Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
8
, No
.
6
,
Decem
ber
201
8
, p
p.
482
9
~
483
5
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v8
i
6
.
pp
482
9
-
483
5
4829
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
Advanced
SOM
& K Me
an Meth
od f
or L
oa
d
Curve
Clust
er
in
g
Pha
n
Thi T
h
anh Bi
nh
1
,
Tr
ong
N
gh
ia
Le
2
,
Nu
i Ph
am
X
u
an
3
1,3
Depa
rt
m
ent of
Elec
tr
ical and
E
le
c
troni
cs
Engi
n
ee
ring
,
HCM
C
Univer
sit
y
of
T
e
chnol
og
y
,
Vie
tn
am
2
Depa
rt
m
ent
of Electrical a
nd
E
l
ec
tron
ic
s
Eng
ineeri
ng,
HCM
C
Univer
sit
y
of Techn
olog
y
and
Ed
uca
t
ion
,
Viet
n
a
m
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Feb
1
, 2
01
8
Re
vised
Jun
3
0
, 201
8
Accepte
d
J
ul
22
, 2
01
8
From
the
loa
d
c
urve
class
ifi
c
at
i
on
for
one
custo
m
er,
the
m
ai
n
f
e
at
ure
s
such
as
the
sea
sonal
fac
tors,
the
we
e
kda
y
f
actors
inf
lue
nc
ing
on
the
el
e
ct
ri
cit
y
consum
pti
on
m
ay
b
e
ex
tra
c
te
d
.
B
y
th
is
wa
y
som
e
utilit
ie
s
c
an
m
ake
de
ci
sion
on
the
ta
r
iff
b
y
s
ea
sons
or
b
y
da
y
in
wee
k.
Th
e
popula
r
cl
ust
eri
ng
te
chn
ique
s
are
the
SO
M
&
K
-
m
ea
n
or
Fuzz
y
K
-
m
ea
n.
SO
M
&K
m
ea
n
is
a
prom
ine
nt
appr
oac
h
for
c
lu
steri
ng
with
a
t
wo
-
le
vel
app
roa
ch:
first,
the
da
t
a
set
will
be
cl
uster
ed
using
the
SO
M
and
in
the
sec
ond
le
v
el,
the
SO
M
will
be
cl
uster
ed
b
y
K
-
m
ea
n.
In
the
first
le
vel,
two
tra
ini
ng
a
lgori
thms
were
exa
m
ine
d:
seque
ntial
and
bat
ch
tra
in
ing.
For
the
sec
ond
le
ve
l,
the
K
-
m
ea
n
has
the
result
s
tha
t
are
strongl
y
d
epe
nd
ed
on
the
initial
val
ues
of
the
c
ent
ers.
To
over
come
thi
s,
t
his
pape
r
u
sed
t
he
subtrac
t
ive
c
l
usteri
ng
appr
o
ach
proposed
b
y
Ch
iu
in
199
4
to
de
te
rm
ine
t
he
c
ent
ers
.
Be
c
ause
th
e
eff
ecti
ve
rad
ius
in
Chiu’s
m
et
hod
has
som
e
infl
ue
nce
on
the
nu
m
ber
of
ce
n
te
rs
,
th
e
pap
er
appl
i
ed
the
PS
O
te
chn
ique
to
fin
d
the
opti
m
um
rad
ius.
To
va
li
d
t
he
proposed
appr
oac
h
,
the
te
st
on
w
el
l
-
k
nown
dat
a
sa
m
ple
s
is
ca
rr
ied
out.
Th
e
appl
i
ca
t
ions f
or daily
loa
d
cur
v
e
s of
one
Souther
n
uti
l
ity
ar
e
pr
ese
nte
d
.
Ke
yw
or
d:
Cl
us
te
r
a
naly
sis
K
-
m
ean
PSO
SO
M
Subtract
ive cl
ust
ering
Copyright
©
201
8
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
:
Ph
a
n
T
hi T
hanh
Bi
nh
,
Dep
a
rt
m
ent o
f El
ect
rical
an
d
Ele
ct
ro
nics
E
nginee
rin
g,
HCMC
Unive
r
sit
y of
Tec
hnol
og
y,
268 Ly
Thu
ong Kiet
steee
t, 10
district
, Ho
c
hi m
inh
cit
y, V
ie
tna
m
.
Em
a
il
: ptt
bin
h@hcm
ut.edu.
vn
1.
INTROD
U
CTION
The
loa
d
cu
rve
cl
assifi
cat
ion
has
one
im
po
rtant
m
eaning
:
the
util
it
y
can
dr
a
w
the
ow
n
featur
e
for
each
gro
up
in
on
e
cl
ass
of
c
on
s
um
er
[1
]
-
[
2]
.
Her
e
the
m
a
in
featu
res
su
c
h
as
the
seas
onal
facto
rs,
the
wee
k
day
factors, infl
uen
ci
ng on the
elec
tric
ity con
su
m
ption
m
a
y be ex
tract
ed.
By
this w
ay
so
m
e u
ti
liti
es can
m
ake
decisi
on
on
th
e
ta
riff
by
sea
so
ns
or
by
da
y
in
a
week
.
So
m
e
util
i
ti
es
will
hav
e
the
diff
e
ren
t
pri
c
es
on
el
ect
rici
ty
fo
r
winter
,
s
umm
e
r.
Othe
rs
will
ta
ke
t
he
pr
ic
es
for
wor
king
da
ys
in
diff
e
re
nc
e
with
th
os
e
f
or
the
week
e
nd
with
the
ve
ry
cl
ear
pur
po
se:
to
s
hi
ft
loads
from
work
i
ng
days
t
o
the
week
e
nd.
Ma
ny
util
it
ie
s
desig
n
their
dem
and
re
spon
se
poli
cy
for
eac
h
c
us
t
om
er g
r
oup ha
vi
ng the
sam
e fo
rm
o
f
loa
d
c
u
r
ves [3
]
.
Loa
d
cu
rv
e
cl
a
ssific
at
ion
is
th
e
cl
us
te
rin
g
wi
th
the
la
rg
e
num
ber
of
input
data.
The
daily
load
cu
rv
e
for
ye
ars
or
m
on
t
hs
m
us
t
be
con
side
re
d.
F
ro
m
the
po
int
of
data
m
ining,
the
way
of
c
lusterin
g
big
da
ta
i
s
necessa
ry
to
extracti
ng
us
e
ful
info
rm
at
ion
.
Ma
ny
auth
or
s
con
ce
ntrate
d
on
data
cl
us
te
ri
ng
basin
g
on
the
K
-
m
ean
al
gorith
m
becau
se
it
is
rathe
r
easy
to
i
m
ple
m
ent
and
app
ly
e
ven
on
la
rg
e
data
set
.
Jung,
et
al
use
d
K
-
m
eans
al
gorith
m
s
co
m
bin
in
g
with
pri
nci
pal
com
po
ne
nt
an
al
ysi
s
to
analy
ze
and
cl
assif
y
us
er
data
ef
fici
ently
[4]
.
B
ut
as
m
e
ntion
e
d
in
[
5
]
-
[
7],
K
-
m
eans
has
t
he
resu
lt
s
that
str
ongly
de
pende
d
on
th
e
init
ia
l
values
of
t
he
centers,
s
o
this
will
influ
ence
on
the
cl
ust
eri
ng
resu
lt
s.
T
o
ov
e
r
com
e
this
dr
a
wb
ac
k,
Be
dboudi,
et
al
use
d
the
com
bin
ing
K
-
m
ean
and
ge
ne
ti
c
al
go
rithm
,
m
eanw
hile
Sa
hu,
et
al
use
d
the
A
dap
ti
ve
K
-
m
ean
[5]
,
[
6]
.
Chiu
pr
ese
nted
the
subtract
ive
m
et
ho
d
t
o
re
m
ov
e
the
in
fluen
ce
of
ce
nt
er
init
ia
li
zat
ion
[
7]
.
For
lo
ad
c
urve
cl
us
te
rin
g,
m
a
ny
works
are
ba
sed
on
dim
ensi
on
al
it
y
red
uc
ti
on
in
orde
r
to
si
m
plify
the
m
od
el
s
or
redu
ce
the
com
pu
ta
ti
on
ti
m
e
su
ch
as
[8
]
-
[
10
]
.
H
ere
the
featur
e
sel
ect
ion
or
c
on
st
ru
ct
ion
is
the
m
a
in
key
fo
r
cl
ust
erin
g.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4828
-
483
5
4830
Fo
r
e
xam
ple,
[10]
propose
d
three
ways
to
co
ns
tr
uct
the
featur
es
,
ex
cepti
on
al
ly
ar
e
su
it
able
for
s
m
art
m
et
ering
:
co
ndit
ion
al
filt
ers
on
ti
m
e
-
reso
luti
on
bas
ed
fe
at
ur
es,
cal
ibrat
ion
a
nd
norm
al
iz
at
ion
,
an
d
us
i
ng
profi
le
e
rrors.
Othe
r
w
orks
con
ti
nue
to
use
the
adv
a
nt
ages
of
K
-
m
ean
al
gorithm
and
c
om
bin
e
with
the
dim
ension
al
it
y
reducti
on
al
gorithm
fo
r
l
oa
d
cu
r
ve
cl
us
te
rin
g.
T
he
popula
r
cl
ust
erin
g
te
chn
i
qu
es
,
ba
sed
on
this
com
bin
in
g,
a
re
the
SOM
&
K
-
m
eans.
W
it
h
the
la
rg
e
num
ber
of
input
data,
S
OM
&
K
-
m
eans
is
a
prom
inent
ap
proac
h
for
cl
us
te
rin
g.
In
[
11]
this
te
ch
nique
i
s
with
tw
o
-
le
ve
l
appr
oach
:
fir
st,
the
data
set
will
be
cl
us
te
red
us
in
g
the
SO
M
by
sequ
e
ntial
trai
nin
g
al
go
rithm
.
The
res
ult
her
e
is
a
set
of
prot
otype
vect
or
s.
In
th
e
seco
nd
le
vel,
t
he
S
OM
will
be
cl
us
te
re
d
by
K
-
m
ean.
But
t
his
m
et
ho
d
c
onta
ins
the
wea
k
poi
nts
of
K
-
m
ean
so
do
e
s
no
t
ha
ve
t
he hig
h
acc
ur
a
cy
. Besides,
th
e seq
ue
ntial
training al
gorith
m
f
or
S
OM is t
i
m
e con
s
um
pti
on.
To
ta
ke
the
f
ul
l
adv
anta
ge
of
SO
M
&
K
-
m
ean
with
the
big
da
ta
,
to
over
c
om
e
i
ts
dr
aw
bac
k,
thi
s
pap
e
r
will
us
e
the
subtract
ive
cl
us
te
rin
g
m
eth
od
f
or
the
sec
on
d
le
vel.
H
oweve
r,
c
hoos
i
ng
the
e
ff
ect
ive
rad
i
us
is
on
e
key
que
sti
on
of
cl
us
te
r
ing
proce
dure.
W
e
pro
posed
app
ly
in
g
the
P
SO
te
c
hn
i
qu
e
to
fi
nd
the
opti
m
u
m
rad
i
us
i
n
order
to
im
pr
ov
e
the
accuracy.
T
he
pap
e
r
al
s
o
us
e
d
a
no
t
her
trai
ni
ng
way
in
SOM
-
the
batc
h
tr
ai
ning
al
gorithm
to
enh
a
nce
t
he
cal
culat
ing
tim
e.
To
validat
e
t
he
pro
posed
m
e
thod,
t
he
Fu
zz
y
K
-
m
ean
al
gorithm
will
b
e als
o
a
ppli
ed
to
g
i
ve
t
he
co
m
par
iso
n.
The
w
ork
is
or
gan
iz
e
d
as
the
fo
ll
owin
g:
so
m
e
m
at
he
m
atics
def
i
niti
on
s
uch
as
SO
M,
K
-
m
ean,
F
uzzy
K
-
m
ean,
PS
O
will
be
m
entioned
i
n
Sect
io
n
2;
the
pr
opose
d
al
gorithm
(d
en
oted
a
s
A
dvance
d
S
OM
&
K
m
eans)
will
be
pr
ese
nte
d
in
S
ect
ion
3
with
s
om
e
te
sts
on
the
fam
ou
s
data
set
;
finall
y,
on
e
case
stu
dy
w
il
l
be
pr
ese
nted
i
n
S
ect
ion
4,
c
ompari
ng
t
he
res
ults
of
dif
fer
e
n
t
al
gorithm
s
su
c
h
as
S
OM
&
K
-
m
eans,
Fu
zzy
K
-
m
ean.
2.
SOM
E
MAT
HEM
ATIC D
EFINITIO
NS
2.1.
SOM
The
SO
M
co
nsi
sts
of
a
regu
la
r,
us
ually
tw
o
-
dim
ension
al
2D
gri
d
of
m
ap
un
it
s.
Data
po
i
nts
ly
ing
near
each
ot
he
r
in
the
i
nput
sp
ace
a
re
m
app
ed
onto
nearby
m
ap
unit
s.
The
SO
M
ca
n
be
i
nter
pr
et
e
d
as
a
topolo
gy prese
rv
i
ng m
app
ing
from
inp
ut s
pa
ce o
nto
t
he 2
-
D gr
i
d of
m
ap un
it
s.
In
our
wor
k,
the
two
al
gori
thm
s
fo
r
trai
ni
ng
of
t
he
m
a
ps
we
re
ca
rr
ie
d
out:
seq
ue
ntial
trai
nin
g
al
gorithm
and
batch
trai
ning.
The
ne
uro
n
whose
weig
ht
ve
c
tor
is
cl
os
est
t
o
the
i
nput
vect
or
is
cal
le
d
t
he
best
-
m
at
ching
unit
(BMU)
denote
d
by
c.
I
n
the
batch
trai
ni
ng
al
gorithm
,
inst
ead
of
us
in
g
a
sing
le
data
vec
tor
at
a
tim
e,
the
wh
ol
e
data
set
is
presented
t
o
the
m
ap
befor
e
a
ny
adjustm
ents
are
m
ade
(h
en
ce
the
nam
e
“batch”
)
.
In
each
trai
ning
ste
p,
the
dat
a
set
is
par
ti
ti
on
ed
acc
ordin
g
to
the
Voron
oi
reg
io
ns
of
the
m
ap
weigh
t
ve
ct
or
s
,
i.e.
each
data
vecto
r
belo
ngs
to
the
data
set
of
the
cl
os
est
m
ap
un
it
.
Af
t
er
this,
the
ne
w
weig
ht
vect
or
s
are
cal
culat
ed
as
f
ollo
w
s:
n
j
ic
n
j
j
ic
i
t
h
x
t
h
t
m
1
1
)
(
)
(
)
1
(
(1)
wh
e
re:
t
denot
es
tim
e;
x
j
is a
n
in
put vect
or
;
h
ci
(
t
)
the n
ei
ghbo
rho
od K
e
rnel
arou
nd the
winner
unit
;
k
j
k
m
x
c
m
i
n
a
r
g
is t
he
i
nd
e
x of t
he
BM
U
of
da
ta
sam
ple, w
it
h
m
k
is sy
nap
ti
c w
ei
ght
vecto
r
k
.
2.2.
The K
-
mea
n
al
go
ri
t
hm
The
K
-
m
ean
-
al
gorithm
is
a
w
el
l
-
known
al
go
rithm
in
cl
us
te
r
ing
fiel
d.
F
or
e
ach
cl
us
te
r
nu
m
ber
K,
t
he
proce
dure
fo
ll
ows a
sim
ple w
ay
to
cl
assify
a
giv
e
n data set
and lo
oks li
ke
t
hat:
m
i
n
2
1
1
k
i
n
j
i
j
z
x
F
(2)
wh
e
re,
.
is
the
Eucli
dea
n
dist
ance
betw
een
x
j
an
d
z
i
.
;
z
i
-
is
the
center
of
the
i
th
cl
us
te
r;
k
-
is
the
num
ber
of
cl
us
te
rs
ce
nter
s;
n
-
nu
m
ber
of
data
.
T
h
e
Da
vies
-
B
ou
l
din
(
DB)
in
de
x
is
a
pp
li
ed
f
or
hard
cl
us
te
ri
ng
[
6
]
.
The
op
ti
m
al
n
um
ber
of
cl
us
te
rs
corres
ponds t
o
th
e
m
ini
m
u
m
v
al
ue of
D
B i
nd
e
x.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Adv
an
ce
d S
om
&
K Mean
Me
thod fo
r L
oad
Curve Cl
us
te
ri
ng (Ph
an T
hi T
hanh Bi
nh
)
4831
2.3.
The sub
tracti
ve
me
thod
Con
si
der
a
c
ollec
ti
on
of
n
da
ta
po
i
nts
{
x
1
,
x
2
…
x
n
}
in
an
M
dim
ension
al
sp
ace
.
I
f
eac
h
data
po
i
nt
is
consi
der
i
ng as
a possible cl
us
t
er ce
nter,
t
he
n t
he
pote
ntial
of d
at
a
po
i
nt
x
i
will
b
e:
n
k
x
x
i
i
k
e
P
1
1
(3)
with
2
/
4
a
r
.
The
c
on
sta
nt
r
a
is
eff
e
ct
ively
the
radi
us
de
fini
ng
a
neig
hborh
ood.
The
data
point
with
th
e
highest
po
te
ntial
is sele
ct
ed
a
s the
first clu
ste
r
ce
nter.
Let
x
1
* be the
locati
on of t
he
fi
rst c
luster ce
nter
a
nd
P
1
*
be
it
s
po
te
ntial
v
al
ue. T
he po
t
entia
l of each
dat
a p
oi
nt
x
i
is
re
vised
b
y t
he
f
or
m
ula:
*
1
*
1
x
x
i
i
k
e
P
P
P
(4)
with
2
/
4
b
r
,
w
he
re
r
b
is
the
e
ff
ect
i
ve
rad
i
us
a
nd
be
e
qu
al
to
1.25
r
a
.
The
dat
a
point
with
t
he
highest
rem
ai
nin
g
pote
ntial
is
sel
ect
ed
as
t
he
sec
ond
cl
us
te
r
cente
r.
T
he
process
is
then
c
onti
nued
f
ur
t
her
un
ti
l
the
rem
ai
nin
g
pote
ntial
o
f
all
d
at
a
points
fall
s b
el
ow so
m
e fr
act
i
on of t
he po
te
nt
ia
l of
the
f
ir
st
cl
us
te
r
ce
nter
P
1
*.
2.4.
The PSO
[1
2]
PSO
was
base
d
on
t
he
ph
e
no
m
eno
n
of
c
ollec
ti
ve
intel
li
gence
insp
i
red
by
the
so
ci
al
beh
a
vior
of
bir
d
floc
king
or
fis
h
sch
ooli
ng.
T
he
fitnes
s
f
un
ct
ion
is
e
valuate
d
f
or
eac
h
par
ti
cl
e
in
the
swa
r
m
and
is
com
par
ed
t
o
the
fitness
of
the
be
st
prev
i
ous
po
sit
io
n
f
or
that
par
ti
cl
e
pbest
t
an
d
to
t
he
fitness
of
t
he
global
best
pa
rtic
le
a
m
on
g
al
l
par
ti
cl
es
in
the
swar
m
gb
est
.
Af
te
r
fin
ding
the
two
best
values
,
the
i
th
par
ti
cl
es
evo
l
ve
by
up
dating
their
velocit
ie
s and
posit
ion
s
accor
ding t
o
th
e f
ollow
i
ng
eq
uations:
)
(
*
)
(
*
2
2
1
1
1
k
i
i
k
i
i
k
i
k
i
s
g
b
e
s
t
r
a
n
d
c
s
p
b
e
s
t
r
a
n
d
c
wV
V
(5)
1
1
k
i
k
i
k
i
V
s
s
(6)
wh
e
re:
s
k
-
cu
rr
e
nt
searc
hing
point;
s
k+1
-
m
od
ifie
d
sear
chin
g
point;
v
k
-
-
cu
rr
e
nt
vel
ocity
;
v
k+
-
-
m
od
i
fied
velocit
y;
ra
nd
1
and
r
and2
-
the
rand
om
values
in
(0,1)
fo
ll
ow
i
ng
a
norm
al
distribu
ti
on
;
c
1
an
d
c2
are
const
ants
cal
le
d
acce
le
rati
on
coeffic
ie
nts;
w
-
so
m
e
weigh
te
d
coe
ff
ic
ie
nt.
The
va
lues
of
c1
an
d
c
2
co
nt
ro
l
the
weig
ht b
al
a
nce
of
pb
est
a
nd
gbest
in
d
eci
ding the
p
a
rtic
le
’s
n
e
xt m
ov
em
e
nt.
2.5.
Fuz
z
y
K
-
mea
ns (
F
K
M)
[
13]
FK
M i
s
one cl
us
te
rin
g
m
et
hod wit
h hig
h flexibil
it
y hav
i
ng
the foll
owin
g o
bj
ect
iv
e f
unct
ion
:
m
i
n
)
,
(
1
1
2
K
i
n
j
j
i
ij
x
z
d
w
F
(7)
wh
e
re
α
is
a
weig
hting
e
xp
on
e
nt;
ij
w
is
the
va
lue
of
m
e
m
b
ersh
i
p
f
unct
io
n
an
d
d(z
i
,
x
j
)
is
the
Eucli
de
an
distance
bet
we
en
x
j
a
nd
the
c
enter
z
i
.
of
i
cl
ust
er.
F
or
deter
m
ining
the
fi
na
l
nu
m
ber
of
c
lusters,
the
re
a
re
m
an
y
crit
eria
are a
pp
li
ed.
T
his pape
r used
the m
et
ho
ds i
n
[
14
]
bas
ed on t
he p
rinc
iples B
el
l
m
and
–
Zade
h.
3.
THE
PROPO
SED
ALGO
R
ITHM
The
pr
opos
e
d
al
gorithm
(d
en
oted
as
Adva
nc
ed
S
OM
&
K
m
eans)
will
be
sh
ow
n
in
Fi
g
ure
1.
T
he
batch
trai
ni
ng
appr
oach
is
use
d
an
d
the
trai
ning
tim
e
wil
l
be
enh
a
nce
d.
Her
e
the
S
ub
t
r
act
ive
cl
us
te
ring
is
app
li
ed
to
fi
nd
ou
t
the
init
ia
l
centers
f
or
K
-
m
eans.
Trad
it
i
on
al
ly
,
the
rad
i
us
r
a
in
(3)
has
the
values
fro
m
0.
15
to
0.8.
Our
e
xa
m
ining
sho
ws
that
the
sm
a
ll
e
r
the
r
a
is,
the
l
arg
e
th
e
num
ber
of
cl
us
te
rs
will
be
receive
d.
S
o,
the
optim
u
m
rad
ius
is
the
on
e
that
will
le
ad
to
the
sm
a
ll
est
value
of
DB
i
nd
e
x.
T
o
fi
nd
ou
t
the
s
uitabl
e
rad
iu
s
r
a
, th
is
p
a
pe
r
a
pp
li
ed
the
PS
O
algorit
hm
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4828
-
483
5
4832
S
t
a
r
t
S
u
b
t
r
a
c
t
i
v
e
c
l
u
s
t
e
r
i
n
g
K
-
m
e
a
n
s
D
B
_
i
n
d
e
x
M
i
n
?
E
n
d
Y
e
s
r
a
P
a
r
t
i
c
l
e
S
w
a
r
m
O
p
t
i
m
i
z
a
t
i
o
n
N
o
D
a
t
a
P
r
o
c
e
s
s
i
n
g
S
O
M
(
B
a
t
c
h
t
r
a
i
n
i
n
g
)
D
a
t
a
c
l
u
s
t
e
r
s
I
n
p
u
t
D
a
t
a
Figure
1. The
pro
po
se
d
al
gor
it
h
m
4.
E
X
PERI
MEN
TAL STU
DIE
S
4.1.
Te
sting
on t
he
wel
l
-
kno
w
n
d
ata
s
amp
le
s
:
Thr
ee
real
a
nd
fam
ou
s
data
s
et
s
(Iris,
WBCD,
Wine
)
a
re
t
aken
T
hese
da
ta
set
s
are
us
e
d
in
m
an
y
works
f
or
te
sti
ng
t
he
cl
us
te
ring
te
c
hniq
ue.
The
I
ris
Pla
nts
Database
[
15
]
con
ta
in
s
15
0
s
a
m
ples
(4
at
tri
bu
te
s
i
n
each
sam
ple)
and
w
as
cl
us
te
red
i
nto
3
cl
as
ses:
Ir
is
S
et
osa
;
Ir
is
Ver
sic
ol
our;
Ir
is
Vir
gi
nica
(
50
sam
pl
es
f
or
each
cl
ass)
.
T
he
W
isc
onsin
Breast
Ca
ncer
Database
[
16
]
was
buil
t
from
the
Un
ive
rsity
of
Ho
s
pi
ta
ls.
It
con
ta
in
s
68
3
te
st
(10
at
trib
utes
in
each
te
st)
and
was
cl
us
te
re
d
i
nto
2
cl
asses:
be
nign
(
65.
5%)
a
nd
m
align
a
nt
(35.5%
).
T
he
la
st on
e [
17
]
is the d
at
a obtai
ne
d
f
ro
m
a ch
em
ic
al
an
al
ysi
s
of
wines grow
n
in the s
am
e reg
io
n
in
Ital
y bu
t der
ive
d
from
thr
ee diff
e
ren
t cult
iva
r
s.
The
analy
sis
d
et
erm
ined
the q
ua
ntit
ie
s o
f
13
c
on
s
ti
tue
nt f
ou
nd
in
each
of
the
three
ty
pes
of
wines
.
Thr
ee
a
lgorit
hm
s:
SO
M
&
K
-
m
ean,
FK
M,
an
d
A
dv
a
nce
d
SO
M
&
K
-
m
ean
are
ap
plied
an
d
the r
esu
lt
s
are
giv
e
n
in
Tabl
e
1
.
From
Table
1,
t
he
c
on
cl
us
io
n
is
th
at
Advan
ce
d
S
OM
&
K
-
m
ean h
as
th
e b
est
res
ult
.
Table
1.
T
est
in
g resu
lt
s
on
we
ll
-
known
d
at
a
sam
ples
Data sa
m
p
l
e
Nu
m
b
e
r
o
f
the
co
rr
ect
clu
ster
Alg
o
rith
m
s
SOM
&
K
-
m
e
an
s
FKM
Ad
v
an
ced
SOM
&
K
-
m
eans
Ir
is
3
2
2
3
W
BC
D
2
3
2
2
W
in
e
3
2
7
3
4.2.
Ap
pli
ca
tio
n fo
r load cur
ve
cl
ust
eri
n
g
4.2.1.
The i
npu
t
d
ata
The
365
daily
load
c
urves
of
on
e
util
it
y
in
the
S
outh
of
Vi
et
nam
are
the
input
data.
Eac
h
loa
d
c
urv
e
is
reg
a
r
ded
as
t
he
vecto
r
of
24
at
tribu
te
s
(
24
hours).
T
he
E
uc
li
dean
distanc
es
bet
ween
tw
o
loa
d
c
urves
j
an
d
k
will
b
e
def
i
ned as:
24
1
2
)
(
i
ik
ij
jk
x
x
d
(8)
wh
e
re,
x
ij
-
loa
d at
i
-
hour
of
j
-
lo
ad
c
urve.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Adv
an
ce
d S
om
&
K Mean
Me
thod fo
r L
oad
Curve Cl
us
te
ri
ng (Ph
an T
hi T
hanh Bi
nh
)
4833
4.2.2.
Extr
act th
e
in
fo
rm
ati
on
Fr
om
the
cl
us
te
rin
g
process
, b
y
lookin
g
into
each
cl
us
te
r,
t
he
m
ai
n
factor
s
char
act
erize
d
each
cl
us
te
r
m
ay
be
ext
racted.
For
e
xam
ple
if
the
loa
d
c
urves
i
n
one
c
luster
a
re
belo
ng
e
d
t
o
t
he
rainy
seas
on,
w
hile
i
n
oth
e
r
cl
us
te
r
-
th
e
dr
y
seas
on,
t
hen
it
ca
n
say
that
there
is
a
ne
cessi
ty
to
for
m
a
season
al
t
ariff.
A
nd
if
th
ere
are
the d
i
ff
e
ren
t cl
us
te
rs
b
y
wee
ke
nd and
w
orki
ng d
ay
,
the
we
ekend
day tari
f
f
m
us
t be
form
ed.
4.2.3.
Implem
ent
ati
on
As
im
ple
m
enta
ti
on
, h
ere
the dai
ly
load
cu
rv
e
s
of
one u
ti
li
ty
in
the
ye
ar of
2012 w
ere u
se
d.
The
ta
ri
f
f
is
TOU
(tim
e
of
use
)
an
d
is
the
sa
m
e
fo
r
al
l
day
in
week
.
All
three
al
go
rithm
s
hav
e
the
sam
e
nu
m
ber
of
cl
us
t
er
(
2
cl
ust
ers)
cal
le
d
ho
li
day
cl
us
te
r
and
norm
al
day
cl
us
te
r.
Th
ere
is
no
sho
w
of
the
rainy
an
d
dry
seaso
n
cl
us
te
r
s.
A
ll
of
S
unday
an
d
public
ho
li
da
ys
are
belo
ng
e
d
t
o
the
holi
day
cl
us
te
r.
This
re
su
lt
is
consi
ste
nt
bec
ause
t
he
Hoch
iM
inh
ci
ty
is
with
t
he
tr
opic
al
cl
i
m
at
e,
and
on
t
he
oth
e
r
hand,
the
re
a
re
m
any
industrial
pa
r
ks
an
d
in
veste
d
abroa
d
e
nterpr
ise
s
so
t
hat
the
dif
fer
e
nce
i
n
l
oad
by
sea
son
s
is
not
cl
early
.
The
Ho
li
day
cl
us
te
r
c
on
ta
ine
d
al
l
of
S
unday
an
d
public
ho
li
da
ys
acco
rd
i
ng
to
Viet
nam
’s
Lab
or
Co
de.
S
o
th
at
there
are
63
da
ys
in
sta
nd
a
r
d
holi
day
cl
ust
er
can
see
in
Fig
ure
2.
It
e
m
ph
asi
zes
the
necessit
y
to
fo
rm
the
diff
e
re
nt
pri
ces
on
el
ect
rici
ty
f
or
w
orki
ng
day
s
an
d
H
olidays
.
But
the r
es
ult
sho
ws
t
hat
the
re
a
re
m
or
e
th
a
n
63
days
in
the
ho
l
iday
s
cl
us
te
r.
Ther
e
a
re
so
m
e
Saturd
ay
s
a
nd
w
orkin
g
day
s
fall
ing
into
the
holi
day
cl
us
te
r
can
see
in
Ta
ble
2.
Figure
2. The
load cu
r
ves of 6
3
sta
nd
a
rd pu
bl
ic
h
olidays
Ther
e
are
di
fferences
in
the
res
ults
of
3
al
gorithm
s
can
see
in
Ta
b
le
2
.
T
o
c
on
si
de
r
the
res
ult
accuracy
of
th
r
ee
al
go
rithm
s,
the
distance
of
each
dif
fer
e
nt
day
to
center
of
the
sta
nd
a
r
d
ho
li
days
(
63
days)
and the
norm
al
days
will
b
e c
al
culat
ed
ca
n
s
ee in Ta
ble
3.
The
re
su
lt
s
of
FK
M
an
d
Adv
anced
SO
M
&
K
-
m
eans
are
c
oin
ci
de
d
e
xce
pt
fo
r
4
days
(Sa
tur
days:
4
-
Feb.,
11
-
Fe
b.
,
18
-
Fe
b,
6
-
Oct.
).
Accor
ding
to
FK
M
,
these
s
days
be
lo
ng
e
d
to
t
he
holi
da
y
cl
us
te
r.
But
from
Tabl
e
3,
these
s
Satur
days
have
the
dista
nce
to
the
ce
nter
of
the
sta
nd
a
r
d
norm
al
day
clu
ste
r
sm
al
le
r
t
han
of
the
sta
ndar
d
holi
day
cl
us
te
r
.
It
m
eans
that
these
4
Satu
r
da
ys
m
us
t
belo
ng
to
the
norm
al
day
cl
us
te
r
.
A
nd
t
hat
m
eans FKM i
s
le
ss accu
rate t
ha
n Adva
nced S
OM & K
-
m
eans.
The
resu
lt
s
of
SO
M
&
K
-
m
e
an
a
nd
A
dvan
ced
S
OM
&
K
-
m
eans
are
c
oin
ci
de
d
e
xcep
t
for
on
e
da
y
(Tu
e
sd
ay
:
31
-
J
an).
This
T
ues
day
has
the
distance
to
the
ce
nter
of
the
sta
ndar
d
norm
al
day
cl
us
te
r
la
rg
er
tha
n
the
sta
nd
a
rd
holi
day
cl
us
te
r
and
m
us
t
be
belonged
to
the
s
ta
nd
a
rd
holi
da
y
cl
us
te
r.
So
,
the
A
dv
a
nce
d
S
OM
&
K
-
m
eans alg
ori
th
m
g
et
s the
be
tt
er accuracy
than
S
OM
&
K
-
m
eans.
Table
2.
N
um
ber
of
wee
kd
ay
s
in Holiday cl
ust
er for 3 al
gor
it
h
m
s
W
eekd
ay
Alg
o
rith
m
s
SOM
&
K
-
m
e
an
s
FKM
Ad
v
an
ced SOM
&
K
-
m
e
an
s
Mon
d
ay
7
7
7
Tues
d
ay
3
4
4
W
ed
n
esd
ay
3
3
3
Thu
rsd
ay
2
2
2
Friday
2
2
2
Satu
rday
4
8
4
Su
n
d
ay
53
53
53
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4828
-
483
5
4834
Table
3.
Dista
nc
e of all
d
i
ff
e
re
nt d
ay
s
to
sta
ndar
d h
oliday cl
us
te
r
’s
ce
nter (
SH
CC
)
a
nd
no
rm
al
d
ay
cluster’s
center
(SNDC
C)
Day
Av
g
.
d
ist.
to
the SHCC
Av
g
.
d
ist.
to
the
SNDCC
31
-
Jan
-
12
1
0
3
0
.4
8
1
9
8
8
.7
0
4
-
Feb
-
12
1
7
4
2
.4
2
1
0
5
2
.6
0
11
-
Feb
-
12
1
7
5
3
.6
9
9
7
7
.21
18
-
Feb
-
12
1
7
4
2
.3
5
1
0
1
4
.1
4
6
-
Oct
-
12
1
8
2
0
.4
8
1
0
4
6
.4
4
This
em
ph
asi
zes
the
fact
tha
t
Advan
ce
d
S
OM
&
K
-
m
ea
ns
al
gorithm
ov
erc
om
e
the
weak
point
of
tho
se
al
go
rith
m
based
on
the
K
-
m
ean,
an
d
the
ch
oosin
g
of
op
ti
m
al
rad
ius
in
Subtract
ive m
et
ho
d
e
nh
a
nc
es
the
accuracy.
4.2.4.
Co
m
pa
r
e in
time calcul
at
i
on d
oma
in
Chan
ging
S
O
M
trai
ning
by
the
batc
h
trai
ning
al
gorit
hm
gr
eat
ly
re
duc
es
trai
ning
ti
m
e.
Be
sides,
app
ly
in
g
the
S
ub
t
racti
ve
cl
ust
ering
al
gorith
m
to
get
init
ia
l
center
i
n
K
-
m
eans
ca
n
le
ad
t
o
quit
e
fast
s
ol
ution
the
perform
ance
te
sts
wer
e
m
ade
in
a
com
pu
te
r
with
4
GB
s
of
m
e
m
or
y
a
nd
2.4
G
Hz
I
nt
el
Core
i3
CPU
an
d
hav
e
the
f
ollo
wing
resu
lt
s:
S
OM & Km
eans
-
15
99(s)
;
A
dvanced S
OM &
K
-
m
eans
-
62(s);
FK
M
-
488 (s
).
5.
CONCL
US
I
O
N
The
data
analy
sis
pr
ese
nted
i
n
this
w
ork
ha
s
been
te
ste
d
a
nd
validat
ed
usi
ng
real
data
of
on
e
util
ity
and
the
well
-
know
n
data
sa
m
ples.
A
m
on
g
three
al
go
rith
m
s
exa
m
ined
in
this
pa
per
,
t
he
pro
posed
A
dv
a
nce
d
SO
M
&
K
-
m
e
ans
has
the
be
tt
er
resu
lt
a
nd
sm
a
ll
est
tim
e
for
cal
c
ulati
ng
.
T
his
al
gorithm
ov
erc
om
es
so
m
e
disad
va
ntages
of
t
rad
it
io
nal
S
OM
&
K
-
m
eans,
FK
M.
I
n
th
e
res
ults,
the
da
il
y
con
su
m
ption
be
ha
vio
r
of
a
r
eal
util
it
y
has
bee
n
analy
zed
by
cl
us
te
rin
g
an
d
it
sh
ow
s
that
it
is
necessary
to
m
ake
diff
ere
nt
el
ect
rici
t
y
pr
ic
es
for
work
i
ng d
ay
s a
nd
for weeke
nds.
T
his alg
or
it
hm
can
al
so
b
e
u
sed f
or
clu
ste
rin
g
dif
fer
e
nt grou
ps
of cust
om
ers
-
the
basic
for
app
ly
in
g
dif
fe
ren
t
ta
riff
f
or
diff
e
re
nt
custo
m
er
cl
asses.
Fo
r
the
f
uture
works,
the
stu
dy
of
po
s
sibil
it
y t
o
a
pp
ly
this
alg
or
i
thm
f
or
detect
ing t
i
m
e zon
es
of Tim
e
-
of
-
Us
e tari
ff will
b
e
carried
out.
ACKN
OWLE
DGE
MENTS
The
a
uthors
w
ou
l
d
li
ke
to
t
ha
nk
t
he
HCMC
Un
i
ver
sit
y
of
Tec
hnol
ogy
and
HCMC
Un
i
ver
sit
y
of
Tech
no
l
og
y a
nd E
ducat
ion f
or their
s
upports
.
REFERE
NCE
S
[1]
G
.
Chic
co,
et
al
.
,
“
Custom
er
cha
racte
ri
zation
o
pti
ons
for
impr
oving
the
ta
r
iff
offe
r
,
”
IEE
E
Tr
ans.
Powe
r
Syst
,
v
ol
/i
ss
ue:
18
(1),
pp
381
-
387
,
200
3.
[2]
D.
Gerbe
c
,
et
a
l
.
,
“
Dete
rm
inatio
n
and
al
lo
cation
of
t
y
p
ical
lo
ad
profil
es
to
th
e
e
l
igi
ble
customers
,
”
in
Proc
IEEE
Bol
ogna
Powe
r
Tech
,
Bo
logna It
aly
,
2003
.
[3]
S.
Vale
ro,
et
a
l.
,
“
Methods
for
customer
and
de
m
and
response
poli
cies
select
io
n
in
new
elec
tr
i
ci
t
y
m
ark
e
ts
,
”
I
ET
Gene
r.
Tr
ansm
.
Distrib
.,
v
ol
/i
ss
u
e:
1(
1
)
,
pp
.
104
-
110,
2007
.
[4]
S
.
H
.
Jung,
et
al
.
,
“
Predi
ct
ion
Da
ta
Proc
essing
Sc
heme
using
an
Artifi
c
ia
l
Neura
l
Network
and
Data
Cluste
r
ing
fo
r
Big
Data
,
”
Inte
r
nati
onal
Journal
of
El
ec
tri
cal
an
d
Computer
Engi
nee
ring
(
IJE
C
E)
,
v
ol
/i
ss
ue:
6
(
1
)
,
pp.
330
-
336
,
2016.
[5]
A
.
Bedboudi,
e
t
al.
,
“
An
Hete
r
ogene
ous
Popul
at
ion
-
B
ase
d
Ge
net
i
c
Algorit
hm
for
Data
Clust
eri
ng
,
”
Indone
si
an
Journal
of
Elec
t
rical
Engi
ne
erin
g
and
Informati
c
s (
IJE
EI)
,
v
ol
/
issue:
5
(
3
)
,
pp
.
275
-
284
,
2017
.
[6]
M
.
Sahu
,
et
al
.
,
“
Para
m
et
ric
Co
m
par
ison
of
K
-
m
ea
ns
and
Adapti
ve
K
-
m
ea
ns
Cl
usteri
ng
P
erf
orm
anc
e
on
Diff
erent
Im
age
s
,
”
In
te
rn
ati
onal
Journal
of
Elec
tric
al
an
d
Computer
En
gine
ering
(
IJE
CE)
,
v
ol
/i
ss
ue:
7
(
2
)
,
pp
.
810
-
817
,
2017.
[7]
S.
L.
Chiu,
“
Fuzz
y
m
odel
ide
nt
ifi
c
at
ion
base
d
on
cl
uster
esti
m
at
ion
,
”
Journal
of
Inte
lligen
t
and
Fuzzy
Syste
ms
,
v
ol
/i
ss
ue:
2
(
3
)
,
1
994.
[8]
N.
Jin,
et
al
.
,
“
Subgroup
discover
y
in
sm
art
e
lect
ric
ity
m
e
te
r
data
,
”
Industrial
Inf
orm
ati
cs,
IE
EE
Tr
ansacti
ons
on
,
vol
/i
ss
ue:
10
(
2
)
,
pp.
1327
-
1336
,
2014.
[9]
I.
Dent
,
et
a
l
.
,
“
Vari
ability
of
b
e
havi
our
in
e
lectr
ic
ity
lo
ad
profile
cl
ust
eri
ng;
who
does
thi
ngs
at
th
e
sam
e
t
ime
e
ach
da
y
,
”
in
Ad
vances
in
Data
Mini
ng.
Appl
i
cat
ions
and
Theoreti
cal
Aspec
ts
,
ser.
Lect
ure
Note
s
in
Computer
Sci
ence
,
P.
Perne
r
,
E
d.
Springer
In
te
rn
at
i
onal
Publishing
,
vol.
8557
,
pp
.
70
-
84,
2014
.
[10]
R
.
Al
-
Otai
bi
,
et
al.
,
“
Feat
ure
C
onstruct
ion
and
Cal
ibr
at
ion
for
Cluste
ring
Daily
Loa
d
Curves
from
Sm
art
Mete
r
Data
,
”
Industria
l
Informatic
s,
IE
EE
Tr
ansacti
ons
on
,
vol
/i
ss
ue
:
12
(
2
)
,
pp
.
645
-
65
4,
2016
.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Adv
an
ce
d S
om
&
K Mean
Me
thod fo
r L
oad
Curve Cl
us
te
ri
ng (Ph
an T
hi T
hanh Bi
nh
)
4835
[11]
S
.
V
.
Verdú,
et
al.
,
“
Cla
ss
ifica
t
i
on,
Filt
er
ing,
an
d
Ide
nti
fica
ti
on
of
El
ectrical
Cu
stom
er
Loa
d
Patt
ern
s
Through
th
e
u
se
of
Self
-
Orga
niz
ing
Maps
,
”
I
EE
E
Tr
ansacti
o
ns on
power
syst
ems
,
v
ol
/
issue:
21
(
4
)
,
2006
.
[12]
M
.
El
-
T
ara
b
ily
,
et
al
.
,
“
A
PSO
–
Based
on
Subtrac
ti
v
e
Dat
a
Cluste
r
ing
Al
gorit
hm
,
”
In
te
rn
ati
onal
Journal
of
Re
search in Co
mputer
Scienc
e,
v
ol
/i
ss
ue:
3
(
2
),
p
p.
1
-
9
,
2013
.
[13]
N.
R.
Pal
and
J.
C.
Bez
d
ek,
“
On
Cluste
r
Vali
d
ity
for
the
Fuzz
y
c
-
m
ea
ns
m
odel
,
”
I
EE
E
Tr
ans,
Fuzzy
syst
.
,
vol
/i
ss
ue:
3
(
3
)
,
pp
.
370
-
37
9,
1995
.
[14]
P
.
T
.
T
.
Binh
,
e
t
al
.
,
“
Dete
rm
in
at
ion
of
Repre
s
ent
a
ti
ve
Lo
a
d
Curve
base
d
on
F
uzzy
K
-
Me
ans
,
”
Proc.
PE
OCO
,
2010
.
[15]
R.
A.
Fis
her
,
“
The
use
of
m
ult
i
ple
m
ea
surem
en
ts
in
ta
xonom
ic
proble
m
s
,
”
Ann
ual
Euge
n
ic
s
,
v
ol.
7,
Part
II
,
pp
.
179
-
188
,
1936
.
[16]
O.
L.
Manga
sar
i
an
and
W
.
H.
W
olbe
rg,
“
Canc
er
dia
gnosis
via
li
n
ea
r
progr
amm
ing
,
”
SI
AM
News
,
v
ol
/i
ss
ue:
23
(
5
)
,
pp.
1
-
18
,
1990
.
[17]
Forina,
e
t
al
.
,
“
An
Ext
endi
ble
Packa
ge
for
Data
Exp
lora
t
io
n
,
”
Classifi
cat
i
on
and
Corr
el
ati
on.
Insti
tute
of
Pharmace
utical and F
ood
Analy
sis and
Technol
o
gie
s,
Vi
a
Br
iga
t
a
Sale
rno
,
16147
Genoa
,
Ital
y
.
BIOGR
AP
H
I
ES
OF
A
UTH
ORS
Phan
Thi
Tha
nh
Binh
rec
ei
ved
Ph.D.
degr
ee
in
el
e
ct
ri
ca
l
eng
ineeri
ng
from
Kiev
Poly
technique
Univer
sit
y
,
Ukrai
ne
in
1995
.
C
urre
ntly
,
she
is
a
As
sos
.
profe
ss
or
and
lectur
er
in
th
e
Facu
l
t
y
El
e
ct
ri
ca
l
and
Elec
tron
ic
s
Eng
ineeri
ng,
HCM
UT.
Her
m
ai
n
are
as
of
rese
ar
ch
int
er
ests
are
power
s
y
stems
stabilit
y,
power
s
y
st
ems
oper
ation
and
co
ntrol
,
loa
d
for
ecasti
ng,
data
m
ining.
Trong
Nghia
Le
rec
ei
v
ed
his
M
.
Sc.
degr
e
e
in
e
le
c
tri
c
al
engi
n
eering
from
Ho
C
hi
Minh
Cit
y
Univer
sit
y
of
T
ec
hnolog
y
and
Educ
a
ti
on
(HCM
UTE)
,
Viet
n
a
m
,
in
2012.
Curre
ntly
,
h
e
is
a
le
c
ture
r
in
the
Facul
t
y
Elec
tri
c
al
and
El
e
ct
ron
i
cs
Engi
nee
r
ing,
HCM
UTE.
His
m
ai
n
are
as
of
rese
arc
h
intere
st
s
are
loa
d
shedd
ing
in
powe
r
s
y
s
te
m
s,
power
s
y
s
t
ems
stabi
li
t
y
,
lo
ad
fore
ca
sting
and
distr
ibut
ion
net
work.
Nui
Pham
Xuan
rec
ei
v
ed
his
M
.
Sc.
degr
ee
in
e
le
c
tri
c
al
engi
n
eering
from
Ho
C
hi
Minh
Cit
y
Univer
sit
y
of
T
ec
hnolog
y
,
Vie
t
nam,
in
2013
.
Curre
ntly
,
he
works
at
Qual
ity
As
suranc
e
and
Te
sting
Center
3
(QU
ATEST
3).
His m
ai
n
ar
ea of
rese
a
rch
intere
st
s is
data
m
ini
ng
.
Evaluation Warning : The document was created with Spire.PDF for Python.