Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
13
,
No.
3
,
Ma
rch
201
9
, p
p.
1136
~
1142
IS
S
N: 25
02
-
4752, DO
I: 10
.11
591/ijeecs
.v1
3
.i
3
.pp
1136
-
1142
1136
Journ
al h
om
e
page
:
http:
//
ia
es
core.c
om/j
ourn
als/i
ndex.
ph
p/ij
eecs
Eva
lu
atin
g
w
ind
owing
-
b
ased
c
on
tinu
ous S
-
t
ra
ns
f
or
m
with
n
eura
l
n
etwork
c
lassifi
er
for
d
ete
ctin
g a
nd
c
lassifyi
ng
p
owe
r
q
uality
d
isturban
ces
K.
Daud
1
,
A.
Farid
Ab
idi
n
2
,
A
. Puad
Ism
ail
3
, M.
Daud
A.
Ha
s
an
4
, M.
A
f
f
an
di
Sh
af
i
e
5
, A. Ism
ail
6
1,
2,3,4,5
Facul
t
y
of Electrical E
ng
in
ee
ring
,
Univ
ersiti
T
eknol
ogi
MA
RA
,
Malay
si
a
6
Facul
t
y
of
Com
pute
r and
Ma
them
at
ic
a
l
Sci
ence
s
,
Univer
si
t
i
Te
kn
ologi
MA
RA,
M
al
a
y
si
a
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
hist
or
y:
Re
cei
ved
Oct
7
, 2
018
Re
vised Dec
6
,
2018
Accepte
d Dec
13
, 201
8
T
he
ai
m
of
thi
s
pape
r
is
to
evalu
at
e
th
e
impleme
nta
ti
on
of
windo
wing
-
base
d
Conti
nuous
S
-
Tra
nsform
(CST)
te
chni
qu
es,
name
l
y
,
one
-
c
y
cle
an
d
hal
f
-
c
y
c
l
e
windowing
with
Multi
-
lay
e
r
Per
ce
pt
ion
(MLP)
Neura
l
Network
cl
assifie
r
.
Both,
the
techni
ques
and
cl
assifi
er
a
re
used
to
d
et
e
ct
and
cl
assif
y
th
e
Pow
er
Quali
t
y
D
isturbance
s
(PQ
Ds
)
into
one
of
poss
ible
c
la
ss
es,
vo
lt
ag
e
sag,
sw
el
l
and
interrupt
di
sturbanc
e
sign
al.
For
rea
l
izing
e
val
ua
ti
on,
we
pr
oposed
the
m
et
hodolog
y
th
at
inc
lud
e
th
e
PQ
D
gene
rat
ion
,
th
e
sign
al
d
etec
t
ion
usin
g
windowing
-
base
d
CS
T,
the
feature
s
ext
ra
ct
ion
from
S
-
cont
our
m
at
ric
es,
PQ
D
cl
assific
ati
on
using
MLP
cl
assifi
er.
The
n
,
we
per
form
t
wo
t
y
p
e
of
assess
m
ent
s.
Firstl
y
,
th
e
a
cc
ura
c
y
assess
m
ent
of
chose
n
class
ifi
er
in
relati
on
to
thr
ee
diffe
r
ent
tra
in
ing
algorithms
.
Seco
ndl
y
,
the
exec
uti
on
t
ime
compari
son
of
the
training
al
gorit
hm
s.
Bas
ed
on
assess
me
nt
result
s,
we
outline
seve
r
al
r
ec
om
m
enda
tions
for
futur
e
w
ork.
Ke
yw
or
ds:
Con
ti
nu
ou
s
S
-
t
ran
s
f
or
m
Mult
i
l
ay
er
p
erceptio
n
n
e
ural
n
et
w
ork
Power
q
ualit
y
Power
q
ualit
y
d
ist
urba
nce
W
i
ndowin
g
t
ec
hn
i
qu
e
Copyright
©
201
9
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed.
Corres
pond
in
g
Aut
h
or
:
K.
Da
ud
,
Faculty
of Elec
tric
al
Engineer
ing
,
Un
i
ver
sit
i Te
knol
og
i M
ARA
,
40450 S
hah A
l
a
m
, S
el
ango
r,
Ma
la
ysi
a.
Em
a
il
:
ka
m
aru
l395@
uitm
.ed
u.
m
y
1.
INTROD
U
CTION
Power
qu
al
it
y
(P
Q
)
is
cru
c
ia
l
in
pr
ovisi
on
i
ng
util
it
ie
s
fo
r
fu
l
fill
ing
the
con
s
um
ers
nee
d
[
1].
Nowa
days,
P
Q
pro
blem
has
beco
m
ing
a
hu
ge
c
halle
ng
e
a
s
m
or
e
co
ns
um
ers
are
dem
a
nd
i
ng
f
or
the
powe
r
qu
al
it
y
.
Ele
ct
ri
cal
de
vices
vu
l
ner
a
ble
to
pow
er
qu
al
it
y
or
la
ck
of
qual
it
y
is
m
or
e
su
it
able
to
be
incl
ud
e
d
in
the
do
m
ai
n
of
power
ap
pea
rs
li
m
it
ed.
All
el
ect
ric
de
vices
a
r
e
dis
po
se
d
to
hav
e
a
pro
ble
m
or
dam
age
wh
e
n
t
hey
are
expose
d
to
on
e
o
r
m
or
e
power
qual
it
y
iss
ues
[
1]
-
[
7].
Ele
ct
ric
m
oto
r,
ge
ner
at
or
,
c
om
pu
te
r,
com
m
un
ic
at
io
n
equ
i
pm
ent,
or
hous
e
hold
a
pp
l
ia
nce
are
t
he
e
xam
ples
of
el
e
ct
rical
dev
ic
es
that
has
a
hi
gh
cha
nces
to
da
m
ag
e
wh
e
n
e
xpos
e
d
to
PQ
d
ist
urba
nces
(
PQDs)
.
To
date
,
t
he
as
set
qu
al
it
y
of
powe
r
is
quit
e
exp
e
ns
i
ve,
s
o
th
ere
is
a
need o
f
m
on
it
ori
ng syst
em
s that can
detect
P
QD act
ivit
ie
s in order
to re
duce costs.
To
im
pr
ove
t
he
power
qua
li
ty
in
the
syst
e
m
,
there
is
a
nee
d
to
de
te
ct
the
pr
e
s
ence
of
t
he
disturba
nces,
i
den
ti
fy
the
sources
of
the
pro
blem
s
and
f
i
nd
the
so
l
utio
n
to
ov
e
rc
om
e
them
.
In
pre
vious
researc
h
a
nd
st
ud
ie
s
,
the
re
se
arch
e
rs
ty
pical
ly
us
e
m
ult
iple
appr
oach
es
t
o
detect
an
d
cl
assify
the
act
ivit
y
of
PQDs.
Am
ong
a
ppr
oach
es
us
e
d
i
n
past
stu
dies
are
S
-
T
ran
s
f
or
m
[2
]
-
[
5]
,
[
8
]
,
Wa
velet
Tra
nsfo
r
m
[
9
]
,
Ne
ural
Netw
ork
,
Discrete
F
ourier
Tra
nsfo
rm
(D
FT),
Fa
st
Four
ie
r
T
ra
ns
f
or
m
(F
F
T),
an
d
Sup
port
Vecto
r
Ma
chine
(
SVM
),
a
com
bin
at
ion
of
any
of
them
or
oth
ers
.
Most
ap
proac
hes
de
scribe
d
befor
e
s
uper
vi
sed
P
Q
pro
blem
s
by
changin
g
from
one
dom
ai
n
to
an
oth
e
r
dom
ai
n
of
m
at
he
m
at
ic
s
wh
ic
h
pro
vid
es
a
dd
it
ion
al
detai
le
d
in
form
at
ion
.
The
m
ai
n
scop
e
of
the
stu
dy
is
cat
ego
rized
into
two
pa
rts
wh
ic
h
are
detect
ion
of
PQ
disturba
nce
base
d
on
the
use
of
S
-
Tra
nsf
or
m
m
a
the
m
atical
te
chn
iqu
es
to
detect
pow
er
qual
it
y
dist
urba
nces
an
d
Ne
ural
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Evalu
atin
g
wi
ndowi
ng
-
base
d con
ti
nu
ous
S
-
tr
an
sf
orm wi
th
ne
ural netw
ork
cl
as
sif
ie
r for
...
(
K. D
au
d)
1137
Netw
ork
nam
e
ly
Multi
Lay
e
r
Perce
ption
Neural
Netw
ork
(ML
PNN)
has
bee
n
ch
ose
n
as
a
cl
assifi
cat
ion
m
et
ho
d
of
cl
as
sific
at
ion
a
nal
ysi
s
perform
ance
f
or
P
Q
disturba
nces.
The
detect
ion
of
P
Q
disturba
nces
ha
ve
been
c
onduct
e
d
base
d
on
tw
o
dif
fer
e
nce
a
ppr
oach
n
am
ely;
On
e
-
Cy
cl
e
W
i
ndowin
g
Te
chn
i
qu
e
(O
C
WT)
a
nd
Half
-
Cy
cl
e W
i
ndowin
g
Tec
hniq
ue
(
HC
WT).
A
m
at
he
m
at
i
cal
cod
es
a
re
c
reated
a
pproac
hes
by
us
in
g
s
of
t
war
e
MATLAB
©
t
o
fin
d
init
ia
l
per
i
od,
the
fi
na
l
per
i
od,
t
he
m
agn
it
ud
e
a
nd
durati
on
of
the
P
Q
distu
r
ban
ce
.
Fu
rt
her
m
or
e
,
t
his
pa
per
will
giv
es
a
br
ie
f
s
umm
ary
of
an
analy
sis
of
th
e
PQ
disturba
nc
es
for
the
detect
ion
and
cl
assifi
cat
ion
base
d
on
CST
f
or
the
di
stribu
ti
on
syst
e
m
by
us
i
ng
the
m
et
ho
d
of
OC
W
T
an
d
H
C
WT
with MLP
N
N.
2.
PQDs
SI
GNA
L GENE
RA
T
ION
The
distu
r
ban
c
es
sign
al
powe
r
qu
al
it
y
are
ge
ner
at
e
d
base
d
on
m
at
he
m
atical
m
od
el
ing
program
m
ing
in
m
-
file
/sc
ript
of
MA
TLA
B©
[
1
0
]
.
T
he
re
a
re
th
ree
ty
pes
of
sig
nal
s
inv
ol
ved,
na
m
el
y,
sag,
swell
,
and
interr
up
t.
T
he param
et
ers
require
d
to
ge
ner
a
te
the signal
s a
re th
e
real t
im
e
o
f
s
ig
nal du
rati
on
a
nd am
plitu
de
of
disturba
nce si
gnal
volt
age
. Ta
ble 1 s
hows
t
he
p
a
ram
et
ers
associat
ed
t
o
ea
ch
ty
pe o
f dist
urba
nce.
T
able
1.
Ma
t
he
m
at
ic
al
PQ
D
Disturba
nce S
i
gn
al
s
Mo
delin
g
Distu
rban
ces
Mod
el equ
atio
n
Para
m
eters
Sag
v
(t)
=1
-
*
(u(t
-
t1
)
-
u
(t
-
t2
))
*
sin
t(ɷt+
ϕ)
No
te:
α=
R
ed
u
ctio
n
level
of rms
volta
g
e in
p
.u.
t = 0.1
:
0.0
0
1
:
0.
1
8
t1
= Time
o
f
V
s
a
g
i
n
itia
tio
n
t2
= Time
o
f
V
s
a
g
r
ecover
y or
clear
a
n
ce
ϕ = P
h
a
se
-
a
n
g
le jump
α= 0
.5
t1
=
4
0
m
s
t2
=
1
0
0
m
s
ϕ
= 90
̊
Swell
v
(t)
=1
+*
(u(t
-
t1
)
-
u
(
t
-
t2
))
*
sin
t(ɷt+
ϕ)
No
te:
α=
I
n
crea
sin
g
level
o
f rms vo
lta
g
e in
p
.u.
t = 0.1
:
0.0
0
1
:
0.
1
8
t1
= Time
o
f
V
s
w
ell
i
n
itia
tio
n
t2
= Time
o
f
V
s
w
ell
reco
very or
clear
a
n
ce
ϕ = P
h
a
se
-
a
n
g
le jump
α= 0
.5
t1
=
5
0
m
s
t2
=
1
1
0
m
s
ϕ
= 45
̊
Interrup
t
v
(t)
=1
-
*
(u(t
-
t1
)
-
u
(t
-
t2
))
*
sin
tɷ
t
No
te:
α=
R
ed
u
ctio
n
level
of rms
volta
g
e in
p
.u.
t = 0.1
:
0.0
0
1
:
0.
1
8
t1
= Time
o
f
V
in
ter
r
u
p
t
in
itia
tio
n
t2
= Time
o
f
V
in
ter
r
u
p
t
reco
very or
clea
ra
n
ce
α= 0
.95
t1
=
5
0
m
s
t2
=
1
1
0
m
s
2.1.
PQDs
Sig
na
l
Det
ec
tion usin
g
O
ne
-
C
ycle
Windowi
n
g
T
echnique
(O
C
WT)
The
cy
cl
es
accor
da
nce
with
windowin
g
te
chn
i
qu
e
of
C
onti
nuous
S
-
Tra
ns
f
or
m
(CST)
is
us
ed
f
or
PQDs
detect
io
n
an
d
feat
ur
e
e
xtracti
on.
Eac
h
cy
cl
e
of
each sam
ple
wind
ow
of
inter
fer
e
nc
e
wav
e
f
or
m
si
gn
al
i
s
analy
zed
acco
r
dan
ce
with
ST
con
t
our
[
3],
[
4].
The
detect
ion
of
PQD
usi
ng
OC
WT
is
perform
ed
fo
r
every
20
m
s
(one
-
cy
cl
e
)
of
ti
m
e
du
r
at
ion
of
sig
nal
s.
T
he
si
gn
al
m
us
t
in
abs
olut
e
conditi
on
to
perform
this
de
te
ct
ion
.
Figure
1
-
3
sho
ws
detect
ion
of
P
Q
Ds
base
d
on
CST
us
i
ng
OC
W
T
.
T
he
re
d
li
ne
is
repres
ent
the
sig
nal
l
ine
of
PQD
in
an
abs
olu
te
co
nd
it
io
n.
The
blu
e
li
ne
rep
rese
nted
de
te
ct
ion
li
ne
of
sign
al
s.
The
n,
S
-
co
ntour
m
atr
ic
es
analy
ze
the
sign
al
us
e
d
to
extract
the
fea
tures
from
the
detect
ion,
f
or
instance;
i.e
m
agn
it
ud
e
,
sta
nd
a
r
d
dev
ia
ti
on, m
ean, fre
quency a
nd phase
. T
hes
e featu
res
a
re t
hen use
d t
o
s
uppo
rt PQD
cl
assifi
cat
ion
proc
ess.
Figure
1
.
O
C
WT base
d on
C
ST
–
V
oltage
Sag
Figure
2
.
O
C
WT base
d on
C
ST
–
V
oltage
Sw
el
l
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
13
, N
o.
3
,
Ma
rc
h 201
9
:
1136
–
1142
1138
Figure
3
.
O
C
WT base
d on
C
ST
–
I
nter
rup
t
2.2.
PQDs
Sig
na
l
Det
ec
tion
u
sin
g
H
alf
-
C
ycle
Windowi
n
g
T
echnique
(H
C
WT)
HC
W
T
re
pr
es
ents
a
hal
f
durati
on
of
on
e
-
cy
cl
e
for
s
uppo
rting
t
he
detect
ion.
A
half
-
cy
cl
e
is
determ
ined
by
10m
s.
Thu
s
,
by
us
i
ng
the
s
a
m
e
PQ
Ds
sig
nal,
the
HC
W
T
is
util
iz
ed
to
lim
i
t
the
sco
pe
of
the
sam
ples
fr
om
t
he
e
ntire
disturbance
sig
nal.
The
n,
CS
T
is
a
pp
li
ed
to
c
reat
e
the
li
ne
dete
ct
ion
wh
ic
h
produce
s
S
-
c
on
t
our
m
at
rices
, as
s
how
n i
n
Fig
ure
4
-
6
.
Figure
4
.
H
C
WT det
ect
ion
base
d on
C
ST
–
Vo
lt
age
Sag
Figure
5
.
H
C
WT det
ect
ion
base
d on
C
ST
–
Vo
lt
age
Sw
el
l
Figure
6
.
H
C
WT det
ect
ion
base
d on
C
ST
–
In
te
r
r
up
t
2.3.
PQDs Si
gnal
Clas
sific
at
i
on
u
sing Neur
al
Net
w
or
k cl
as
s
ifie
r
In
t
his
pap
e
r,
Mult
i
-
la
ye
r
Perce
ptron
(
MLP)
is
us
e
d
as
N
N
cl
ass
ifie
r
to
cl
assi
fy
from
PQ
disturba
nces
sign
al
[
1
1
],
[1
2
].
An
MLP
co
m
pr
ise
s
of
m
ulti
ple
la
ye
rs
of
nodes
in
a
dire
ct
ed
gr
a
ph
,
wit
h
each
la
ye
r
fu
ll
y c
onnected t
o
t
he n
ext one.
Fig
ure
7
s
how
s the
st
ru
ct
ur
e a
rc
hitec
ture of
MLP
f
or this
p
ape
r.
3.
RESU
LT
S
A
ND AN
ALYSIS
Fo
r
t
he
da
ta
set
pr
e
par
at
io
n,
100
dataset
is
us
ed
a
s
in
pu
t
s
to
cl
assify
the
volt
age
s
w
el
l
in
PQ
D
s
sign
al
.
T
he
in
pu
ts
a
re
pa
rtit
i
on
i
nto
th
ree
par
ts,
wh
ic
h
a
re
trai
ning,
va
li
dation
an
d
te
sti
ng
dataset
s
.
Eigh
t
hidden
la
ye
rs
a
re u
se
d
to t
rain
the MLPNN a
nd
1000
it
erati
on
is set
for
cl
assifi
cat
ion
. F
ur
therm
or
e, this
pape
r
us
es
th
ree
dif
f
eren
t
ty
pes
of
trai
ning
al
gor
it
h
m
s
fo
r
eval
uating
the
cl
as
sific
at
ion
perf
or
m
ance,
w
hic
h
are;
Gr
a
dient
Desc
ent
wi
th
M
ome
ntu
m
and
A
da
ptive
LR
‘t
raingd
x’
[1
3
]
,
Le
venbe
rg
-
Ma
r
quar
dt
‘trai
nlm
’
[1
2
]
-
[1
4
]
and B
FG
S
Quasi
-
New
t
on ‘
trai
nb
fg’ [1
2
],
[1
5
]
.
Table
2
s
hows
the
cl
assifi
cat
ion
of
Sam
ple
1
us
i
ng
ML
PNN
cl
assifi
er
ba
sed
on
CST
w
it
h
OC
W
T
.
The
res
ults
ha
ve
s
how
n
t
ha
t
trai
ning
al
gorithm
Gr
adi
ent
De
scent
with
M
om
ent
um
and
Ad
a
pt
ive
LR
‘train
gdx’
cl
as
sifie
d
98%
of
a
ccur
at
e
cl
assifi
cat
ion
.
Me
a
nwhile
for
Le
venberg
-
Ma
rqua
rdt
‘trainlm
’,
the
resu
lt
of
P
Q
distu
rb
a
nces
cl
assifi
cat
ion
is
100%
,
wh
il
e
f
or
al
gor
it
h
m
BFGS
Qu
asi
-
New
t
on
‘train
bfg’
;
it
pr
ovi
de
d
97%
of
cl
assi
ficat
ion
acc
urat
eness.
T
her
e
fore
,
al
gorith
m
‘trainl
m
’
has
produce
d
the
hi
gh
e
r
acc
ur
acy
of
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Evalu
atin
g
wi
ndowi
ng
-
base
d con
ti
nu
ous
S
-
tr
an
sf
orm wi
th
ne
ural netw
ork
cl
as
sif
ie
r for
...
(
K. D
au
d)
1139
cl
assifi
cat
ion
com
par
ed
to
ot
her
al
gorithm
s.
As
for
the
volt
age
swell
cl
assifi
cat
ion
,
L
evenbe
rg
-
Ma
r
qu
a
r
dt
‘trainlm
’
cl
assifi
ed
100%
of
cl
assi
ficat
ion
accu
racy,
m
or
e
hig
he
r
com
par
ed
Gr
a
dient
Descen
t
with
Mom
entu
m
and
Ad
a
ptive
L
R
‘train
gdx’
a
nd
BF
GS
Q
ua
si
-
Ne
wton
‘tra
inbfg
’
wh
e
re
t
hey
pr
oduce
d
98%
of
accuracy
pe
rce
ntage.
Figure
7
.
Archi
te
ct
ur
e
of
ML
P
-
NN
T
able
2
.
Cl
assi
ficat
ion
base
d on OC
WT
–
S
a
m
ple 1
Ty
p
e of
PQD
Test Set
Tr
ain
in
g
Algo
rithm
traing
d
x
(
%)
trainl
m
(%)
trainb
f
g
(
%)
Interrup
t
10
90
100
90
Sag
40
100
100
9
7
.5
Swell
50
98
100
98
Accurac
y
100
98
100
97
Table
3
s
hows
the
cl
assifi
cat
ion
of
Sam
ple
2
us
in
g
MLP
N
N
cl
assifi
er
wi
th
OC
WT.
Fro
m
the
PQD
s
cl
assifi
cat
ion
,
it
was
f
ound
that
trai
ning
al
gorithm
Gr
adi
ent
Desce
nt
w
it
h
Mom
entum
and
Ad
a
ptive
L
R
‘train
gdx’
cl
as
sifie
d
97%
of
accuracy.
Me
anwhil
e
f
or
Le
venbe
rg
-
Ma
r
quar
dt
‘t
rainlm
’
,
the
res
ult
of
PQDs
cl
assifi
cat
ion
is
99%,
w
hile
f
or
al
go
rithm
B
FG
S
Qu
a
si
-
N
e
wton
‘trai
nbfg’;
it
pr
ovide
d
96%
o
f
cl
assifi
cat
ion
accurate
ness.
The
‘trainlm
’
pro
du
ce
d
t
he
highest
acc
ur
a
cy
of
cl
assi
ficat
ion
c
om
par
ed
to
ot
her
al
gorithm
s.
As
f
or
vo
lt
age
swell
cl
assifi
cat
ion
,
Le
ve
nb
e
r
g
-
Ma
rquardt
‘t
rainlm
’
cl
assified
10
0%
acc
urat
e
of
cl
assifi
c
at
ion,
m
or
e
hig
he
r
c
om
par
ed
Gr
a
di
ent
Desce
nt
with
Mom
entu
m
and
A
da
ptive
LR
‘trai
ngdx’
a
nd
BF
GS
Qu
asi
-
New
t
on ‘
trai
nbfg’ whe
re t
hey
pro
du
ce
d 9
7.8
% of acc
ur
acy
per
c
e
ntage
.
T
able
3
.
Cl
assi
ficat
ion
base
d on OC
WT
–
S
a
m
ple 2
Ty
p
e of
PQD
Test Set
Tr
ain
in
g
Algo
rithm
traing
d
x
(
%)
trainl
m
(%)
trainb
f
g
(
%)
Interrup
t
10
100
100
90
Sag
45
9
5
.6
9
7
.8
9
5
.6
Swell
45
9
7
.8
100
9
7
.8
Accurac
y
100
97
99
96
Table
4
s
hows
the
cl
assifi
cat
ion
of
Sam
ple
3
us
in
g
MLP
N
N
cl
assifi
er
wi
th
OC
WT.
Fro
m
the
PQD
s
cl
assifi
cat
ion
,
it
was
f
ound
that
trai
ning
al
gorithm
Gr
adi
ent
Desce
nt
w
it
h
Mom
entum
and
Ad
a
ptive
L
R
‘train
gdx’
cl
a
ssifie
d
98%
of
accu
rate
c
la
ssific
at
ion
.
Me
anwhil
e
f
or
Le
ve
nb
e
rg
-
Ma
rquardt
‘tr
ai
nl
m
’,
the
PQDs
cl
as
sific
at
ion
is
99%,
wh
il
e
f
or
al
go
rithm
BFGS
Q
uasi
-
Ne
wton
‘trai
nbfg’;
it
pr
ovide
d
98%
of
accuracy.
S
o,
al
gorithm
‘tr
ai
nl
m
’
pro
duc
ed
t
he
higher
accu
racy
of
cl
assifi
cat
ion
com
par
ed
to
othe
r
al
gorithm
s.
As
f
or
volt
age
s
well
cl
assifi
cat
ion
,
al
l
ty
pe
of
trai
ni
ng
al
go
rithm
s
pr
od
uc
e
d
100%
of
ac
cur
acy
per
ce
ntage
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
13
, N
o.
3
,
Ma
rc
h 201
9
:
1136
–
1142
1140
T
able
4
.
Cl
assi
ficat
ion
base
d on OC
WT
–
S
a
m
ple 3
Ty
p
e of
PQD
Test
Set
Tr
ain
in
g
Algo
rithm
traing
d
x
(
%)
trainl
m
(%)
trainb
f
g
(
%)
Interrup
t
10
100
100
100
Sag
50
96
98
96
Swell
40
100
100
100
Accurac
y
100
98
99
98
Table
5
s
hows
the
c
om
par
iso
n
of
accu
racy
Sam
ple
1
f
or
the
cl
assifi
cat
io
n
of
P
QD
s
us
i
ng
MLP
N
N
cl
assifi
er
wit
h
dif
fer
e
nt
trai
ning
al
go
rith
m
s
accord
in
g
to
HC
WT.
F
r
om
the
analy
sis,
trai
ni
ng
al
gorithm
Gr
a
dient
De
sc
ent
with
Mom
entum
and
A
da
ptive
LR
‘tra
ingdx’
pr
ov
i
de
d
98%
of
an
accuracy
an
d
BFGS
Qu
asi
-
Ne
wton
‘train
bfg
’
al
s
o
pr
ov
i
ded
98
%
for
the
cl
assifi
cat
ion
of
PQDs.
Wh
il
e
cl
assifi
cat
ion
us
in
g
Leve
nb
e
rg
-
Ma
rquardt
‘trai
nlm
’
a
lgo
rithm
pro
vid
e
d
the
hi
gh
est
acc
ur
ac
y
com
par
ed
th
e
oth
er
s
with
99%
of
correct
cl
assifi
cat
ion
of
PQD
s.
As
f
or
vo
lt
a
ge
swell
cl
assifi
cat
ion
,
al
l
ty
pe
of
trai
ning
al
gorithm
s
pr
oduc
e
d
98%
of
acc
ura
cy
p
erce
ntage
.
T
able
5
.
Cl
assi
ficat
ion
base
d on HC
WT
–
S
a
m
ple 1
Ty
p
e of
PQD
Test
Set
Tra
in
in
g
Algo
rithm
traing
d
x
(
%)
trainl
m
(%)
trainb
f
g
(
%)
Interrup
t
10
90
100
90
Sag
40
100
100
9
7
.5
Swell
50
98
98
98
Accurac
y
100
98
99
98
In
Ta
ble
6
shows
t
he
com
par
iso
n
of
accu
racy
Sam
ple
2
f
or
the
cl
as
sific
at
ion
of
P
QD
s
us
in
g
MLPN
N
cl
assi
fier
with
dif
fere
nt
trai
ni
ng
al
gorithm
s
base
d
on
CST
a
cco
rd
i
ng
to
HC
WT.
F
ro
m
the
a
na
ly
sis,
trai
ning
al
gori
thm
Gr
adient
Desce
nt
with
Mom
entu
m
and
A
da
ptive
L
R
‘train
g
dx
’
pro
vid
e
d
95%
of
a
n
accuracy
an
d
BFGS
Q
ua
si
-
Ne
wton
‘tr
ai
nbfg’
pro
vi
ded
96%
f
or
the
cl
assif
ic
at
ion
of
P
QD.
Wh
il
e
cl
assifi
c
at
ion
us
in
g
Le
venbe
rg
-
Ma
r
quar
dt
‘t
rainlm
’
trai
ning
al
gori
thm
pr
ovide
d
t
he
highest
acc
ur
acy
com
par
ed
the
oth
ers
with
97%
of
c
orre
ct
cl
assifi
cat
io
n
of
P
Q
Ds.
As
f
or
volt
ag
e
swell
cl
assif
ic
at
ion,
Leve
nb
e
rg
-
Ma
rquardt
‘trai
nlm
’
c
la
ssifie
d
100%
of
cl
assif
ic
at
ion
accu
rac
y,
m
or
e
highe
r
com
par
ed
G
ra
dient
Desce
nt
with
Mom
entu
m
and
A
da
ptive
L
R
‘train
gdx’
a
nd
B
FGS
Q
ua
si
-
Ne
wton
‘tr
ai
nbfg’
w
he
re
they
pro
du
ce
d 9
7.8
% accu
racy.
T
able
6
.
Cl
assi
ficat
ion
base
d on HC
WT
–
S
a
m
ple 2
Ty
p
e of
PQD
Test
Set
Tr
ain
in
g
Algo
rithm
traing
d
x
(
%)
trainl
m
(%)
trainb
f
g
(
%)
Interrup
t
10
90
90
90
Sag
45
9
3
.3
9
5
.6
9
5
.6
Swell
45
9
7
.8
100
9
7
.8
Clas
sif
icatio
n
acc
u
racy
100
95
97
96
In
Ta
ble
7
shows
t
he
com
par
iso
n
of
accu
racy
Sam
ple
3
f
or
the
cl
as
sific
at
ion
of
P
QD
s
us
in
g
MLPN
N
cl
assi
fier
with
dif
fere
nt
trai
ni
ng
al
gorithm
s
base
d
on
CST
a
cco
rd
i
ng
to
HC
WT.
F
ro
m
the
a
na
ly
sis,
trai
ning
al
gori
thm
Gr
adient
Desce
nt
with
Mom
entu
m
and
A
da
ptive
L
R
‘
trai
ngdx
’
pro
vid
e
d
97%
of
a
n
accuracy
an
d
BFGS
Q
uas
i
-
Ne
wton
‘tra
inbfg
’
al
so
pro
vid
e
d
97%
fo
r
the
cl
assifi
cat
ion
of
PQ
D.
Wh
il
e
cl
assifi
c
at
ion
us
in
g
Le
venbe
rg
-
Ma
r
quar
dt
‘t
rainlm
’
trai
ning
al
gori
thm
pr
ovide
d
t
he
highest
acc
ur
acy
com
par
ed
the
oth
ers
with
98%
of
c
orre
ct
cl
assifi
cat
io
n
of
P
Q
Ds.
As
f
or
volt
ag
e
swell
cl
assif
ic
at
ion,
Leve
nb
e
rg
-
Ma
rquardt
‘trai
nlm
’
cl
assifi
ed
97.5%
accu
rate
of
cl
assifi
c
at
ion
,
higher
com
par
ed
Gr
a
dien
t
Desce
nt
with
Mom
entu
m
and
A
da
ptive
L
R
‘train
gdx’
a
nd
B
FGS
Q
ua
si
-
Ne
wton
‘tr
ai
nbfg’
w
h
e
re
they
pro
du
ce
d 9
5% of ac
cu
racy p
e
rcen
ta
ge.
In
re
gards
t
o
t
h
e
sam
ples
as
sh
ow
n
i
n
Fi
gure
8,
t
he
highe
st
accuracy
of
cl
assifi
cat
ion
f
or
in
div
id
ual
disturba
nces
is
100%
by
us
in
g
CST
OC
WT
for
S
am
ple
1,
Sam
ple
2
an
d
Sam
ple
3,
whil
e
cl
assifi
cat
ion
by
us
in
g
CST
H
C
W
T
,
N
N
cl
assifi
er
reac
h
98%
f
or
Sam
ple
1,
100%
for
Sa
m
ple
2
and
97
.
5%
f
or
Sa
m
ple
3.
On ther
h
a
nd, Fi
gure
9
s
hows
the c
om
par
iso
n
in
term
o
f
th
e eff
ect
ive
ness
operati
ng ti
m
e
taken usi
ng d
i
ff
e
rent
trai
ning
al
gorithm
s
to
com
pl
et
e
the
PQ
Ds
cl
assifi
cat
ion
.
By
us
i
ng
ei
ght
nodes
of
hidd
en
la
ye
r
for
the
PQ
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Evalu
atin
g
wi
ndowi
ng
-
base
d con
ti
nu
ous
S
-
tr
an
sf
orm wi
th
ne
ural netw
ork
cl
as
sif
ie
r for
...
(
K. D
au
d)
1141
disturba
nces cl
assifi
cat
ion
, t
he
Lev
e
nb
e
r
g
-
Ma
rquardt ‘trai
nlm
’
al
go
rithm
co
m
plete
d
the
cl
assifi
cat
ion
proces
s
faster c
om
par
e
d
to
o
t
her al
gor
it
h
m
s an
d he
nc
e i
m
pr
ovem
ent in ove
rall
ef
fici
ency.
T
able
7
.
Cl
assi
ficat
ion
base
d on HC
WT
–
S
a
m
ple 3
Ty
p
e of
PQD
Test
Set
Tr
ain
in
g
Algo
rithm
traing
d
x
(
%)
trainl
m
(%)
trainb
f
g
(
%)
Interrup
t
10
100
90
100
Sag
50
98
100
98
Swell
40
95
9
7
.5
95
Clas
sif
icatio
n
acc
u
racy
100
98
98
97
Figure
8
.
Anal
ysi
s o
f v
oltage
swell
s classi
fic
at
ion
perform
a
nce
Figure
9
.
Oper
at
ing
ti
m
e o
f
P
Q dist
urba
nce
cl
assifi
cat
ion
98
97.8
100
98
97.8
95
100
100
100
98
100
97.5
98
97.8
100
98
97.8
95
0
20
40
60
80
1
0
0
1
2
0
Sam
p
le
1
Sam
p
le
2
Sam
p
le
3
Sam
p
le
1
Sam
p
le
2
Sam
p
le
3
ST
On
e
C
y
cle
W
T
-
ML
P
NN
ST
Half
C
y
cle
W
T
-
ML
PNN
P
e
rc
e
ntag
e
of a
c
c
ura
c
y
(%
)
t
r
a
i
n
g
d
x
t
r
a
i
n
l
m
t
r
a
i
n
b
f
g
Ave
ra
g
e
C
lasssi
fic
a
ti
on =
99.1%
Ave
ra
g
e
C
lassific
a
ti
on =
97.8%
(98.7
%
)
4
1
7
5
2
9
3
1
6
4
1
10
4
1
4
4
1
5
0
2
4
6
8
10
12
tra
i
ngdx
tra
i
nlm
tra
i
nbf
g
tra
i
ngdx
tra
i
nlm
tra
i
nbf
g
tra
i
ngdx
tra
i
nlm
tra
i
nbf
g
S
a
mpl
e
1
S
a
mpl
e
2
S
a
mpl
e
3
T
i
m
e
t
ak
en
(s
eco
n
d
s
)
(T
y
p
e
o
f
T
rai
n
i
n
g
A
l
g
o
ri
t
h
m
)
On
e-
C
y
cle
W
T
Half
-
C
y
cle
W
T
(98.5
%
)
(95.8
%
)
(100%
)
(98%
)
(98.5
%
)
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
13
, N
o.
3
,
Ma
rc
h 201
9
:
1136
–
1142
1142
4.
CONCL
US
I
O
N
We
hav
e
pres
ented
t
he
e
valuati
on
of
t
he
pro
po
se
d
m
et
ho
dolo
gy
for
de
te
ct
ing
a
nd
c
la
ssifyi
ng
of
PQD
sig
nals.
The
detect
ion
i
s
base
d
on
CS
T
with
ei
the
r
OC
W
T
or
HCWT.
Me
a
nwhi
le
,
the
cl
assifi
c
at
ion
is
i
m
ple
m
ented
us
ing
MLP
N
N.
Fu
rt
her
m
or
e,
S
-
co
ntour
m
at
rices
are
util
iz
ed
to
extract
the
releva
nt
featu
r
es
of
PQDs
t
hat
ser
ver
as
a
in
pu
t
for
e
valuati
ng
the
P
QD
cl
assi
ficat
ion
.
T
hr
ee
dif
fer
e
nt
trai
ni
ng
al
gorithm
s
we
re
us
e
d
to
eval
uate
the
accuracy
of
the
P
Q
D
cl
assifi
cat
ion
.
T
he
res
ults
ha
ve
sh
ow
n
that,
th
e
trai
nin
g
al
g
ori
thm
of Leve
nber
g
-
Ma
rquad
t
‘trai
nlm
’
ou
tpe
rform
ed
oth
e
rs
e
spe
ci
al
ly
f
or
clas
sifyi
ng
t
he vo
lt
age s
well
.
ACKN
OWLE
DGE
MENTS
The
a
utho
r
ac
knowle
dges
t
he
fina
ncial
s
uppo
rt
gi
ven
by
Mi
nistry
of
Higher
E
du
cat
ion
(MO
HE
)
Ma
la
ysi
a fo
r s
ponsori
ng this
researc
h
i
n
the
form
o
f gr
a
nt
-
in
-
ai
d 600
-
RM
I
/FR
GS
5/3 (
0103
/
2016)
.
REFERE
NCE
S
[1]
C
.
S
anka
r
an
,
“
Pow
er
Qualit
y
,
Unite
d
State
of
Am
eri
ca
,”
CRC
PR
ESS
LL
C,
200
2.
[2]
K.
Daud
,
et
al
.
,
“
Cla
ss
ifi
cation
of
Pow
er
Quali
t
y
Disturb
ance
base
d
on
Conti
nuous
S
-
Tra
nsform
-
W
indowing
Tech
n
ique
(CST
-
W
T)
and
AN
OV
A a
s a
Fe
at
ur
e S
el
ection
,
”
vo
l.
785,
pp
.
368
–
37
2,
2015
.
[3]
K.
Daud,
et
al
.
,
“
New
W
indow
ing
Te
chn
ique
Dete
c
ti
on
of
Sa
gs
and
Sw
el
ls
Based
on
Conti
nuous
S
-
Tra
nsform
(CST)
,
”
vo
l
/i
ss
u
e:
2
(
4
)
,
2012
.
[4]
C.
Venka
t
esh,
e
t
al
.
,
“
Cla
ss
ific
at
ion
of
V
olt
ag
e
S
ag,
S
well
an
d
H
armonics
us
ing
S
-
T
ran
sform
base
d
M
odul
ar
N
eur
al
N
et
work
,
”
ICHQ
P
2010
-
14th
In
t. Conf
.
Har
mon.
Qual.
Powe
r
,
v
o
l
.
1,
p
p.
1
–
7
,
2010
.
[5]
E.
S
y
stems
,
“
An
E
xper
t
S
y
stem
B
ase
d
on
S
-
T
ran
sform
and
N
eur
al
N
et
work
for
A
utomati
c
C
la
ss
if
ic
a
ti
on
of
P
ower
Q
ual
ity
D
isturbance
s
,
”
2009.
[6]
P.
Fei,
et
al
.
,
“
A
Novel
Po
wer
Quali
t
y
Disturbances
Dete
ction
and
Clas
sific
at
ion
Me
t
hod
,
”
vol.
737
,
pp.
193
–
198
,
20
15.
[7]
N.
H.
T.
Huda,
e
t
al.
,
“
Pow
er
Q
u
al
ity
S
ignals
D
et
ec
t
ion
using
S
-
T
ran
sform
,
”
Proc.
2013
IEE
E
7th
Int.
Powe
r
Eng.
Optim.
Conf
.
PEO
CO 2013
,
pp.
552
–
557,
2013
.
[8]
G.
U.
Jin
-
hong
,
et
a
l.
, “
Sim
ula
tion of
Pow
er
Qua
li
t
y
Us
ing
S
-
Trans
form
,
”
vol. 42
9,
pp
.
172
–
178
,
2012.
[9]
Z.
Za
n
ja
n,
“
New Met
hod
to
Ana
l
y
z
e
Pow
er
Qualit
y
Phenom
ena u
sing Wave
let
an
d
S
-
Tra
nsform
Matri
x
.
”
[10]
M.
A.
Dandwat
e
,
et
al
.
,
“
Gene
ration
of
Mathe
m
at
i
ca
l
Mod
el
s
for
var
ious
PQ
Signal
s
u
sing
MA
TL
AB
,
”
pp.
47
–
51
,
2014
.
[11]
V.
Nati
onal,
e
t
al.
,
“
Detect
ion
of
Pow
er
Quali
ty
Disturban
ce
s
Us
ing
W
ave
le
t
Tra
nsform
And
Artifi
c
ia
l
Neur
a
l
Network
,
”
pp.
1
–
5,
2014
.
[12]
C.
Eng
,
“
Th
e
E
ff
ec
ts of
Tr
ai
ning
Algorit
hm
s in
M
LP
Network
on
I
m
age
Cla
ss
ifica
t
ion
.
”
[13]
H.
P.
Gavin,
“
The
Le
venb
er
g
-
Marqua
rdt
M
et
hod
for
N
onl
ine
ar
L
ea
st
S
q
uar
es
C
urve
-
F
itting
P
roble
m
s
,”
pp.
1
–
19
,
2017
.
[14]
M.
T.
Haga
n,
et
al.
,
“
Tra
in
ing
F
ee
dforward
N
e
tworks
with
the
Marqua
rdt
A
lgori
thm
-
Neura
l
Networks
,
”
IE
EE
Tr
ansacti
ons on
,
vol
/i
ss
ue:
5
(
6
)
,
pp.
2
–
6
,
1994
.
[15]
A.
Li
kas
,
et
al.
,
“
Tr
ai
ning
the
R
andom
N
eur
al
N
et
w
ork
using
Q
uasi
-
Newton
M
e
thods
,
”
vol
.
1
26,
pp.
331
–
339
,
20
00.
Evaluation Warning : The document was created with Spire.PDF for Python.