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.
489
2
~
490
1
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v8
i
6
.
pp
489
2
-
490
1
4892
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
Appl
ying of Dou
ble Se
asonal AR
IMA Mo
del for E
lectri
ca
l
Powe
r Demand F
orecasti
ng at PT
. PLN G
resik Ind
onesia
Ismit
Mado
1
, Adi S
oepri
jan
to
2
,
Suhar
ton
o
3
1
,2
Depa
rt
m
ent of
Elec
tr
ical
Engi
n
ee
ring
,
Inst
it
ut
T
eknol
ogi
Sepulu
h
Nopem
ber
,
Ind
onesia
3
Depa
rtment of
Stat
isti
cs
,
Inst
it
u
t
T
eknol
ogi
Sep
uluh
Nopem
ber
,
Indone
sia
1
Depa
rtment of
El
e
ct
ri
ca
l
Eng
in
ee
ring
,
Univ
ersitas Borne
o
Ta
r
ak
an
,
Indon
esia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
J
a
n
11
, 2
01
8
Re
vised
Ju
l
8
,
201
8
Accepte
d
J
ul
29
, 2
01
8
The
pre
d
ic
t
ion
of
the
use
of
e
l
ec
tr
ic
power
is
ver
y
important
t
o
m
ai
nta
in
a
bal
an
ce
be
tween
the
suppl
y
an
d
demand
of
elec
tr
ic
power
in
the
power
gene
ra
ti
on
s
y
s
tem
.
Due
to
a
f
lu
ct
ua
ti
ng
of
el
e
ctrical
power
d
emand
in
th
e
el
e
ct
ri
ci
t
y
lo
ad
c
ent
er
,
an
accurat
e
fore
c
asti
ng
m
et
hod
is
req
uire
d
to
m
ai
nta
in
the
eff
icien
c
y
an
d
reliab
i
li
t
y
of
p
ower
generation
s
y
stem
cont
inu
ousl
y
.
Such
condi
ti
ons
gre
at
l
y
aff
ec
t
the
d
y
n
amic
stabi
l
ity
of
power
gene
rati
on
sy
st
ems
.
The
objecti
v
e
of
thi
s
rese
arc
h
is
to
propose
Doub
le
Seasona
l
Aut
ore
gre
ss
ive
Inte
gra
te
d
Movi
ng
Avera
ge
(DS
ARIM
A)
to
pre
dict
el
e
ct
r
ic
i
t
y
loa
d.
Half
hourl
y
lo
ad
da
ta
for
of
thr
ee
y
e
ar
s
per
iod
at
PT
.
P
LN
Gresik
Indo
nesia
powe
r
pla
nt
uni
t
ar
e
used
as
c
ase
stud
y.
The
p
ara
m
eter
s
of
DS
A
RIMA
m
odel
are
esti
m
at
ed
b
y
us
ing
least
squar
e
s
m
et
hod.
Th
e
result
show
s
th
at
th
e
b
est
m
odel
to
pre
dic
t
th
ese
d
a
ta
is
subs
et
DS
ARIMA
with
orde
r
(
[
1
,
2
,
7
,
16
,
18
,
35
,
46
]
,
1
,
[
1
,
3
,
13
,
21
,
27
,
46
]
)
(
1
,
1
,
1
)
48
(
0
,
0
,
1
)
336
with
MA
PE
about
2.
06%
.
T
hus,
future
r
ese
a
rch
coul
d
b
e
do
ne
b
y
using
the
s
e
pre
di
ct
iv
e
result
s a
s m
ode
l
s of
opti
m
al c
on
t
rol
par
amete
rs o
n
the power
s
y
st
em side
.
Ke
yw
or
d:
DSARIMA
m
od
el
E
le
ct
rical
p
ow
er
dem
and
F
oreca
sti
ng
L
east
squar
e
s
m
et
ho
d
T
i
m
e
-
series pa
tt
ern
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
:
Ism
i
t M
ado
,
Dep
a
rt
m
ent o
f El
ect
rical
En
gi
neer
i
ng,
In
sti
tut Te
knol
og
i
Sepulu
h N
op
em
ber
,
Ar
ie
f
Rahm
an Hakim
Road
,
Keputi
h
Ca
m
pu
s, Su
koli
lo
60111, S
ur
a
baya,
Indo
nesia
.
Em
a
il
:
is
m
itm
a
do@
gm
ail.co
m
1.
INTROD
U
CTION
Pr
e
dicti
on
of
el
ect
rical
po
wer
dem
and
is
an
i
m
po
rtant
first
ste
p
in
plann
i
ng
of
po
wer
pl
ant
syst
e
m
op
e
rati
on
[1]
.
Plann
i
ng
of
po
wer
pla
nt
syst
e
m
op
erati
on
is
about
buil
ding
a
pla
n
for
t
he
pr
e
par
at
io
n
of
powe
r
plant
syst
e
m
op
erati
on
f
or
a
certai
n
tim
e
pe
rio
d.
Ba
sed
on
the
issues
to
be
addresse
d,
th
e
pow
er
pla
nt
syst
e
m
op
e
rati
on
pla
n
is
div
ide
d
into sev
eral
ty
pes
of
tim
e
per
iod
pl
an,
i.e.
a
nnual p
la
n,
quarterly
p
la
n,
m
on
thly
p
la
n,
week
ly
plan
a
nd
daily
pla
n.
The
operati
on
plan
of
the
powe
r
plant
syst
e
m
al
so
inclu
des
m
anag
erial
,
m
ai
ntenan
ce,
a
nd
ope
rati
on
s
in
syst
em
s
and
equ
i
pm
ent
to
e
ns
ure
good
ec
onom
ic
value
of
fina
ncin
g,
s
yst
e
m
reli
abili
ty
, an
d serv
ic
e
qual
it
y.
Re
fer
ri
ng
t
o
the
pro
blem
s
s
olv
in
g
in
pow
er
syst
e
m
op
er
at
ion
,
power
predict
io
n
is
cl
assifi
ed
int
o
three
cat
e
gori
es,
i.e.
lo
ng
-
t
erm
,
m
ediu
m
-
te
rm
a
nd
s
hort
-
te
rm
pr
e
dicti
on
s
.
L
ong
te
rm
el
ect
rical
powe
r
pr
e
dicti
on
sa
re
r
equ
i
red
for
pe
ak
loa
d
capaci
ty
plann
in
g
a
nd
syst
em
m
a
intenance
sc
he
dule
[2]
,
m
ediu
m
-
ter
m
pr
e
dicti
on
s
a
re
required
for
pl
ann
i
ng
a
nd
operati
on
of
the
powe
r
plant
syst
e
m
[3]
,
and
s
hort
-
te
rm
pr
edi
ct
ion
s
are
need
e
d for
con
t
ro
ll
in
g
a
nd sch
e
duli
ng the
pow
e
r plant sy
stem
[4]
.
On
e
of
t
he
c
oncer
ns
in
t
he
powe
r
plant
s
yst
e
m
op
erati
on
is
t
he
qual
it
y
of
el
ect
rical
pow
er.
T
he
e
le
ct
ric
power
gen
e
rated
s
hall
be
al
ways
e
qual
to
the
el
ect
ric
powe
r
co
nsu
m
ed
by
the
el
e
ct
ric
power
use
r.
I
f
the
powe
r
that
is
distribu
te
d
is
gr
eat
er
tha
n
r
equ
i
red,
then
t
her
e
will
be
wastage
of
e
nergy.
A
nd
if
the
pow
e
r
pro
du
ce
d
is
sm
al
le
r
than
re
quired
,
it
will
occ
ur
over
loa
d
w
hich
will
af
fect
the
occ
urren
c
e
of
powe
r
ou
t
ages
.
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
Ap
plying of
D
ouble
Se
asonal
ARIMA
Model
for Elect
ric
al
Power
De
man
d
F
or
ec
as
ti
ng ... (Is
mit Ma
do)
4893
In
fact,
t
he
us
e
of
el
ect
rical
energy
te
nds
t
o
change
at
a
ny
tim
e
in
accor
da
nce
with
the
needs
of
c
onsum
ers.
Ther
e
f
or
e,
it
is
necessary
to
predict
the
use
of
el
ect
rical
po
wer
that
is
abl
e
to
m
ai
ntain
a
balance
betwe
en
the
su
pply
a
nd con
su
m
ption
of
el
ect
rical
p
owe
r i
n
the
powe
r pl
ant syst
em
.
Fluctuati
ons
i
n
loads
m
ean
sm
al
l
distur
ba
nc
es
in
the
dy
na
m
ic
al
sta
bility
stud
ie
s
of
po
w
er
ge
ne
rati
on
syst
e
m
s.
This
pro
blem
occu
rs
after
the
first
s
wing
w
he
n
c
on
t
ro
l
e
qu
i
pme
nt
su
c
h
as
govern
or
a
nd
e
xc
it
at
ion
syst
e
m
s
hav
e
worked
.
Dyna
m
ic
sta
bili
ty
st
ud
ie
s
in
plant
syst
e
m
s
are
sti
ll
bein
g
dev
el
oped
[5
]
,
[6]
.
A
naly
sis
of
sm
al
l
sign
al
sta
bili
ty
of
powe
r
syst
em
s
with
pro
bab
il
ist
ic
un
ce
rtai
nt
y
on
re
ne
wab
l
e
ene
rg
y
ge
ne
rati
on
equ
i
pm
ent b
ec
om
es the top
ic
of the c
urre
nt s
tud
y
[7]
.
The
tim
e
serie
s
pr
e
dicti
on
m
od
el
is
an
acc
ur
at
e
ch
oice
and
grow
i
ng
co
ntinuo
us
ly
to
this
day
for
powe
r
pr
e
dicti
on
an
d
f
or
eca
sti
ng
[8]
-
[
10
]
.
The
stu
dy
of
tim
e
series
-
ba
sed
forecast
in
g
of
el
ect
rical
pow
e
r
consum
ption
e
vo
l
ved
i
nto
tw
o
pa
rts,
i.e.
pre
dicti
on
m
od
el
s
base
d
on
sta
ti
sti
cal
m
a
the
m
a
ti
cal
m
od
el
s
suc
h
as
m
ov
ing
ave
ra
ge
,
ex
pon
e
ntial
s
m
oo
thi
ng,
re
gr
essi
on,
an
d
ARIMA
B
ox
-
J
enk
i
ns
;
an
d
ar
ti
fici
al
intelli
gen
ce
-
base
d
pre
dicti
on
m
od
el
s
suc
h
as
ne
ur
al
netw
orks
,
ge
netic
al
gorithm
s,
si
m
ulated
a
nn
eal
in
g,
gen
et
ic
pro
gr
am
m
ing
, classi
ficat
ion
, an
d
hybri
d.
Som
e
tim
e
series
researc
h
based
o
n
s
ta
ti
sti
cal
m
at
hem
atics
as
i
n
[
8]
,
[11]
-
[
15]
.
The
app
li
cat
io
n
of
the
Box
-
Je
nk
i
n
s
ARIM
A
m
od
el
is
de
velo
ped
th
r
ough
a
seaso
nal
patte
r
n
[
16
]
-
[22]
.F
or
m
od
el
s
based
on
arti
fici
al
intel
li
ge
nce
has
al
s
o
be
com
e
the
at
ten
ti
on
of
resear
cher
s
as
in
[23
]
-
[27]
.
The
hybri
d
m
od
el
has
al
s
o
been
de
velo
pe
d
to
obta
in
t
he
best
data
in
e
le
ct
rical
load
pr
e
dicti
o
n
stu
dy
as
in
[28]
-
[
36]
.
This
stu
dy
pro
po
s
es
a
D
oubl
e
Seaso
nal
AR
IMA
(
DSARI
MA)
m
et
ho
d
f
or
pr
e
dicti
ng
or
f
or
ecast
i
ng
powe
r
dem
and
m
od
el
in
PT.
PLN
Gr
esi
k
I
ndonesi
a
ba
sed
on
t
hr
ee
ye
ars
load
trai
ni
ng
and
te
sti
ng
data
(d
ai
ly
data
eve
ry
half
hour)
.
The
pr
e
dicti
on
res
ults
are
us
e
d
as
ref
e
ren
ce
pa
ram
et
ers
fo
r
op
ti
m
u
m
con
trol
in
i
m
pr
ovin
g
the
sta
bili
ty
of
el
ect
rical
ly
gen
erated
syst
em
s
in
our
res
ear
ch
.
Tim
e
seri
es
m
et
ho
d
bas
ed
on
sta
ti
sti
cal
m
a
the
m
at
ic
al
m
od
el
is
cho
sen
bec
ause
of
it
s
su
p
erior
it
y
to
proc
ess
data
w
hich
is
no
t
sta
ti
on
a
ry
an
d
no
t
li
nea
r.
Sta
ti
sti
cal
m
at
hem
at
ic
al
m
od
els
are
al
so
ca
pa
ble
of
ge
ne
ra
ti
ng
data
t
hat
is
no
t
incl
ud
e
d
in
t
he
trai
ning
proces
s.
2.
FORE
CASTI
NG MET
HO
DS
2.1.
AR
I
M
A Mod
el
ARIMA
m
et
ho
d
or
c
omm
on
ly
ref
err
e
d
to
a
s
Box
-
Jen
ki
ns
m
et
ho
d
is
a
m
od
el
inte
ns
ivel
y
dev
el
ope
d
by G
eo
r
ge
Bo
x and
Gwil
yn Je
nk
i
ns
in
19
70. Th
is f
or
ecast
in
g
m
od
el
sti
ll d
om
inate
s
m
any areas o
f resear
ch
to
date.
U
nfor
t
unat
el
y,
ARIMA
m
od
el
can
only
be
app
li
ed
for
sta
ti
on
ary
tim
e
-
series
data.
If
the
data
i
s
no
t
sta
ti
on
ary,
the
n
to
m
ake
the
data
beco
m
es
sta
ti
on
ary
it
is
n
ecessa
ry
to
do
the
diff
e
re
ntiat
ion
pro
cess
[
37]
.Th
e
auto
regressive
(A
R)
m
od
el
ind
ic
at
es a con
ne
ct
ion
betwee
n a value at th
e present ti
m
e
(
)
with a v
al
ue
in t
he
pr
e
vious
ti
m
e
(
−
)
,
plu
s
a
rand
om
value
.
Wh
il
e
t
he
m
ov
in
g
a
ve
rag
e
(MA
)
m
od
el
sho
ws
t
he
dep
e
ndence
of
the curre
nt ti
m
e v
al
ue
(
)
w
it
h
t
he
resid
ual v
al
ue
at
the
pre
vio
us t
im
e
(
−
)
with
=
1
,
2
,
…
.
The
ARIM
A
m
od
el
(
,
,
)
is
a
com
bin
at
ion
of
AR
(
)
and
M
A
(
)
m
od
el
s,
wi
th
d
th
-
order
diff
e
re
ntiat
ion
process
w
he
n
t
he
data
patte
r
n
is
no
t
sta
ti
on
a
ry.
Gen
e
ral
for
m
of
the
AR
I
MA
m
od
el
(
,
,
)
is
as f
ollows:
∅
(
)
(
1
−
)
̇
=
(
)
(
1)
wh
e
re
B
is
the
backshift
opera
tor
a
nd
is t
he r
andom
p
r
ocess
v
al
ue
s,
a
nd
∅
(
)
=
(
1
−
∅
1
−
.
.
.
−
∅
)
(
)
=
(
1
−
1
−
.
.
.
−
)
Gen
e
rali
zat
ion
of
the
ARI
MA
m
od
el
fo
r
data
that
has
a
seas
on
a
l
patte
rn
is
e
xpresse
d
by
ARIMA
(
,
,
)
(
,
,
)
an
d f
orm
ulate
d
as foll
ow
s
[
37
]
:
∅
(
)
Φ
(
)
(
1
−
)
(
1
−
)
̇
=
(
)
Θ
(
)
(
2)
wh
e
re
s
is t
he
s
easo
nal p
e
rio
d, an
d
∅
(
)
=
1
−
∅
1
−
∅
2
2
−
.
.
.
−
∅
Φ
(
)
=
1
−
Φ
1
−
Φ
2
2
−
.
.
.
−
Φ
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
:
489
2
-
490
1
4894
(
)
=
1
−
1
−
2
2
−
.
.
.
−
Θ
(
)
=
1
−
Θ
1
−
Θ
2
2
−
.
.
.
−
Θ
Shor
t
-
te
rm
po
wer
c
on
s
um
pti
on
data
by
co
ns
um
ers
has
a
doub
le
seas
onal
patte
rn,
i.e.
daily
and
week
ly
seas
on.
T
he
AR
I
MA
m
od
el
with
a
doub
le
season
al
patte
rn
is
e
xpress
ed
by
ARIMA
(
,
,
)
(
1
,
1
,
1
)
1
(
2
,
2
,
2
)
2
an
d has t
he
f
ol
lowing
ge
ner
al
for
m
[38]
:
∅
(
)
Φ
1
(
1
)
Φ
2
(
2
)
(
1
−
)
(
1
−
1
)
1
(
1
−
2
)
2
̇
=
(
)
Θ
1
(
1
)
Θ
2
(
2
)
(3)
wh
e
re
1
an
d
2
are
d
if
fer
e
nt seas
onal
p
e
rio
ds
.
2.2.
AR
I
M
A
B
ox
-
Jenk
ins
Mo
d
e
l Proced
ure
The pre
dicti
on
proce
dure
of AR
IMA
B
ox
-
Je
nk
i
ns
m
od
el
th
rou
gh f
ive
stag
es of it
erati
on, a
s foll
ows:
i.
Pr
e
par
at
io
n of
data, incl
udin
g chec
king
of d
a
ta
stat
ion
ary.
ii.
Id
e
ntific
at
ion o
f
AR
IMA
m
od
el
throug
h
a
utoc
orrelat
ion f
unct
ion
a
nd p
a
rtia
l autoc
orrelat
ion f
unct
ion.
iii.
Estim
at
ion
o
f
ARIMA
m
od
el
p
a
ram
et
ers:
p
,
d
,
and
q
.
iv.
Determ
inati
on
of A
RIM
A
m
od
el
equati
ons.
v.
Pr
e
d
ic
ti
on
.
2.3.
Le
as
t
Sq
u
ares
Estima
tio
n
On
e
m
et
ho
d
th
at
can b
e u
se
d
to
est
im
a
te
ARIMA
m
od
el
p
a
ram
et
ers
is
the
le
ast
sq
ua
res
m
et
hod
[39]
.
Fo
r
AR
(
1),
c
a
rr
ie
d ou
t by
i
nc
lud
in
g
no
n
-
ze
ro
m
ean
par
am
et
er,
μ.
This
pa
ram
et
er
is
further
est
im
a
te
d
by
le
ast
sq
ua
res
.
Co
ns
i
der
t
he
first
-
or
der
case
w
her
e
−
=
∅
(
−
1
−
)
+
:
The
eq
uatio
n
is
a
regressio
n
m
od
el
with
−
1
as the p
re
dictor varia
ble
and
as th
e re
s
pons
e
var
ia
ble.
L
east
squar
es
est
i
m
at
ion
the
n processe
d by
m
ini
m
iz
ing
the
su
m
o
f
s
quare
s of
diff
e
re
nce
s
(
−
)
−
∅
(
−
1
−
)
(
4)
Since
on
ly
1
,
2
,
…
,
are
observe
d,
t
hen w
e ca
n o
nly s
um
f
ro
m
=
2
to
=
. L
et
it
(
∅
,
)
=
∑
{
(
−
)
−
∅
(
−
1
−
)
}
2
=
2
(
5)
This
eq
uatio
n
is
cal
le
d
the
co
nd
it
io
nal
su
m
sq
ua
re
d
f
un
ct
i
on.
Acc
ordin
g
to
the
basic
pri
nciple
of
the
le
ast
sq
uar
e
s
m
et
ho
d,
we
c
an
est
i
m
at
e
∅
and
by
the
resp
ect
ive
val
ues
th
rou
gh
pa
ram
et
er
values
1
,
2
,
…
,
.
Co
ns
ide
r
t
he e
qu
at
io
n
⁄
=
0
. We
hav
e
=
∑
2
[
(
−
)
−
∅
(
−
1
−
)
]
(
−
1
+
∅
)
=
2
=
0
(
6)
or, sim
plifyi
ng
and s
olv
i
ng fo
r
,
=
1
(
−
1
)
(
1
−
∅
)
[
∑
=
2
−
∅
∑
−
1
=
2
]
(
7)
Now, f
or lar
ge
n
,
1
−
1
∑
=
2
=
1
−
1
∑
−
1
=
2
=
̅
Th
us
, re
gardless o
f
th
e
value of
∅
, equati
on (
7
)
re
duces t
o
̂
=
1
1
−
(
̅
−
∅
̅
)
=
̅
(
8)
then
it
ca
n be
wr
it
te
n,
̂
=
̅
Nex
t
we rec
onsider
m
ini
m
iz
i
ng the e
quat
io
n
(
∅
,
̅
)
. W
e
h
a
ve
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
Ap
plying of
D
ouble
Se
asonal
ARIMA
Model
for Elect
ric
al
Power
De
man
d
F
or
ec
as
ti
ng ... (Is
mit Ma
do)
4895
∅
=
∑
2
[
(
−
̅
)
−
∅
(
−
1
−
̅
)
]
(
−
1
−
̅
)
=
2
(
9)
Sett
ing
this
equal t
o
ze
r
o
a
nd so
l
ving fo
r
∅
yi
el
ds
∅
̂
=
∑
(
−
̅
)
(
−
1
−
̅
)
=
2
∑
(
−
1
−
̅
)
2
=
2
(
10)
Exce
pt
f
or
one
te
rm
m
issi
ng
i
n
the
denom
inator
(
−
̅
)
2
,
this
is
th
e
sam
e
as
1
.
T
he
lone
m
issi
ng
te
rm
is
neg
li
gi
ble
f
or
sta
ti
on
ary
prosse
s,
a
nd
th
us
the
le
as
t
sq
ua
res
a
nd
m
et
ho
d
-
of
-
m
ome
nts
est
i
m
at
o
rs
ar
e
near
ly
ide
ntica
l, especial
ly
fo
r
la
r
ge
sam
ples.
Fo
r
the
gener
al
AR(
p
)
proc
e
ss,
the
m
e
thods
us
e
d
to
obt
ai
n
e
qu
at
io
ns
(
7
)
an
d
(
8
)
ca
n
easi
ly
be
exten
ded
to
yi
el
d
the
sam
e
r
esult,
nam
el
y
̂
=
̅
.
T
o
ge
ner
al
iz
e
the
est
im
a
ti
on
∅
’s
,
we
ca
n
be
co
ns
i
der
e
d
thr
ough a sec
ond
-
order eq
uation.
S
o, we
re
pl
ac
e
by
̅
in t
he c
onditi
on
al
sum
-
of
-
squa
res f
un
ct
io
n,
(
∅
1
,
∅
2
,
̅
)
=
∑
[
(
−
̅
)
−
∅
1
(
−
1
−
̅
)
−
∅
2
(
−
2
−
̅
)
]
2
=
3
(
11)
Sett
ing
∅
1
⁄
=
0
, w
e
h
a
ve
−
2
∑
(
−
̅
)
−
∅
1
(
−
1
−
̅
)
−
∅
2
(
−
2
−
)
=
3
=
0
(
12)
Wh
ic
h we ca
n rew
rite
as
∑
(
−
̅
)
(
−
1
−
̅
)
=
3
=
(
∑
(
−
1
−
̅
)
2
=
3
)
∅
1
+
(
∑
(
−
1
−
̅
)
(
−
2
−
̅
)
=
3
)
∅
2
(
13)
The
s
um
of
t
he
la
gg
e
d
pro
du
ct
s
by
∑
(
−
̅
)
(
−
1
−
̅
)
=
3
is
ve
ry
near
ly
the
num
erator
of
1
.
W
e
can
div
ide
both
sides
of
the
equ
at
io
n
by
∑
(
−
̅
)
2
=
3
then
,
e
xcep
ts
f
or
e
nd
ef
fects,
wh
ic
h
ar
e
ne
gligible
unde
r
the
stat
ion
a
ry
ass
um
pti
on, we
ob
ta
in
1
=
∅
1
+
1
∅
2
(
14)
W
it
h
the
sam
e appr
oach f
or d
i
ff
e
ren
ti
at
ion eq
uations
∅
2
⁄
=
0
, obtain
ed
2
=
1
∅
1
+
∅
2
(
15)
These
tw
o
e
qu
at
ion
s
are
cal
le
d
Y
ule
-
Walker
eq
ua
ti
ons
f
or
the
AR(
2)
m
od
el
.
F
or
MA
(1),
=
−
−
1
;
if the
M
A
m
od
el
is in
ver
ti
ble then
=
−
−
1
−
2
−
2
−
3
−
3
−
⋯
+
(
16)
T
he
n
the
least
sq
ua
res
can
b
e
done by sel
ect
ing
a
value
that
m
ini
m
iz
es.
(
)
=
∑
(
)
2
=
∑
[
+
−
1
+
2
−
2
+
3
−
3
+
⋯
]
2
(17)
It
is
ob
vi
ous
that
the
le
ast
s
qu
a
res
pr
ob
le
m
in
equ
at
ion
(17)
is
nonli
ne
ar
in
the
par
a
m
et
er.
W
e
will
no
t
be
able
t
o
m
ini
m
iz
e
(
)
by
ta
kin
g
a
de
riv
at
ive
with
re
spe
ct
to
θ,
set
ti
ng
it
to
zer
o,
a
nd
so
l
ving.
T
o
address
these i
ssu
es,
consi
der evaluati
ng
(
)
for
a sin
gle
giv
e
n value
of
. Rew
rite
first
-
orde
r e
qu
at
io
n
a
s
=
+
−
1
(
18)
Using
this
eq
ua
ti
on
,
1
,
2
,
…
,
can
be
cal
culat
ed
rec
ur
si
vely
if
we
hav
e
t
he
i
niti
al
val
ue
0
.
A
com
m
on
ap
pr
oxim
a
ti
on
is to
s
et
0
=
0
.
W
e c
an
obt
ai
n
1
=
1
2
=
2
+
1
3
=
3
+
2
⋮
=
+
−
1
}
(
19)
Fo
r
h
i
gher
-
ord
er m
ov
in
g
a
verage m
od
el
s, w
e can
c
om
pu
te
=
(
1
,
2
,
…
,
)
rec
ur
si
vely
f
r
om
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
:
489
2
-
490
1
4896
=
+
1
−
1
+
2
−
2
+
⋯
+
−
(
20)
W
it
h
0
=
−
1
=
⋯
=
−
=
0
.
Th
e
s
um
of
squares
i
s
m
ini
m
iz
ed
jointl
y
in
1
,
2
,
…
,
us
in
g
a
m
ul
ti
var
ia
te
num
erical
m
e
tho
d.
F
or
a
ge
ne
ral
ARM
A
m
od
el
:
the
c
ondit
ion
al
s
um
of
s
quares
can
be
de
fine
d
m
uch
as in
t
he
MA
case,
w
e
c
on
si
der
=
(
∅
,
)
and w
i
sh
t
o
m
ini
m
iz
e
(
∅
,
)
=
∑
2
, so
=
−
∅
−
1
+
−
1
(
21)
To
obta
in
1
,
we
now
ha
ve
a
n
a
dd
it
io
nal
pro
ble
m
,
nam
el
y
0
.
On
e
a
ppro
ac
h
is
t
o
set
0
=
0
or
to
̅
if
our
m
od
el
c
on
ta
in
s
a
no
nz
ero
m
ean.
H
oweve
r,
a
bette
r
appr
oach
is
t
o
be
gin
t
he
re
cur
si
on
at
=
2
,
th
us
avo
i
ding
0
al
to
gethe
r,
an
d
si
m
pl
y
m
ini
m
ize,
(
∅
,
)
=
∑
2
=
2
.
F
or
ge
neral
ARMA
(
,
)
m
od
el
,
we
com
pu
te
=
−
∅
1
−
1
−
∅
2
−
2
−
⋯
−
∅
−
+
1
−
1
+
2
−
2
+
⋯
+
−
(22)
W
it
h
=
−
1
=
⋯
=
+
1
−
=
0
a
nd
th
en
m
ini
m
iz
e
(
∅
1
,
∅
2
,
…
,
∅
,
1
,
2
,
…
,
)
num
erical
ly
to
ob
ta
in
the
co
ndit
ion
al
least
s
qu
a
res
e
stim
ates
of all
p
a
ram
et
ers.
2.4.
Measuri
n
g Ac
curac
y
Le
vel
of
Predi
cti
on
Results
Ba
sic
al
ly
,
m
ea
su
ri
ng
t
he
acc
ur
acy
of
pre
dicti
on
res
ults
can
be
done
by
var
io
us
sta
ti
sti
cal
analy
sis
m
et
ho
ds
;
s
uch
as
the
r
oo
t
m
ean
of
s
qu
a
re
error
(RMSE)
,
the
m
ean
of
a
bs
ol
ute
er
ror
(
MAE)
value
a
nd
t
he
m
ean
of
abs
ol
ute
per
ce
ntage
err
or
(MA
PE).
In
this
resea
r
ch,
we
use
d
MAPE
as
sta
ndar
d
m
easur
em
ent
of
pr
e
dicti
on r
es
ul
ts acc
ur
acy
.
MAPE
is
d
e
fine
d as f
ollo
ws
[40
]
:
=
∑
|
−
̂
|
=
1
×
100%
(
23
)
wh
e
re
an
d
̂
are
the act
ual
valu
e an
d pr
e
dicti
on
value, w
hile
n
is t
he nu
m
ber
of
predict
io
n values
.
2.5.
Data Se
t
This
st
ud
y
us
e
s
el
ect
rical
po
wer
dem
and
i
n
th
e
loa
d
c
en
te
r
data
(take
n
eve
ry
hal
f
hour)
at
pow
e
r
plant
un
it
of
P
T.
PL
N
Gresi
k
Ind
onesi
a
on
Jan
uar
y
1,
2009
-
De
cem
ber
31,
2011.
Whe
re,
data
f
r
om
J
anu
a
r
y
1,
20
09
-
Dec
e
m
ber
24,
2011
is
us
ed
f
or
f
or
ecast
in
g
an
d
data
fr
om
Dece
m
ber
25
-
31,
20
11
is
us
e
d
for
te
sti
ng
.
3.
RESU
LT
S
A
ND
DI
SCUS
S
ION
3.1.
Model
Fit
ting
an
d
Iden
tific
ati
on Par
ame
te
r
Ex
plo
rati
on
of
el
ect
rical
power
co
nsum
pti
on
data
is
do
ne
thr
ough
ti
m
e
series
plo
t
on
Jan
uar
y
1,
2009
-
Decem
ber
24,
2011.
The
data
patte
rn
is
ver
y
fluct
uate
as
s
how
n
in
Fig
ur
e
1.
T
his
co
ndit
ion
m
ay
be
influ
e
nce
d
by
the
integrated
powe
r
distri
bu
ti
on
syst
em
in
Java
-
Ma
dura
-
Ba
li
interconnecti
on
syst
e
m
in
Ind
on
esi
a.
F
rom
the p
ic
tu
re it sh
own
t
hat the
d
at
a
has n
ot been stat
io
nar
y.
Figure
1. Ele
ct
rical
pow
e
r dat
a eve
ry h
al
f
ho
ur on Ja
nuary
1,
2009
-
Dece
m
ber
2
4,
2011
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
Ap
plying of
D
ouble
Se
asonal
ARIMA
Model
for Elect
ric
al
Power
De
man
d
F
or
ec
as
ti
ng ... (Is
mit Ma
do)
4897
Figure
2
s
hows
that
the
ACF
coe
ff
ic
ie
nt
is
sign
ific
a
ntly
diff
ere
nt
from
zero
and
the
PA
CF
coeffic
ie
nt
is
cl
os
e
to
zer
o
aft
er
the
first
la
g.
Ba
sed
on
these
two
points,
it
sh
ows
that
the av
era
ge
data
ha
s
no
t
been
sta
ti
onar
y
and
there
i
s
a
season
al
patte
rn
e
d
tre
nd.
T
her
e
fore,
it
is
necessary
to
process
1
st
le
vel
diff
e
re
ncin
g
(
=
1
)
.
Figure
2
.
The
ACF a
nd P
AC
F of
The
plo
ts
of
ACF
a
nd
PA
C
F
in
Fig
ure
3
hav
e
t
hro
ugh
a
diff
e
re
ntiat
ion
pr
ocess
s
o
t
hat
the
non
-
seaso
nal
co
ndit
ion
s
hav
e
be
en
sta
ti
on
a
ry
i
n
the
m
ean
va
lue.
Ba
se
d
on
the
ACF
pl
ot
it
shows
th
at
the
autoc
orrelat
ion
values
of
th
e
sta
ti
on
ary
data
go
dow
n
to
ze
ro
a
fter
the
sec
ond
la
g
a
nd
th
e
third
la
g,
w
hi
le
for
the
seaso
nal
da
ta
it
is
still
no
t
sta
ti
on
ary
in
th
e
m
ean
value.
Th
e
ACF p
lot
al
so
shows
t
he
existe
nce
of
an
oth
e
r
seaso
nal p
at
te
r
n wh
ic
h
is a
we
ekly
seasonal
patte
rn on la
g 336, 6
72 a
nd s
o on.
Figure
3
.
The
ACF a
nd P
AC
F of
after
=
1
,
1
=
1
and
1
=
48
Figure
4: L
oad d
em
and
se
ries
after
=
1
,
1
=
1
,
1
=
48
,
2
=
1
, a
nd
2
=
336
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
:
489
2
-
490
1
4898
The
ACF
plo
t
of
t
he
loa
d
data
patte
rn
in
Fi
gure
5
ha
s
bee
n
sta
ti
on
a
ry
in
the
m
ean
value
after
go
i
ng
thr
ough
a
seco
nd
orde
r
dif
fe
r
entia
ti
on
pr
oce
ss
(w
it
h
2
=
1
).
For
non
-
seaso
nal
patte
rn
s
base
d
on
ACF
a
nd
PA
CF
plo
ts,
th
e
load
te
nds
to
decr
ea
se
gra
dual
ly
esp
eci
al
ly
after
the
first
la
g
s
o
that
the
non
-
seas
on
al
m
od
el
assum
ption
is
the
ARMA
(
1
,
1
)
m
od
el
.
Wh
il
e
the
daily
seaso
nal
patte
rn
(
1
=
48
)
sho
ws
that
t
he
plo
t
of
th
e
ACF dat
a p
at
te
rn
te
nd
s t
o
be
interru
pted
a
fter th
e lag
48 and the
P
ACF
pa
tt
ern
in
dicat
es the d
at
a
patte
rn
te
nd
s
to
decr
ease
gradu
al
ly
,
t
hen
t
he
as
su
m
ption
of
t
he
daily
s
easo
nal
m
od
el
is
MA
(
1
)
48
m
od
el
.
Fo
r
the
week
l
y
seaso
nal
patte
r
n
(
2
=
336
)
it
sh
ows
th
at
ACF
an
d
P
ACF
plo
ts
al
ong
la
g
336,
67
2,
a
nd
so
on
t
end
t
o
f
al
l
gr
a
dual
ly
, th
en
the ass
um
ption
of the
wee
kly seaso
nal m
od
el
is
MA
(
1
,
1
)
336
m
od
el
.
Figure
5
.
The
ACF a
nd P
AC
F of
after
=
1
,
1
=
1
,
1
=
48
,
2
=
1
and
2
=
336
Ba
sed
on
the
i
den
ti
ficat
io
n
of
AC
F
a
nd
P
A
CF
patte
r
ns
,
th
e
assum
ption
of
an
ap
pro
pr
i
at
e
ARIMA
m
od
el
is
a
double
seas
onal
A
RIMA
m
od
el
(
1
,
1
,
1
)
(
0
,
1
,
1
)
48
(
0
,
0
,
1
)
336
.H
owe
ver,
if
w
hite
noise
is
de
te
ct
ed
in
the
te
st
data,
it
is
necessary
to
ad
d
or
substi
tute
the
order
of
diff
e
re
ntiat
i
on
process
.
I
n
this
stud
y,
Stat
ist
ic
al
An
al
ysi
s Syste
m
(
SA
S)
pro
gra
m
m
ing
to
ols
are
us
ed
to
a
na
ly
ze load data
of do
ub
le
sea
s
on
al
ARIM
A m
od
e
ls.
3.2.
DSARI
M
A M
od
el
P
ar
amet
e
r Estima
tio
n
The
AR
an
d
MA
coe
ff
ic
ie
nt
s
in
the
DSA
RIMA
m
od
el
are
est
im
a
te
d
by
the
le
ast
squares
m
et
hod.
The
i
niti
al
estim
at
e o
btaine
d has
been u
sed
as the initi
al
v
a
lue of the
it
erati
ve
est
im
a
ti
on
m
et
ho
d. T
hrough
the
double
seas
o
na
l
ARIM
A
m
od
el
(
1
,
1
,
1
)
(
0
,
1
,
1
)
48
(
0
,
0
,
1
)
336
,
the
init
ia
l
data
of
the
A
R
an
d
M
A
co
eff
ic
ie
nt
s
wer
e
obtai
ne
d as f
ollows:
Ba
sed
on
Tabl
e
1,
t
o
m
eet
the
w
hite
no
ise
c
rite
rion,
p
-
val
ue
m
us
t
be
great
er
t
han
fa
ult
t
ol
eran
ce
=
5%
,
wit
h
al
pha
sign
i
ficance
le
ve
l
le
ss
tha
n
0.0
001.I
n
a
ddit
ion
,
the
m
od
el
has
an
im
pr
ov
em
ent
patte
rn
with
3
MA
pa
ram
et
er
s
i.e.
MA
(
1
,
1
)
,
MA
(
2
,
1
)
dan
M
A
(
3
,
1
)
,
so
t
hat
these
th
ree
par
am
et
ers
shou
l
d
be
inclu
de
d
in
the
m
od
el
est
i
m
at
ion
.
Wh
il
e
the
resid
ual
a
ssu
m
ption
te
st
that
inclu
des
the
assum
ption
of
w
hite
no
is
e
m
us
t
m
eet
the cr
it
eria o
f
in
de
pende
nt and
norm
al
d
ist
rib
ution
(
0
,
2
)
.
The
Lju
ng
-
Bo
x
te
st
is
us
e
d
to
c
hec
k
t
he
resi
du
al
i
nd
e
pende
nce
as
sum
pt
ion
with
t
he
fo
ll
owi
ng
hypothesis:
0
∶
1
=
2
=
.
.
.
=
=
0
1
∶
at
least
one
1
w
hich
is
not e
qual
to
zer
o for
=
1
,
2
,
…
,
.
with
a
fau
lt
tol
eran
ce
of
5%
t
hen
0
is
rej
ect
e
d
if
p
-
val
ue
<
,
wh
ic
h
m
eans
the
re
sid
ual
do
es
not
m
eet
th
e
wh
it
e
no
ise
ass
um
ption
.
Ba
sed
on
the
final
re
su
lt
of
AR
a
nd
MA
coe
ff
ic
ie
nt
pa
ram
et
er
est
i
m
a
ti
on
in
Ta
ble
2,
it
ca
n
be
plo
tt
ed
the
res
idu
al
norm
al
p
roba
bili
ty
to
determ
ine
wh
et
her
resi
du
al
ha
s
fu
lfil
le
d
w
hi
te
no
ise
assu
m
pt
ion
with
li
m
i
t
of
±
1
.
96
√
⁄
≈
±
0
.
009
.
By
it
erati
ng
t
he
ad
diti
on
of
AR
a
nd
M
A
pa
ram
et
ers,
the
best
it
erati
on
value
has
be
en
obta
ined
wh
ic
h
has
f
ul
fill
ed
w
hite
no
ise
ass
um
pt
ion
,
i.e.
do
uble
seaso
nal
A
RIMA
(
[
1
,
2
,
7
,
16
,
18
,
35
,
46
]
,
1
,
[
1
,
3
,
13
,
21
,
27
,
46
]
)
(
1
,
1
,
1
)
48
(
0
,
0
,
1
)
336
.
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
Ap
plying of
D
ouble
Se
asonal
ARIMA
Model
for Elect
ric
al
Power
De
man
d
F
or
ec
as
ti
ng ... (Is
mit Ma
do)
4899
Table
1
.
A
n O
utput S
AS
of
m
od
el
w
it
h
CL
S ite
rati
ve
The ARI
NA
P
roce
d
u
re
Co
n
d
itio
n
al L
east Sq
u
ares
Esti
m
atio
n
Para
m
ater
Esti
m
at
e
Stan
d
ard Erro
r
t Value
Ap
p
rox
Pr
>
|
t
|
Lag
MA1, 1
-
0
.35
1
8
4
0
.01
8
9
9
-
1
8
.53
<.00
0
1
1
MA2, 1
0
.95
7
3
4
0
.00
1
3
0
0
7
7
3
6
.02
<.00
0
1
48
MA3, 1
-
0
.04
5
2
6
0
.00
4
5
1
0
3
-
1
0
.03
<.00
0
1
336
AR1
,
1
-
0
.14
5
7
8
0
.02
0
0
6
-
7
.27
<.00
0
1
1
Variance
Esti
m
ate
1
3
0
.7053
Std
E
rr
o
r
E
sti
m
ate
1
1
.43
2
6
4
AIC
4
0
1
9
4
6
SBC
4
0
1
9
8
1
Nu
m
b
e
r
o
f
Resid
u
als
5
2
1
2
7
*
AIC and
SBC d
o
no
t inclu
d
e log
det
er
m
in
an
t
Au
to
co
rr
elatio
n
Ch
eck o
f
Resid
u
als
To Lag
Ch
i
-
Sq
u
are
DF
Pr>Ch
iSq
Au
to
co
rr
elatio
n
s
6
1
5
3
.39
2
<.00
0
1
-
0
.00
2
-
0
.01
9
-
0
.04
1
-
0
.01
7
-
0
.02
3
-
0008
12
2
7
4
.15
8
<.
0001
-
0
.03
3
-
0
.02
7
-
0
.01
4
-
0
.00
9
-
0
.01
4
-
0
.00
7
18
3
4
2
.13
14
<.00
0
1
-
0
.00
9
-
0
.00
9
-
0
.00
8
-
0
.01
7
-
0
.01
6
-
0
.02
3
24
4
2
2
.74
20
<.00
0
1
-
0.
023
-
0
.01
8
-
0
.02
0
-
0
.01
1
-
0
.01
3
-
0
.00
3
30
4
7
3
.05
26
<.00
0
1
-
0
.00
9
-
0
.00
8
-
0
.01
7
-
0
.00
8
-
0
.01
7
-
0
.01
4
36
4
8
9
.03
32
<.00
0
1
-
0
.01
1
-
0
.00
9
-
0
.00
2
0
.00
0
-
0
.01
0
0
.00
0
42
4
9
7
.60
38
<.00
0
1
-
0
.00
7
-
0
.00
8
0
.00
2
-
0
.00
5
-
0
.00
4
0
.00
3
48
8
0
4
.03
44
<.00
0
1
0
.00
1
0
.00
2
0
.00
6
0
.01
8
0
.00
4
0
.06
0
Table
2
.
A
n O
utput S
AS
of
m
od
el
Co
n
d
itio
n
al L
east
Sq
u
ares E
sti
m
atio
n
Para
m
eter
Esti
m
at
e
Stan
d
ard Erro
r
T
Valu
e
Ap
p
rox
Pr
>
|
t
|
Lag
MA1,1
0
.95
1
3
1
0
.01
2
9
2
7
3
.64
<.00
0
1
1
MA1,2
-
0
.07
2
6
9
0
.00
6
4
8
2
6
-
1
1
.21
<.00
0
1
3
MA1,3
0
.01
3
5
7
0
.00
3
0
7
8
6
4
.41
<.00
0
1
13
MA1,4
0
.01
0
5
3
0
.00
2
8
6
2
0
3
.68
0
.00
0
2
21
MA1,5
0
.01
7
2
7
0
.00
2
5
8
8
4
6
.
67
<.00
0
1
27
MA1,6
0
.05
0
7
4
0
.00
4
5
4
7
8
1
1
.16
<.00
0
1
46
MA2,1
0
.97
1
5
5
0
.00
1
0
8
9
0
8
9
2
.14
<.00
0
1
48
MA3,1
-
0
.05
3
2
0
0
.00
4
4
7
7
4
-
1
1
.88
<.00
0
1
336
AR1
,1
1
.13
3
0
0
0
.01
4
0
2
8
0
.80
<.00
0
1
1
AR1
,2
-
0
.28
0
5
6
0
.00
7
8
4
2
2
-
3
5
.78
<.00
0
1
2
AR1
,3
-
0
.01
8
6
9
0
.00
2
5
3
3
2
-
7
.38
<.00
0
1
7
AR1
,4
-
0
.01
0
5
2
0
.00
3
2
9
8
3
-
3
.19
0
.00
1
4
16
AR1
,5
-
0
.01
0
2
4
0
.00
3
2
2
6
0
-
3
.18
0
.00
1
5
18
AR1
,6
-
0
.00
6
8
8
9
3
0
.00
2
1
6
0
9
-
3
.19
0
.00
1
4
35
AR1
,7
0
.07
3
9
5
0
.00
4
9
7
6
8
1
4
.86
<.00
0
1
46
AR2
,1
0
.04
0
5
1
0
.00
4
9
4
8
9
8
.19
<.00
0
1
48
Variance
Esti
m
ate
1
2
7
.2317
Std
E
rr
o
r
E
sti
m
ate
1
1
.27
9
7
AIC
4
0
0
5
5
4
SBC
4
0
0
6
9
5
.7
Nu
m
b
e
r
o
f
Resid
u
als
5
2
1
2
7
*
AIC and
SBC d
o
no
t inclu
d
e log
det
er
m
in
an
t
Au
to
cp
rr
elatio
n
Ch
eck o
f
Resid
u
als
To Lag
Ch
i
-
Sq
u
are
DF
Pr>Ch
iSq
Au
to
co
rr
elatio
n
s
6
.
0
.
-
0
.00
0
-
0
.00
0
0
.00
0
0
.00
5
-
0
.00
4
0
.00
4
12
.
0
.
-
0
.00
6
-
0
.00
2
0
.00
6
0
.00
3
-
0
.00
7
-
0
.00
4
18
1
2
.54
2
0
.00
1
9
0
.00
4
0
.00
0
-
0
.00
2
-
0
.00
2
-
0
.00
3
-
0
.00
2
24
1
6
.38
8
0
.03
7
3
-
0
.00
4
-
0
.00
4
0
.00
1
0
.00
5
-
0
.00
1
0
.00
4
30
2
3
.11
14
0
.05
8
6
-
0
.00
4
-
0
.00
6
0
.00
1
0
.00
8
-
0
.00
3
-
0
.00
2
36
2
7
.18
20
0
.13
0
4
-
0
.00
3
-
0
.00
3
0
.00
1
0
.00
1
-
0
.00
3
0
.00
7
42
3
1
.53
26
0
.20
9
3
-
0
.00
4
-
0
.00
4
0
.00
6
-
0
.00
2
0
.00
7
-
0
.01
1
48
3
9
.26
32
0
.17
6
5
0
.00
3
0
.00
3
0
.00
6
-
0
.00
6
0
.00
7
-
0
.00
1
3.3.
Model
Testin
g an
d
Measuri
ng
of
F
orecast
ing Le
vel A
cc
urac
y
Fo
r
acc
ur
acy
m
easur
em
ent,
te
sti
ng
us
es
the
MAPE
proce
dure.
Ba
sed
on
the
com
par
ison
of
pr
e
dicti
on loa
d data
with act
ua
l l
oad
data,
th
e ave
rag
e
d
at
a
accuracy
of 2,0
6%
is
obta
ined
.
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
:
489
2
-
490
1
4900
3.4.
Predi
ction
Ba
sed
on
the
final
resu
lt
of
par
am
et
er
est
i
m
ation
in
Table
1
we
ge
t
the
ARIMA
coe
ff
ic
ie
nt
par
am
et
ers
as
fo
ll
ows:
AR
(
1
,
1
)
=
1.1
33,
AR
(
1
,
2
)
=
-
0.2
8056,
AR
(
1
,
3
)
=
-
0.0
1869,
A
R
(
1
,
4
)
=
-
0.010
52,
AR
(
1
,
5
)
=
-
0.010
24,
AR
(
1
,
6
)
=
-
0.0
0688
93,
AR
(
1
,
7
)
=
0.0
7395,
AR
(
2
,
1
)
=
0.040
51
MA
(
1
,
1
)
=
0.9
5131,
MA
(
1
,
2
)
=
-
0.0
7269,
M
A
(
1
,
3
)
=
0.01357,
MA
(
1
,
4
)
=
0.010
53,
MA
(
1
,
5
)
=
0.017
27,
MA
(
1
,
6
)
=
0.0
5074,
MA
(
2
,
1
)
=
0.971
55, MA
(
3
,
1
)
=
-
0.0
532.
Thro
ugh
the
par
am
et
er
of
this
pr
e
dicti
on
m
od
el
we
ge
t
the
e
quat
io
n
for
D
SA
R
I
MA
m
od
el
(
[
1
,
2
,
7
,
16
,
18
,
35
,
46
]
,
1
,
[
1
,
3
,
13
,
21
,
27
,
46
]
)
(
1
,
1
,
1
)
48
(
0
,
0
,
1
)
336
as foll
ows:
(
1
−
)
(
1
−
48
)
(
1
−
1
.
133
+
0
.
281
2
+
0
.
019
7
+
0
.
011
16
+
0
.
010
18
+
0
.
007
35
−
0
.
074
46
)
(
1
−
0
,
041
48
)
̇
=
(
1
−
0
.
951
+
0
.
073
3
−
0
.
014
13
−
0
.
011
21
−
0
.
017
27
−
0
.
051
46
)
(
1
−
0
.
972
48
)
(
1
+
0
.
053
33
6
)
Figure
6.
The
ou
t
-
sam
ples o
f
actual
dat
a an
d on
e
-
ste
p ahe
ad ou
t
-
sam
ple f
oreca
sts
4.
CONCL
US
I
O
N
Stat
ist
ic
al
analy
sis
base
d
on
DS
ARIMA
m
od
el
is
s
uitable
with
el
ect
rical
powe
r
c
ha
racteri
sti
cs
with
con
ti
nu
ous
an
d
fluctuati
ng
l
oa
d
patte
rn
s
.
T
he
load
cha
nges
are
al
ways
un
exp
ect
e
d
at
a
ny
tim
e
dep
en
din
g
on
el
ect
rical
po
w
er
dem
and
in
the
load
ce
nt
er.
W
it
h
t
he
s
ta
ti
sti
ca
l
analysis
m
od
el
,
pre
dicti
on
s
a
re
a
ble
to
gen
e
rate
data
that
is
not
incl
uded
in
the
data
trai
ning
proce
ss.
T
hro
ugh
th
e
best
m
od
el
a
ssu
m
ption
,
t
he
m
od
el
in
this
stud
y
w
as
able
to
pr
e
di
ct
with
the
aver
age
acc
ur
acy
of
MAP
E
of
2.06%.
F
ur
the
r
researc
h
that
can
be
dev
el
op
e
d
is
the
patte
r
n
of
el
ect
rical
po
wer
dem
and
on
a
la
rg
e
-
sca
le
area
su
ch
as
Java
-
Ma
dur
a
-
Ba
li
,
Ind
on
esi
a ele
ct
rici
ty
n
et
wor
k i
ntercon
necti
on.
REFERE
NCE
S
[1]
G.
J.
Tsekour
as,
et
al.
,
“
A
non
-
li
ne
ar
m
ult
iva
ri
a
ble
reg
ression
m
odel
for
m
idt
erm
ene
rg
y
fore
c
asti
ng
of
power
s
y
stems
,
”
Elec
tr.
Powe
r S
yst. Res.
,
vol
/i
ss
ue
:
77
(
12
)
,
pp
.
1560
–
15
68,
2007
.
[2]
P.
E.
McSharr
y
,
et
al.
,
“
Probabil
isti
c
for
ecasts
of
the
m
agni
tude
and
ti
m
ing
of
pea
k
elec
tri
c
ity
demand,
”
IE
E
E
Tr
ans.
Powe
r Sy
st.
,
vo
l
/i
ss
ue:
20
(
2
)
,
pp
.
1166
–
11
72,
2005
.
[3]
E.
G
.
Rom
era,
“
Monthl
y
El
e
ct
r
i
c
Ene
rg
y
Dem
a
nd
Forec
asti
ng
Based
on
Tre
nd
Ext
ra
ct
ion
,
”
I
E
EE
Tr
ans.
Power
Syst.
,
vol
/i
ss
ue:
2
1(
4
)
,
pp
.
1946
–
1953,
2006
.
[4]
J.
W
.
Tay
lor
an
d
P.
E.
McShar
r
y
,
“
Short
-
t
erm
loa
d
fore
c
asti
ng
m
et
hods:
An
eva
luation
base
d
o
n
Europe
an
da
ta,”
IEE
E
Tr
ans.
Po
wer
Syst.
,
vo
l
/i
ss
ue:
22
(
4
)
,
pp
.
22
13
–
2219,
2007
.
[5]
F.
W
.
Berna
l,
et
al.
,
“
Sm
al
l
-
signa
l
ana
l
y
sis
of
power
sy
st
em
sw
in
g
m
odes
as
aff
ec
te
d
b
y
wind
turb
ine
-
gen
era
to
rs
,
”
2016
IEEE Pow
er
and
En
ergy
C
onfe
renc
e
at
Il
linois
,
P
ECI
2016
,
2016
.
[6]
K.
K.
Al
-
Dalwi
and
A.
M.
Vur
al,
“
Modeli
ng
and
sim
ula
ti
on
of
B
azy
an
Gas
Pow
er
Plant
,
”
2017
4
t
h
Inte
rnationa
l
Confe
renc
e
on
E
le
c
tric
al
and
El
e
ct
ronics
Engi
ne
e
ring,
ICE
EE 201
7
,
2017
.
[7]
Y.
Zh
ou
,
e
t
al
.
,
“
The
Stocha
st
ic
Response
Surfac
e
Method
for
S
m
al
l
-
signal
Stab
il
ity
S
tud
y
of
Pow
er
S
y
st
em
wit
h
Probabil
isti
c
Unce
rt
ai
nt
ie
s
in
Corre
lated
Photovolt
aic
and
Loa
d
,
”
IEE
E
Tr
ansacti
ons
on
Powe
r
S
yste
ms
,
vo
l.
99
,
p
p.
1
,
2017
.
[8]
H.
Hahn,
et
a
l.
,
“
El
ectric
loa
d
fore
ca
st
ing
m
ethods
:
Tool
s
for
dec
ision
m
aki
n
g,
”
Eur.
J
.
Ope
r.
Re
s.
,
vol
/i
ss
ue:
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
Ap
plying of
D
ouble
Se
asonal
ARIMA
Model
for Elect
ric
al
Power
De
man
d
F
or
ec
as
ti
ng ... (Is
mit Ma
do)
4901
199
(
3
)
,
pp
.
902
–
907,
2009
.
[9]
M.
M.
Nia,
et
al.
,
“
Stocha
sti
c
appr
oac
h
to
a
r
ai
n
attenuation
ti
m
e
seris
sy
n
th
esiz
er
for
he
av
y
rai
n
r
egi
ons,”
Inte
rnational
Jo
urnal
of El
e
ct
ri
c
al
and
Comput
er
Engi
n
ee
ring
(
IJE
CE)
,
vol
/
issue:
6
(
5
)
,
pp
.
2379
–
2
386,
2016
.
[10]
S.
Kam
ley
,
et
al.
,
“
Perform
an
ce
fore
ca
sting
of
share
m
ark
et
using
m
ac
hine
le
arn
ing
techni
ques:
A
rev
ie
w,
”
Inte
rnational
Jo
urnal
of El
e
ct
ri
c
al
and
Comput
er
Engi
n
ee
ring
(
IJE
CE)
,
vol
/
issue:
6
(
6
)
.
pp
.
3196
–
3
204,
2016
.
[11]
A.
T.
Lor
a,
e
t
al
.
,
“
Ti
m
e
-
seri
es
pre
diction:
Appli
ca
t
ion
to
the
sho
rt
-
te
rm
elec
tr
ic
e
ner
g
y
d
emand,
”
Curr
.
Top.
Arti
f.
Inte
ll.
,
pp.
577
–
586,
2004
.
[12]
H.
M.
Al
-
Ham
adi
and
S.
A.
Solim
an,
“
Lo
ng
-
te
rm
/mid
-
te
r
m
el
ectric
loa
d
fore
c
asti
ng
b
a
sed
on
short
-
term
cor
relati
on
and
a
nnual
growth
,
”
El
e
ct
r.
Powe
r S
y
st.
R
es.
,
vol
/i
ss
ue:
74
(
3
)
,
pp
.
353
–
361,
2005
.
[13]
H.
M.
Al
-
Ham
adi
and
S.
A.
Solim
an,
“
Short
-
term
el
ec
tr
ic
lo
ad
fore
ca
st
ing
base
d
on
Kalman
fil
te
r
ing
al
gor
it
h
m
with
m
oving
wi
ndow wea
ther and l
oad
m
odel
,
”
El
e
ct
r.
Powe
r S
y
st.
R
es.
,
vol
/i
ss
ue:
68
(
1
)
,
pp
.
47
–
59,
2004
.
[14]
J.
W
.
T
a
y
lor,
“
Densit
y
for
ec
ast
in
g
for
th
e
eff
ic
i
en
t
balanc
ing
of
th
e
gen
era
t
ion
and
consum
pti
on
of
el
e
ct
ri
ci
t
y
,
”
Int
.
J.
Forec
ast
.
,
vol
/
issue:
22
(
4
)
,
pp.
707
–
724,
2006
.
[15]
S.
A.
Sol
iman
an
d
A.
M.
Al
-
Kan
dar
i,
“
El
e
ct
ri
cal Load
Fore
ca
st
in
g
,”
2010
.
[16]
S.
N.
Hass
an
,
e
t
al.
,
“
A
compari
son
of
the
fore
c
ast
per
form
an
ce
of
double
sea
so
nal
ARIM
A
and
double
sea
sona
l
ARF
IMA
m
odel
s of
elec
tri
c
ity
lo
ad
demand
,
”
Ap
pl.
Ma
th. Sc
i
.
,
v
ol
/i
ss
ue:
6
(
133
–
136
)
,
pp
.
6705
–
6712,
2012
.
[17]
N.
Moham
ed,
e
t
al.
,
“
Double
Seasona
l
ARIM
A
Model
for
Forec
ast
ing
Loa
d
Dem
and,
”
Matem
ati
ka
,
vol
/is
sue:
26
(
2
)
,
pp
.
217
–
2
31,
2010
.
[18]
N.
Moham
ed,
e
t
al.
,
“
Im
proving
Short
Te
rm
Load
Forec
asti
ng
Us
ing
Double
Se
asona
l
Arim
a
Mod
el
,
”
World
App
l.
Sci
.
J
.
,
vol
/
issue:
15
(
2
)
,
pp.
223
–
231,
2011
.
[19]
Suhartono,
“
Tim
e
Serie
s
Forec
asti
ng
b
y
usin
g
Seasona
l
Aut
ore
gre
ss
ive
Int
e
gra
te
d
Moving
Avera
ge:
Subs
e
t
,
Multi
plicative
or
Additi
v
e
Model
,
”
J. M
ath
.
S
tat
.
,
vol
/i
ss
ue:
7
(
1
)
,
p
p.
20
–
27
,
2011
.
[20]
J.
D.
J.
Deng
a
nd
P.
Jirut
it
i
ja
r
oen,
“
Short
-
t
er
m
loa
d
fore
ca
st
ing
u
sing
t
ime
serie
s
an
aly
sis:
A
ca
se
stud
y
f
or
Singapore
,
”
C
yb
ern.
Int
ell.
Syst.
(
CIS
)
,
2010
IEEE
Conf
.
,
v
o
l
.
1
,
pp.
1
–
6
,
2010
.
[21]
C.
Nata
r
aja,
et
a
l.
,
“
Short
T
erm
Loa
d
Forec
asti
n
g
Us
ing
Ti
m
e
Serie
s
Anal
y
sis:
A Ca
se
Stud
y
fo
r
Karna
ta
k
a,
Ind
ia,”
Int.
J. E
ng
.
S
ci
.
I
nnov.
Te
chnol.
,
vol
/i
ss
ue:
1
(
2
)
,
p
p.
45
–
53
,
2012
.
[22]
N.
Moham
ed,
et
al.
,
“
Im
proving
Short
T
erm
Lo
ad
Forec
asti
ng
Us
ing
Double
Seasona
l
Arim
a
Model,
”
Appl.
S
ci.
,
vol
/i
ss
ue:
15
(
2
)
,
pp.
223
–
231
,
20
11.
[23]
C.
M.
Pere
ira,
et
al.
,
“
Fuzz
y
m
odel
ing
to
fore
ca
st
an
el
e
ct
ri
c
loa
d
t
ime
serie
s,”
in
P
roce
dia
Computer
Sci
ence
,
vol
.
55,
pp
.
395
–
404
,
2015.
[24]
R.
Mam
look,
et
al.
,
“
A
fuz
z
y
inf
ere
nc
e
m
odel
for
short
-
te
rm
loa
d
fore
ca
sting
,
”
En
ergy
Pol
i
cy
,
vol
/
issue:
37
(
4
)
,
pp.
1239
–
1248,
200
9.
[25]
S.
R
amos
,
et
al.
,
“
Short
-
te
rm
loa
d
fore
ca
st
ing
base
d
on
loa
d
profil
ing
,
”
in
201
3
IEE
E
Powe
r
&
E
nergy
Soci
e
ty
Gene
ral M
e
eting
,
pp
.
1
–
5
,
2013
.
[26]
N.
Kandil,
et
a
l.
,
“
An
eff
ic
i
ent
a
pproa
ch
for
sho
rt
te
rm
loa
d
for
ec
ast
ing
using
a
rti
fi
ci
a
l
neur
al
n
et
works
,
”
Int
.
J
.
El
e
ct
r.
Powe
r
E
nergy
Syst
.
,
vol
/is
sue:
28
(
8
)
,
pp.
525
–
530,
2006
.
[27]
M.
R.
AlRashi
di
and
K.
M.
EL
-
Nagga
r
,
“
Long
te
rm
el
e
ct
ri
c
loa
d
for
e
ca
sting
bas
ed
on
par
ti
c
le
sw
arm
opti
m
iz
ation,
”
A
ppl.
Ene
rgy
,
vo
l
/
issue:
87
(
1
)
,
pp.
320
–
326,
2010
.
[28]
A.
A.
E
l
Desou
k
y
and
M.
M.
El
kateb,
“
H
y
bri
d
ada
p
ti
ve
tech
nique
s
for
el
e
ctric
-
lo
ad
for
ec
ast
using
AN
N
an
d
ARIM
A,”
Gene
r
.
Tr
ansm
.
Distrib.
IE
E Proceedi
ngs
,
vol
/
issue:
1
47
(
4
)
,
pp
.
213
–
2
17,
2000
.
[29]
H.
Nie,
et
al
.
,
“
H
y
brid
of
ARIM
A
and
S
VM
s
for
short
-
te
rm
loa
d
fore
c
asti
ng
,
”
in
En
ergy
Proce
dia
,
vol
/i
ss
u
e:
16
(
PA
RT
C
)
,
pp
.
1455
–
1460
,
20
11.
[30]
R.
H.
L
ia
ng
and
C.
C.
Ch
eng,
“
Com
bine
d
reg
re
ss
ion
-
fuz
z
y
appr
oac
h
for
short
-
term
loa
d
fore
c
ast
ing,
”
I
EE
Proc.
-
Gene
r.
Tr
ansm
.
Distrib.
,
vo
l
/i
ss
u
e:
147
(
4
)
,
p
p
.
26
1,
2000
.
[31]
Suhartono,
et
al
.
,
“
Two
-
le
ve
l
se
a
sonal
m
odel
bas
ed
on
h
y
brid
ARIM
A
-
ANFIS
for
fore
c
asti
ng
shor
t
-
te
rm
e
le
c
tri
c
ity
loa
d
in
Indone
si
a,
”
ICSS
BE
201
2
-
Proce
edi
ngs,
2012
Inte
rnatio
nal
Confe
renc
e
on
Stat
isti
cs
in
Sci
en
ce,
Busine
s
s
and
Engi
n
ee
ring
:
“
Empowering Deci
sion M
a
ki
n
g
wit
h
Statis
ti
ca
l
Scienc
es,
”
pp.
6
34
–
638
,
2012
.
[32]
D.
Niu,
e
t
al
.
,
“
Com
bina
ti
on
fore
ca
st
ing
m
odel
for
m
id
-
long
t
erm
loa
d
base
d
on
le
ast
squa
re
s
support
vec
to
r
m
ac
hine
s
and
a
m
ende
d
par
ti
c
l
e
sw
arm
opti
m
iz
ation
al
gor
it
h
m
,
”
in
Proceed
ings
-
2009
Inte
rnational
Jo
int
Confe
renc
e
on
B
ioi
nformatic
s
,
Sy
stems B
iol
og
y
a
nd
Intelli
g
ent Co
mputing,
I
JCB
S
2009
,
pp
.
525
–
5
28
,
2009
.
[33]
S.
Sunar
y
o,
e
t
al
.
,
“
Double
Seasona
l
Rec
urr
en
t
Neura
l
Netwo
rks
for
Forec
asti
ng
Short
Te
rm
El
ectricit
y
Load
Dem
and
in
Indo
nesia
,
”
Re
curr.
Neural
Ne
tworks
Tempor
al
Data
Proce
ss
.
,
2011.
[34]
E.
Sriniv
as
and
A.
Jain,
“
A
Met
hodolog
y
for
Short
Te
rm
Loa
d
Forec
asti
ng
Us
i
ng
Fuzz
y
Logic
and
Sim
il
arit
y
,
”
Syst.
Re
s.
,
pp
.
1
–
6,
2009
.
[35]
H.
H
.
Cev
ic
and
M.
Cunkas,
“
Short
-
te
rm
lo
ad
f
ore
ca
st
ing
using
fuz
z
y
logic
and
AN
FIS
,
”
Neura
l
Comput.
Appl.
,
vol
/i
ss
ue:
26
(
6
)
,
pp.
1355
–
1367
,
2015.
[36]
P.
Bunnoon,
“
El
ec
tr
ic
i
t
y
Pe
ak
L
oad
Dem
and
using
De
-
noising
W
ave
le
t
Tr
ansfo
rm
int
egr
at
ed
wi
th
Neura
l
Netwo
r
k
Methods,
”
I
nt. J.
Elec
tr
.
Comput.
Eng.
,
vo
l
/i
ss
ue:
6
(
1
)
,
p
p
.
12
,
201
6.
[37]
W
.
W
ei
,
“
Ti
m
e
Serie
s
Anal
y
sis
:
Univar
i
at
e
and
Multi
var
ia
t
e
Methods
,
”
2nd
ed
.
New
York:
Pe
arson,
Addison
W
esley
,
2006
.
[38]
G.
E
.
P.
Box,
et
al.
,
“
Ti
m
e
ser
ie
s
anal
y
s
is:
Fore
cas
ti
ng
and
Contro
l
,”
2008.
[39]
J.
D.
Cr
y
e
r and
K
.
S.
Ch
an,
“
Tim
e
Serie
s Ana
l
ysis: With
Appli
c
at
ions
in
R
,
”
200
8.
[40]
N.
Moham
ec
l,
e
t
al.
,
“
Forec
asti
n
g
short
te
rm
loa
d
demand
using d
ouble
sea
sonal a
rima
m
odel
,
”
W
orld A
ppl.
Sci.
J
.
,
vol
/i
ss
ue:
13
(
1
)
,
pp.
27
–
35
,
2011
.
Evaluation Warning : The document was created with Spire.PDF for Python.