TELKOM
NIKA
, Vol. 11, No. 12, Decem
ber 20
13, pp.
7021
~70
2
6
e-ISSN: 2087
-278X
7021
Re
cei
v
ed
Jun
e
29, 2013; Revi
sed Aug
u
st
3, 2013; Accepted Augu
st
18, 2013
Resear
ch in Residential Elect
ricity Characteristics and
Short-Term Load Forecasting
Haixia Fen
g
*
,
Zhongfen
g Wang, Weic
hun Ge, Yingnan Wang
Shen
ya
n
g
Institute of Automation C
h
in
ese Ac
adem
y
of Scie
nces, Univ
ersit
y
of Ch
ines
e Academ
y of
Scienc
e, Chin
a
Lia
o
-Ni
ng Elect
r
ic Po
w
e
r lim
ite
d
-lia
bi
lit
y
C
o
m
pan
y. She
n
y
a
n
g
Po
w
e
r Su
pp
l
y
Com
p
a
n
y
, C
h
in
a
F
a
x: 02
4-23
97
043
3
*Corres
p
o
ndi
n
g
author, em
ail
:
fenghai
xias
hi
w
o
@1
63.com
A
b
st
r
a
ct
In this pap
er
w
e
make r
e
se
arch in R
e
si
de
ntia
l sh
ort-term load for
e
casti
ng. Differe
nt applic
atio
n
scenes
hav
e d
i
fferent affectin
g factors
of sh
ort-term
loa
d
, s
o
w
e
sh
oul
d sp
ecifical
ly a
n
a
l
y
s
is of factors t
hat
affect the lo
ad
of the res
i
d
enti
a
l e
l
ectricity.
We use SPSS
(Statistic Pack
age f
o
r Soci
al
Scienc
e) to fig
u
r
e
out the re
lati
on
ship
betw
een t
he d
a
ily
loa
d
a
nd te
mp
er
atur
e, w
eather co
n
d
itio
ns
an
d oth
e
r factors, findi
n
g
the main factor
s among th
e i
m
p
a
ctin
g facto
r
s, and an
aly
z
i
ng resi
de
ntial
electricity co
ns
umptio
n ha
bits
an
d
loa
d
ch
aracter
i
stics. T
hen, th
e p
aper
i
n
trod
uces
th
e co
mmo
n
pred
ictio
n
metho
d
s. Co
mb
ini
n
g
w
i
th th
e
abov
e an
alysis
to choose s
h
o
r
t-term lo
ad for
e
castin
g me
th
ods for resi
den
tial users, w
e
create a
u
to
mat
i
c
line
a
r re
gress
i
on
mo
de
l a
nd
artificial
n
eura
l
netw
o
rk mod
e
l to predict the f
u
ture
el
ectric
ity loa
d
, ca
lcul
ati
n
g
the res
i
du
al
be
tw
een the
pre
d
i
cted v
a
lu
es a
n
d
the
actu
al v
a
l
ues
and
me
an
squar
e d
e
vi
atio
n of th
e va
lu
es
,
and
eval
uati
n
g
the accur
a
cy of the lo
ad
for
e
castin
g. T
he
results prov
e t
hat auto
m
atic l
i
ne
ar regr
essio
n
mo
de
l is effective in resi
denti
a
l shor
t-term e
l
e
c
tricity load for
e
castin
g.
Ke
y
w
ords
:
res
i
de
ntial e
l
ectric
ity, short-term l
oad forec
a
st
in
g, line
a
r regres
sion, artifici
al n
eura
l
netw
o
rk
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
The po
we
r lo
ad fore
ca
stin
g begi
ns by
consi
der
i
ng th
e kn
own ele
c
tricity histo
r
ical data,
the eco
nomi
c
, so
cial influen
ce, climat
e and user
s’
electri
c
ity habits. To ma
ke a re
ason
able
electri
c
ity loa
d
fore
ca
sting
of the future,
we
su
m
up all
the influen
ce
s ab
ove an
d
make
a
suita
b
le
model. Sho
r
t-term loa
d
foreca
sting i
s
to
predi
ct
the
next perio
d time (several
hours, a d
a
y or
several days) of load or its chan
ging tre
nds [1
]. Fore
ca
st re
sults d
epen
d on foreca
st date type
(holid
ays, we
ekd
a
ys
o
r
weekend
s),
te
mperature,
a
nd
weathe
r condition
s
et
c. Its accu
ra
cy is
importa
nt to
sched
uling, t
he o
p
timal
combinatio
n
of
the
unit, e
c
o
nomic di
spat
ch
and
ele
c
tricity
mark
et trading [2].
In the gene
ration sid
e
of
large p
o
wer sy
stem
s, in
cludi
ng conv
entional hyd
r
opo
wer,
thermal
po
wer pl
ants a
nd pu
mpe
d
storage
po
wer statio
n. Aiming to
sh
ort-te
rm
load
optimizatio
n, many expe
rts a
nd
scho
lars propo
se
different
op
timized
algo
rithms, such
as
dynamic p
r
og
rammin
g
, ge
netic
algo
rith
m, sh
ort-
te
rm
load
optimi
z
ation b
a
sed
on d
a
ta
stora
ge
[3]. On the d
e
mand
side o
f
the grid, load forecast
in
g can reflect th
e entire con
s
umption level
of
the grid, so it has a
n
impo
rtant sig
n
ifica
n
ce
to the d
e
v
elopment of
powe
r
ge
neration plan
nin
g
.
To urban
co
mpre
hen
sive
sho
r
t-te
rm l
oad fo
re
ca
sting with
co
m
p
lex influen
ci
ng facto
r
s, some
authors p
r
e
s
ent the met
hod
s that combine
d
th
e
relevant lite
r
ature ing
r
ed
ients b
a
sed
on
prin
cipal
com
pone
nt analy
s
is m
e
thod
with the BP neural net
wo
rk
[4]. The prin
cipal com
pon
e
n
t
analysi
s
met
hod can redu
ce the dim
e
n
s
ion
a
lity of
th
e inputs d
a
ta of BP neural
netwo
rk, ma
king
the model m
o
re effe
ctive. In the aspe
ct of i
ndust
r
i
a
l electri
c
ity, the literature describe
s
the
comm
only used artificial
n
eural n
e
two
r
k, autor
eg
re
ssive model, gray model, illu
strating
ho
w to
improve the p
r
edi
ction a
c
cura
cy and co
mpen
sate
for
the predi
ction
erro
r by examples [5].
In re
cent ye
ars,
elect
r
icit
y con
s
umpti
on
is
gro
w
in
g sig
n
ifica
n
tly, and the
distan
ce
betwe
en pe
a
k
load
and v
a
lley load i
s
increa
sing
da
y by day and playing an i
m
porta
nt role
in
fueling [6]. T
h
erefo
r
e, it n
e
eds mo
re
an
d mo
re
a
ttention h
o
w to im
prove th
e p
r
e
d
iction
a
c
curacy
of the re
sid
ential ele
c
tri
c
ity con
s
u
m
ption. Ho
we
ver, most of
the cu
rrent
sho
r
t-te
rm l
oad
forecastin
g a
nd optimi
z
ati
on meth
od
s
are
discu
s
se
d for la
rg
e in
dustri
a
l u
s
e
r
s or fo
r a
wh
ole
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 12, Dece
mb
er 201
3: 702
1 – 7026
7022
so
ciety or re
g
i
on [4, 5].The predi
ction
system abov
e is
compl
e
x and
is not suita
b
le
for resi
dential
power loa
d
forecastin
g.
In this pa
per,
based o
n
th
e analy
s
is
of l
oad
cha
r
a
c
teristi
cs
and i
n
fluen
cing fa
ctors of
comm
unity resid
ents, we
compa
r
e th
e sho
r
t-term load fore
ca
sti
ng method
s and mod
e
ls, and
then
choo
se
simple
an
d e
ffective meth
ods for
re
sid
ential
po
we
r con
s
um
ption sho
r
t-te
rm
lo
ad
forecas
t
ing.
2.
Reside
ntial
Electricit
y
Charac
teris
t
ic
s Analy
s
is
2.1. Outline
With re
side
ntial
ele
c
tri
c
ity con
s
um
ption
and th
e n
u
m
ber of u
r
ba
n
popul
ation
co
ntinue
risin
g
, whi
c
h
cau
s
e
s
great chall
enge
s to
the grid
po
wer gen
eration
,
transmi
ssi
o
n
and di
stribu
tion
se
ctor. The
demand of
resid
ential
electri
c
it
y co
nsum
ption lo
ad fore
ca
sting and el
ectric
manag
eme
n
t on the dema
n
d
side i
s
incre
a
sin
g
urg
entl
y
.
To
Shan
ghai resi
dential el
ectri
c
ity
con
s
umption,
for
example, in t
he late nin
e
ties of the
last centu
r
y, the average
power
l
oad
gro
w
th rate
is ab
out 7
%
, and the
rate of g
r
o
w
t
h
in
con
s
um
ption
is a
r
ou
nd
5%. After 200
0,
with th
e
deve
l
opment t
r
en
d of the
e
c
on
omy be
comi
n
g
better as
well
as som
e
oth
e
r clim
atic factors, t
he growth rate of power load a
nd p
o
we
r ri
sing ra
te
are n
ear to
doubl
e-di
git. The re
sid
enti
a
l elect
r
ic
ity prop
ortio
n
in
cre
a
sed
signi
ficantly, sho
w
ing
the feature
s
that peak-to-valley difference in
crea
se
s and living l
oad propo
rtio
n gro
w
s
hea
vily
than the ele
c
tri
c
ity con
s
u
m
ption prop
ortion [7].
To this en
d, the re
side
nts load ele
c
tri
c
ity
forecastin
g a
nd mana
gem
ent comp
ared
with power
manag
eme
n
t is even mo
re
importa
nt.
2.2. Tempera
t
ure Influe
nc
e on the Res
i
dential Electricit
y
Load
Acco
rdi
ng to the cha
ngin
g
of load comp
ositi
on, re
sid
ential elect
r
ici
t
y load can be divided
into two p
a
rt
s, base l
oad
compon
ent an
d se
asonal l
o
ad compo
n
e
n
t [8]. Base l
oad
comp
on
ent
inclu
d
e
s
lig
hting, ele
c
tri
c
heaters, tel
e
vision
se
t
s
, kitch
en ele
c
trical equi
pme
n
t, refrigerators,
wa
shin
g machine
s
etc. Seaso
nal load
compon
ent
is
mainly cau
s
e
d
by the air conditionin
g
a
nd
heating. T
he
transfo
rmatio
n of the tem
perat
u
r
e i
n
flu
ences
huma
n
body comfort, and then t
he
cha
nge of hu
man body co
mfort influen
ces ele
c
tri
c
ity con
s
um
ption.
Unde
r norma
l circum
stan
ces
,
whe
n
the
ro
om temp
erat
ure i
s
high
e
r
than
26
°C or
belo
w
1
0
°C, th
e el
e
c
tri
c
ity load
will
signifi
cantly chang
es
with tempe
r
ature changi
ng.
Among all of ele
c
tri
c
ity equip
m
ent above,
the
air conditioni
ng is mo
st ca
sual, an
d sh
a
r
es th
e mo
st
signifi
cant pa
rt in pow
er lo
ad. Du
ring
wi
nter
in so
uthe
rn
Chin
a, peo
pl
e use ai
r
co
nditioning
to heating and dehumi
d
ificati
on,
so
ele
c
tri
c
ity
load in
crea
ses by the te
mperature. I
n
sum
m
er
, g
enerally no
rth-sout
h hi
gh
-temperature
air-
con
d
itioning
load
va
rie
s
with
tem
pera
t
ure.
In
sp
ri
ng a
nd
autu
m
n, outd
oor tempe
r
atu
r
e
is
suitabl
e, and
the turn-on t
i
me of
tem
p
e
r
ature-reg
u
lat
i
ng d
e
vice
s a
r
e
sho
r
ter,
so the im
pa
ct
of
temperature
on the load
shows wea
k
ly.
2.3. Holida
y
s
Electrical Charac
teris
t
ic
s
Chin
a's m
a
jo
r holid
ays in
clud
e The L
abor
Day, National Day and Spri
ng
Festival.
Duri
ng the L
a
bor
Day and
Nation
al Day
Festival, t
he tempe
r
ature is neither too h
o
t nor too col
d
,
and they’re lo
nger h
o
liday
s. Many peopl
e cho
o
se to
go out for part
y
, tourism or
other vari
etie
s of
lifestyles, m
a
kin
g
the el
ectri
c
ity load
more
disp
ersed.
Duri
n
g
the Spri
n
g
Fe
stival, the
temperature
s
are lo
w, an
d peopl
e visi
t friends
and
relatives to
hold family g
a
therin
gs,
wh
ich
make
s the living ele
c
tricity load he
avier.
2.4. Life Hab
i
ts
Comm
unity resid
ential ele
c
tri
c
ity con
s
u
m
pti
on in
clud
es lightin
g, cooki
ng, temp
eratu
r
e
regul
ation
an
d so
on. Sm
a
r
t meteri
ng
can read
ea
ch
hou
seh
o
ld'
s
electri
c
ity con
s
umptio
n eve
r
y
15 min
u
tes,
a
nd eve
n
som
e
mete
rs can
do real
-t
ime
data readi
ng.
The
wi
rele
ss meter,
rea
d
t
h
e
load data a
n
d
sen
d
it to the con
c
e
n
trator, and th
e
n
they are sent to the ba
ckgro
und of t
h
e
Electri
c
ity Authority. The data use
d
in
this arti
cle a
r
e ele
c
tricity con
s
um
ption
values on
ce
an
hour. Th
e differen
c
e b
e
twe
en adja
c
e
n
t data con
s
titute
s re
side
nts hi
stori
c
al loa
d
seque
nces.
Figure 1
is a
distri
ct of
da
ily load
cu
rve
.
You can
se
e the
pea
k
p
e
riod
con
c
ent
rates in
the 10:00 am
-12:00
pm an
d 16:00 to 20
:00. It has
two pea
k pe
rio
d
s at noo
n a
nd night. Du
ri
ng
these
pe
riod
s p
eople
are
co
okin
g, wa
tching
TV, la
undry
and
d
o
othe
r a
c
tivities. It brin
g
s
sub
s
tantial g
r
owth of the lo
ad.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Re
sea
r
ch in Re
side
ntial Electri
c
ity
Cha
r
acteri
stics an
d Short-T
e
rm
Load… (Hai
xia Feng
)
7023
Figure 1. Dail
y Load Cu
rve
In addition, the load level
is also affe
ct
ed by wo
rking days/
r
est
days, the weather
con
d
ition
s
(sunny/rainy
). Selecting
th
e
sampl
e
of
March
el
ectricity con
s
u
m
ption
per
day,
together with
the daily m
a
ximum temp
erature, mi
nimu
m tempe
r
atu
r
e, we
ather co
ndition
s (rain
y
or sunny), it analyzes the correl
ation of these factors
by SPSS.
Daily co
nsum
ption, daily m
a
ximum temp
eratu
r
e, mini
mum tempe
r
ature, rainy o
r
su
nny
status Pe
arson co
rrelatio
n are
sho
w
n
in Figure 2. The daily co
nsum
ption an
d daily maximum
temperature
Pearson co
rrel
ation
coeffi
cient i
s
0.5
8
8
,
the daily mi
nimum tem
p
e
r
ature correla
t
ion
coeffici
ent is
0.237, and th
e correl
ation
coe
ffici
ent of the weath
e
r condition
s is 0.
527.
Table 1. Co
rrelation
s
of Da
ily C
onsumpti
on and Mete
o
r
ologi
cal F
a
ct
ors
Daily
electricity
load
(L)
Max
i
mum
Temperature
(Max
T)
Minimum
temperatu
re
(Mi
n
T)
Sunn
y
/
rain
y
(SR)
Pearson
correlation
L 1.000
-.237
-.588
-.527
MaxT
-.237
1.000
.683
-.052
MinT -.588
.683
1.000
.581
SR -.527
-.052
.581
1.000
Sig. L
.113
.000
.002
MaxT
.113
.000
.397
MinT .000
.000
.001
SR .002
.397
.001
N L
28
28
28
28
MaxT
28
28
28
28
MinT 28
28
28
28
SR 28
28
28
28
The impo
rtan
ce of predi
ctor varia
b
le
s Sort
re
sults
are
sho
w
n in
Figure
3. Th
e factors
affect the ele
c
tri
c
al loa
d
in
seq
uen
ce i
s
daily maximum tempe
r
at
ure, weathe
r
con
d
ition
s
(ra
i
ny
or su
nny) a
n
d
the daily minimum tempe
r
ature.
Figure 2. The
Influence Fa
ctors in Sequ
ence
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 12, Dece
mb
er 201
3: 702
1 – 7026
7024
Duri
ng the po
wer lo
ad fore
ca
sting, we
can
add o
r
del
ete the input para
m
eters b
a
se
d on
the importa
nce of influenci
ng facto
r
s ap
prop
riatel
y. It
can o
p
timize
the netwo
rk a
nd red
u
ce dat
a
redu
nda
ncy.
3.
Prediction S
c
heme Simulation Analy
s
is
Re
side
ntial el
ectri
c
ity syste
m
is a sm
all grid
. The l
a
ws of ele
c
tricity
are obvio
us,
and the
power lo
ad i
s
rel
a
tively stable. Co
nsi
d
ering
th
e tim
e
and
sp
ace
compl
e
xity with a
c
curacy
requi
rem
ents,
we use line
a
r
reg
r
e
ssi
on a
nalysi
s
metho
d
and BP neural network method for loa
d
forecastin
g a
nalysi
s
and
compa
r
ison.
In this pape
r, we take the
March ele
c
tricity
load dat
a in a resi
de
ntial distri
ct in Anhui
Province a
s
experim
ental
sam
p
les. T
he raw
dat
a
sho
u
ld b
e
pre
-
processe
d, finding o
u
t
the
default data a
nd maki
ng su
ppleme
n
t to the data.
3.1. Linear Regres
sion M
odel
Reg
r
e
ssi
on a
nalysi
s
is a
method that
predi
cts futu
re data tre
n
d
s
ba
se
d on
histori
c
al
data variatio
n. It requires less data, a
nd the cal
c
ul
ation spe
ed i
s
faster [9]. It is suitable for
descri
b
ing
sp
ecific i
s
sue
s
whi
c
h sequ
en
ce is
stable a
nd may achi
e
v
e good re
sul
t
s.
Powe
r the load reg
r
e
ssi
on
model fore
casting
te
chni
que
s is ba
se
d on the histo
r
ical d
a
ta
of the load, a
nd e
s
tabli
s
he
s mathe
m
atical analy
s
is
m
odel
s to pred
ict the future
of the load. I
n
this pap
er, two wee
k
s loa
d
data before the fore
ca
stin
g day are i
n
p
u
tted. Corre
s
pondi
ng varia
b
le
data from Ma
rch 1 to 16 l
oad seque
nce are the inp
u
ts
12
1
6
,
,
...,
x
xx
. Regre
ssi
on model o
u
tput
for the 17th lo
ad se
que
nce
y
.And March 1
7
actual lo
ad
is the mea
s
u
r
ed value
17
x
.
Then:
11
2
2
...
nn
Ya
b
x
b
x
b
x
(
1
)
16
,
n
12
,
,
,
...,
n
ab
b
b
are multiple
reg
r
e
ssi
on p
a
ram
e
ters, it automaticall
y
determined
by SPSS according to the load sequence.
Then the correlation on the
predi
cted value and t
he actual measure
m
ent value is tested,
and the correl
ation co
efficie
n
t and the me
an and vari
an
ce pa
ram
e
ters are
cal
c
ulat
ed.
22
11
11
()
nn
ii
i
ii
M
SE
E
Y
Y
nn
(
2
)
The predi
cte
d
value and
the actual va
lue of
the linear reg
r
essio
n
model o
u
tp
uts are
sho
w
n i
n
T
able 2.
We
cal
c
ulate th
e co
rr
elation
coeffici
ent
and the m
e
an an
d vari
ance
para
m
eters
u
s
ing E
quatio
n (2
). Th
e m
ean
squ
a
re
error i
s
0.00
16 an
d the
root mea
n
sq
uare
error is 0.0
4
.
Table 2. Co
m
pari
s
on of the
Actual Value
s
and Predi
ct Values u
s
in
g Linea
r Re
gre
ssi
on Mod
e
l
Sequence electrical
data
3-17(actual
value)
0.21 0.16 0.16 0.17 0.17 0.21
0.43
0.7
9
0.71 0.67 0.83
1.11 0.51
0.59
0.5
5
0.54 0.64 0.82 0.85 1.15
0.82
0.5
8
0.44
0.26
3-17(fo
recast
res
u
l
t
s
）
0.25 0.18 0.16 0.16 0.15 0.23
0.43
0.8
1
0.72 0.7
0.89
1.05 0.5
0.61
0.44
0.55 0.64 0.82 0.88 1.11
0.84
0.5
7
0.43
0.25
Analyzing th
e relative re
gularity of the load
seq
uen
ce, the i
m
pact of re
side
ntial
electri
c
ity loa
d
facto
r
is
rel
a
tively small, and
the
use
of automatic l
i
near
re
gre
s
si
on mod
e
l ba
sed
on hi
stori
c
al
l
oad val
ue
ca
n get
a b
e
tter pre
d
ictio
n
re
sult. Th
e reg
r
ession
mod
e
l
will
be
save
d
as
an XML file, and then we can pre
d
ict oth
e
r data by usi
ng this mod
e
l
file.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Re
sea
r
ch in Re
side
ntial Electri
c
ity
Cha
r
acteri
stics an
d Short-T
e
rm
Load… (Hai
xia Feng
)
7025
3.2. BP Neur
al Net
w
o
r
k
Model
The a
r
tificial
neural n
e
two
r
k
can
con
s
id
er th
e
no
n-lin
ear
ch
aracte
ristics of th
e l
oad. Th
e
forwa
r
d BP n
eural n
e
two
r
k is widely u
s
e
d
in
the electricity load fore
ca
st [10, 11].
Figure 3. Dail
y Consumptio
n and Tem
p
e
r
ature
Curv
e
s
Figure 4. BP
Network Trai
ning Perfo
r
m
ance
Neu
r
al n
e
two
r
k A
r
tificial
Neural
Net
w
o
r
k (A
NN) p
r
ed
iction te
chn
o
l
ogy, it can m
i
mic the
human b
r
ai
n to do the intelligent proce
ssi
ng of
non
-linear, no
n-d
e
termini
s
tic l
a
ws of adapt
ive
function. In
most
ca
se
s, t
he tem
peratu
r
e i
s
ta
ke
n a
s
a
n
im
porta
nt facto
r
affec
t
s
the
s
h
ort-term
load, and
oth
e
r clim
atic fa
ctors a
r
e ge
nerally
ig
nored,
su
ch
a
s
clou
d
cove
r, wind spe
ed and
load, in ord
e
r
to avoid network st
ru
cture bei
ng
to
o cumb
erso
me [12]. Figure 5 is the
daily
con
s
um
ption temperature curve
in
M
a
rch 2012. The t
e
mpe
r
ature o
n
the daily load co
rrelation
is
not obviou
s
, and to simplif
y the artificial neural network model, it ca
n be igno
red.
Figure 5. Test Sample Output Va
lue of the Predi
ction
Erro
r Cu
rve
In this pap
er, we co
nst
r
u
c
t a thre
e la
yer BP neural netwo
rks,
inclu
d
ing in
p
u
t layer,
hidde
n
laye
r and output
layer. We use
the
previo
us
day'
s
2
4
h
histo
r
i
c
al l
o
ad a
s
input
s and
sele
ct March
1 to March 16, 2012, 16
days’ dat
a as sam
p
le
s. The whol
e poi
nt load value
as
training
input
s P, the
whol
e point
of time load
va
lue
s
from M
a
rch
2 to Ma
rch 17
every 24
h a
s
the
training o
u
tp
uts. Enter the numbe
r of input
layer n1 equal
s 2
4
, output layer m equal
s 24.
Acco
rdi
ng to t
he Kolmo
g
o
r
ov theorem, t
he n
u
mbe
r
of
hidde
n laye
r
neuron
s, sco
p
ing, try meth
od
to adjust, set the numbe
r
of hidden lay
e
r ne
uro
n
s n
2
equal
s 31.
Till now
we create a n
e
u
r
al
netwo
rk. Set
the maximum
numb
e
r of it
eration
s
fo
r 2
0
00, the ta
rg
et error fo
r 0.
05.The traini
ng
function i
s
t
r
aindgx L
e
a
r
n
i
ng Algo
rithm
s
which a
r
e
gradi
ent d
e
scent m
o
ment
um metho
d
and
adaptive le
arning rate. Training i
s
com
p
leted
setting
up a te
st th
e vecto
r
P_test a
s
Ma
rch
17
24ho
urs’ hi
st
orical lo
ad
d
a
ta a
s
the
n
e
twork
i
nput
s. The ta
rget
outputs Y a
r
e 24
hou
rs’
lo
a
d
values o
n
M
a
rch 18. Fig
u
r
e 6
sho
w
s the pe
rfor
m
a
n
c
e of BP net
w
ork training.
The num
be
r o
f
iteration
s
epo
ch eq
ual
s 17
06 to achi
eve
the requi
rem
ents of netwo
rk
setting
s.
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 12, Dece
mb
er 201
3: 702
1 – 7026
7026
The test sam
p
le output an
d the actual value of
the error
curve a
r
e
sho
w
n in Fig
u
re 7. It
can
be
se
en
that the maxi
mum e
r
ror i
s
about
0.3,
a
nd
calculated
mean
squa
re erro
r MSE
is
0.1202.
It can
be
seen from th
e an
alysi
s
above, fo
r distri
ct
resi
d
ents’ sh
ort-t
e
rm
l
oad
forecastin
g, a
simple lin
ear regressio
n
a
nalysi
s
meth
od ca
n achie
v
e more a
ccura
cy pre
d
icti
on
results. In ad
dition, linear
reg
r
e
ssi
on a
nalysi
s
meth
od is sim
p
le
r than BP neural net
wo
rk,
and
easi
e
r to impl
ement.
4. Summar
y
This
pap
er
a
nalyze
s
th
e chara
c
te
risti
c
s of re
sid
ential
ele
c
tricity
co
nsum
ption, a
nd u
s
e
s
economi
c
analysis
software SPSS to
filter the factors that affect
people in
residential electri
c
i
t
y
load. The
re
sult sho
w
s th
a
t
the daily maximum
temp
eratu
r
e
an
d weath
e
r con
d
itions
(rainy o
r
sun
n
y)
have
larger
corre
l
ation. Con
s
i
derin
g th
at
the
re
sidentia
l ele
c
tri
c
ity load i
s
rel
a
tively
stable, influ
e
n
cin
g
facto
r
s
and oth
e
r
ch
ara
c
teri
st
ics, we sele
ct
the
linear reg
r
e
s
sion
m
odel a
n
d
BP neural ne
twork mod
e
l to do re
sidenti
a
l electri
c
it
y short-te
rm loa
d
forecastin
g. By comparin
g
the disp
arity betwee
n
the predi
cted
values a
nd
the actu
al values, we find that using
th
e
reg
r
e
ssi
on m
odel
ca
n a
c
hi
eve hig
her p
r
edictio
n a
c
cu
racy
than
u
s
i
ng BP n
e
u
r
al
network m
o
d
e
l.
Therefore,
we can u
s
e
the regressio
n
mod
e
l to
mak
e
short-term res
i
dential electri
c
ity load
forecas
t
ing.
Ackn
o
w
l
e
dg
ements
This work was suppo
rted
by
National
sci
en
ce and
techn
o
logy m
a
jor proje
c
ts
funded
proje
c
ts zx03
006
(2
010
-2
0
10-0
1
).
Referen
ces
[1]
Kang C
h
o
ngq
i
ng, Xi
a Qing,
Liu Mei. El
ectric po
w
e
r s
y
ste
m
load forec
a
sting. Beij
ing.
Chin
a Electric
Po
w
e
r Press. 2
007; 4-5.
[2]
Niu Xia
odo
ng, Gao
Sh
uh
ua, Z
hao Lei.
P
o
w
e
r
lo
ad
fore
c
a
s
t
ing tec
h
n
o
lo
g
y
an
d its
ap
plic
ation. B
e
ij
in
g.
Chin
a Electric
Po
w
e
r Press.1
998; 3-9.
[3]
Z
hang
Re
ngo
ng, W
a
n
g
Pe
ipin
g, W
a
n
g
W
anlia
ng.
B
a
sed o
n
d
a
ta
stored
in s
hort-term lo
ad
optimiz
ation d
e
s
ign.
Infor
m
ati
on an
d Co
ntrol
. 2011; 40
06: 8
25-8
57+
84
6.
[4]
Du Li, Z
h
a
ng
Jian-j
un. N
eur
al net
w
o
rk a
p
p
licatio
ns i
n
po
w
e
r lo
ad forec
a
sting, Sim
u
lat
i
on. 2
011; 2
8
(10): 297-
30
0.
[5]
Yuan Z
h
a
ng
w
e
i. Stando
uts i
ndustr
y an
d ot
her users
of th
e simulati
on of
short-term loa
d
forecastin
g.
Co
mp
uter Si
mulati
on
. 20
05; 22(9): 19
2-2
0
0
.
[6]
Sil, Ari
ndam
K
u
mar D
eb,
Nir
mal K
u
mar; M
a
itra,
Ash
o
k K
u
mar. En
erg
y
savin
g
b
y
ch
o
ppi
ng
off pe
ak
dema
nd usi
ng
da
y li
ght.
Te
l
k
om
ni
ka
. 20
11; 9
(
2): 395-4
00.
[7]
Gu Jie, Z
hang
Chao C
h
e
n
, Z
heng Ji
azhi.
Po
w
e
r s
y
stem
load forec
a
sti
ng bas
ed o
n
fuzz
y
th
eor
y,
Shan
gh
ai Jia
o
tong Un
iversit
y
: 1999; 33(
12) :153
7-15
39.
[8]
Lia
o
Ni
hua
n, Hu Z
h
ih
on
g, Ma Ying
yi
n
g
. Short T
e
rm Load F
o
r
e
casti
ng Meth
ods.
Power system
protectio
n
an
d control
. 20
11;
39(1); 14
7-1
5
2
.
[9]
Z
hang
Xin. P
o
w
e
r s
y
stem
short-term lo
a
d
foreca
sti
ng
intell
ig
ent pr
e
d
ictio
n
meth
o
d
. Cho
n
g
q
in
g
Univers
i
t
y
. 2
0
1
1
; 10.
[10]
Guo Ji
ng, Su
n
W
e
iju
an.
Neu
r
al
n
e
t
w
ork
th
eor
y
and
MAT
L
AB7. Be
iji
ng.
Electro
n
ic In
d
u
str
y
Pr
ess,
200
5; 3.
[11]
T
anoto, Yuask. Ongsakul,
W
eerakor
n Marpa
ung, Ch
a
r
les OP. Levenb
erg-Mar
q
u
a
rdt recurre
nt
net
w
o
rks for lo
ng-term el
ectri
c
it
y
p
eak lo
ad f
o
recasti
ng.
T
e
l
k
omnik
a
.
201
1; 9(2): 257-2
66.
[12]
Kand
i, N W
a
mkeue, R
Saa
d
M, et al.
A
n
efficient
ap
pr
oach f
o
r sh
ort-term l
o
a
d
for
e
castin
g us
ing
artificial
n
eura
l
netw
o
rks
. Industrial
Electronics. IEEE Inter
nat
ional S
y
mposium.
M
ontreal, Quebec,
Can
ada. 2
006;
3: 1928-1
9
3
2
.
Evaluation Warning : The document was created with Spire.PDF for Python.