TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 15, No. 1, July 2015, p
p
. 20 ~ 25
DOI: 10.115
9
1
/telkomni
ka.
v
15i1.807
0
20
Re
cei
v
ed Ma
rch 4, 2
015;
Re
vised
Ma
y 13, 2015; Accepted Ma
y 28
, 2015
NARX Based Short Term Wind Power Forecasting
Model
M. Nandan
a
J
y
othi*, V. Dinakar, N. S S Rav
i
Teja, K. Nanda Kis
h
ore
Dep
a
rtment of Electrical
and
Electron
ics En
gin
eeri
ng, K L Univers
i
t
y
, An
d
h
ra Prad
esh, Indi
a-52
25
02
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: tj
y
o
th
i@kl
uni
versit
y
.
i
n
A
b
st
r
a
ct
Now
adays, w
i
th the grow
in
g nee
ds
of the consu
m
ers ther
e is a hu
ge d
e
man
d
for the
electri
c
pow
er but the f
uel res
e
rves ar
e also
dep
leti
n
g
at t
he sa
me
pace. So, this
has creat
ed th
e nee
d to de
pe
nd
up on the re
ne
w
able en
ergy r
e
sourc
e
s to meet the req
u
ire
d
pow
er de
ma
nd. Since the
pow
er gen
erat
e
d
throug
h ren
e
w
abl
e reso
urces
is eco frien
d
ly
in natur
e an
d distrib
u
ted, this
is
an ad
ded
a
d
vanta
ge. Of all
the ren
e
w
able
energy r
e
sou
r
ces solar
and
w
i
nd pla
ys th
e most crucia
l
part in the p
o
w
e
r gen
erati
o
n
beca
u
se
of the
i
r w
i
de spr
ead
avail
a
b
ility. But
the w
i
nd
ener
gy is vo
latil
e
a
nd i
n
ter
m
ittent
by natur
e, du
e
to
this interco
n
n
e
c
ting the p
o
w
e
r gener
ated to
grid b
e
co
mes
a hectic task. So in this p
a
per a w
i
nd p
o
w
er
forecastin
g mo
del w
i
th the he
l
p
of artificial n
e
u
ral
n
e
tw
orks (ANN) is dev
elo
ped so th
at the w
i
nd pow
er can
be forec
a
sted
w
e
ll in
pro
g
res
s
, w
h
ich hel
ps
in
ma
inta
i
n
i
n
g
and op
eratin
g
grid interco
n
n
e
ction an
d
als
o
sched
uli
ng of u
n
its. T
he dev
el
ope
d mod
e
l is
base
d
on th
e n
on-li
ne
ar auto r
egress
i
ve w
i
th exog
en
ous i
n
p
u
t
(narx) too
l
w
h
i
c
h trains th
e A
NN for the ti
me seri
es
. T
he i
nput p
a
ra
mete
rs taken i
n
to c
onsi
derati
on
ar
e
w
i
nd sp
ee
d, te
mp
eratur
e, pr
e
ssure, a
i
r
dens
i
t
y and
th
e
o
u
tp
ut par
a
m
eter
is
ge
ner
ated
po
w
e
r. T
he re
quir
e
d
data is col
l
ect
ed fro
m
the E
nergy
D
epart
m
ent of KLUn
iv
ersity, Andhr
a
Prades
h w
h
ich
consists of 720
hours
d
a
ta fro
m
t
hat 6
7
2
h
o
u
rs d
a
ta
is us
e
d
for tra
i
ni
ng
a
nd
48
ho
urs
d
a
ta is
us
ed f
o
r
pre
d
ictio
n
. Me
an
squar
e error a
nd root
mea
n
s
quar
e error are
ca
lcul
ated fro
m
the pr
edicte
d
and kn
ow
n results.
Ke
y
w
ords
:
AN
N, narx, netc, hybrid
me
th
od, w
i
nd pow
er forecastin
g
Copy
right
©
2015 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
In the pre
s
e
n
t
day scena
ri
o the ne
ed fo
r ele
c
tri
c
al p
o
we
r is
growi
ng in a
n
exp
onential
manne
r. Th
e
fuel re
se
rves
for the el
ect
r
i
c
al
p
o
wer
ge
neratio
n thro
ugh
conve
n
tional m
e
thod
s are
depletin
g at a very faster rate and a
r
e cau
s
in
g se
vere ha
rmful
effect on the environ
me
nt.
Acco
rdi
ng to the Global Wind Energy O
u
tlook
20
14
by Global Wi
nd Ener
gy Coun
cil (G
WEC)
power secto
r
is the sole e
m
itter of abo
ut 40%
of the ca
rbon di
o
x
ide and 25
% of all the
gree
n
hou
se ga
se
s.
The b
e
tter
solution that
a
ddre
s
se
s mo
st of
the p
r
o
b
lems that a
r
ise
s
be
cau
s
e of the
fossil fu
els is
the usage
of
rene
wa
ble e
n
e
rgy. Am
on
g
all the rene
wable e
nergie
s
wind
ene
rgy
is
the mo
st p
r
o
m
ising
an
d
cheap
er to
op
erate.
GWE
C
in its re
ce
nt
publi
c
ation
st
ated that th
e
wind
energy co
uld
rea
c
h 2
000
GW by 2
0
3
0
. But
the major
hind
ra
nce fo
r the
expan
sion a
n
d
integratio
n of
the win
d
po
wer to the g
r
i
d
is the
high volatile
and
i
n
termittent
n
a
ture of
the wind
power. Be
ca
use
of the
s
e
nature
s
of
wi
nd p
o
wer it i
s
very diffic
u
lt to inte
gr
a
t
e to
th
e g
r
id
an
d
sched
ule th
e
po
we
r. To
overcome
th
e state
d
pr
o
b
lems
wind
power fore
ca
sting i
s
the
very
helpful.
Wind
po
wer forecasting
mod
e
l
help
s
the
po
wer sy
stem
o
perato
r
s in
p
o
we
r
sched
ul
ing,
disp
atch a
nd
maintainin
g the re
se
rve ca
pacitie
s.
Mostly emplo
y
ed method
s for wind p
o
w
er fo
re
ca
sting are Pe
rsistence
i
s
empl
oyed by
makin
g
an a
s
sumption th
at the wind speed an
d wi
nd
power at
a certai
n time in future wi
ll be
s
a
me as
it is
when the forec
a
s
t
is made
[1]. Let
the
wi
nd p
o
wer
and
win
d
spe
ed
at t are P(t) a
n
d
v(t), then the
wind
power
and
wind
sp
eed at t +
∆
t
can
be form
ulated a
s
[2]. This m
e
thod
is
more
accu
rat
e
than oth
e
r forecastin
g
method
s in
ca
se of ultra
-
sh
ort-t
e
rm f
o
re
ca
sting. T
h
e
accuracy of t
h
is method
will decreas
e
rapidly with the increase
of ti
me-scal
e
of
forecasting. [3].
Phys
ic
al
met
hod
uses th
e la
ws on
whi
c
h th
e at
mosp
he
ric b
ehavior de
p
end
s u
pon,
for
estimating th
e wind flo
w
arou
nd the
wind turbi
n
e
s
and the
wind
powe
r
corre
s
po
ndin
g
to the
wind flo
w
obt
ained by the
estimation,
can be
kno
w
n
by the turbine
cha
r
acte
ri
stics [4].
Statis
tic
a
l
method
uses
a mo
del
whi
c
h give
s a
rela
tion bet
wee
n
meteorologi
cal pa
ram
e
ters and
the
po
wer
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
NARX Based
Short Term
Wind Po
we
r Fore
ca
sting
Model (M. Na
ndan
a Jyothi)
21
gene
rated. T
h
is mod
e
l is
develop
ed by studying t
he
histori
c
al d
a
ta. From this
model wi
nd p
o
we
r
can b
e
pre
d
icted [4].
Hy
brid
method is t
he com
b
inati
on of physi
cal
and statisti
ca
l methods.
The p
r
e
d
icati
on
carried
ou
t in this
pap
e
r
is
by u
s
in
g
the hybri
d
m
e
thod. Statisti
cal
data
is coll
ecte
d from the Energy D
epa
rtme
nt of KL University, Andh
ra Prade
sh
which
con
s
i
s
ts of
720 ho
urs d
a
ta from that
672 ho
urs d
a
ta is u
s
ed f
o
r trai
ning a
nd 48 h
o
u
r
s
data is u
s
e
d
for
predi
ction a
n
d
also the p
h
ysical laws are tak
en i
n
to accou
n
t for the cal
c
u
l
ation of po
wer
gene
ration. T
he develo
ped
model is
ba
sed o
n
t
he n
on-lin
ea
r aut
o reg
r
e
ssive
with exoge
no
us
input (na
r
x) tool whi
c
h tra
i
ns the ANN for t
he time serie
s
. The
input parame
t
ers ta
ken int
o
con
s
id
eratio
n
are wi
nd
sp
eed, tempe
r
a
t
ure, pre
s
su
re, air den
sity and the out
put paramete
r
is
gene
rated
p
o
we
r. Mean
squ
a
re
error and root m
ean squa
re error
a
r
e cal
c
ulate
d
from
the
predi
cted a
n
d
kno
w
n re
sult
s.
2. Res
earc
h
Method
2.1. Calculati
ons from th
e
Data
The
wind
po
wer ge
nerate
d
by the tu
rbi
nes
dep
end
s upon
the fa
ctors li
ke
win
d
spe
ed,
ambient tem
peratu
r
e,
win
d
pre
s
su
re,
air de
ns
ity. Among all th
e facto
r
s
win
d
sp
eed
and
air
den
sity dominates the p
o
w
er g
ene
rate
d.
Wind p
o
wer g
enerated is
known by:
(1)
Whe
r
e P: Wi
nd po
wer g
e
n
e
rated
ρ
: Air
den
sity at the given temperature
A: Area swept by the turbine bl
ad
es
V
:
Wind s
p
eed
Wind p
o
wer
gene
rated i
s
highly affecte
d
by
the air d
ensity and th
e wind
spe
e
d
,
as the
area
swept b
y
the turbines blade
s rem
a
i
n
con
s
tant for a taken turbine [5].
The data i
s
collecte
d
from
Energy Depa
rtm
ent of K L University, vadde
swaram
area fo
r
a time
spa
n
of on
e mo
n
t
h whi
c
h
co
mpri
se
s of
wind
speed,
ambie
n
t te
mperature,
a
nd ai
r
pre
s
sure. Th
e air d
e
n
s
ity in the co
nsi
dere
d
area i
s
not kno
w
n.
For the d
e
n
s
ity cal
c
ulati
ons
vapour p
r
e
ssure is
req
u
ire
d
according t
o
the formula
[5].
∗
.
∗
.
(2)
Whe
r
e
ρ
: Air den
sity at the given temperature
D: Air
den
sity at absolute temp
erature
T:
Giv
en temperature
B
:
Barome
tr
ic
(
a
tmos
p
h
e
r
ic
)
pr
ess
u
re
e: Vapour p
r
e
s
sure
of the air at the given temp
eratu
r
e
All the data required for th
e density cal
c
ulation
is pre
s
ent exce
pt t
he vapou
r pressure. In
a clo
s
ed
syst
em the pressure exe
r
ted
by a v
apour
in thermo
dyn
a
mic e
quilib
ri
um at a give
n
temperature
i
s
the
vapo
ur
pre
s
sure. Va
pour p
r
e
s
sure is calculate
d
u
s
ing
the
Clausi
s
-cla
peu
ron
relation i.e.
ln
(3)
Whe
r
e P
1
, P
2
: The vapou
r pre
s
sures at tempe
r
atures
T1, T2 re
spe
c
tively
H
va
p
: Enthalpy of vapou
rization
of liquid
R: Real gas
c
o
ns
tant (8.314J
)
T
1
: Temperature at
whi
c
h the vap
our p
r
e
s
sure i
s
kn
own
T
2
: Temperature at
whi
c
h the vap
our p
r
e
s
sure
to be cal
c
ulat
ed
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 15, No. 1, July 201
5 : 20 – 25
22
Take
T1
and
P1 at STP condition
s a
n
d
P2 is th
e va
pour pressu
re to be
cal
c
u
l
ated at
whi
c
h the
te
mperature
is T2. By this
formula
vapo
ur P2 i.e, ‘
e
’
in the
den
s
i
t
y calculation
is
obtaine
d. Air density is t
hen calculat
ed by t
he st
ated formul
a
.
The data required for t
h
e
cal
c
ulatio
n of
power g
e
n
r
a
t
ion is o
b
tain
ed an
d
the p
o
we
r that
ca
n be g
e
ne
rat
ed by u
s
ing
all
these p
a
ra
me
ters i
s
cal
c
ula
t
ed by ignorin
g the operatio
nal losee
s
of turbine.
2.2. Artificial
Neural Net
w
ork (ANN)
Artificial ne
u
r
al net
wo
rks are th
e ne
ural n
e
two
r
ks de
rived from the in
spi
r
ation of
biologi
cal ne
ural net
works (animal cent
ral nervo
us
system). The
s
e artificial ne
ural net
w
orks are
use
d
to ta
ke
logical de
ci
si
ons
ba
sed
o
n
the in
puts.
ANN
ca
n de
a
l
with n
on-li
n
e
ar
and
co
m
p
lex
probl
em
s in
terms of cl
assificatio
n
o
r
fore
ca
sting
by extractin
g
the dep
en
den
ce bet
w
e
en
variable
s
thro
ugh the traini
ng pro
c
e
s
s. So the ANN based metho
d
is an app
ro
priate metho
d
to
apply to the probl
em of forecastin
g win
d
power be
ca
use it is di
re
ctly proportio
n
a
l to wind sp
eed
whi
c
h i
s
hig
h
ly intermitte
nt in nature. Among
the
available m
e
thod
s u
s
ing
artificial n
e
u
ra
l
netwo
rks the
NARX, a dyn
a
mic recu
rren
t method, is u
s
ed to solve the time se
rie
s
pro
b
lem.
2.2.1. ANN T
r
aining
One
of the key element
s
of neura
l
net
works i
s
their ability to
learn.
A neural
net
work is a
compl
e
x ada
ptive system,
which mea
n
s
it can
cha
n
ge its internal
stru
cture
ba
sed o
n
the in
puts
and targets.
These ANNs need to b
e
trained fo
r doi
n
g
a pa
rticul
ar
task. Th
ere a
r
e three type
s of
training p
a
rad
i
gms to train t
he artifici
al n
eural n
e
two
r
k and are a
s
fo
llows:
a) Su
perv
is
ed training:
It
is th
e
pro
c
e
s
s of
provi
d
ing th
e n
e
t
w
ork with
a
seri
es o
f
sampl
e
in
put
s a
nd
com
p
a
r
ing th
e o
u
tp
ut with
the
e
x
pected
re
sp
onse. The
training
co
ntin
ues
until the network is a
b
le to provide th
e expec
te
d resp
on
se. Th
e prop
osed
work is
sup
e
r
vised
training
with back propa
ga
tion techni
qu
e.
b) Unsu
perv
ised
tr
aining
:
In thi
s
m
e
thod
of traini
n
g
, the in
put
vector
and
th
e targ
et
output is
not
kno
w
n. Th
e n
e
twork m
a
y modify in su
ch a way that the mo
st simil
a
r inp
u
t vecto
r
i
s
assign
ed to the sam
e
outp
u
t unit.
c)
Rein
for
c
e
m
ent trainin
g
:
It is the proce
s
s of train
i
ng th
e netwo
r
k in the p
r
e
s
ence of a
teach
e
r
but i
n
the ab
se
nce of targ
et vector.
T
he te
ach
e
r
gives
only the an
swer whethe
r
it is
cor
r
e
c
t (1
) or
wro
ng (
0
).
2.2.2. Non Linear Auto
Regres
s
iv
e
w
i
th Exogen
o
u
s
Input (NARX)
The nonli
nea
r autoregressive network
with
exogen
o
u
s inp
u
ts (NARX) is a re
curre
n
t
dynamic network, with
feed
back con
n
e
c
tions en
clos
i
n
g
several layers of the n
e
twork. T
he
NARX
model is b
a
sed on the line
a
r ARX mod
e
l
, which i
s
co
mmonly used
in time-se
r
ie
s modeli
ng.
The definin
g equatio
n for the NARX mo
del is:
1
,
2
,.
...,
,
1
,
2
,....,
(4)
a) Series par
a
llel architec
ture
Used
whe
n
the outp
u
t of
the
NA
RX n
e
twork i
s
co
nsid
ere
d
to
be an
estim
a
te of the
output of
so
me no
nlinea
r dynami
c
sy
stem. The
o
u
tput is fe
d
back to th
e i
nput of the f
eed
forwa
r
d n
e
u
r
al netwo
rk as part of the standard NA
RX architectu
re. Becau
s
e t
he true
outpu
t
is
available
du
ring the t
r
aini
n
g
of the
net
work, you
co
uld
cre
a
te a
seri
es-p
arall
e
l archite
c
ture
, in
whi
c
h the true output is used in
stea
d of feeding
back the estimated out
put. This ha
s two
advantag
es
whi
c
h a
r
e th
e first is that
the input
to
the feed forward n
e
two
r
k i
s
mo
re a
c
curate.
The seco
nd i
s
that the re
sulting network ha
s a pu
rel
y
feed forwa
r
d architectu
re
, and static b
a
ck
prop
agatio
n can be u
s
ed fo
r trainin
g
.
Figure 1. Seri
es Parallel Archite
c
ture
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
NARX Based
Short Term
Wind Po
we
r Fore
ca
sting
Model (M. Na
ndan
a Jyothi)
23
b) Parallel architecture
Later thi
s
archite
c
ture i
s
conve
r
ted int
o
parallel archite
c
ture fo
r the predi
c
ti
on. The
predi
ction
of
the next valu
e dep
end
s o
n
the inp
u
ts
and p
r
eviou
s
outputs to the net
work.
The
depe
nden
ce
on the p
r
evio
us outp
u
t ca
n
be adju
s
ted
by using
dela
ys, input dela
y
s and fee
d
b
a
ck
delays.
Figure 2. Parl
lel Archite
c
ture
3.
Resul
t
s and
Error Analy
s
is
3.1. Performance o
f
the
ANN
The coll
ecte
d
data is given as input
s whi
c
h are temperature
s
, pre
s
sure, air density,
spe
ed a
nd
calcul
ated p
o
w
er ge
neration i
s
given
a
s
o
u
tput to t
h
e
NARX too
l
box for trai
ning.
After training,
the neural n
e
twork is rea
d
y for the pr
e
d
iction. The l
a
st 2 days da
ta is given to
the
neural n
e
two
r
k
and th
e p
r
edicte
d
outp
u
t
is obtai
ned.
The
pre
d
icte
d output i
s
compa
r
ed to
the
cal
c
ulate
d
po
wer a
nd the
perfo
rman
ce
is monito
r
ed
by calculating
the erro
rs by
various m
e
a
n
s
of erro
r calcul
ations.
Figure 3. Plot betwee
n
temperatu
r
e
s
, pressu
re, air de
nsity, spee
d vs time (ho
u
rs)
a) Mea
n
Erro
r (ME):
It is the ba
sic type
of erro
r cal
c
u
l
ation. It is th
e averag
e of the
er
rors
.
∑
(5)
b) Mean Squ
a
re Error (M
SE):
It is one of the basi
c
types of error
cal
c
ulatio
n. It
is the
averag
e of the squ
a
re
s of the errors.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 15, No. 1, July 201
5 : 20 – 25
24
∑
(6)
c) Root M
e
an Square
Error (RMS
E):
RMSE is the sta
n
d
a
rd d
e
viatio
n of the
differen
c
e
s
b
e
twee
n predi
cted valu
es a
nd a
c
tual va
l
ues. It is the
squ
a
re
ro
ot o
f
the averag
e
of
squ
a
re
s of the errors.
∑
(7)
Whe
r
e N:
No.
of sample
s
T: Ac
t
ual Output
P:
Predic
t
ed Output
Predi
ction is
carrie
d out by varying the del
ays of the input and
also the nu
m
ber of
neuron
s in
th
e hid
den
lay
e
r. Th
e e
r
rors at
differe
nt
delays an
d di
fferent n
u
mb
er
of ne
uro
n
s in
the hidde
n la
yer are
Table 1. The
Perform
a
n
c
e
of the predi
cted ANN m
o
d
e
l
ANN DELA
Y
ME
MSE
RMSE
4-3-1
4-3-1
4-3-1
4-3-1
2
4
6
8
0.17734
0.54563
0.323178
0.0919
4.942264
23.30557
8.445927
4.925418
1.771711
4.339749
2.561335
1.362869
4-5-1
4-5-1
4-5-1
4-5-1
2
4
6
8
0.29408
0.03643
0.03457
0.16524
11.84562
4.728486
4.2436
5.371358
2.808699
1.436289
1.7854
1.867995
3.2. Perform
a
nce Plots o
f
the
ANN
The traine
d n
eural n
e
two
r
k is employed
for the predi
ct
ion of the power g
ene
rate
d for the
last 48
days
by giving the
input pa
ram
e
ters
by
varyin
g the inp
u
t time delay
s an
d also chan
ging
the numbe
r
of neurons i
n
the hidde
n
layer. In
ca
se of 3 ne
urons in the h
i
dden laye
r the
minimum e
rro
r is attaine
d
whe
n
the time delay given
is 8 and in
case of 5 n
eurons in the hi
d
den
layer the mini
mum error a
r
e attained wh
en
the delay
given is 6 are
sho
w
n bel
ow.
Figure 4. The
ANN predi
cted output
plot
when 3 hi
dd
en layer ne
urons a
nd time delay of 8
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
NARX Based
Short Term
Wind Po
we
r Fore
ca
sting
Model (M. Na
ndan
a Jyothi)
25
Figure 5. The
ANN predi
cted output
plot
when 5 hi
dd
en layer ne
urons
with time delay of 6
4. Conclu
sion
To ad
dre
s
s th
is p
r
obl
em a
wind
po
wer fore
ca
sting m
odel
with the
help of
artifici
al neu
ra
l
netwo
rks (ANN) is develo
p
ed
so th
at th
e wi
nd
po
wer ca
n b
e
fo
re
casted
well in
advan
ce,
whi
c
h
help
s
in
main
taining
grid
in
terco
nne
ction
and
al
so
sch
edulin
g of u
n
i
t
s. The
devel
oped
mod
e
l i
s
based on the
non-lin
ea
r a
u
to reg
r
e
ssiv
e
with exoge
nou
s input (n
arx) tool which train
s
the ANN
for the time
seri
es.
Wind
powe
r
g
ene
ration
s
de
pe
nds on
the para
m
eters
li
ke wind spe
ed,
temperature,
pre
s
sure, air
den
sity. So these pa
ram
e
ters
are
give
n as input to
the ANN mo
del
develop
ed
an
d after trai
nin
g
of th
e m
o
d
e
l wi
nd
po
we
r i
s
p
r
e
d
icte
d
.
Predi
ction
i
s
ca
rrie
d
out
b
y
varying the n
u
mbe
r
of neu
ron
s
in the hi
dden laye
r
an
d also the d
e
l
a
y given to the netwo
rk. F
r
o
m
the results, it came to kno
w
that if the
delay
is incre
a
se
d the erro
r is going to
be red
u
ced a
n
d
also if the n
u
mbe
r
of hid
den laye
r ne
uron
s in
crea
se
s the com
putational
p
o
w
er
of the m
odel
increa
se
s whi
c
h al
so re
du
ces the e
rro
rs i
n
the predi
cte
d
output.
Ackn
o
w
l
e
dg
ements
Authors thank to Mr. R. B.
R. Prakash, EEE Departm
ent, KL
University for providing the
wind
data
u
s
ed fo
r the
m
odelin
g of A
N
N mod
e
l a
n
d
al
so th
e m
anag
ement
o
f
KLUnive
r
sit
y
for
their su
ppo
rt for this
work.
Referen
ces
[1]
Z
hao
X, W
a
ng
SX,
Li T
.
Revi
e
w
of
eva
l
uati
on crit
eria
an
d
main
meth
ods
of
w
i
nd
po
w
e
r
forecasti
n
g
.
Energy Proc
ed
ia
. 201
1; 12: 7
61-7
69.
[2]
W
en-Yea
u
Ch
ang. Short-T
e
rm W
i
nd Po
w
e
r F
o
recasti
ng Usin
g the
Enhanc
ed P
a
rticle S
w
ar
m
Optimizatio
n
Based H
y
b
r
id M
e
thod.
En
ergi
e
s
. 2013; 6: 487
9-48
96.
[3]
Catal
ão JPS,
Osório GJ. Pousi
nho
HMI.
Short-T
e
rm
W
i
nd Pow
e
r
F
o
recastin
g U
s
ing a
Hybri
d
Evoluti
onary
In
tellig
ent A
ppr
o
a
ch
. In Pr
oce
e
d
in
gs of th
e
1
6
th Intern
atio
n
a
l C
onfer
enc
e
on Inte
lli
ge
nt
S
y
stem Ap
plic
ation to Po
w
e
r
S
y
stems (I
SAP
). Hersoniss
os, Greece. 201
1: 1-5.
[4]
Pan Z
hao, Jia
ngfen
g W
ang,
Junron
g Xia,
Ying
xi
n Shen
g, Jie Yue. Pe
rformanc
e ev
aluti
on an
d
accurac
y
e
nha
nceme
n
t of a
da
y-a
hea
d
w
i
n
d
po
w
e
r forec
a
sting i
n
chin
a
.
Renew
abl
e Energy.
2
012;
43: 234-
24
1.
[5]
M Carol
i
n Ma
b
e
l, E F
e
rna
nde
z. Anal
ysis
of w
i
nd
po
w
e
r g
e
nerati
on a
nd
p
r
edicti
on us
ing
ANN: A cas
e
study
.
R
enew
a
b
le En
ergy
. 20
08
;
3
3
: 986-
99
2.
Evaluation Warning : The document was created with Spire.PDF for Python.