494
0.227
18
0.2627
0.3066
-0.3378
0.3919
-0.0183
-0.2393
-0.4729
0.2443
-0.1695
0.237
19
0.2112
0.2379
-0.285
-0.098
0.0677
-0.1407
-0.1072
0.2452
-0.1276
0.2671
20
0.2426
-0.0986
-0.1024
-0.0595
0.3378
-0.1097
-0.2674
0.1727
-0.1293
0.2651
21
0.0374
0.0604
0.1749
0.2784
0.3204
0.1918
0.1211
0.011
0.0581
-0.0266
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fault diagn
osis of Electri
c
Powe
r Grid B
a
se
d
on Im
prove
d
RBF Ne
ural Netwo
r
k (Luo Yi-Pi
ng)
6741
6. Conclusio
n
In this pa
per, a new
hyb
r
id fuzzy
clu
s
terin
g
alg
o
ri
thm is p
r
op
o
s
ed to
optim
ize the
para
m
eters
o
f
the RBF
NN, and it
is ap
plied to
the f
ault dia
gno
si
s of
po
we
r g
r
id. Simulatio
n
results
sh
ow t
hat the meth
o
d
in this pa
pe
r re
du
ce
s
the
influen
ce of t
he cl
uste
ring
initial ch
oice to
the diag
no
stic results, an
d im
proves the conve
r
ge
nce
sp
eed
a
nd a
c
curacy
of RBFNN. It ha
s
validity for power g
r
id fault
diagno
si
s, espe
cially
for the noi
se di
sturba
nce, su
ch as switchin
g or
prote
c
ting m
a
lfunction, it h
a
s hi
gh
rob
u
s
tne
ss. T
h
is method ha
s a
practi
cal si
gnifica
nce
for
the
fast and a
c
cu
rate fault diag
nosi
s
, and th
e enha
nceme
n
t of supply reliability.
Referen
ces
[1]
W
ang Jia
lin,
Xia Li, W
u
Z
h
e
ngg
uo, et al. State
of arts of fault diag
nosi
s
of po
w
e
r s
ystem.
Power
System Protect
i
on a
nd C
ontrol
. 2010; 38(
18): 210-
216 (i
n Ch
ines
e)
[1]
Card
osog Jr, R
o
limi
g
. App
licat
ion
of neur
al-n
et
w
o
rk
mod
u
l
e
s to s
y
stem fa
ult sectio
n esti
mation. IEEE
T
r
ans on Po
w
e
r Deliver
y. 2
0
0
4
; 19(3): 10
34-
104
1.
[2]
Guo C
hua
ng
xi
n, Z
hu C
h
u
a
n
bai, C
a
o
Yij
i
a,
et al. St
at
e of
arts of fau
l
t d
i
ag
nosis
of p
o
w
e
r s
y
stems.
Autom
a
tion of Electric Power System
s
. 20
06
; 30(8): 102-1
0
7
(in Ch
ines
e)
[3]
Bi T
i
anshu, Ni
Yixin, Ya
ng Qi
xu
n. An ev
al
ua
tion of ar
tific
i
al
intell
ig
ent tech
nol
ogi
es for fa
ult di
agn
osi
s
in po
w
e
r net
w
o
rk.
Automati
on
of Electric Pow
e
r Systems.
20
00; 24(2): 1
1
-1
6 (in Ch
ines
e)
[4]
Lin S
h
e
ng, H
e
Z
heng
yo
u, Qia
ng Qin
g
q
u
a
n
. Revie
w
a
n
d
de
velo
pment
on f
ault d
i
a
gnos
is i
n
po
w
e
r
gri
d
.
Po
w
e
r S
y
stem
Protection a
nd
Contro
l.
201
0; 38(4): 14
0-1
5
0
(in Chi
nese)
[5]
Che
n
Yul
i
n, Ch
en Yu
npi
ng, S
un Ji
nli
n
, et al.
A surv
e
y
of p
o
w
e
r
s
y
stem fau
l
t dia
gnos
is. El
ectric Po
w
e
r.
200
6; 39(5): 27
-31 (in C
h
in
ese
)
[6]
Li Xia
oqu
an, Z
hua
ng De
hu
i, Z
hang Qia
ng. A ne
w
f
ault di
a
gnos
is mode
l of electric po
wer grid b
a
se
d
on ro
ugh r
a
d
i
cal b
a
sis functi
o
n
ne
ural
net
w
o
rks. Po
w
e
r S
y
s
t
em Protection
and
Contro
l. 2
009; 3
7
(1
8):
20-2
4
(in Ch
in
ese)
[7]
Bi T
i
anshu,
Ni
Yi
xi
n, W
u
F
u
l
i
, et a
l
.
A
nov
el
neur
al
netw
o
rk a
ppro
a
ch
for fau
l
t secti
o
n esti
mation
.
Procee
din
g
s of
the CSEE. 2002; 22(2): 7
3
-7
8 (in Ch
ines
e)
[8]
Antoni
os D. Ni
ros, George E
T
s
ekouras.
A nove
l
traini
ng
alg
o
rith
m for R
B
F
neural n
e
t
w
ork using a
hybri
d
fu
zz
y
cl
usterin
g
ap
pro
a
ch
. F
u
zz
y
Set
s
and S
y
stems
.
2011; 8: 1-20.
[9]
SB Roh, T
C
Ahn, W
Pedr
y
c
z
.
T
he desig
n method
ol
og
y of r
adi
al b
a
sis fun
c
tion ne
ura
l
ne
t
w
orks
base
d
on fuzz
y
K-
ne
a
r
est neig
h
b
o
rs appr
oach.
Fu
z
z
y Sets Syst.
2010; 16
1(1
3
): 1803
–1
822.
[10]
Sun
Dan,
W
a
n L
i
min
g
, Su
n
Yanfe
ng,
et a
l
. An
improv
ed
h
y
br
id
le
arni
n
g
a
l
gor
ithm for
RBF
n
eur
al
net
w
o
rk.
Jour
n
a
l of Jili
n Univ
e
r
sity (Science
Editio
n).
201
0; (5): 17-82
2 (in
Chin
ese)
[11]
W
ang Li
ang,
W
ang Shito
ng.
Merc
er kernel
based h
y
b
r
id
C-means fuz
z
y
cluster
i
ng a
l
gorit
hm
w
i
t
h
d
y
nam
ic w
e
i
g
h
t.
Applicatio
n R
e
searc
h
of Co
mp
uters.
201
1;
28(8): 285
2-2
855 (i
n Chi
nes
e)
[12]
H Sarimve
i
s, oga
nis, Ale
x
an
dridis. A cl
assi
fica
tion tec
hni
que
base
d
o
n
radia
l
b
a
sis functio
n
ne
ura
l
net
w
o
rks.
Adv.
Eng. Softw
are.
2006; 7(4): 21
8–2
21.
[13]
Ge
w
e
n
i
ger T
,
Z
u
lke D, Ham
m
er B, et al.
Medi
an fuzz
y
c-means for cl
usterin
g
dissi
milarit
y
data.
N
e
u
r
o
c
om
pu
ti
ng
. 2010; 7
3
(1): 110
9-11
16.
[14]
Li Pei
ngq
ia
ng, Li Xi
nr
an, Che
n
Hui
hua, et al
.
T
he character
i
stics classifica
tion an
d synth
esis of pow
e
r
loa
d
bas
ed on
fu
zz
y
clust
e
rin
g
. Proceee
di
ng
s of the CSEE. 2005; 2
5
(24):
73-7
8
(in Ch
in
ese)
[15]
GE T
s
ekouras, D
Darze
n
tas,
I. Drakou
laki,
et al.
F
a
st
fuz
z
y
v
e
ctor
qu
an
tization. IEEE
Internatio
na
l
Confer
ence
on
F
u
zz
y
S
y
st
ems, Barcelon
a. 2
010.
[16]
Li Gua
n
li
n, Ma
Z
hanh
on
g, Hu
ang
Cho
ng, et
al. Segm
entati
on of co
lo
r
ima
ges of gr
ap
e di
seases
usin
g
K_mea
n
s clust
e
rin
g
alg
o
rithm
.
T
r
ansactions
of
the CSAE. 2010; 26(S
u
p
p
.2): 32-37 (i
n C
h
in
ese)
[17]
Jian
g H
u
il
an,
Liu
X
iao
jin, G
u
anYi
ng, et
al.
Short-term l
o
ad for
e
castin
g
bas
ed
on
H
a
rd-C m
e
a
n
clusteri
ng al
gor
ithm and su
pp
ort vector
machin
e. Po
w
e
r S
ystem
T
e
chnol
o
g
y
.
[18]
AD Niros, GE
T
s
ekouras. On trainin
g
ra
dial
basis fun
c
tion ne
ural n
e
t
w
o
r
ks usi
n
g
optimal fuzz
y
clusteri
ng. 17th
IEEE Mediterrane
an C
onfere
n
ce
on C
ontro
l and Autom
a
tio
n
, Mediterra
ne
an. 200
9.
[19]
Qian T
ao. Ap
pl
icatio
n a
n
d
R
e
search
in
D
i
stri
but
io
n N
e
t
w
ork
F
ault
Di
agn
osi
s
b
y
Ro
ugh
Se
t T
heor
y
an
d
Neur
al Net
w
o
r
k. Nanji
ng: Na
njin
g Un
iversit
y
of Science a
n
d
T
e
chnolo
g
y
.
200.
Evaluation Warning : The document was created with Spire.PDF for Python.