TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 12, No. 11, Novembe
r
2014, pp. 79
1
2
~ 791
9
DOI: 10.115
9
1
/telkomni
ka.
v
12i11.65
18
7912
Re
cei
v
ed
Jul
y
25, 201
4; Revi
sed Septe
m
ber
11, 201
4; Acce
pted
Septem
ber 2
6
, 2014
Optimization of Support Vector Regression using
Genetic Algorithm and Particle Swarm Optimization for
Rainfall Prediction in Dry Season
Gita Ad
hani*
1
, Agus Buo
n
o
1
, Akhmad
Faqih
2
1
Departme
n
t of Computer Sci
ence, F
a
cult
y o
f
Mathematics and N
a
tural Sc
ienc
es,
Bogor Agr
i
cult
ural U
n
ivers
i
t
y
,
Bogor 16
68
0, Indon
esi
a
2
Departme
n
tof Geoph
ys
ics an
d Meteoro
l
o
g
y
,
F
a
cu
lt
y
of Mat
hematics a
nd
Natura
l Scienc
es,
Bogor Agr
i
cult
ural U
n
ivers
i
t
y
,
Bogor 16
68
0, Indon
esi
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: adha
ni.g
ita@
gmail.
c
o
m, pud
esha
@gma
il.c
o
m, akhmadfa
q
ih@
g
ma
il.co
m
A
b
st
r
a
ct
Supp
ort Vector
Regress
i
o
n
(SVR) is Sup
port
Vect
or Machi
n
e (SVM) is use
d
for regress
i
o
n
case.
Regr
essio
n
me
thod is on
e of pred
iction se
as
on me
tho
d
has
been co
mmon
l
y used. SVR p
r
ocess req
u
ire
s
kerne
l
functi
on
s to transfor
m
the n
on-l
i
n
ear
inp
u
ts in
to
a
h
i
gh
di
mensi
o
n
a
l fe
ature s
pac
e. T
h
is r
e
sear
c
h
w
a
s conducted
to predict rainf
a
ll in the dry s
easo
n
at
15 w
eather statio
ns
in Indr
a
m
ay
u district. T
he basic
meth
od
used
i
n
this study w
a
s Supp
ort Vector Regr
e
ssio
n
(SVR) opti
m
i
z
e
d
by a hy
brid
a
l
gorit
hm GAPS
O
(Genetic Alg
o
ri
thm an
d Partic
le Sw
arm Opti
mi
z
a
t
i
o
n
).
SVR mod
e
ls create
d
usin
g Rad
i
al
Basis F
unctio
n
(RBF
) kernel. T
h
is hybri
d
techni
que i
n
cor
p
o
r
ates c
once
p
ts from GA and
PSO and creat
es ind
i
vid
uals
ne
w
gen
eratio
n not
only by cross
o
ver an
d muta
tion op
erati
on
in GA, but also throug
h the
process of PS
O.
Predictors us
e
d
w
e
re India
n
Ocean Di
pol
e (IOD) and
NIN
O
3.4 Sea Surf
ace T
e
mper
atu
r
e Ano
m
a
l
y (SST
A)
data. T
h
is res
earch
obtai
ne
d
an SVR
mo
d
e
l w
i
th t
he h
i
g
hest correl
a
tio
n
coeffici
ent of
0.87 a
nd N
R
M
SE
error
va
lue of 11.53 at
Bul
a
k stati
on. Cik
ed
u
ng statio
n h
a
s
the low
e
st
NM
RSE error v
a
lu
e of 0.7
8
a
nd t
h
e
correlati
on co
e
fficient of 9.01.
Ke
y
w
ords
:
rainfa
ll i
n
dry
seaso
n
, g
e
n
e
tic al
gorit
hm,
particl
e sw
a
r
m o
p
ti
mi
z
a
t
i
o
n
, sup
port ve
ctor
regressi
on
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
Indone
sia i
s
cou
n
try state
d
between
2
cont
in
ents, Asia
an
d
Au
stralia, and 2
oce
a
n
s
,
Pacific an
d Hindia. The
r
ef
ore Indo
ne
si
a climate
an
d weath
e
r a
r
e signifi
cantly
affected by both
oce
a
n
s
condi
tion.
Climate and weath
e
r factor ha
s
im
portant
role i
n
many a
s
p
e
c
t of hum
an
ki
nd.
Beside
s, rai
n
fall as varia
b
le whi
c
h d
e
termini
ng th
e climate
co
ndition is di
rectly linke
d to
agri
c
ultu
re a
n
d
plantatio
n
succe
ss. A
s
a
g
ra
rian
co
unt
ry, Indone
sia
depe
nd
s o
n
agri
c
ultu
re a
n
d
plantation circum
stan
ce. High rate rai
n
fall
wo
ul
d
cause floodi
ng
indicating g
r
eat probabilit
y of
failed crop
s. As bad a
s
too
long dro
ught
that
would le
ad to not gro
w
n an
d dea
d plants.
Extreme we
a
t
her can be
related to
cli
m
ate deviati
on whi
c
h d
e
fined a
s
ano
maly of
weath
e
r a
nd
climate
com
p
ared t
o
no
rm
al enviro
n
me
nt in pa
rticula
r
time rang
e.One exa
m
ple
of
the deviation
s is o
c
curren
ce of E
N
SO
phen
omen
on
namely El
Ni
no an
d L
a
Ni
na. El Nin
o
case
gene
rally is
con
n
e
c
ted to
long time d
r
oug
ht or
d
r
y
sea
s
o
n
be
cause of de
crease in rainf
a
ll;
otherwise L
a
Nina i
s
relate
d to floodi
ng.
La
Nina l
ead
s to ove
r
loa
d
ed a
c
cumul
a
tion of ai
r ma
ss
that contain
s
of a lot water
vapor so that in
crea
se the
poten
cy of rain clou
d forma
t
ion.
Climatic p
h
e
nomen
on in
Pacific O
c
ean c
an be
seen o
n
e
x
istence of Southern
Oscillation Index (SOI) and S
ea Surf
ace T
e
mperature Anom
aly (SSTA) of NINO. Clim
ate
con
d
ition in
Hindi
a on
the
other ha
nd
can be
viewed
on Indi
an
O
c
ea
n Di
pole
(IOD). Be
sid
e
o
f
ENSO phe
no
menon in Pa
cific O
c
ea
n, IOD al
so affe
ct signifi
cantl
y
sea su
rfa
c
e
and atmo
sp
here
status. IOD a
nd SSTA NINO3.4 play rol
e
as indi
cato
rs to monitor t
he ENSO ph
enome
non.
Indram
ayu is one of Indon
esia
n distri
ct whi
c
h
is cent
er of agri
c
ult
u
ral produ
cts such a
s
rice [1]. Thi
s
place is very
vulnera
b
le to
drou
ght and
flooding, e
s
p
e
cially when
ENSO hap
pe
ns
in Indo
nesi
a
.
Based
on
dat
a collect
ed in
Annual
Re
port of Indra
m
ay
u Depa
rtment
of Agri
cultu
r
e
,
it is
k
n
ow
n tha
t
in
pr
e
v
io
us
ye
ars
w
h
en El
Ni
no and
La Nina
en
su
ed,
Ind
r
amay
u
en
co
untere
d
plenty food plants (rice
)
da
mage
s [2]. Accordi
ng to
Estiningtya
s [3], the main fa
ctor of the crop
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Optim
i
zation of Support Ve
ctor
Reg
r
e
ssi
on usi
ng Ge
n
e
tic Algorithm
and… (Gita Adhani
)
7913
failure is d
r
o
ught (79.8%
), pest org
ani
sm
s (15.6%
) and floodin
g
(5.6%)
whi
c
h a
r
e stron
g
ly
influen
ced by
climate devia
tion.
The re
se
ar
ch
was f
o
cu
sed
on rainf
a
ll f
o
r
e
ca
st
ing in d
r
y season in Indra
m
ayu. Predicto
r
s
use
d
we
re va
riable
s
relate
d to dry sea
s
on rainfall. T
hose we
re In
dian O
c
ea
n Dipol
e (IOD)
and
Sea Surfa
c
e
Tempe
r
atu
r
e Anomaly (SSTA) in NI
NO3.4
are
a
. Method a
p
p
lied wa
s Sup
port
Vector
Reg
r
e
ssi
on (SVR)
optimize
d
by Geneti
c
Algorithm and Part
icle Swa
r
m O
p
timization.
Suppo
rt Vector Ma
chi
n
e
(SVM) ch
o
s
en
in
reg
r
ession
case
is Su
ppo
rt
Vecto
r
Reg
r
e
ssi
on (SVR). The re
sea
r
ch ado
pted previ
ous
SVR method
applied by
Adhani [4] a
bout
rainfall
pre
d
iction in d
r
y se
aso
n
u
s
ing S
O
I dat
a a
nd
NINO
3.4
se
a su
rfa
c
e te
mperature. S
V
R
pro
c
e
ss
nee
d
s
kern
el to tra
n
sform no
n-li
nier in
put to h
i
gh dime
nsi
o
n feature
roo
m
. The re
se
a
r
ch
only ap
plied
the RBF
kernel b
e
cau
s
e
in former
st
udy by Adh
a
n
i[4] that ha
s
sho
w
n
hig
her
correl
ation value an
d sm
aller
NRMSE
erro
r of RB
F kernel
com
pare
d
to Lini
er or Polyn
o
m
ial
kernel
s. Moreover, RBF
kernel is t
he s
i
mp
le
on
e b
y
its
p
a
r
ame
t
e
r
C
an
d
γ
. Kernel h
a
s
para
m
eter va
lue that
have
to be
d
e
termined
at
first. The
re
sea
r
ch imple
m
ente
d
me
rge
r
of two
optimation m
e
thod in o
r
d
e
r to define t
he optimal
kernel fu
nctio
n
paramete
r
whi
c
h is
Gen
e
tic
Algorithm (GA) and Parti
c
l
e
Swarm
Opti
mizati
on
(PSO) with abbreviation of GAPSO [5].
2. Rese
arch
Metho
d
Flowcha
r
t of rese
arch meth
ods
can b
e
seen in Figu
re
1.
Figure 1. Re
search Meth
od
Flowcha
r
t
2.1. Data a
n
d Predictor
s
Selection
Indian O
c
ea
n
Dipole (I
OD) and Sea Su
rface
T
e
mp
erature Anom
al
y (SSTA) NINO a
r
e
indicators to
monitor the
ENSO p
hen
o
m
enon.
ENS
O
ha
s
great
role i
n
extre
m
e rain va
ria
b
ility
con
d
ition. Flu
c
tuation of E
N
SO in Pa
cific o
c
ean i
s
hi
ghly relate
d to rainfall in In
done
sia [6]. IOD
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 11, Novem
ber 20
14: 79
12 – 791
9
7914
is
sea
phe
n
o
meno
n follo
wed
by atm
o
sp
here p
h
e
nomen
on i
n
Hin
d
ia o
c
e
an e
quato
r
t
hat
influen
ce
s the climate of Australi
a and o
t
her co
unt
ry surroun
ding Hi
ndia o
c
ea
n cavity [7]. IOD is
identified
a
s
deviation
of physi
cal con
d
ition
in
atmosp
here–
sea
intera
ction i
n
tropi
cal
Hi
ndi
a
oce
an that is assu
med ca
n lead to dro
ught in
Indon
esia [8]. NINO Sea Surfa
c
e Temp
erature
Anomaly (NI
N
O
SSTA
)
i
s
index
of se
a surfa
c
e
te
mp
eratu
r
e i
n
so
me regio
n
s.
There a
r
e
4
NINO
areas
acc
o
rding to IRI [9],
s
u
c
h
as
NINO1+2, NI
NO
3
,
NINO3.4, a
nd NINO4. NI
NO 3.4 is
sta
t
ed
betwe
en eq
u
a
tor latitude
of 5°
S–5°N
d 170°–12
0°E and ha
s hi
gh varia
b
ility in El Nino time
scale.
NINO
3.4 is
co
mm
only used i
n
global
c
lim
a
t
e variability
that ha
s b
r
o
ad imp
a
ct. S
e
a
surfa
c
e
temp
eratu
r
e va
ria
b
ility in this
area
ha
s stronge
st imp
a
c
t on
rainfall
frictio
n
o
n
We
st
Pac
i
fic
[9].
IOD d
a
ta fro
m
197
9 to 2
008
wa
s o
b
tained
by cal
c
ulatin
g difference of Sea
Surfa
c
e
Tempe
r
atu
r
e
(SST) b
e
twe
en west
and
east e
nd
of Hindi
ao
cean.
The d
a
ta wa
s colle
cted from
IRI site by o
p
ening E
R
SST
data lin
k o
n
IRI Data
Lib
r
a
r
y (IRI
DL).
O
n
that lin
k IO
D data
(in
pa
rt o
f
data sel
e
ctio
n) was
cho
s
en ba
sed
on
desi
r
ed time
rang
e and
a
r
ea. NI
NO3.4
data ha
s sa
me
year
ran
ge
a
s
IO
D. Thi
s
data
can
al
so be
gai
ned
from IRI
site
s by
applyin
g
the
sa
me
way.
Observation
data
wa
s rain
fall data
ran
g
ed fro
m
1
979
–200
8 in1
5
weather stat
io
n
s
in
Indramay
u
.
IOD an
d NI
NO3.4 SSTA were u
s
ed
as predi
cto
r
s
otherwise rainf
a
ll data in d
r
y
sea
s
o
n
on
May,
Jun
e
, July an
d August we
re ones p
r
edi
cted. Tho
s
e rainfall data were divide
d into 15 weath
e
r
station
s
n
a
m
e
ly: Bang
kir,
Bula
k, Bon
d
an, Ci
dem
pe
t, Cike
dun
g,
Juntinyu
at, Kedo
kanBu
n
d
e
r,
Krang
ke
ng, L
o
sa
ran
g
, Loh
bene
r, Suka
d
ana, Sumu
rwatu, Sudima
mpir, Tugu a
nd Uju
nga
ris.
Data coll
ectin
g
was o
b
je
cte
d
to gain train
i
ng and testin
g data. Traini
ng data wa
s
use
d
to
build SVR m
odel, wh
ere
a
s
testing d
a
ta
to count accuration of fini
she
d
SVR m
odel. Te
sting
data
use
d
we
re on
ly in period of
1 year. Re
se
arch metho
d
flowcha
r
t can
be se
en bel
o
w
in Figu
re 1.
2.2. Support
Vector
Regr
ession (SV
R
) Proces
s
Traini
ng data
was p
r
o
c
e
s
sed usin
g SVR trainin
g
to obtain mod
e
l
which u
s
in
g
rainfall
data in dry season a
s
inp
u
t for the training. Ke
rn
el function a
ppli
ed in SVR p
r
ocessin
g
wa
s
Radi
al Basi
s Function (RBF). This fun
c
tion ha
s parameter value
that must be determin
e
d
at
first, su
ch as
para
m
eter
C and
γ
. Tho
s
e
values affe
ct signifi
cantly the re
sulted S
V
R model. M
o
re
optimal the para
m
eter le
ads to better built m
odel. Search of the ke
rnel fu
nction o
p
timum
para
m
eter was
a
ssi
sted b
y
merge
r
ing optimizatio
n
algorith
m
s of
Geneti
c
Algorithm and Part
icle
Swarm
optimization (GAPSO). SVR is application
of
Support Vector M
a
ch
ine (SVM) in term
of
reg
r
e
ssi
on. In regression
case, outp
u
t is real or
contin
ous nu
mbe
r
. SVR method
is able to sett
le
the over-fitting (con
dition
whe
n
mod
e
l t
u
rnin
g t
oo
co
mplex then
causi
ng b
ad p
r
edi
ction
re
su
lts)
so can ge
nerate great pe
rf
orma
nce [10].
SVR u
s
e
s
ke
rnel
fun
c
tion
to tran
sform
non-li
nea
r i
n
put into
featu
r
e
roo
m
with
high
er
dimesi
on b
e
cause ge
nerall
y
real
worl
d
probl
em i
s
ra
rely linea
r
se
para
b
le. Ke
rn
el functio
n
ca
n
solve non-linear sepa
rabl
e cases li
ke thi
s
. Afterthat, S
V
R
w
ill do linearcal
c
ulatio
n to find opti
m
al
hyperpl
ane i
n
the feature room. Ke
rnel project
s
data into hig
h
dimen
s
ion
feature roo
m
to
increa
se
com
puting a
b
ility of linea
r stu
d
ying ma
chin
e.
Equation (1)
of Radi
al Basis
Fun
c
ti
on
(RBF
)
ke
rnel function can be
se
en
bel
o
w
:
x
,
y
= exp(-
γ
x
-
y
2
(
1
)
2.3. Op
timization
of Su
pport Vec
t
o
r
Regre
ssio
n
(SV
R
) usi
ng G
e
ne
tic
Algorithm
a
n
d
Particle S
w
a
r
m Optimiz
a
tion (GAPSO)
The
re
sea
r
ch
implem
ented
mergeri
ng
of two
optimi
z
a
t
ion metho
d
s to ge
ne
rate
optimal
kernel fun
c
tio
n
para
m
eter,
such as G
e
netic Al
go
rith
m (GA) an
d Particle Swarm Optimizati
on
(PSO)
with abbreviation of GAPSO. Previous
study rel
a
ted to m
e
thod
optimized by
GA and
PSO (GAPS
O)
wa
s
co
nd
ucted
by Ka
o an
d Z
aha
ra [5] an
d
Ririd [11]. Kao
and
Zaha
ra
[5]
adju
s
ted GA
PSO optimization to multi
m
odal fun
c
tio
n
.
This hybrid
techni
que co
mbined co
ncept
s
of GA and PSO and ge
ne
rated n
e
w g
e
neratio
n i
ndiv
i
dual, not onl
y by GA crossover
ope
rati
on
and
mutation but
also P
S
O pr
ocessi
ng. The
result showed advantage
of
GAPSO solution
quality and
h
y
brid ap
proa
ch
conve
r
ge
nce
com
p
a
r
e
d
to 4 othe
r
approa
che
s
whi
c
h a
pplie
d 17
multimodal fu
nction
s obtai
ned from lite
r
ature. Figu
re
2 describe
s
concept of the
two algo
rith
ms
merger.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Optim
i
zation of Support Ve
ctor
Reg
r
e
ssi
on usi
ng Ge
n
e
tic Algorithm
and… (Gita Adhani
)
7915
Jua
ng[12] ob
serve
d
optimi
z
ation of re
curr
ent neu
ral
and fuzzy n
e
tworks de
si
gn. This
s
t
udy compared performanc
es
of algorithms
optimized by GAPSO with GA and PSO. Result
indicated bett
e
r GAPSO performance
among others whi
c
h used GA or PSO.
Figure 2. GA and PSO Met
hod Cl
assification Dia
g
ra
m[5]
2.4. Analy
s
is
and Ev
aluation
Analysis and
evaluation
step to
pre
d
i
c
tion
re
sult
wa
s cond
uct
ed after th
e
testing.
Testing
data
wa
s u
s
ed
a
s
input in
SVR mo
del to
a
c
hieve
output
in form
of p
r
edi
ction val
u
e
.
Accu
ra
cy a
n
d
e
rro
r m
e
a
s
urem
ent of
p
r
edi
ction
re
su
lt whi
c
h
wa
s
obtaine
d fro
m
SVR m
o
d
e
l to
testing data u
s
ed
Norm
alized Root
M
e
a
n
Squa
re
Error (NRMSE) (2)
an
d correl
ation coeffici
ent
(3). Erro
r a
p
p
lication
wa
s o
b
jecte
d
to det
ermin
e
deviat
i
on of p
r
edi
ct
ed value
co
m
pare
d
to a
c
tu
al
value. Erro
r cal
c
ulatio
n u
s
ed
NRMSE.
Correlatio
n
coeffici
ent (R) define
s
con
nectio
n
stren
g
th
between two
variables. Model suitability can
be
achi
eved if R value co
m
e
s
near 1 and NRM
S
E
come
s ne
ar
0. Beside
s, a
nalysi
s
and e
v
aluation al
so can b
e
pe
rformed u
s
in
g
Taylor diag
ra
m
[13]. This dia
g
ram
is abl
e t
o
evalu
a
te
se
veral a
s
p
e
ct
s from
a
com
p
l
e
x model
o
r
a
s
sess
relia
bili
ty
of some mo
dels at on
ce.
Taylor diagram wa
s
built from Root Mean Squa
re Error
(RM
SE),
stand
ard d
e
viation and
co
rrelation bet
we
en pre
d
ictio
n
and ob
se
rvation.
NRMSE=
1
n
∑
y
oi
-
y
pi
2
N
p
i
=1
σ
y
2
= observatio
n
data in peri
o
d
i
to
n
= pre
d
iction
result in pe
rio
d
i
to
n
n
= numbe
r of data
= stand
ard d
e
viationof pre
d
iction
R
n
∑
x
i
y
i
n
i
=1
-
∑
x
i
n
i
=1
∑
y
i
n
i
=1
n
∑
x
i
2
-
∑
x
i
n
i
=1
2
n
i
=1
[
n
∑
y
i
2
- [
∑
y
i
n
i
=1
]
2
n
i
=1
]]
(3)
= observatio
n
datainp
erio
d
i
to
n
= pre
d
iction
resultin
peri
od
i
to
n
3. Results a
nd Analy
s
is
3.1. Data a
n
d Predictor
Selection
Predi
ctor
sel
e
ction in
orde
r
to pre
d
ict d
r
y
sea
s
o
n
rainfa
ll on May,
Ju
ne, July and
August
(MJJA)
wa
s
condu
cted
by correl
ating IO
D an
d SST
A NINO
3.4 ea
ch month to d
r
y sea
s
on
rain
fall
MJJA data from observed
weathe
r stat
ions. Out
put
resulted by correlating tho
s
e two value
s
sho
w
e
d
mont
hs with IOD
and SSTA NINO3.4 si
gnif
i
cantly relatin
g
each oth
e
r to dry season
rainfall. Not a
ll the IOD an
d NINO3.4 p
r
edicto
r
month
s
were
used,
there
were o
n
ly ones
whi
c
h
had high
est
correl
ation u
s
ing Pea
r
so
n method wi
th used d
r
y sea
s
on rain
fall sele
cted
as
predi
cto
r
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 11, Novem
ber 20
14: 79
12 – 791
9
7916
Figure 3
(
a
)
shows
high
est
IOD co
rrelat
ion
o
c
curre
d
on
Octo
be
r with val
ue
of 0.50,
se
con
d
highe
st wa
s on No
vember
with value of
0.47 and third wa
s on Septem
ber with valu
e of
0.40. Figure
3(b
)
de
scrib
e
s
NINO
3.4 p
r
edicto
r
s’
co
rrelation value
s
. High
est
co
rrel
a
tion valu
e in
NINO
3.4 wa
s obtained o
n
February whi
c
h is 0.
2
4
, Ja
nuary of 0.20
and Septem
b
e
r of 0.20.
Correl
ation v
a
lue
of IOD
and
NINO3.4
to d
r
y sea
s
o
n
rainfall by
Pearson
met
hod
ha
s
negative
an
d po
sitive value.
Negati
v
e value
in
IOD data correl
ation mean
s
inve
rsely
prop
ortio
nal
relation
ship.
High
er th
e I
O
D val
ue,
low
e
r dr
y s
eas
o
n
r
a
in
fa
ll.
Positive value in
NINO
3.4 correlation me
an
s di
re
ctlyprop
ortional
rel
a
tionship. Hi
gh
er the
NINO3
.
4 value, hig
her
dry se
ason
ra
infall. Based
on correlatio
n
result, the re
sea
r
ch is
hel
d on Septe
m
ber, O
c
tob
e
r
and
Novemb
er a
s
IOD pre
d
icto
rs an
d Septe
m
ber,
Janu
ary and Februa
ry as NI
NO3.
4 predi
cto
r
s.
3.2. Model Performa
nce
Bas
e
d on O
p
timization
Algorithms
The re
se
arch
was
co
ndu
ct
ed to trainin
g
data
of 20 years. Perform
ance of SVR ke
rn
el
function
can
be se
en on correlation lev
e
l and predi
ction error val
ue com
p
a
r
ed
to observati
on
data. Mod
e
l
perfo
rman
ce
is a
s
sesse
d
well if the
co
rrel
a
tion level
is hi
gh a
nd
predi
ction
error
value is low.
Traini
ng u
s
in
g SVR need
s pa
ramete
r
fits with its kernel. In ord
e
r to obtain
optimal
kernel, when training occured,
optimi
z
ation was conducted us
ing GAPSO (Genetic Algorit
h
m
and Pa
rticle
Swarm
Opti
mization
) hyb
r
id al
gorit
hm
s. Para
meter optimized in
RBF
kernel
wa
s
para
m
eter C
and
γ
(gamm
a
).
Table 1. Co
rrelation an
d NRMSE Value
in Rain Statio
n of Indramay
u
Station Correlation
NRMSE
Bangkir
0.72 13.93
Bulak
0.87 11.53
Bondan
0.72
9.47
Cidempet
0.67 16.02
Cikedung
0.78
9.01
Juntinyuat
0.72 16.56
KedokanBunder
0.82 15.16
Krangkeng
0.13 32.55
Losarang
0.32 15.10
Lohbener
0.78 12.22
Sukadana
0.57 20.85
Sumurw
atu
0.73 15.60
Sudimampir
0.49 18.49
Tugu
0.87 10.43
Ujungaris
0.49 17.98
Figure 3. Correlation Val
u
e
of (a) IOD an
d
(b)
NINO3.4 with MJJ
A
dry s
e
as
on rainfall
(a)
(b
)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Optim
i
zation of Support Ve
ctor
Reg
r
e
ssi
on usi
ng Ge
n
e
tic Algorithm
and… (Gita Adhani
)
7917
Based o
n
opt
imization
cal
c
ulation re
sult
of
RBF kerne
l
paramete
r
u
s
ing 24 p
opul
ations,
100 times
of iterations each pr
ocess in
GA and PSO, 10 iterations
in GAPSO processing, val
u
e
of min_c= 0.1
,
max_c=50, min_ga
mma
= 0 and ma
x_
gamma
=1
0, correlation val
ue and
NRM
SE
of dry sea
s
on rai
n
fall p
r
edi
ction M
J
JA ea
ch sta
t
ions were obtaine
d. Ta
ble 1 de
scri
bed
correl
ation a
nd NRMSE value e
a
ch weather
statio
ns in In
dra
m
ayu. More
de
tail descri
p
tio
n
of
RBF ke
rnel f
unctio
n
perfo
rman
ce in S
V
R model ca
n be see
n
in comp
ari
s
o
n
chart in Figu
re
4
.
The
cha
r
t
sh
owe
d
con
n
e
c
tion bet
wee
n
ob
servatio
n
value
and
d
r
y se
ason
ra
infall predi
ction
result on Ma
y, June, July
, August (CHMK M
JJA
). Bold co
nne
ct
ion between
observation
and
predi
ction
i
n
dicate
s stron
ger co
rrel
a
tion
an
d lo
we
r erro
r bet
ween o
b
serve
d
and
pre
d
i
c
ted
values. Fig
u
re 4 define
s
o
b
se
rvation va
lue and
CH
M
K
MJJA p
r
ed
iction result of Bulak statio
n
has highe
st correl
ation coefficient
val
ue,
wh
i
c
h i
s
0.87, and
Cikedu
ng st
ation ha
s lo
we
st
NRMSE
erro
r value,
whi
c
h
is
9.01. S
c
at
ter plot
in
F
i
gu
r
e
5
s
h
ow
ed c
o
n
n
e
c
t
io
n pa
tte
r
n
b
e
t
w
e
e
n
observation v
a
lue an
d pred
iction result. Linea
r con
n
e
c
tion that forms st
raig
ht line indi
cate
s there
is firm co
nne
ction bet
wee
n
observation
and pre
d
ictio
n
result.
There we
re e
x
treme rainfal
l
value in som
e
points in ob
servatio
n dat
a of year 200
1/2002,
2004/2
005 a
nd 200
7/200
8 and oth
e
r
extreme p
o
in
ts
whi
c
h a
r
e
minimum from ob
servati
o
n
rainfall. A
s
su
mption
usi
n
g
IOD an
d
NINO3.4
data
i
n
tho
s
e
extreme
point
s h
ad n
o
t resulted
optimal predi
ction value ye
t becau
se its i
n
se
ns
itiven
ess in re
spo
ndi
ng the extrem
e pattern.
3.3. Analy
s
is
and Ev
aluation
Figure 4. Co
mpari
s
o
n
Ch
art of Observ
ation and Pre
d
iction
CHMK
MJJA in (a)
Bulak statio
n whi
c
h ha
s
highe
st co
rrel
ation co
efficie
n
t value (b)
Ciked
ung
station whi
c
h h
a
s
lowe
st NRMS
E erro
r value
(
a
)
(
b
)
Figure 5. Sca
tter Plot of O
b
se
rvation wit
h
Predi
ction (a) Bula
k and
(b)
Cikedu
ng
station
(
a
)
(
b
)
Figure 6.Valu
e Cha
r
t of (a) correlatio
n coeffi
cient an
d
(b) NRMSE
error of predi
ction an
d
observation
result in dry season rai
n
fall
predi
ction in
each station
s
(
a
)
(
b
)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 11, Novem
ber 20
14: 79
12 – 791
9
7918
Figure 7. Taylor Dia
g
ram of each
weath
e
r
stat
ion in In
dram
ayu (a
) Bangki
r
, (b
) Bulak, (c)
Cidem
pet, (d) Cike
dun
g, (e
) Losarang, (f
) Suka
dan
a,
(g)Sumu
r
watu
, (h) Tug
u
, (i)
Ujun
gari
s
, (j)
Lohb
ene
r, (k) Sudimampi
r
,(l) Juntinyuat,
(m
) Kedo
ka
n
Bunder, (n
)K
reng
ke
ng, (o
) Bondan
(
a
)
(
b
)
(
c
)
(
d
)
(
e
)
(
f
)
(g
)
(h
)
(i)
(k
)
(j)
(l)
(m
)
(n
)
(o
)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Optim
i
zation of Support Ve
ctor
Reg
r
e
ssi
on usi
ng Ge
n
e
tic Algorithm
and… (Gita Adhani
)
7919
Predi
ction of
rainfall in
dry
sea
s
o
n
u
s
ing
SVR re
sulte
d
varied
correlation
coeffi
cient a
nd
NRMSE e
r
rorvalue. Based
on RBF kernel functio
n
,
Bulak
station
has hig
h
e
s
t correl
ation va
lue
whe
r
ea
s Ci
kedun
g
station
ha
s lo
we
st
NRMSE
er
ro
r
value. Co
rrel
a
tion coeffici
ent
an
d NRM
S
E
error value b
e
twee
n pre
d
i
c
tion re
sult a
nd ob
serv
atio
n data of dry
sea
s
on
rainf
a
ll in Indram
ayu
are compl
e
tel
y
describe
d
in Figure 6.
Correl
ation
coefficient val
ue
sho
w
ed
con
n
e
c
tion p
a
ttern b
e
twe
en ob
se
rvati
on an
d
predi
ction. B
u
lak
station h
a
s hig
h
e
s
t co
rrel
a
ti
on valu
e of 0.87 whi
c
h me
an
s 87
% of obse
r
va
tion
value total diversity can b
e
defined by
its linear
co
nne
ction with
predi
ction value. Figure 7 is
Taylor Dia
g
ra
m that shows model re
sult
ed in th
is re
search ge
nera
t
ed va
ried ou
tputs. The be
s
t
SVR model i
n
ea
ch
statio
ns i
s
mod
e
l
with Taylo
r
di
agra
m
po
sitio
n
stated
clo
s
est to ob
se
rv
ation
point, by loo
k
ing
at sta
n
d
a
rd
deviation
, RMSE an
d correl
ation. Observation
point
is
sta
n
dard
deviation poi
nt in an observed locatio
n
[13].
4. Conclusio
n
The re
se
arch has
succe
ssfully built m
odel of Support Ve
ctor Regressio
n
(SVR)
optimiz
ed by
GAPSO hybrid algori
thm in predic
t
ing
rainfall in dry
seas
on with highes
t
c
o
rrelation
coeffici
ent value and lo
we
st NMRSE value usi
ng IOD an
d SSTA NINO3.4 d
a
ta. That SVR
model was o
b
tained u
s
ing
Radial Ba
sis Function
(RBF) ke
ren
e
l with 24 po
pul
ations, 10
0 time
s
iteration
each GA and PSO, 10
iterations of
GAPSO, min_c
= 0.1, max_c=50, min_gam
m
a
=0
and max_
ga
mma=10. Station weath
e
r
of Bulak ha
s
highe
st co
rrel
ation co
effici
ent value am
ong
others,
whi
c
h
is 0.87,
and
NRMSE
erro
r value
of 1
1
.53. Ci
ke
dun
g
station
ha
s l
o
we
st
NM
RSE
error val
ue,
whi
c
h i
s
9.01
, and
correl
ation coefficie
n
t
value of 0.7
8
. It was
ca
u
s
ed
by fun
c
tion
form that
wa
s not m
a
tch
e
d
with
data,
or
wr
o
ng p
a
rameter
ra
nge
colle
cted
wh
en optimi
z
ati
on
occurre
d
.
Referen
ces
[1]
Z
e
in. Pemod
e
la
nBackpr
o
p
a
gatio
n Neur
al
Net
w
o
r
ks d
an Prob
abi
list
i
c Neura
l
Ne
t
w
ork u
n
tuk
Pend
ug
aan
A
w
a
l
Musim
Hu
j
an B
e
rd
asarka
n Ind
e
ks
Ikl
i
m
Globa
l. PhD
T
hesis. B
o
g
o
r:
Postgrad
uate
IPB; 2014.
[2]
Sucia
n
tini, B
o
er R, Hi
da
yat
R. Evalu
a
si
Prakira
an C
u
r
ah H
u
ja
n BM
G: Studi Kas
u
s Kab
u
p
a
ten
Indrama
y
u.
J. Agrom
e
t
. 200
6
;
20(1): 34–
43.
[3]
Estinin
g
t
y
as W
.
Peng
emba
ng
an Mo
de
l Asur
ansi In
deks Ikl
i
m
untuk M
eni
n
g
katkan
Keta
h
ana
n Peta
n
i
Padi d
a
lam Me
ngh
ad
api Per
u
bah
an Iklim. Phd Dis
s
e
rtation
.
Bogor: Po
stgradu
ate IPB; 2012.
[4]
Adha
ni G, Bu
ono A, F
a
qih
A.
Supp
ort Ve
ctor Regr
essio
n
Mod
e
ll
in
g For Ra
infal
l
Pre
d
ictio
n
in
Dr
y
Seaso
n
Bas
e
d
on So
uther
n
Oscillatio
n
Ind
e
x an
d Ni
no
3.4
. In Adva
n
c
ed C
o
mputer
Scienc
e an
d
Information S
ystems (ICACSIS). Bali. 2013:
315
–3
20.
[5]
Kao YT
,
Z
aha
ra E. A H
y
br
i
d
Genetic Al
g
o
rithm
an
d Pa
rticle S
w
arm
Optimizatio
n
F
o
r Multimo
dal
F
unctions.
Ap
p
lied S
o
ft Comp
uting.
20
08; 8: 849
–8
57.
[6]
Aldria
n E,
LD
Gates, F
H
W
i
d
odo. S
eas
on
al
Varia
b
il
it
y of In
don
esia
n
Rai
n
fall
in
ECHAM
4
Simul
a
tio
n
s
and i
n
T
he Reana
l
y
ses: T
he
role of ENSO.
T
heoretic
al an
d Appl
ied C
l
i
m
atolo
g
y
. 200
7; 87: 41–
59.
[7
]
Sa
j
i
N
H
,
Go
sw
ami
BN
, Vinay
a
c
ha
nd
ra
n
PN
, Ya
mag
a
t
a T
.
A D
i
p
ol
e
Mo
d
e
i
n
th
e
T
r
o
p
i
ca
l
In
di
an
Ocean.
Nature.
1999; 4
01: 36
0-36
3.
[8]
Ashok K, Gu
a
n
Z
,
Yama
gat
a T
.
A Look
a
t
t
he R
e
lati
ons
hip
Bet
w
e
e
n
T
he E
N
SO a
n
d
T
he Indi
a
n
Ocean Di
pol
e.
J Meteorol
ogic
a
l Soci
ety
. 200
3; 18(1): 41-5
6
.
[9]
[IRI]
T
he Internatio
nal R
e
se
arch Institute
fo
r Climate a
n
d
Societ
y(US).
2007. Mo
nito
ring ENSO.
Avail
abl
e:
http://iri.columbia.e
du/climate/ENS
O/background/monitoring.ht
ml
. [do
w
n
l
oad
ed 20
12 No
v
25].
[10]
Smola AJ, Schölk
opf B. 200
4. A
T
u
torial on Supp
ort Vector Regressi
on
.
Statistics and Computin
g
.
200
4; 14: 199-
222.
[11]
Ririd
ART
H
, Arifin AZ
, Yu
ni
a
r
ti A. 20
10. O
p
timasiMeto
de Discrimi
native
l
y Re
gul
arize
d
Least
S
q
u
a
r
e
Classific
a
tio
n
Den
ganA
lg
orit
maGenetik
a.
JITI
. 2010; 5(3): 166-
174.
[12]
Juan
g CF
. A H
y
brid
of Genet
ic Algor
ithm and Partic
le S
w
a
rm Optimiz
a
tion for R
e
cu
rrent Net
w
or
k
Desig
n
.
IEEE Trans. Syst., M
an. Cyber
n., B
. 2004; 3
4
(2): 9
97-1
006.
[13]
T
a
y
l
or KE. 20
01. Summar
i
zi
ng Mu
ltipl
e
A
s
pect of Mo
d
e
l Perform
anc
e in
a Si
ngl
e
Dia
gram.
J
Geophys
i
cal R
e
searc
h
: Atmo
spher
es
. 200
1; 106(D
7
): 718
3
-
719
2.
Evaluation Warning : The document was created with Spire.PDF for Python.