TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 12, No. 8, August 201
4, pp. 6423 ~ 6430
DOI: 10.115
9
1
/telkomni
ka.
v
12i8.619
5
6423
Re
cei
v
ed Ap
ril 1, 2014; Re
vised J
une 3,
2014; Accept
ed Ju
ne 15, 2
014
Downsc
aling Modeling Using Support Vector
Regression for Rainfall Prediction
Sanusi*
1
, Ag
us Buon
o
2
, Imas S Sitanggang
3
, Akh
m
ad Faqih
4
1,2,
3
Department
of Computer S
c
ienc
e, F
a
cult
y of Mathematic
s and Natur
a
l
Scienc
es,
Bogor Agr
i
cult
ural U
n
ivers
i
t
y
,
1668
0 Bog
o
r,
Indon
esi
a
, Ph/F
ax. +
62-2
51-6
284
48/6
229
61
4
Departme
n
t of Geoph
ysics a
nd Meteor
ol
og
y, F
a
cult
y
of Mathematics a
n
d
Natural Sci
enc
es,
Bogor Agr
i
cult
ural U
n
ivers
i
t
y
,
1668
0 Bog
o
r,
Indon
esi
a
, Ph/F
ax. +
62-2
51-6
284
48/6
229
61
Corresp
on
din
g
author, e-mai
l
: sanusi
u
marh
a
s
an@
gmai
l.co
m
*1
, pudesha
@
y
ah
oo.co.id
2
,
imas.sitan
gga
n
g
@gma
il.com
3
, akhmadfa
@
ip
b.ac.id
4
A
b
st
r
a
ct
Statistical d
o
w
n
scali
ng
is a
n
effort to link
gl
oba
l
scale to local scale var
i
able. It uses GCM m
o
del
which usually
used
as a prime instrument in lear
ning
system
of various clim
ate. The pur
pose
of this st
udy
is as a SD
mo
del by
usin
g S
V
R in or
der to
pred
ict t
he rai
n
fall in
dry seas
on; a case stu
d
y at Indra
m
ay
u.
T
h
roug
h the mode
l of SD, SVR is created w
i
th line
a
r
kerne
l
and RBF
kern
el. T
he results
show
ed that th
e
GCM mo
dels
can b
e
us
ed t
o
pre
d
ict rai
n
fall i
n
the
dry
seaso
n
. T
he b
e
st SVR mod
e
l is
obtai
ne
d
at
Ciked
un
g r
a
in
station
i
n
a
li
near
kern
el
fu
nction
w
i
th co
rrelatio
n
0.74
4
an
d RMSE
2
3
.937, w
h
ile
th
e
mi
ni
mu
m
pred
i
c
tion resu
lt is g
a
in
ed at Ci
de
mpet rain
stati
o
n
w
i
th correlatio
n
0.40
1 an
d R
M
SE 36.96
4. T
h
is
accuracy
is still not high, the
select
i
on of p
a
r
ameter va
lues
for each k
e
rn
el functi
on n
e
e
d
to be
do
ne w
i
th
other opti
m
i
z
a
t
i
on techn
i
q
ues.
Ke
y
w
ords
:
st
atistical
dow
ns
calin
g, g
e
n
e
ral
circul
asi
mod
e
ls, su
pport v
e
ctor re
gressi
o
n
, rai
n
fal
l
i
n
dry
seaso
n
Copy
right
©
2014 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 some recent years a
g
o
, many efforts
have al
re
ady done to
explore the
effect of
climate vari
ety whethe
r in
a big scal
e o
r
clim
at
e ch
a
nge toward the varia
b
ility of rainfall in t
he
worl
dwi
de [1]
.
The clim
ate
variety esp
e
c
ially rai
n
fall
in Indon
esi
a
mostly influe
nce
d
by glo
b
a
l
phen
omen
on
su
ch a
s
El
-Nin
o an
d Southern O
sci
llation (ENS
O), ENSO i
s
conventio
n
a
lly
identified a
s
oce
an temp
e
r
ature wa
rmi
ng in ea
st
ern
Pacific [2]. Indian O
c
ea
n Dipol
e (IO
D
),
IOD
as a mo
du
s
of tropical p
h
ysic in In
di
an Ocean i
s
stron
g
ly beli
e
ved as
a m
a
in effect wh
ic
h
cau
s
e
s
dryn
ess in Indo
nesi
a
[3]. Madde
n Juli
a
n
Oscillation
(MJO
), MJO as a glo
bal
phen
omen
on
influen
ce
s the climate in western of
Indo
nesi
a
[4]. This phe
nome
n
o
n
also ha
ppe
ns
in Indramayu.
It is one
of Indone
sia distri
ct
whi
c
h ha
s monsoon rai
n
and
a
s
a
cen
t
ral
produ
ctio
n
of agri
c
ulture
particula
rly ri
ce [5]. The m
a
in fa
cto
r
s
ca
use
crop failu
res i
n
Indram
ayu are d
r
yne
ss
(79.8%), pe
st
attack (15.6
%
) and float (5.6%) [6].
One of inst
ruments
whi
c
h
can be used
to obser
ve the indication of climate vari
ability is
Gene
ral
Circulation Mo
de
[7]. It can be kn
own that
GCM
ha
s a
n
inten
s
e rel
a
tionship b
e
twee
n
big scal
e cli
m
ate an
d wh
ether
on lo
ca
l scale fo
r
rainfall predic
tion [8], [9]. Si
mulated
rainf
a
ll
pattern from
the variou
s model
s of GCM is able
t
o
give basi
c
informatio
n that neede
d to the
future develo
p
ment [10]. Ho
wever, G
C
M data is
co
nsid
ere
d
to the low of re
solution and gl
obal
scale which d
i
fficult to be used in doi
ng
predi
ction b
e
c
au
se lo
cal cl
imate need
s
high re
sol
u
tio
n
,
but GCM i
s
st
ill can be used if it
mixed to the downscaling technique.
Many model
s that al
rea
d
y use
d
to
pre
d
i
ct
clim
ate in
GCM
and
SD su
ch
as Buo
no
et a
l
(201
0) [11]
statistical d
o
w
n
s
caling
m
odelin
g u
s
in
g Artificial
Neu
r
al
Net
w
orks (ANN)
for
predi
ction
mo
nthly rainfall
i
n
Ind
r
amayu.
In ad
dition, Wige
na (20
0
6
)
[1
2]
stati
s
tical do
wn
scali
ng
model with
Regre
s
sion Projectio
n
Persuit (PPR) to
fore
ca
st the rainfall (month
l
y rainfall ca
se in
Indram
ayu). This study uses
Su
ppo
rt V
e
ctor Regres
sion on do
wn
scaling mod
e
l
to
predict
th
e
rainfall in dry season.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 8, August 2014: 642
3 –
6430
6424
Statistical d
o
wn
scaling i
s
define
d
a
s
tran
sfe
r
functio
n
that
descri
b
e
s
functio
nal
relation
shi
p
of global atmosp
he
ric
circul
ation
with
local clim
ate element
s [13]. Figure
1 i
s
process illustration of dow
nscaling statist
i
cal.
Y
,
f
X
,,,
(1)
Where,
Y = local
clim
ate variable
X = GCM out
put variable
t =
time period
p = many of Y variable
q = many of X variable
s = ma
ny of atmosph
e
re la
yer
g = GCM do
main
Figure 1. Statistical
Do
wn
scalin
g Illustrat
i
on
1.2.
Support Vector Regr
ession
Suppo
rt Vector Re
gressio
n
(SVR) i
s
the
expan
sion
of Support Vector M
a
chin
e (SVM).
SVM use
d
to
solve
cla
r
ification p
r
obl
em
, while SV
R
use
d
to
reg
r
e
ssi
on
ca
se. S
V
R is a m
e
th
od
that can overcome overfitting, so that
it will result better perform
a
nce [14].
Suppo
se we
have a set of data as much a
s
ℓ
set training d
a
ta in a formula:
χ
x
i
,
y
i
w
i
t
h
i
1
,
…
,
ℓ
, by x in
put data = {x
1
, x
2
, x
3
, .
..,n}
⊆
N
a
nd the co
rrespon
ding outp
u
t as
y
y
,…,y
ℓ
⊆
. When
ε
value is equal as
0, we will get
a perfe
ct regression. Supp
ose we
have a functi
on as reg
r
e
s
sion line belo
w
:
f
x
w
∙
ϕ
x
b
(
2
)
By
ϕ
x
sho
w
s a
point in fe
ature
sp
ace F t
he ma
pping
result of x in
i
nput spa
c
e.
Coeffici
ent of
w and b a
r
e e
s
timated by
minimizi
ng th
e risk fun
c
ti
on that describ
es in the follo
wing formulat
ion:
min
∥w
∥
C
ℓ
∑
L
∈
y
,
f
x
ℓ
(
3
)
Dep
end
s on
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Do
wn
scaling
Modelin
g Usi
ng Suppo
rt Vector
Reg
r
e
s
sion for
Rainf
a
ll Predi
ction
(Sanu
si)
6425
y
w
ϕ
x
b
ε
w
ϕ
x
by
ε
,i
1
,2
,3
,…,
ℓ
With,
L
y
,
f
x
|
y
f
x
|
ε
,
|
y
f
x
|
ε
0
,t
o
the
o
thers
By minimizing
∥w
∥
will make t
he function as thin as
possible, as a
result the
capacity
function can
be cont
rolle
d.
ε
-inse
n
sitive loss function required to minimize n
o
rm from w a
c
hiev
e
better ge
nera
lization to re
gre
ssi
on fun
c
tion f(x). Tha
t
is why we
have to solv
e the followi
ng
probl
em:
min
∥w
∥
(4)
Dep
end
s on:
y
w
ϕ
x
b
ε
w
ϕ
x
by
ε
,i
1
,2
,3
,…,
ℓ
Assu
me the f
unctio
n
of f(x) whi
c
h
ca
n
approximate
to all of these point
s
x
,y
. Then,
we will get a cylinder as describe in Figure 2.
Figure 2. Reg
r
essio
n
Fun
c
t
i
on at SVR [15]
A
ccu
ra
cy
of
ε
in this case we assum
e
that all points in the range f
ε
(feasible). In the
case of i
neligi
b
ility, where there are
som
e
point
s that
may be out of
range f
ε
, we
need to
ad
d
variable of sl
ack
ξ
,
ξ
∗
. Furthe
rmore, the opti
m
ization p
r
o
b
l
em can u
s
e t
he followi
ng formul
a:
min
‖
w
‖
C
∑
ξ
ξ
∗
ℓ
ℓ
(5)
Dep
end
s on:
y
w
ϕ
x
ξ
b
ε
,i
1
,2
,3
,…,
ℓ
w
ϕ
x
y
ξ
∗
b
ε
,i
1
,2
,3
,…,
ℓ
ξ
,
ξ
∗
0
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 8, August 2014: 642
3 –
6430
6426
The co
nsta
nt of C > 0 dete
r
mine
d the b
a
r
gai
ning b
e
twee
n the thin
ness of functi
on f and
the uppe
r li
mit of deviation that more than
ε
was
still tolerated.
ε
was
com
para
b
le t
o
t
h
e
accuracy of the app
roxima
tion of the
tra
i
ning data. The highe
st value of
ε
was re
lated to
ξ
∗
that
has
small a
n
d
low ap
proximation a
c
curacy. The hig
h
e
st value for
variable
ξ
∗
will make
empi
rical
errors which
have a
con
s
i
dera
b
le influ
e
n
ce
on t
he
re
gulari
z
atio
n factor. In SV
R sup
port ve
ctor
there was the
training data
whi
c
h lo
cated
out of f from
the deci
s
io
n functio
n
.
By C was
determi
ned b
y
user,
Kx
,x
was dot-pro
du
ct kernel
that identified a
s
Kx
,x
ϕ
x
ϕ
x
, by using Lagrang
e multipliers and
optimalizati
on con
d
ition,
The
reg
r
e
ssi
on fu
nction
wa
s formulate
d
expl
icitely in the followin
g
form
ula:
f
x
∑
α
α
∗
ℓ
K
x,
x
b
(
6
)
Before d
o
ing
training
and t
e
st of SVR, it
is bette
r for
us to d
e
ci
de
para
m
eter value of C,
ε
to the function of Linea
r Kernel a
nd C
para
m
eter,
ε
, and
γ
to RBF
k
e
rnel func
tion.
2. Res
earc
h
Method
This stu
d
y was
und
erta
ke
n in
seve
ral
pha
se
s. All o
f
those
ph
ases
ca
n b
e
se
en in
th
e
followin
g
figure Figure 3.
Figure 3. Re
search Flo
w
ch
art
The begi
nnin
g
of this stu
d
y was litera
t
ure
revie
w
; it used in ord
e
r to unde
rstand all
probl
em
s tha
t
will be
re
se
arched. T
he
data u
s
ed i
n
this research
is
se
cond
ary
data divide
d
to
GCM hin
d
ca
st data re
sult
(used a
s
cla
r
ify vari
able)
and data of rainfall ob
serv
ation (u
sed a
s
respon
d vari
able).
Re
sult
of GCM
hin
d
ca
st d
a
ta
wa
s a
c
qui
red
from the
Cli
m
ate Inform
a
t
ion
Tool Kit (CLI
K) APEC
Climate C
enter
(APCC) as
the rainfall dat
a and type of
ASCII file whic
h
con
s
i
s
ts
of 6
mod
e
ls with
a
re
solutio
n
grid
of latitud
e
an
d lo
ngitu
de 2.5
0
x2.5
0
,
data
acce
ssed
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Do
wn
scaling
Modelin
g Usi
ng Suppo
rt Vector
Reg
r
e
s
sion for
Rainf
a
ll Predi
ction
(Sanu
si)
6427
from the webs
ite CLIK APCC
(http://c
lik.apcc
21.or
g), as
well as
two models
of
GCM hindc
a
s
t
rainfall
obtain
ed fro
m
the
web
s
ite
of th
e Intern
at
ion
a
l Research
Institute Data
Library
(IRI
DL)
(http://iridl.ldeo.c
o
lumbia.edu), as
data
of Climate P
r
edic
t
ion C
ent
er(CPC) Unifi
ed Gauge-Bas
ed
Analysis of G
l
obal Daily Precipitatio
n
fro
m
The
Intern
ational Re
se
arch Institute for Climate a
nd
Society (IRI)
and TSV file type with a gri
d
resolution o
f
latitude and longitud
e
0.5
0
x0.5
0
. Hindc
as
t
GCM data used
to build predictio
n
mo
d
e
l
in 3
diffe
re
nt month
s
: M
a
y, Jun
e
, an
d
July
(M
JJ) from
the year of 1982-200
8 (27
years) ev
ery
model at every rainfall stat
i
on. In this study, there are 8
GC
M hindcast
rainfalls to b
u
ild pre
d
ictio
n
model a
s
shown in Tabl
e 1.
The d
a
ta of
rainfall
ob
servation (re
s
po
nd va
ri
able
)
i
s
the
averag
e value
of seasona
l
rainfall
at ev
ery rainfall
station in In
d
r
amayu
by l
ongitudi
nal p
o
sition
of10
7
o
52
’
-10
8
o
36
’
BT
and 6
o
15
’
-6
o
40
’
LS, it wa
s obtaine
d from the m
e
asu
r
em
ent a
nd te
st that perfo
rme
d
by
Meteorology Dep
a
rtme
nt in Indramayu.
There we
re
15 observatio
n
stations u
s
ed as sho
w
n
in
Table 2. The
data of
rainfall obse
r
vatio
n
wa
s use
d
3 months: M
a
y, June, Jul
y
(MJJ) from
the
year of 1982
-2008 (27 yea
r
s) at every rainfall station.
Data of GCM
was
cro
ppe
d
in grid of 7x7 and
then m
a
ke all of GCM data mode
l to the
line ve
ctor;
Next, averag
e rainfall
of
data G
C
M a
nd observati
ons to be the annual rai
n
fall.
Furthe
rmo
r
e, distrib
u
te trai
ning and te
st data by us
ing
9-fold cross
Validation, 9 is divided du
e
to
the numbe
r o
f
year and re
done in nine t
i
mes. T
he dat
a PCA is necessary to be done be
ca
use it
can avoi
d the
double lin
ea
r data in GCM
model an
d to save
comp
uting time during trainin
g
a
n
d
testing the SVR model. Redu
ction
process is held by
taking one o
r
more majo
r compon
ents
with
diver
s
ity of
≥
98%. Finally the SVR traini
ng and te
stin
g can b
e
don
e.
Tabel 1. The
Data of GCM Hind
ce
st Rai
n
fall and its F
ound
ers
No
Model
Name
Ensemble Institution
Sources
References
1 GCPS
T6
3T21
4
Korea
http://clik.apcc21
.
org
[16]
2 GDAPS
T1
06L2
1
20
Korea
http://clik.apcc21
.
org
[16]
3
CMC1-C
anCM3
120
Columbia
http://iridl.ldeo.columbia.edu
[17], [19]
4 CanCM3-A
GCM
3
10
Canada
http://clik.apcc21
.
org
[16]
5
GFD
L
-CM2P
1
120
Columbia
http://iridl.ldeo.columbia.edu
[17], [19]
6 NASA-GS
FC
L3
4
8
U.S.A
http://clik.apcc21
.
org
[16]
7
METRI AGCM L
17
10
Korea
http://clik.apcc21
.
org
[16]
8 PNU
5
Korea
http://clik.apcc21
.
org
[16]
Tabel 2. The
Name a
nd Lo
cation of the
15 Ra
i
n
fall O
b
se
rvation Stations in Ind
r
amayu
Y
Station
Name
LS
BT
Y
Station
Name
LS
BT
Y
1
Bangkir -6.336
108.325
Y
9
Ujungaris
-6.457
108.287
Y
2
Bulak
-6.338
108.116
Y
10
Loh
berne
r
-6.406
108.282
Y
3
Cidempet
-6.354
108.246
Y
11
Sudimampir
-6.402
108.366
Y
4
Cikedung
-6.492
108.185
Y
12
Juntinyuat
-6.433
108.438
Y
5
Losarang
-6.398
108.146
Y
13
Krangkeng
-6.503
108.483
Y
6
Sukadana
-6.535
108.300
Y
14
Bondan
-6.606
108.299
Y
7
Sumur
w
atu
-6.337
108.325
Y
15
Kedokan
-6.509
108.424
Y
8
Tugu
-6.433
108.333
Bunder
3. Resul
t
s
and
Analy
s
is
Do
wn
scaling
model by u
s
ing SVR to
predi
ct the
rainfall in d
r
y season
with cla
r
ify
variable in m
odel of GCM
and ob
se
rvation of rainfa
ll
as re
sp
ond v
a
riabl
e, All of
those d
a
ta were
use
d
at every 15 rainfall
station
s
in In
dram
ayu.
He
re a
r
e the
re
sults
of the predi
ction
of the
model G
C
M rainfall avera
g
ed as
sho
w
n i
n
Table 3.
Based on the predi
ction result on Tabl
e 3, it c
an be
said that the result will be better if it
has
a high
correl
ation while RMSE
i
n
lo
w
valu
e. On
the
ke
rnel lin
ear fu
nction
the
hi
gh
correl
ation v
a
lue
wa
s ob
tained at
Ci
ked
ung
ra
inf
a
ll station.
On the
othe
r ha
nd, the
low
correl
ation va
lue wa
s g
o
tten at Cida
mpe
t
rainfall st
ati
on. Overall, i
t
c
an be
co
ncluded that
re
sult
prod
uctio
n
by using kern
el linear fun
c
tio
n
was
b
e
tter than RBF kernel function. It was marke
d
by
the correl
atio
n value or RMSE value in every rainfall
station.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 8, August 2014: 642
3 –
6430
6428
Tabel 3. The
Average
Correlation of the Predi
ction Re
sult by usin
g GCM Mo
del
Data an
d
RMSE Value
s
between
Ra
infall Observa
t
ion in Indram
ayu
N
o
St
a
t
i
on
Kernel Li
near
Kernel R
B
F
Correlati
on
RMSE
Correlati
on
RMSE
1 Bangkir
0.578
62.269
0.562
67.799
2 Bulak
0.684
26.052
0.345
30.298
3 Cidempet
0.401
36.964
0.241
35.353
4 Cikedung
0.744
23.937
0.538
42.483
5 Losarang
0.721
26.955
0.556
32.823
6 Sukadana
0.419
30.517
0.528
31.287
7 Sumur
w
atu
0.670
36.918
-0.053
42.855
8 Tugu
0.651
28.449
0.472
32.258
9 Ujungaris
0.515
29.653
0.422
32.261
10 Lohbener
0.675
32.349
0.579
35.478
11 Sudimampir
0.514
55.424
0.472
57.634
12 Juntinyuat
0.611
44.384
0.648
49.783
13 Kedokan
Bunder
0.726
39.267
0.696
43.202
14 Krangkeng
0.655
43.335
0.414
49.422
15 Bondan
0.681
24.730
0.208
27.580
The be
st G
C
M mo
del
was in T
a
ylor cha
r
t that close
r
to the
observation
point. By
looki
ng at
sta
ndard deviati
on, RM
SE an
d co
rrelation,
observation
p
o
int is the
sta
ndard deviati
on
of data poi
nt at a pa
rticula
r
locatio
n
[20]. Ther
e a
r
e 8
explanation
o
f
GCM mo
del
s we can find
at
Taylor
chart, they are: 1.
CMC1-CanCM3,
2. GDAPS T106L21, 3. GFDL-CM2P1, 4.
GCPS
T63T21, 5. CanCM3-AGCM3, 6. METRI AGCM L
17, 7. NASA-GSF
C
L34, 8. PNU. Here is
Taylor chart f
o
r G
C
M mod
e
l at Cike
dun
g and Ci
dem
pet rainfall st
ation as
sho
w
n in Figure 5.
Figure 5. Taylor Ch
art for G
C
M Mod
e
l
Based
on th
e cha
r
t in Fi
gure
5, it wa
s kn
own that
Cike
dun
g ra
infall station
wa
s at
stand
ard d
e
viation about ±44 and RMS
E
value ±30.
The 1 model
wa
s potentiali
ty to be
the best
model
in thi
s
locatio
n
if it
compa
r
ed
to a
nother mo
del
whil
e
Cidem
pet rainfall
st
ation
wa
s at
±36
stand
ard d
e
viation. The 1
model be
ca
me the bes
t
model in this location if it compa
r
ed
to
anothe
r mo
d
e
l. But, the 1 model at
Cid
e
mpet
stat
ion
wa
s not a
s
b
e
tter a
s
1 mo
del at Ci
ke
du
ng
station, it wa
s ca
used
by the 1 mod
e
l a
t
Cidemp
e
t station ha
s ±3
2
RMSE value
.
The overall
of
linear
ke
rnel f
unctio
n
wa
s b
e
tter than RB
F kernel fun
c
tion.
Cikedu
ng Station
Cidem
pet Station
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Do
wn
scaling
Modelin
g Usi
ng Suppo
rt Vector
Reg
r
e
s
sion for
Rainf
a
ll Predi
ction
(Sanu
si)
6429
4. Conclu
sion
To sum it up,
the models which were resulted
to predict the
rainfall
in dry
season will be
better if it loo
k
ed f
r
om the
averag
e of p
r
edictio
n
re
sul
t
or the e
r
ror
averag
e. The
best
co
rrel
a
tion
value wa
s ob
tained at Cikedun
g rainfal
l
stat
ion in 0.744 correlati
on value and
23.937 RMS
E
while
the l
o
we
st line
a
r
kernel fu
ncti
on
wa
s g
a
in
ed at
Cid
e
m
pet rainfall
station in
0.4
0
1
correl
ation va
lue an
d 3
6
.9
64
RMSE. T
he
kernel
fu
n
c
tion
of RBF
wa
s n
o
t incl
u
ded to t
he b
e
st
function b
e
ca
use the
re
sult
predi
ction
was lo
we
r
than
linear
ke
rnel
function. It ca
n be seen fro
m
the correl
atio
n value or RMSE on RBF kernel fun
c
tio
n
.
Sugge
stion
to the
next
rese
arch,
do
wn
scaling
m
odel
of G
C
M mod
e
l
dat
a can
be
applie
d in order to p
r
edi
ct
the rainfall i
n
dry sea
s
on
by using Su
pport Ve
ctor
Reg
r
e
ssi
on.
The
utilization
of GCM
gri
d
ca
n
be
u
s
e
d
b
e
s
ide
s
gri
d
of
7x7.
The
a
c
cura
cy wa
s no
t high yet, and
then the
sele
ction
of pa
ra
meter val
u
e
s
for
ea
ch
ke
rnel fun
c
tion
need
s to
be
perfo
rmed
with
other optimi
z
ation tech
niq
ues.
Referen
ces
[1]
Karamo
uz M,
F
a
lla
hi M, N
a
zif S, F
a
raha
n
i
,
RM. Lo
ng
L
ead
Ra
infal
l
P
r
edicti
on Us
in
g Statistica
l
Do
w
n
sc
al
ing
a
nd
Artifici
al Ne
ural Net
w
ork
Mode
lin
g.
Arch
ive of SID. S
h
arif
Un
iversity of
Techn
o
lo
gy
.
200
9; 16( 2): 1
65-1
72.
[2]
Che
n
T
S
, Yang CT
, Kuo MC,
Kuo
HC, Yu
S
P
. Proba
bi
l
i
stic
Drou
ght F
o
r
e
c
a
sting
in
South
e
rn T
a
i
w
a
n
Using E
l
Niñ
o-
Souther
n Oscill
ation Ind
e
x
.
T
e
rr. Atmos. Ocean. Sci.
201
3; 24(5): 91
1-9
2
4
.
[3]
Ashok K, Guan
Z
,
Yamagata T
.
A Look at th
e Rel
a
tions
hi
p betw
een the E
N
SO and the Indi
an Ocea
n
Dipo
l
e
. J Mete
orol
ogic
a
l Soci
et
y
.
20
03; 18(
1
)
: 41-56
[4]
Evana L, Effen
d
y
S, Herm
a
w
an E. Peng
em
ban
ga
n Mode
l Prediksi
M
add
en Jul
i
an Oscil
l
atio
n
(MJO)
Berbas
is p
a
d
a
Hasil
An
alis
is
Data
Rea
l
Ti
me Mu
ltivari
a
te
MJO
(RMM1 dan RMM2).
J. Agrom
e
t.
20
08;
22 (2): 144-
15
9.
[5]
Z
e
in. Pemod
e
l
an Back
prop
agati
on Ne
ura
l
Net
w
orks d
an Prob
abi
listi
c Neura
l
Net
w
o
r
k untu
k
Pend
ug
aan A
w
a
l
Mus
i
m Hu
j
an Ber
dasark
a
n Indeks Ik
lim
Globa
l. T
hesis. Bogor: Postgr
adu
ate Bo
gor
Agricult
ure Un
i
v
ersit
y
; 20
06.
[6]
Estinin
g
t
y
as W
.
Peng
emb
ang
an M
o
d
e
l As
ur
ansi
Ind
e
ks Ikli
m untuk
Me
nin
g
katkan
Keta
h
ana
n P
e
tan
i
Padi
da
lam
Meng
had
ap
i P
e
rub
aha
n Ikl
i
m
. Dissertat
i
o
n. Bog
o
r: Pos
t
gradu
ate B
o
g
o
r Agric
u
ltur
e
Univers
i
t
y
; 20
12
[7]
Villa
ges RJ, Ja
rvis A. Do
w
n
sc
alin
g Glob
al Ci
rc
ulati
on Mod
e
l
Outputs:
T
he
Delta Meth
od
Decisi
on a
n
d
Polic
y An
al
ysi
s
W
o
rking.
J
Centro Inter
n
a
t
iona
l de A
g
ri
cultura T
r
op
ic
al Intern
ation
a
l
Center fo
r
T
r
opical A
g
ricu
lture
. 201
0; 1: 1-18.
[8]
Liu Y, Fan K.
A Ne
w
Statisti
cal Do
w
n
sca
lin
g Mode
l for Autumn Preci
p
it
ation i
n
Chi
na.
J. Climato
l
.
2012; DOI: 10.100
2/joc.3514.
[9]
Kann
an S, Gh
osh S. Pred
icti
on of d
a
il
y ra
in
fall stat
e in
a ri
ver basi
n
usi
n
g
statistical do
wnscali
ng from
GCM output. Sprin
ger, Dep
a
rtment of Civil E
ngi
neer
in
g,
Indian Institute of T
e
chnolog
y Bo
mba
y
. In
dia
,
Spin
ger. 20
10; DOI 10.100
7/s004
77-0
10-
04
1
5
-
y
.
[10]
F
aqih
A. R
a
inf
a
ll
Vari
abi
lit
y i
n
the
Austra
l-I
ndo
nesi
a
n
Re
gio
n
a
n
d
the
Role
of I
ndo-P
a
cific C
lim
at
e
Drivers. Disser
t
ation Un
iversit
y
of South
e
rn
Queens
la
nd. 2
010.
[
1
1
]
Buon
o
A, F
a
qih
A, Boe
r
R
,
San
t
i
k
a
y
asa
IP,
R
a
madha
n
A,
Mu
ttaq
ien
MR
,
Asy
har A. A
N
e
u
r
al
N
e
t
w
o
r
k
A
r
c
h
it
e
c
ture
for
S
t
atist
i
ca
l
D
o
w
n
scali
n
g T
e
chni
que: A C
a
s
e
S
t
ud
y
in I
n
d
r
a
m
a
y
u
D
i
st
rict
. Public
atio
n i
n
Inte
rnati
ona
l C
onference
,
T
h
e Qual
it
y
I
n
fo
rmation for
C
o
mpe
t
i
t
i
v
e Agr
i
cultur
al
Bas
ed Prod
ucti
o
n
System a
nd C
o
mme
r
ce (
A
F
I
T
A
)
. 2010.
[12]
W
i
gen
a HA. P
e
mod
e
la
n stati
s
tical d
o
w
n
scal
i
ng
de
ng
an r
e
g
r
essio
n
pr
ojecti
on
persu
it unt
u
k
peram
al
a
n
curah
huj
an (k
asus cura
h h
u
j
an b
u
l
ana
n di
Indram
a
y
u).
Dissertati
on. B
ogor: Postgr
ad
uate Bo
gor
Agricult
ure Un
i
v
ersit
y
. 2
006.
[13] Sutikno.
Statistical Dow
n
sc
ali
n
g
Lu
ara
n
GCM dan Pe
ma
nfaatan
nya u
n
tu
k Peramal
an P
r
oduks
i
Padi
.
Dissertati
on. Bogor: Postgra
d
uate Bo
g
o
r Agr
i
cultur
e Univ
er
sit
y
. 20
08.
[14]
Smola
A, Schö
lkopf B. A
T
u
torial
on
Sup
port
Vector R
egres
sion. N
eur
oCO
L
T
,
Technical
Rep
o
rt
NC-
T
R
-98-0
3
0
, Ro
ya
l Hol
l
o
w
a
y
C
o
lle
ge, Un
ivers
i
t
y
of Lo
nd
on, UK. 2004; 1
9
9
–22
2.
[15]
Arampo
ngsa
n
u
w
at
S,
Me
e
s
ad
P. Pred
i
c
ti
on o
f
PM
10
usi
n
g
Sup
port V
e
cto
r
Regr
essi
on. I
n
ternati
o
n
a
l
Confer
ence
on
Information a
n
d
Electron
ics Engi
neer
in
g, IACSIT
Press. Si
nga
pore. 2
011;
6.
[16]
An KH,
Heo Jin Y, Hameed
SN.
CLIK
2.0
CLi
m
at
e Infor
m
ati
o
n
tool
Kit
User M
a
n
ual
. APEC
Climat
e
Center. 20
10.
[17]
Xi
e P, Yata
ga
i
A, Chen
M, Ha
yas
a
ka T
,
F
u
kushim
a Y, Li
u
C, Yang
S. A g
aug
e-bas
ed an
al
ysis of
da
il
y
preci
p
itatio
n ov
er East Asia.
Journ
a
l of Hydr
ometeor
olo
g
y.
200
8; (8): 607-
627.
[18]
Chen MW, Shi P, Xie VBS,
Silva VE, K
o
usk
y
R, Higgins W, Jano
w
i
ak
JE. Assess
ing
objective
techni
qu
es for gaug
e-b
a
sed
anal
ys
es
of glo
bal d
a
i
l
y
pr
ecipit
ation.
J. Geophys
. 20
08
. R
e
s. 1
1
3
,
D04
110, d
o
i:10
.1029/2
0
0
7
JD0
091
32.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 8, August 2014: 642
3 –
6430
6430
[19]
Che
n
M, Xie
P.
CPC Un
ifie
d Gaug
e-b
a
se
d Ana
l
ysis
of Globa
l Da
ily P
r
ecipit
ation
. W
e
stern Pac
i
fi
c
Geoph
ys
ics Meetin
g, Cair
ns, Australi
a.
200
8
.
[20]
T
a
y
l
or KE. S
u
mmarizin
g
mu
l
t
iple
asp
e
ct of
m
ode
l p
e
rfor
mance
in
a s
i
ngl
e d
i
agr
am.
J Geop
h
y
sica
l
Rese
arch: Atmosph
e
res. 20
0
1
; 106(D
7
): 71
83-7
192.
Evaluation Warning : The document was created with Spire.PDF for Python.