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.
10
,
No.
4
,
A
ugus
t
2020
,
pp.
3911
~
39
17
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v10
i
4
.
pp
3911
-
39
17
3911
Journ
al h
om
e
page
:
http:
//
ij
ece.i
aesc
or
e.c
om/i
nd
ex
.ph
p/IJ
ECE
Sh
or
t
-
t
erm lo
ad fo
recasti
ng with
usin
g multip
le
lin
ea
r r
egressi
on
Bhatti D
hav
al
,
A
nur
ad
h
a D
eshpande
Depa
rtment
o
f
E
le
c
tri
c
al E
ngi
n
eering,
Fa
cul
t
y
of
Te
chno
log
y
a
nd
Engi
ne
eri
ng,
Maha
raja
Sa
y
a
jirao
U
nive
rsi
t
y
o
f
Baroda
,
Ind
ia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
J
un
25
, 2
019
Re
vised Feb
2
5
,
2020
Accepte
d
Ma
r
3
, 2
020
In
thi
s
pape
r
short
te
rm
loa
d
fore
ca
sting
(STLF)
i
s
done
with
usin
g
m
ult
ipl
e
li
ne
ar
reg
ression
(MLR)
.
A
da
y
ahe
ad
lo
ad
for
ec
ast
ing
is
obta
i
ned
in
this
pape
r.
Reg
ressi
on
coe
fficie
nts
were
found
out
with
the
hel
p
of
m
et
hod
of
le
ast
square
estim
at
ion.
Lo
ad
in
el
ectri
ca
l
powe
r
sy
st
em
is
depe
ndent
on
te
m
per
at
ur
e,
du
e
point
and
s
ea
sons
and
al
s
o
loa
d
h
as
co
rre
lation
to
the
pr
evi
ous
lo
a
d
consum
pti
on
(
Historic
a
l
da
ta
)
.
So
the
input
v
a
ria
bl
es
are
t
empera
tur
e,
d
u
e
point,
lo
ad
of
prior
da
y
,
hour
s,
an
d
lo
ad
of
prior
wee
k
.
To
validate
th
e
m
odel
or
check
the
accura
c
y
o
f
the
m
odel
m
ea
n
absolute
per
ce
n
ta
g
e
err
or
is
used
and
R
s
quar
ed
is
check
ed
which
is
show
n
in
result
sec
ti
on
.
Us
ing
d
a
y
ah
ea
d
fore
ca
s
te
d
d
at
a
wee
k
l
y
fore
ca
st
is
al
so o
bta
in
ed.
Ke
yw
or
d
s
:
Mult
iple l
inear
r
e
gr
essi
on
Re
gr
essi
on
Shor
t t
erm
loa
d forecast
in
g
Copyright
©
202
0
Instit
ute of
Ad
v
ance
d
Engi
ne
eri
ng
and
Sc
ie
n
ce
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Anurad
ha Des
hp
a
nde
,
Dep
a
rtm
ent o
f El
ect
rical
an
d
En
gin
eeri
ng,
Ma
har
a
j
a
Saya
j
ira
o Un
i
ver
sit
y
of Bar
od
a
,
Kala b
ha
van,
Near B
adam
adiBa
ug
,
V
a
doda
ra, G
uj
a
rat,
Ind
ia
.
Em
a
il
:
a.s.d
esh
pande
-
eed
@m
su
ba
r
od
a
.ac.in
1.
INTROD
U
CTION
In
to
day’s
w
orl
d
co
ntin
uity
in
el
ect
rical
s
upply
is
neces
sar
y.
Ele
ct
rici
ty
i
s
an
esse
ntial
need
in
daily
li
fe
of
pe
ople
.
Ele
ct
rical
util
ities
try
to
pr
ov
i
de
c
on
ti
nu
ous
powe
r
s
upply
to
their
cu
stom
ers.
T
o
e
nsure
this
load
forecast
in
g
is
a
n
im
po
rtant
aspect
.
L
oa
d
f
oreca
sti
ng
is
ab
ou
t
est
i
m
at
ing
f
utu
re
consum
ption
s
base
d
on
va
rio
us
dat
a
an
d
i
nfor
m
ation
a
vaila
ble
a
nd
as
per
c
on
s
um
er
beh
a
vior
.
L
oad
f
or
eca
s
ti
ng
has
a
n
e
ffec
ti
ve
ro
le
in
eco
nom
ic
op
e
rati
on
of
powe
r
util
it
ie
s.
Sho
rt
-
te
rm
load
f
oreca
sti
ng
ref
e
rs
to
foreca
sti
ng
loa
d
for
hour,
day
or
week
a
head.
T
her
e
ar
e
m
any
too
ls
and
te
ch
niques
avail
able
f
or
load
f
or
eca
sti
ng
s
uch
as
sim
ple
tim
e
series,
re
gressi
on,
ne
ural
net
works
et
c.
L
oa
d
f
or
ec
ast
ing
m
ean
forecast
ing
a
ve
rag
e
l
oa
d
in
K
W
or
tot
al
load
in
K
Wh
for
pe
r
io
ds
or
bl
oc
ks
of
15
m
inu
te
s,
30
m
inu
te
s,
1
ho
ur,
day
,
wee
k,
m
on
th
or
a
ye
ar
f
or
daily
forecast
, wee
kly forecast
, m
on
thly
for
eca
st
or ye
arly
.
Ther
e
a
re
m
any
factor
s
[
1]
w
hich
in
flue
nce
the
accuracy
of
load
forecast
ing
li
ke
weathe
r
var
ia
bles,
ho
li
days
,
fes
ti
vals
or
e
ven
ts,
ta
riff
str
uctu
re
s,
avail
able
his
torical
data,
ti
m
e
of
the
ye
ar
,
day
of
the
we
ek
an
d
hour
of
the
da
y.
W
eat
he
r
va
r
ia
ble
include
s
tem
per
at
ur
e
,
hum
idity,
rain
a
nd
wind.
Tem
per
at
ur
e
a
nd
hum
idit
y
has
a
co
ns
ide
r
able
eff
ect
on
power
c
onsum
pt
ion
b
eca
use
as
the
tem
per
at
ur
e
rises
pe
op
le
will
tur
n
on
ai
r
conditi
on
e
rs
a
nd
if
te
m
per
at
ur
e
is
lo
w
the
n
ai
r
heaters
will
get
turn
e
d
on,
wh
ic
h
causes
increase
in
el
e
ct
rici
ty
dem
and
. If
t
here i
s f
est
ival,
th
e ele
ct
rici
ty
d
em
and
will
r
ise
du
e
to
li
ghti
ng
s.
Data
u
se
d
for
STLF
is
m
os
tl
y
histor
ic
al
hourl
y
load
a
nd
weathe
r
pa
ram
et
ers.
Hou
rly
load
data
is
su
bject
to
error
s
m
ade
by
instr
um
ents
an
d/o
r
oper
at
or
s
a
nd
al
s
o
un
wan
te
d
or
un
plan
ned
c
hange
s
in
the
power
s
yst
e
m
.
Paper
has
co
ns
ide
red
five
diff
e
re
nt
cases
of
days
and
A
bsolute
no
rm
al
iz
ed
residu
al
m
et
ho
d
is
use
d
for
im
pr
oper
data
m
ark
in
g.
A
bsolute
nor
m
al
iz
ed
resid
ua
l
m
e
tho
d
is
use
d
for
i
den
ti
fi
cat
ion
and m
od
ific
at
ion o
f
im
pr
oper
data
[2
]
.
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.
10
, No
.
4
,
A
ugus
t
2020
:
3911
-
3917
3912
Goo
d
qua
li
ty
of
data
is
esse
ntial
in
powe
r
syst
e
m
bef
or
e
ta
kin
g
decisi
on.
S
o
data
m
u
st
be
cl
eane
d
and
filt
ered
be
fore
op
e
rato
r
t
akes
a
ny
decis
ion
from
the
da
ta
.
Othe
r
wise
it
will
cause
hazar
dous
co
ndit
ion.
Poor
qual
it
y
of
data
aff
ect
s
de
ci
sion
m
aking
without
knowl
edg
e
of
ope
rator.
Data
sm
oo
thin
g
is
necessa
ry
fo
r
so
m
e app
li
cat
i
on
fo
r
e.g. load fo
reca
sti
ng
in powe
r
syst
e
m
.
Tech
niques u
s
ed
f
or
this are St
at
ist
ic
al
te
chn
iq
ues
su
c
h
as
O
WA
(
opti
m
a
ll
y
w
ei
gh
te
d
ave
rage
)
an
d
M
A
(
m
ov
ing
a
ver
a
ge
)
.
Data
sm
oo
thin
g
a
nd
filt
eri
ng
is d
on
e
on s
m
art m
et
er d
at
a wi
th m
easur
em
e
nts
of
15 m
inu
te
i
nter
val of
D
ONWOD
r
e
gion
of N
e
w
Y
or
k [
3].
In
Pa
pe
r
[
4]
Sh
ort
te
rm
load
forecast
will
no
t
ass
um
e
th
e
con
tri
bu
ti
on
of
a
gr
ic
ultur
al
load
w
hic
h
is
seaso
n
dep
e
nd
e
nt
a
nd
th
us
is
assum
ed
to
be
c
on
sta
n
t
in
the
sp
a
n
of
fe
w
days.
T
he
tr
ia
ngular
m
e
m
ber
shi
p
functi
on
is
us
e
d
w
her
e
the
s
uppo
rt
of
the
m
e
m
ber
sh
i
p
f
unct
ion
is
decide
d
on
the
basis
of
the
c
ollec
te
d
data
.
Re
la
ti
on
sh
i
p
of
load
with
w
eat
her
va
riable
s,
tem
per
at
ure
and
hu
m
idit
y
is
sh
own
.
Th
ese
par
am
et
e
rs
are
m
od
el
ed
in
the
do
m
ai
n
of
f
uzzy,
to
obta
in
reali
sti
c
value
of
the
fore
cast
ed
load
w
hen
c
om
par
ed
with
the
act
ual
load
.
In
pap
e
r
[
5]
ANN
(
a
rtfici
al
neural
netw
ork
)
te
ch
nique
is
us
ed
f
or
STL
F.
Th
ree
m
od
el
s
are
consi
der
e
d,
1.
Pr
e
dicti
on
of
t
he
nex
t
ho
ur
load
,
2.
Pr
e
dicti
on
of
the
nex
t
day
l
oad
pr
of
i
le
,
an
d
3.
P
redi
ct
ion
of
th
e
ne
xt
da
y
peak
loa
d.
ANN
m
od
el
s
us
es
loa
d
prof
i
le
and
weathe
r
sit
uation
as
i
nput
la
ye
r
an
d
resu
lt
s
in
f
or
eca
ste
d
nex
t
hour
l
oad,
ne
xt
day
loa
d
prof
il
e,
a
nd
nex
t
day
peak
load
.
Me
a
n
a
bs
ol
ute
er
r
or
i
s
use
d
to ev
al
uate the
perform
ance o
f
Forecast
in
g.
In
pa
per
[
6]
m
ul
ti
ple
li
near
regressi
on
an
al
ysi
s
te
chn
iq
ue
is
us
e
d
t
o
f
oreca
st
dam
ping
in
N
ordiac
powe
r
syst
em
.
A
sta
ti
c
MLR
m
od
el
is
devel
op
e
d
to
e
xp
l
ai
n
the
va
riabi
li
ty
of
the
da
m
pin
g
of
t
he
0.35
-
H
z
inter
-
a
rea
m
od
e
in
the
Nordi
c
syst
e
m
.
In
[7
]
five
s
hort
te
rm
load
fore
cast
ing
te
ch
niq
ue
s
are
disc
usse
d.
These
five
te
chn
i
qu
e
s
are
(i)
Mult
iple
li
nea
r
re
gressi
on,
(i
i)
Stoc
hastic
ti
m
e
series,
(iii
)
Ge
ner
al
e
xponentia
l
sm
oo
thing,
(iv
)
Stat
e
s
pace
a
nd
K
al
m
an
filter,
an
d
(
v)
Knowled
ge
-
base
d
appr
oach.
Algorithm
s
and
ne
cessary
equ
at
io
ns
f
or
these
te
c
hn
i
qu
e
s
are
disc
us
se
d.
F
or
ot
her
pa
r
a
m
et
ers,
m
et
h
od
s
of
short
te
rm
load
f
or
eca
sti
ng,
con
t
ro
ll
in
g
fac
tors
et
c,
discu
s
sed
in
[8
-
18
]
.
In
c
om
par
iso
n
to
[
19
-
2
9
]
,
propose
d
m
et
ho
d
cal
culat
es
th
e
short
te
rm
load
f
or
e
cast
us
in
g
m
ul
ti
ple
reg
re
ssio
n
m
et
ho
d
by
c
al
culat
ing
the
r
andom
err
or
(
5%
)
w
hich
is
s
howi
ng
the
diff
e
ren
c
e
betwee
n
o
bs
er
ved
a
nd
fitt
ed
m
od
el
.
T
he
tr
ai
ned
m
od
el
is
95%
a
ccu
rate.
Als
o
t
he
e
ff
e
ct
of
var
ia
bles is inc
lud
e
d.
2.
RESEA
R
CH MET
HO
D
Me
thod
use
d
for
S
hort
-
te
rm
Loa
d
F
or
eca
s
ti
ng
is
m
ulti
pl
e
li
near
re
gres
sion.
Re
gressi
on
a
naly
sis
helps
to
kn
ow
the
relat
ion
s
hip
betwee
n
loa
d
and
va
riables
aff
ect
in
g
load
consum
ption
s
li
ke
weathe
r
va
riable,
seaso
nal
ef
fec
ts
and
previ
ous
loa
d
c
ons
um
ption
data.
This
m
et
ho
d
is
base
d
on
sta
ti
sti
cal
app
r
oac
h.
Re
gr
essi
on
a
na
ly
sis
helps
to
f
ind
ou
t
unknown
values
of
pa
ram
et
ers
on
the
basis
of
give
n
data
set
(a
va
il
able
histor
ic
al
data
).
T
he
cal
cula
ti
on
of
un
known
values
of
pa
ram
et
ers
can
be
e
xpla
in
ed
by
ta
ki
ng
si
m
ple
regressio
n
m
odel
.
Sim
ple li
near
r
e
gressi
on
m
od
el
:
=
0
+
1
+
(1)
Her
e
, i =
1, 2,
n
y i
s elec
tric
al
load
0
an
d
1
are Pa
ram
et
ers
(Re
gr
es
sion coe
ff
ic
ie
nts)
x
is e
xpla
nato
r
y varia
ble
is rand
om
e
rr
or
In
this
m
od
el
relat
ion
sh
i
p
bet
ween
loa
d
y
and
var
ia
ble
x
i
s
expresse
d
with
the
help
of
par
am
et
ers.
We
can
obta
in
the
value
of
l
oad
y
f
or
kn
own
value
of
va
riable
x,
if
w
e
hav
e
t
he
β
va
lues
of
pa
ra
m
et
ers.
So
,
if
we
kn
ow
the
β
val
ue
s
of
par
am
et
ers,
we
c
an
obta
in
the
value
s
of
loa
d
at
an
y
giv
en
values
of
x.
is
rand
om
err
or
w
hic
h
sho
ws
how
the
f
it
te
d
m
od
el
is
dif
fer
e
nt
f
rom
the
act
ual
m
od
el
.
It
ref
l
ect
s
the
dif
fer
e
nce
betwee
n
ob
se
r
ved
an
d
fitt
ed
m
od
el
.
Her
e
this
m
od
el
is
for
only
one
e
xp
la
nato
ry
va
r
ia
ble,
bu
t
if
we
ha
ve
m
or
e
than
one
ex
planato
ry
var
ia
bles,
m
ulti
ple
li
near
reg
ressi
on
m
od
el
is
us
ed.
I
n
m
ulti
ple
li
near
r
e
gressi
on m
or
e tha
n o
ne
f
act
ors a
re c
o
ns
ide
red f
or af
fecti
ng the
res
pons
e
(
L
oa
d)
.
[
y
1
y
2
⋮
y
n
]
=
[
1
x
11
1
x
21
…
x
1k
…
x
2k
⋮
⋮
1
x
n1
…
⋮
…
x
nk
]
[
β
0
β
1
⋮
β
k
]
+
[
ε
1
ε
2
⋮
ε
n
]
(2)
Abo
ve rep
rese
ntati
on
is
b
as
e
d on
m
at
rix
dim
ension
s
as s
how
n
belo
w
,
and s
olv
a
ble in
(3)
[
Y
]
=
[
X
β
]
+
[
ε
]
(3
)
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
Shor
t
-
te
r
m
l
oad
fo
recasti
ng
wi
th u
sin
g m
ulti
ple li
nea
r
reg
ression
(
Bh
atti
Dha
v
al
)
3913
Y: nX
1 vecto
r of o
bs
e
rv
at
io
ns o
n
st
ud
y
var
ia
ble or
re
s
pons
e
v
a
riable
X: nX
k
m
at
rix
of n o
bs
e
rv
at
io
ns
eac
h of t
he k in
de
pende
nt
var
ia
bles
β
=
(
β
1
,
β
2
…
,
β
k
)
T
: kX1 vect
or
of r
e
gr
essi
on co
e
ff
ic
ie
nts as
so
ci
at
ed
with
x
1
,
x
2
,
…
x
n
ε
=
(
ε
1
,
ε
2
,
…
ε
n
)
T
: nX1 vect
or
of r
a
ndom
err
ors
In
te
rce
ptor te
r
m
ta
kes
first
co
lum
n
of
X
t
o b
e
(
1
,
1
,
…
1
)
Assum
ption
s:
-
E
(
ε
)
=
0
-
V
(
ε
)
=
E
(
ε
ε
′
)
=
σ
2
I
n
=
[
σ
2
0
0
:
σ
2
:
0
.
.
σ
2
]
-
Ra
nk(x)=
K:fu
l
l colum
n
m
at
ri
x
-
ε
~
N
(
0
,
σ
2
,
I
n
)
If
we
hav
e
th
e
values
of
ve
ct
or
s
f
or
regr
ession
c
oeffici
ents
we
can
predict
the
loa
d
for
gi
ven
ob
s
er
vations.
To
fin
d
the
va
lues
of
regress
ion
coe
ff
ic
ie
nts
m
et
ho
d
of
le
ast
sq
ua
res
is
us
e
d.
It
is
bas
ed
on
the princi
ple of
m
axi
m
a/mi
nim
a. H
ere
m
ini
m
iz
at
ion
o
f
s
um
o
f
square
d
e
rror
s
is c
on
si
de
red.
min
imi
ze
∑
ε
i
2
N
i
=
1
So
from
(
3) w
e
can w
rite
,
S
(
β
)
=
∑
ε
i
2
=
ε
T
ε
=
(
Y
−
X
β
)
T
(
Y
−
X
β
)
n
i
=
1
(4)
S
(
β
)
is
a
real
value
d,
c
onve
x,
di
fferentia
ble
f
un
c
ti
on
wh
e
re
t
he
m
ini
m
u
m
wil
l
al
ways
exist.
Acc
ordin
g
t
o
pr
i
nciple
of
m
inim
a,
∂
S
(
β
)
∂
β
=
0
β
̂
=
(
X
T
X
)
−
1
X
T
Y
(5)
So
v
a
l
u
e
s
o
f
p
a
r
a
m
e
t
e
r
s
c
a
n
b
e
f
o
u
n
d
o
u
t
b
y
t
h
i
s
e
q
u
a
t
i
o
n
.
V
a
r
i
a
n
c
e
o
f
r
e
s
i
d
u
a
l
a
l
s
o
n
e
e
d
s
t
o
b
e
c
o
n
s
i
d
e
r
e
d
s
o
,
Va
r
(
ε
i
)
=
σ
2
σ
̂
2
=
SS
res
n
−
k
=
MS
res
SS
res
is sum
o
f
s
quar
ed
e
rror
s
due t
o resid
uals
MS
res
is m
ean sq
ua
re
d
e
rror
On
ce
the
valu
e
s of
par
am
et
ers
are
know
n
fitt
ed
m
od
el
ca
n b
e obtai
ned.
Fit
te
d
m
od
el
:
Y
̂
=
X
β
̂
=
X
(
X
T
X
)
−
1
X
T
Y
=
HY
(6)
H
=
X
(
X
T
X
)
−
1
X
T
(7)
Re
sidu
al
s
(
)
can
be ob
ta
ine
d fro
m
the d
if
fer
e
nc
e b
et
w
ee
n o
bs
e
rv
e
d
a
nd f
it
te
d
values
.
ε
̂
=
Y
−
Y
̂
=
(
1
−
H
)
Y
=
H
̅
Y
(8)
Total
d
e
viati
on and Sum
o
f
s
quare
d resid
ual
is give
n
as
,
y
i
−
y
̅
=
(
y
i
−
y
i
̂
)
+
(
y
i
̂
−
y
̅
)
(9)
y
̅
is avera
ge val
ue
and
̂
is est
im
at
ed
val
ue.
SS
T
=
SS
reg
+
SS
res
(10)
∑
(
y
i
−
y
̅
)
2
n
i
=
1
=
∑
(
y
i
̂
−
y
̅
)
2
n
i
=
1
+
∑
(
y
i
−
y
i
̂
)
2
n
i
=
1
(11)
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.
10
, No
.
4
,
A
ugus
t
2020
:
3911
-
3917
3914
SS
reg
is
su
m
of
sq
ua
red
e
rror
s
due
to
regressi
on
a
nd
su
m
of
squ
ared
e
rrors
due
to
resid
uals.
T
o
che
c
k
how g
ood
t
he
m
od
el
is o
r h
ow
well
the m
od
el
f
it
s,
c
oe
ff
ic
ie
nt of
determ
i
nation ca
n b
e
use
d.
R
2
=
1
−
SS
res
SS
T
=
SS
re
g
SS
T
3.
RESU
LT
S
A
ND AN
ALYSIS
Historical
data
of
Ne
wyo
rk
c
it
y
of
9
ye
ars
sta
rting
f
ro
m
Ma
y
20
07
to
April
2016
is
us
e
d
to
trai
n
the
m
od
el
.
This
data
ha
d
m
e
asur
em
ent
interval
of
one
hour
s
o
the
re
are
m
or
e
than
se
ven
ty
ei
ght
th
ou
s
an
d
ob
s
er
vations.
Fo
r
ta
ki
ng
acc
ount
of
the
ef
fect
of
weathe
r
var
ia
bles,
dry
bu
l
b
te
m
perat
ur
e
a
nd
due
point
is
co
ns
ide
red.
The
pur
pose
of
ta
ki
ng
lots
of
in
f
orm
ation
is
due
t
o
t
he
fac
t
that,
if
m
or
e
data
is
us
e
d
t
o
buil
d
the
m
od
el
,
good
acc
ur
acy
i
n
predic
ti
on
of
the
m
od
el
is
obta
ined
.
T
his
data
c
on
ta
in
s
bad
data
li
ke
m
issi
ng
data,
m
easur
em
ent
err
ors
an
d
sud
den
s
pike
s,
b
ut
load
f
oreca
st
nee
ds
fi
ne
gran
ularit
y
of
data.
S
o
th
e
data
needs to
be
cl
e
aned o
r fil
te
red befo
re it i
s use
d
to
m
ake p
re
di
c
ti
on
s.
T
his is
ob
ta
ine
d
i
n our
p
re
vi
ou
s
work
[
3].
So
t
he fil
te
red
data is
us
ed
h
e
re fo
r
loa
d fore
cast
ing
.
Firstl
y
to
fo
rec
ast
the
load,
re
gr
essi
on
m
od
el
need
s
t
o
be
bui
lt
based
on
av
ai
la
ble
inform
a
ti
on
w
hic
h
is
pr
e
vious
ye
a
r’
s
data
of
loa
d
co
nsum
ption
and
te
m
per
at
ure
data.
Softwa
re
he
re
use
d
for
trai
ning
t
he
m
od
el
and
f
oreca
st
th
e
data
is
MAL
AB
2018b.
In
order
to
buil
d
an
acc
ur
at
e
m
od
el
,
us
ef
ul
predict
or
s
are
ne
eded.
A
com
m
on
te
chn
i
qu
e
with
te
m
po
ral
pr
e
dictors
requires
breakin
g
them
in
to
their
sepa
rat
e
par
ts
so
that
they
can
be
var
ie
d
ind
e
pe
ndentl
y
of
each
othe
r.
For
t
hat
three
pre
dictor
s
are
create
d,
hour,
wee
kd
ay
and
is
week
e
nd
or
no
t.
T
he
loa
d
data
it
sel
f
can
be
us
ed
as
a
pr
e
dictor
.
Lo
ad
can
be
us
e
d
as
predict
or
or
no
t,
can
be
ver
ifie
d
by
c
orrelat
ion
with
loa
d.
T
he
peaks
in
the
a
ut
ocorr
el
at
io
n
at
the
24
an
d
16
8
hour,
a
re
la
ggin
g,
so
,
la
gged
pred
ic
tors
of
1
a
nd
7
days
can
be
us
e
d.
S
o,
tw
o
pr
e
dictor
s
are
create
d
one
a
s
load
of
pri
or
da
y
and
oth
e
r
as
pri
or
week.
Re
s
ponse
va
riable
is
l
oa
d,
a
s
loa
d
is
to
be
forecast
ed
an
d
e
xp
la
nato
ry
va
riables
a
r
e
hour
,
day of
wee
k,
te
m
per
at
ur
e,
due
point,
weath
er
it
is w
ee
kend or
not, l
oad of
pr
i
or
day a
nd prior wee
k.
As
s
how
n
i
n
r
esearc
h
m
et
ho
d,
m
od
el
is
t
r
ai
ned
us
i
ng
m
ulti
ple
li
near
r
egr
es
sio
n
a
nd
le
ast
sq
ua
re
est
i
m
ation
.
O
nc
e
the
m
od
el
i
s
trai
ne
d,
it
is
check
e
d
how
the
m
od
el
will
perform
on
dat
a.
Data,
f
or
th
e
ye
ar
2012
f
ro
m
data
set
is
re
m
oved
an
d
loa
d
f
oreca
st
for
ye
ar
2012
is
obta
in
ed
from
trai
ned
m
od
el
.
To
check
the ef
fecti
ve
ne
ss of
forecast
in
g
m
et
ho
d
m
ean a
bsolute
per
c
entage e
rro
r
(
MAPE)
ca
n be
u
se
d.
=
100
×
(
−
)
(13)
MAPE
c
onside
rin
g
weathe
r
da
ta
,
f
or
trai
ni
ng
data
is
5.1
1%
an
d
f
o
r
te
st
data
or
data
it
had
not
see
n
befor
e
is
5.1
5%
.
MAPE
without
co
ns
ide
ri
ng
weathe
r
da
ta
fo
r
trai
ning
data
is
5.
05
%
and
for
te
st
data
is
5.06%.
S
o
the
trai
ne
d
m
od
el
is
nea
rly
95%
acc
ur
at
e
in
pre
dic
ti
on
of
a
day
ahe
ad
loa
d
a
nd
m
et
hod
is
accurate.
T
he
reason
beh
i
nd
dec
rease
in
MAPE
wh
e
n
we
do
no
t
c
onsider
weathe
r
data
is
for
day
ahea
d
load
f
or
e
cast
ing
because
it
is
no
t
m
uch
dep
e
nd
e
nt
on
weather
var
i
able.
So
whe
n
we
do
no
t
consi
der
tem
per
at
ur
e
a
nd
due
point
as
exp
la
nato
ry
va
riable,
M
AP
E
decr
ease
s.
F
or
fu
rthe
r
valid
at
ion
of
m
od
el
2
c
an
be
c
hec
ked
a
nd
it
is
0.8
9
a
nd
a
dju
ste
d
2
is
al
so
0.89,
w
hich
m
eans
m
odel
is
able
to
e
xp
la
in
near
ly
90%
of the
var
at
i
on
in loa
d.
He
re T
-
sta
i
sti
c and P
-
values
are sh
own
in
T
able
1
.
Estim
at
ed
c
oeffici
ents:
Table
1.
T
-
sta
i
sti
c and P
-
valu
es
Predicto
rs
Esti
m
at
e
SE
tStat
P
-
v
alu
e
(I
n
terce
p
t
)
6
7
1
.84
9
.12
5
7
7
3
.62
1
0
Ho
u
r
4
.48
1
4
0
.26
9
7
1
1
6
.61
6
6
.84
3
4
e
-
62
Day
o
f
W
eek
-
7
7
.79
5
0
.83
4
3
5
-
9
3
.24
1
0
is W
eeken
d
-
3
8
3
.48
3
.74
5
-
1
0
2
.4
0
Prior Da
y
0
.69
2
5
8
0
.00
2
1
8
1
5
3
1
7
.47
0
Prior
W
eek
0
.25
8
2
4
0
.00
2
2
1
2
2
1
1
6
.73
0
If
t
he
T
-
sta
t
is
great
er
t
han
2(>2
)
or
le
ss
t
ha
n
2(
<
-
2)
of
e
ach
var
ia
ble,
it
is
acce
ptable.
I
n
ta
ble
al
l
the
val
ues
of
T
-
sta
ts
are
gr
eat
er
tha
n
2
or
le
s
s
tha
n
2,
m
eaning
th
a
t
al
l
the
va
riables
are
sta
ti
sti
cal
l
y
sign
ific
a
nt.
P
-
values
in
dicat
e
that
the
popula
ti
on
par
am
et
er
is
e
qual
to
ze
ro.
I
f
it
is
le
ss
than
0.1
,
it
in
di
cat
es
a
sign
ific
a
nt
r
egr
es
sio
n.
As
sh
ow
n
in
ta
ble
m
os
t
p
-
values
are
zer
o
an
d
on
e
is
al
s
o
bel
ow
0.1,
t
her
e
f
or
e
al
l
the
repress
ors
are
sig
nificant.
So
the
trai
ne
d
m
od
el
is
good
and
ca
n
be
use
d
to
m
ake
pr
e
dic
ti
on
.
Her
e
a
day
ahead
loa
d
f
oreca
sti
ng
is
ob
t
ai
ned
for
25
th
Ju
ne
on
hourl
y
b
asi
s
by
us
i
n
g
trai
ned
m
od
el
as
giv
e
n
i
n
Ta
ble
2
.
Table
2
disp
la
ys
forecast
ed
load o
f
25
th
Jun
e 2019
on
hour
ly
b
asi
s.
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
Shor
t
-
te
r
m
l
oad
fo
recasti
ng
wi
th u
sin
g m
ulti
ple li
nea
r
reg
ression
(
Bh
atti
Dha
v
al
)
3915
Table
2
.
Day a
head l
oa
d fore
cast
f
or
25
th
June
2019
Date
Load
'2
5
-
Ju
n
e
-
2
0
1
9
00
:
0
0
:0
0
'
6
3
7
2
.5
8
7
8
1
1
3
8
6
6
3
'2
5
-
Ju
n
e
-
2
0
1
9
01
:
0
0
:0
0
'
6
3
0
0
.3
9
9
1
0
5
3
7
6
2
4
'2
5
-
Ju
n
e
-
2
0
1
9
02
:
0
0
:0
0
'
6
2
2
0
.5
1
7
4
0
3
4
4
3
8
6
'2
5
-
Ap
r
-
2
0
1
9
03
:0
0
:0
0
'
6
1
3
1
.4
7
8
1
3
4
1
3
4
8
4
'2
5
-
Ap
r
-
2
0
1
9
04
:0
0
:0
0
'
6
0
3
9
.6
2
4
5
1
3
8
4
0
3
3
'2
5
-
Ap
r
-
2
0
1
9
05
:0
0
:0
0
'
5
9
5
1
.2
9
6
0
2
4
6
4
2
5
6
'2
5
-
Ap
r
-
2
0
1
9
06
:0
0
:0
0
'
5
8
9
0
.9
0
2
6
1
4
9
4
9
9
5
'2
5
-
Ap
r
-
2
0
1
9
07
:0
0
:0
0
'
5
8
8
3
.0
9
6
0
8
4
8
8
8
2
8
'2
5
-
Ap
r
-
2
0
1
9
08
:0
0
:0
0
'
5
9
2
7
.7
8
8
0
4
8
5
7
7
9
2
'2
5
-
Ap
r
-
2
0
1
9
09
:0
0
:0
0
'
5
9
9
9
.0
0
9
9
3
0
2
9
7
2
1
'2
5
-
Ap
r
-
2
0
1
9
10
:0
0
:0
0
'
6
0
9
0
.8
4
8
6
6
8
5
9
0
2
8
'2
5
-
Ap
r
-
2
0
1
9
11
:0
0
:0
0
'
6
2
0
9
.3
7
6
9
0
1
0
8
2
7
9
'2
5
-
Ap
r
-
2
0
1
9
12
:0
0
:0
0
'
6
3
6
5
.7
3
8
2
9
2
7
7
1
7
4
'2
5
-
Ap
r
-
2
0
1
9
13
:0
0
:0
0
'
6
5
5
7
.6
3
2
4
5
6
2
6
9
2
7
'2
5
-
Ap
r
-
2
0
1
9
14
:0
0
:0
0
'
6
7
6
9
.7
5
4
7
6
0
4
7
5
3
0
'2
5
-
Ap
r
-
2
0
1
9
15
:0
0
:0
0
'
7
0
0
0
.6
3
0
9
8
6
9
5
2
8
2
'2
4
-
Ap
r
-
2
0
1
9
16
:0
0
:0
0
'
7
2
3
4
.8
4
2
7
2
4
0
7
5
6
1
'2
4
-
Ap
r
-
2
0
1
9
17
:0
0
:0
0
'
7
4
6
4
.4
4
9
7
0
7
2
5
8
0
5
'2
4
-
Ap
r
-
2
0
1
9
18
:0
0
:0
0
'
7
6
6
2
.5
2
8
8
3
5
8
4
3
2
8
'2
5
-
Ap
r
-
2
0
1
9
19
:0
0
:0
0
'
7
7
8
9
.9
0
5
0
5
9
0
9
8
1
6
'2
5
-
Ap
r
-
2
0
1
9
20
:0
0
:0
0
'
7
8
5
3
.0
0
9
0
3
5
9
1
5
8
6
'2
5
-
Ap
r
-
2
0
1
9
21
:0
0
:0
0
'
7
8
7
4
.4
3
5
9
0
8
9
2
4
3
3
'2
5
-
Ap
r
-
2
0
1
9
22
:0
0
:0
0
'
7
8
5
8
.5
8
7
2
7
4
2
7
3
9
2
'2
5
-
Ap
r
-
2
0
1
9
23
:0
0
:0
0
'
7
7
9
5
.3
0
2
7
9
3
3
9
8
9
2
In
Ta
ble
2
da
y
ahead
l
oad
f
or
ecast
in
g
is
obta
ined
an
d
w
it
h
the
hel
p
of
day
ahea
d
lo
ad
f
or
e
cast
,
week
ahea
d
l
oa
d
f
oreca
st
is
al
so
obta
ine
d.
Week
ahea
d
l
oa
d
f
oreca
st
is
ob
ta
ine
d
by
c
onside
rin
g
day
ahead
load
f
or
ecast
a
s
act
ual
loa
d
da
ta
and
use
d
f
or
ne
xt
day
lo
ad
forecast
.
Th
is
process
is
done
f
or
w
hole
week.
Her
e
forecast
of
wee
k
ahea
d
is
as
show
n
i
n
Fig
ure
1
.
Fi
gure
1
dis
play
s,
W
ee
k
days
25
th
J
une
to
1
st
Ju
ly
on
X
-
axis
an
d
Loa
d
in
10
00
to
80
00
MW
on
Y
-
axis
.
We
ekly
f
or
ecast
c
annot
be
ob
ta
i
ned
di
rectl
y
be
cause
forecast
m
od
el
is
dep
en
de
nt
on
data
of
pr
e
vious
day.
S
o
first
day
ahea
d
load
f
or
ec
ast
is
ob
ta
ine
d
an
d
it
is
us
e
d
as
data, t
he
n process
is re
peated a
nd
for
ecast
f
or
ne
xt day
is obtai
ned
Figure
1
.
Figure
1.
W
ee
k
a
head
loa
d
fore
cast
fro
m
2
5t
h
J
une to
1
st
J
uly
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.
10
, No
.
4
,
A
ugus
t
2020
:
3911
-
3917
3916
4.
CONCL
US
I
O
N
Shor
t
Term
load
f
or
eca
sti
ng
for
a
day
a
hea
d
f
or
ecast
is
obta
ined
us
i
ng
m
ul
ti
ple
li
near
regressio
n.
U
se
of
m
ulti
ple
li
near
reg
re
s
sion
is
becaus
e
load
is
dep
e
nd
e
nt
on
m
any
var
ia
bles
li
ke
histor
ic
al
load
an
d
weathe
r
var
ia
bl
e.
Mult
iple
l
i
near
re
gr
e
ssio
n
m
od
el
fo
r
s
hort
t
erm
load
fo
recasti
ng
is
easy
to
dev
el
op
a
nd
MATLAB
is
use
d
to
de
velo
p
the
m
od
el
.
F
or
day
a
hea
d
lo
ad
f
oreca
st
the
re
is
no
m
uch
eff
ect
of
te
m
per
at
ur
e
and
due
point
on
l
oad
a
s
th
e
m
ean
abs
olut
e
per
ce
ntage
error
with
we
at
her
var
ia
ble
and
with
ou
t
w
ea
ther
var
ia
ble
is
nea
rly
sam
e.
Mode
l
is
95%
acc
ur
at
e
i
n
pr
e
dic
ti
on
.
So
the
da
y
ahead
loa
d
f
or
ecast
in
g
is
done
with
the
de
vel
op
e
d
m
od
el
a
nd
with
t
he
he
lp
of
day
ah
ead
loa
d
forec
ast
data,
week
ahea
d
loa
d
f
or
ecast
is al
so
ob
ta
in
e
d.
ACKN
OWLE
DGE
MENTS
Tha
nk
s
t
o
the
el
ect
rical
eng
ineeri
ng
dep
a
rtm
ent,
of
Facul
ty
of
Technolo
gy
and
E
ng
i
ne
erin
g
M.S.
U
.
of
B
ar
od
a
.
REFERE
NCE
S
[1]
Harsh
Pate
l,
M
ahe
sh
Pand
y
a,
Mohan
Aw
are
,
"S
hort
Te
rm
Loa
d
Forec
ast
ing
of
India
n
S
y
st
em
Us
ing
Li
nea
r
Regre
ss
ion
and
Artifi
ci
a
l
Neura
l
Network
,
"
5t
h
Nirma
Unive
rs
it
y
Inte
rnatio
nal
Confe
renc
e
on
Engi
nee
rin
g
(
NUiCONE
)
,
pp.
1
-
5,
2015.
[2]
P.
Ans
ari
m
ehr
,
S
.
Barghi
nia,
Z
.
Mirsepa
ss
i,
H.
Habibi
,
"Iden
ti
fi
ca
t
ion
and
Modi
fic
a
ti
on
of
Im
prope
r
Loa
d
Dat
a
Us
ed
in
Short
-
T
erm L
oad
Forec
asti
ng
,
"
IEEE Ru
ss
ia
Powe
r Tec
h
,
St.
,
pp
.
1
-
5
,
20
05.
[3]
Dhava
l
Bhat
t
i,
Anuradha
Deshpande
,
"Bad
Dat
a
det
e
ct
ion
and
Data
Filt
er
ing
in
Pow
er
Sy
st
em
,
"
in
Inte
rnational
Journal
of
Computer
App
li
ca
ti
on
s
,
vol.
182,
no.
3
0,
pp
.
36
-
39
,
De
c
ember
2018
.
[4]
P.
Mukhop
adh
y
a
y
,
G.
Mitr
a,
S.
Bane
rj
ee,
G.
Mukherjee,
"El
ectri
ci
t
y
lo
ad
fore
ca
s
ti
ng
using
fuz
z
y
logi
c:
Short
t
er
m
loa
d
for
ec
ast
ing
factori
ng
we
ather
par
amete
r
,
"
2017
7th
In
te
r
nati
onal
Con
fe
r
enc
e
on
Pow
er
Syste
ms
(
ICPS)
,
pp.
812
-
819
,
20
17.
[5]
M.
Rameza
ni
,
H.
Fala
gh
i,
M.
-
R.
Haghif
am,
G.A.
Shahr
y
ar
i
,
"S
ho
rt
-
Te
rm
Elec
tr
ic
Loa
d
Forec
asti
n
g
Us
ing
Neura
l
Networks
,
"
EUR
OCO
N
2005
-
The
Int
ernati
onal
Confe
renc
e
on
"Computer
as a
T
ool
"
,
pp
.
1525
-
1
528,
2005
.
[6]
Franc
esc
o
Sull
a
,
Matt
iKoiv
isto,
Janne
Seppäne
n
,
JukkaTur
unen
,
Li
isa
C.
Ha
arla,
Olof
Sam
ue
ls
son
,
"S
ta
ti
sti
ca
l
Anal
y
sis
and
Fo
rec
ast
ing
of
D
a
m
ping
in
the
N
ordic
Pow
er
S
ystem
,
"
in
IE
EE
Tr
ansacti
ons
on
Powe
r
S
yste
ms
,
vol.
30
,
no
.
1
,
pp
.
306
-
315
,
Jan
.
2
015.
[7]
I.
Moghram
,
S.
Rahman,
"
Anal
ysis
and
Eva
lu
ati
on
o
f
Five
Shor
t
-
Te
rm
Lo
ad
Forec
ast
ing
T
ec
hni
q
ues
,
"
in
IE
EE
Tr
ansacti
ons on Power
Syst
ems
,
vol.
4
,
no
.
4
,
pp
.
1484
-
1491,
Nov.
1989.
[8]
N.
Yu
and
S.
Shah
and
R.
Joh
nson
and
R.
Sheric
k
and
M.
H
ong
and
K.
Lop
aro
,
“
Big
d
ata
a
naly
tics
in
pow
er
distri
buti
on
s
y
st
ems
,
”
2015
IE
EE
Pow
er
&
E
nergy
Soc
ie
t
y
I
nnovat
i
ve
Smar
t
Gr
id
Technol
o
gie
s
Confe
ren
c
e
(
ISG
T
)
,
W
ashingt
on,
DC,
pp.
1
-
5,
2015
.
[9]
Hos
sein
Akhava
n
-
Heja
z
i,
Ham
e
dMohs
eni
an
-
Ra
d
,
“
Pow
er
sy
ste
m
s
big
dat
a
ana
l
y
t
ic
s:
An
assessment
of
par
adi
g
m
shift
bar
r
ie
rs
and
prospec
ts,
”
Ener
gy
Re
ports
,
vol
.
4,
pp.
91
-
100,
Novem
ber
2018.
[10]
T.
Hong
and
S.
Fan,
“
Probabil
isti
c
e
le
c
tric
loa
d
fore
ca
st
i
ng:
A
tut
ori
al
rev
ie
w,
”
Int
ernati
onal
Journal
of
Forec
asting
,
vol
.
32
,
no
.
3
,
pp
.
9
14
-
938,
Jul
y
201
6.
[11]
P.
W
ang,
B.
Li
u
,
and
T
.
Hong,
“
El
e
ct
ri
c
loa
d
for
ec
ast
ing
with
re
ce
nc
y
eff
e
ct
:
A
big
dat
a
a
pproa
c
h,
”
Int
ernati
ona
l
Journal
of
Fore
c
asting,
vo
l. 32, n
o.
3
,
pp
.
585
-
59
7,
2016
.
[12]
T.
Hong,
P.
Pi
nson,
and
S.
F
an,
“
Global
en
erg
y
fore
ca
sting
competi
t
ion
2
012,
”
I
n
te
rnatio
nal
Journal
o
f
Forec
asting
,
vol
.
30
,
no
.
2
,
pp
.
3
57
-
363,
June
20
14.
[13]
Chongqing
Kang,
Y
i
W
ang,
Yu
shengXue,
Gang
Mu,
and
Ruij
in
Li
ao
,
“
Big
Data
Anal
y
ti
cs
in
China
’s
El
ectr
i
c
Pow
er
Industr
y
,
”
in
IE
EE
Powe
r
and
En
ergy
Ma
gazine
,
vo
l. 16,
no.
3
,
pp
.
54
-
65
,
Ma
y
-
June
2018
.
[14]
F.
L.
Quilumba,
W
.
-
J.
L
ee,
H.
Huang,
D.
Y.
W
ang,
and
R.
L.
Sz
aba
dos,
“
Us
ing
sm
art
m
et
er
d
at
a
to
improve
the
a
cc
ur
acy
of
int
rad
a
y
lo
ad
f
ore
ca
st
ing
consi
der
ing
customer
beha
vior
sim
il
a
rit
ie
s
,
”
in
IEEE
Tr
ansacti
ons
on
Smar
t
Gr
id
,
vol
.
6,
no
.
2
,
pp
.
911
-
918,
Mar
ch
201
5.
[15]
N.
Ding,
Y.
B´esa
nger
,
a
nd
F.
W
urtz
,
“
Next
-
da
y
MV
/L
V
subs
ta
ti
on
loa
d
for
ecaste
r
using
ti
m
e
serie
s
m
et
hod,
”
El
e
ct
ric
Pow
er
S
yste
ms
Re
sear
ch
,
vol
.
119
,
pp
.
34
5
–
354,
Febru
ar
y
2015.
[16]
T.
Hong,
P.
W
ang,
and
L.
W
hite,
“
W
ea
the
r
sta
tion
sele
ction
for
el
e
ct
ri
c
loa
d
for
ec
ast
ing,
”
In
te
rn
ati
onal
Journal
of
Forec
asting
,
vol
.
31
,
no
.
2
,
pp
.
2
86
–
295,
June
20
15.
[17]
Hos
sein
Akhava
n
-
Heja
z
ia,
Ham
e
d
Mohs
eni
an
-
Ra
d,
“
Pow
er
s
y
st
e
m
s
big
dat
a
an
aly
t
ic
s:
An
assess
m
ent
of
par
adi
g
m
shift
bar
r
ie
rs
and
prospec
ts,
”
Ener
gy
Re
ports
,
vol
.
4,
pp.
91
-
100
,
Novem
ber
2018
.
[18]
A
hm
ed
Yous
uf
Saber
,
AfrozAl
a
m
,
“
Short
te
rm
loa
d
forc
ast
ing
u
sing
m
ult
ipl
e
l
in
ea
r
reg
r
ession
fo
r
big
Data,
”
I
EEE
Symposium Se
ri
es
on
Computat
i
onal
Int
el
l
ige
nc
e
(
SSCI)
,
pp.
1
-
6
,
2017
.
[19]
Saee
d
M
.
Badr
an,
Os
asa
m
a
B.
Abouela
t
ta
,
“
Forec
ast
ing
Elec
tr
ic
a
l
loa
d
using
AN
N
co
m
bine
d
with
Multi
pl
e
Regre
ss
ion
Meth
od,
”
Th
e
res
earc
h
bulleti
n
of
Jor
dan
ACM
,
vol
.
2
,
no
.
2
,
pp
.
52
-
58
,
April
2012
.
[20]
Benj
amín
Zay
as
,
Alfre
do
Espinosa,
Vick
y
Sanch
e
z,
Javi
er
Pere
z
,
“
Gett
ing
re
ad
y
fo
r
dat
a
an
aly
t
ic
s
o
f
elec
tric
power
distri
buti
on
s
y
st
ems
,
”
Int
ernati
o
nal
Journal
of
C
omputers
,
vol
.
2
,
pp.
179
-
186,
20
17.
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
Shor
t
-
te
r
m
l
oad
fo
recasti
ng
wi
th u
sin
g m
ulti
ple li
nea
r
reg
ression
(
Bh
atti
Dha
v
al
)
3917
[21]
Mande
ep
Singh,
Raman
Maini,
“
Vari
ous
El
e
ct
ri
c
ity
Lo
ad
Forca
st
ing
Techni
ques
with
pros
and
co
ns,
”
Inte
rnat
ion
al
Journal
of
R
ec
e
nt
Techno
logy a
nd
Engi
n
ee
ring
(
IJR
TE)
,
vol
.
8
,
n
o.
6
,
pp
.
220
-
22
9,
Mar
ch
2020
.
[22]
Ta
o
Hang,
Min
G
ul,
Mesut
E
Ba
ran
,
H L
e
e
W
il
ls,
“
Modeli
ng
and forec
ast
ing
hourl
y
e
le
c
tric
loa
d
b
y
m
ult
iple
li
n
e
ar
reg
ression
with
i
nte
ra
ct
ions,
”
I
E
EE
PE
S
Gene
ral
Me
e
ti
ng
,
Provid
enc
e
,
RI
,
pp
.
1
-
8
,
2010
.
[23]
Tre
vor
Hast
ie,
Robert
Ti
bshi
ran
i,
Jer
om
e
F
rie
dm
an,
“
The
Elem
ent
s
of
Stat
ist
ical
Learni
ng,
”
Spring
er,
2
nd
edition
,
2016
.
[24]
S.Sara
vana
n
,
S.
K
anna
n
and
C.
T
hanga
ra
j,
“
India’s
El
ec
tr
ic
i
t
y
D
emand
Fore
ca
st
Us
ing
Regre
ss
i
on
Anal
y
sis
a
nd
Artifi
c
ia
l
Neur
al
Network
s
Based
on
Princi
pal
C
om
ponent
s,
”
ICTACT
Journal
On
Soft
Computing
,
vol.
2
,
no.
4,
pp.
365
-
370
,
Jul
y
2012.
[25]
Li
dong
Dua
n
,
Dongxiao
Niu
and
Zh
ihongGu,
“
Long
and
te
r
m
power
loa
d
fore
ca
st
ing
with
m
ult
i
-
reg
r
essiv
e
Regre
ss
ion
anal
y
sis,
”
IITA
Proc
ee
dings
of
sec
on
d
Inte
rnational
Symposium
on
int
el
l
ige
nt
In
formation
Technol
og
y
App
li
ca
ti
on
,
pp
514
-
518,
De
ce
m
ber
2008.
ht
tps:/
/
doi.
org/10
.
1109/
IIT
A.2008.
397
.
[26]
Heng
Shi,
Ming
hao
Xu,
R
an
L
i,
“
Dee
p
Learni
n
g
f
or
Hous
ehol
d
Loa
d
Forec
ast
i
ng
-
A
Novel
Pooling
De
ep
RNN
,
”
IEE
E
Tr
ansacti
o
ns on
Smar
t
Gr
id
,
no
.
99
,
2017
.
[27]
Mos
ad
Alkhat
hami,
“
Introduc
tion
to
El
e
c
tri
c
Loa
d
Forec
ast
in
g
Methods,
”
Columbia
Inte
rna
ti
onal
Publishin
g
Journal
of
Ad
va
nce
d
El
e
ct
ri
cal
a
nd
Computer
E
ngine
ering
,
v
ol
.
2
, n
o
.
1
,
pp
.
1
-
12
,
2015
.
[28]
Iss
ac
Adekunie.
S,
Ade
y
ink
a
Ajao.
A,
Chihu
rum
an
y
a
Fe
lly
.
N,
A
y
okunl
e
Aw
el
ewa
,
“
Me
dium
-
Te
rm
Load
Forec
asti
ng
o
f
Covena
nt
Univ
e
rsit
y
Us
ing
The
Regre
ss
ion
Anal
y
s
is
Methods,
”
Journal
of
En
ergy
Technol
og
i
es
and
Policy
,
v
ol
.
4,
n
o
.
4,
pp
.
10
-
16
,
2014
.
[29]
Enri
co
Ferre
r
a,
Ric
c
ard
o
Tomasi,
Xiaolei
Hu
&
Cla
udio
Pastrone,
”
Eva
lu
ation
of
Short
Te
rm
Loa
d
Forec
ast
in
g
Te
chn
ique
s
Ap
pli
ed
for
Sm
art
Micro
Grids
,”
W
orld
a
ca
de
m
y
of
sci
ence,
Engi
ne
eri
ng
a
nd
Te
c
hnolo
g
y
,
Inte
rnational
Jo
urnal
of
Compu
te
r,
El
e
ct
ri
cal
,
Aut
omation
,
Co
ntrol
&
Informa
ti
on
Eng
ineeri
n
g,
vol
9,
n
o
.
11
,
pp.
1942
-
1947
,
2014.
BIOGR
AP
H
I
ES
OF
A
UTH
ORS
D
h
a
v
a
l
B
h
a
t
t
i
R
e
s
e
a
r
c
h
s
t
u
d
e
n
t
2
0
1
8
-
2
0
1
9
,
D
e
p
t
.
o
f
E
l
e
c
t
.
E
n
g
g
.
,
F
T
E
,
M
S
U
o
f
B
a
r
o
d
a
,
V
a
d
o
d
a
r
a
3
9
0
0
0
1
.
E
m
a
i
l
:
d
h
a
v
a
l
.
b
h
a
t
t
i
4
9
9
5
@
g
m
a
i
l
.
c
o
m
A
n
u
r
a
d
h
a
D
e
s
h
p
a
n
d
e
A
s
s
o
c
i
a
t
e
P
r
o
f
e
s
s
o
r
o
f
e
l
e
c
t
r
i
c
a
l
E
n
g
i
n
e
e
r
i
n
g
,
D
e
p
t
t
o
f
E
l
e
c
t
E
n
g
g
,
F
a
c
u
l
t
y
o
f
T
e
c
h
n
o
l
o
g
y
a
n
d
E
n
g
i
n
e
e
r
i
n
g
,
M
S
U
o
f
B
a
r
o
d
a
,
V
a
d
o
d
a
r
a
3
9
0
0
0
1
.
S
h
e
h
a
s
M
E
(
E
P
E
)
i
n
E
l
e
c
t
r
i
c
a
l
E
n
g
i
n
e
e
r
i
n
g
.
S
h
e
h
a
s
b
o
o
k
p
u
b
l
i
s
h
e
d
,
m
a
n
y
N
a
t
i
o
n
a
l
a
n
d
I
n
t
e
r
n
a
t
i
o
n
a
l
j
o
u
r
n
a
l
p
a
p
e
r
s
p
u
b
l
i
s
h
e
d
i
n
h
e
r
c
r
e
d
i
t
s
c
o
r
e
a
n
d
r
e
v
i
e
w
e
r
,
t
e
a
c
h
e
r
a
n
d
g
u
i
d
e
t
o
s
t
u
d
e
n
t
s
.
E
m
a
i
l
:
a
.
s
.
d
e
s
h
p
a
n
d
e
-
e
e
d
@
m
s
u
b
a
r
o
d
a
.
a
c
.
i
n
Evaluation Warning : The document was created with Spire.PDF for Python.