Optimization of Support Vector Regression using Genetic Algorithm and Particle Swarm Optimization for Rainfall Prediction in Dry Season

lflf

LO,N
ＺＮｊｬ｟ｾ＠

'-b\'-«Ji
ｈ］ｃｬ｟ｾｔｴｒ＠

MUf.E:'
ｃｕｦｩｾｔ＠

A'l-L_UCfMNf',
H ml'

USlll

Tflltn1t '1nly ｰＱｾｴＧｩ＠

t 1bt ｨｩｬｾＢＧ＠
been c.:tr"fvrjv i..uoohetld ::i110 pot !.hed lO &1tO•(J hav 1119 to
thr IT'Onu,.., pt ldt. r 1n 111" 11•11lt•w ＱｲｵｦＧＬｾ＠

Ayrhn•\ •hvuld ｰＱｾｮｲ＠
ｴｾ＠
r wor'• ｨｯｮｾ＼ｴＱﾷ＠
without f.3br1e ｾﾷ＠ o i. ｦ｡ｴｾ＠
f C'i'IUC'n, o 19q dr ｾＮＬ＠
or fl.30D'"OP1 •ate ｊＺ｡ｴｾ＠
f113 ''Pulati(Jn "°1ll 1uscr "ts 1u e
•ccttpt.c.-d wnh ttie u•ld "!fttbnd ng lhJt thrl dr .... ')rrc,1P1ttl or r•t11 nd,..d \·f , .. ,on of prttvlou' v ｰＺ［Ｌ｢ｕｾｍ､＠
cai:cl'$ n ｣ｯｾｦｴｲｳ＠
｡ｾ､Ｏｯｲ＠
iourna • ll"d Lhat. 1t !'le war. rccclVcO a:'I ():'l:cral s:ionsc sit D l has
:iu,
:: for op
a t ".: ..1f\fl
to r dJ.r s rt: t'la: tour oaprr ｾ＠
ＬＮｾ＠
.. re•
,
ｾ＠ J l"'" ,
•tern

t
d 1 c. liltc> ﾷＮｾｲｳ＼Ｐｮ＠
ｾｬｮｃＺ＠
y p •ase

ｲ･ＢﾰＧｾ＠

1,.,.'

download the Journ1I tcmpl1t1 CHS Word l
t rt f JtJ
... n._, j'.: ne
f ,tt · I
•t. ,.,..

LAUX ycrslon This
•m,.

on tf

.., v""'C !.CJ.

ｒ｣｡ｾｯｮ＠

to pubhsh (AutlHH ｾ＠

｢･ｮｬＡｦｩｴｾＩ＠

ｾ＠

wi

ｾ＠

ｾＱＮｯｲ｣＠

to

ｾ＠ p otesstng

ldt-1 l!C ttt dUtho s "" 11 r

DC

fr•r puhl1.:O
-Obftf\·#1

I
｡ｾ＠

so

-

i

-

Pr•cll.1'

T- ,

(a)
(b)
Figure 4. Comparison Chart of Observation and Prediction CHMK MJJA in (a) Bulak station which has
highest correlation coefficient value (b) Cikedung station which has lowest NRMSE error value

1:
:..:oo

../

•

ｾＷＭ

;toe

- -

Ｍﾷ＠
0

j=-· -.-- .

/_,,,

ｾ＠

:CC

ｬｾｏ＠

°"""'... ,,

::00

/

://

ｾ＠

0

Ｈｾ＠

ｾ＠

100

:5C

::00

ｾｏ＠

Olowr·u. ,/

ｾＩ＠

Figure 5. Scatter Plot of Observation with Prediction (a) Bulak and (b) Cikedung station
Based on optimization calculation result of RBF kernel parameter using 24 populations,
100 times of iterations each process in GA and PSO, 10 iterations in GAPSO processing, value
of min_c= 0.1, max_c=50, min_gamma= O and max_gamma=10, correlation value and NRMSE
of dry season rainfall prediction MJJA each stations were obtained. Table 1 described
correlation and NRMSE value each weather stations in lndramayu. More detail description of
RBF kernel function performance in SVR model can be seen in comparison chart in Figure 4.
The chart showed connection between observation value and dry season rainfall prediction
result on May, June, July, August (CHMK MJJA). Bold connection between observation and
prediction indicates stronger correlation and lower error between observed and predicted
values. Figure 4 defines observation value and CHMK MJJA prediction result of Bulak station
has highest correlation coefficient value, which is 0.87, and Cikedung station has lowest
NRMSE error value, which is 9.01. Scatter plot in Figure 5 showed connection pattern between
observation value and prediction result. Linear connection that forms straight line indicates there
is firm connection between observation and prediction result.

There were extreme rainfall value in some points in observation data of year 2001/2002,
2004/2005 and 2007/2008 and other extreme points which are minimum from observation
rainfall. Assumption using 100 and NIN03.4 data in those extreme points had not resulted

optimal prediction value yet because its insensitiveness in responding the extreme pattern.

3.3. Analvsis and Evaluation

li:t ii I. i1··1ii Ii.,
.
(a)
Cb)
Figure 6.Value Chart of (a) correlation coefficient and (b) NRMSE error of prediction and
observation result in dry season rainfall prediction in each stations

Optimization of Support Vector Regression using Genetic Algorithm and... (Gita Adhani)

7918

ISSN: 2302-4046

•

(c)

(b)

(a)

110

0 1

70 .. ｾ＠ . •• . · ·' ,.
;

60

'

.! ../..ｾﾷＧＯ＠

so

ﾷ ﾷＢＧｾｦ＠

0

I

•

.
7 ; .,

t'r·.// '•//

,,

. ·. ,,, .. _,

- 40 ,J.......
I/ ''Ii
1!
.( • • ,,•,