Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
14
,
No.
1
,
A
pr
il
201
9
, p
p.
443
~
449
IS
S
N: 25
02
-
4752, DO
I: 10
.11
591/ijeecs
.v1
4
.i
1
.pp
443
-
449
443
Journ
al h
om
e
page
:
http:
//
ia
es
core.c
om/j
ourn
als/i
ndex.
ph
p/ij
e
ecs
Reservoi
r wate
r level f
orecastin
g usin
g norm
alizati
on and
mu
ltip
le r
eg
re
s
sion
Siti R
af
id
ah
M
-
D
awam
1
,
Ku
Ru
h
ana
K
u
-
Ma
h
am
ud
2
1
Facul
t
y
of
Com
pute
r and
Ma
them
at
ic
a
l
Sci
ences,
Univer
si
ti Te
kn
ologi
MA
RA Ke
dah,
Ma
lay
si
a
2
School
of
Com
puti
ng,
Coll
ege
of
Arts a
nd
Science
s,
Univ
ersiti
Utar
a
Ma
lay
si
a,
Sintok,
Ked
ah,
Malay
s
ia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Oct
6
, 2
018
Re
vised N
ov
1
9
, 2
018
Accepte
d
D
ec
27
, 201
8
Man
y
non
-
par
a
m
et
ric
techniqu
e
s
such
as
Neura
l
Network
(NN
)
are
used
to
fore
ca
st
cur
r
e
nt
re
servoir
water
le
v
el
(RW
Lt
)
.
How
eve
r,
m
ode
ll
ing
using
the
se
t
ec
hn
ique
s
ca
n
be
esta
bl
ished
without
kno
wledge
of
the
m
at
hemat
ic
a
l
re
lationship
b
etw
ee
n
th
e
inpu
t
s
and
th
e
co
rre
sponding
outputs
.
Another
important
issue
to
be
consid
er
ed
which
is
re
l
at
ed
to
for
ecast
ing
is
th
e
pre
proc
essing
st
age
where
m
ost
non
-
par
ametr
i
c
te
chni
qu
es
norm
al
iz
e
da
ta
int
o
discr
et
i
ze
d
dat
a
.
Da
ta
nor
m
al
iz
a
ti
on
ca
n
infl
uen
ce
the
th
e
re
sults
o
f
fore
ca
st
ing.
Th
i
s
pape
r
pre
sent
s
re
servoir
wate
r
le
ve
l
(RW
L)
fore
ca
sting
using
norm
al
izat
i
on
and
m
ultiple
re
gre
ss
ion.
In
th
is
stud
y
,
con
ti
nu
ous
dat
a
o
f
ra
infall
(RF)
and
cha
nges
of
re
ser
voir
wate
r
le
ve
l
(W
C)
are
norm
al
ized
using
two
diffe
re
n
t
no
rm
al
iz
a
ti
on
m
ethods
,
Min
-
Max
and
Z
-
Score
tec
hnique
s.
Its
compara
ti
v
e
stu
die
s
and
for
ecasti
ng
proc
ess
are
ca
rri
ed
out
usi
ng
m
ult
ipl
e
re
gre
ss
ion.
Three
input
sce
n
arios
for
m
ult
iple
re
gre
ss
ion
wer
e
design
ed
which
comprise
of
te
m
pora
l
pa
tt
ern
s
of
W
C
and
RF
,
in
which
the
slidi
ng
window
te
chni
q
ue
has
bee
n
app
li
ed
.
The
expe
r
i
m
ent
al
re
sults
show
ed
tha
t
the
b
est
i
npu
t
sc
ena
rio
for
for
ec
a
sting
the RWLt em
plo
y
s bot
h
th
e
RF
and th
e
W
C,
in
which
th
e
best
pre
di
ct
ors
are
three
d
a
y
’s
del
a
y
of
W
C
an
d
two
da
y
s’
del
a
y
of
RF
.
Th
e
findi
ngs
al
so
suggested
that
th
e
per
form
anc
e
o
f
the
RW
L
fore
ca
st
ing
m
odel
using
m
ul
ti
ple
re
gr
essi
on
was
depe
nd
e
nt
on
th
e
norm
al
iz
a
ti
on
m
et
hods.
Ke
yw
or
d
s
:
Fo
r
ecast
ing m
od
el
Re
servoir m
odel
li
ng
Re
servoir
wate
r
release
Sli
din
g wi
ndow
Tem
po
ral d
at
a
m
ining
Copyright
©
201
9
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
po
nd
in
g
Aut
h
or
:
Sit
i R
afidah M
-
Da
wam
,
Faculty
of Com
pu
te
r
an
d
Ma
them
a
ti
cal
Scie
nces,
Un
i
ver
sit
i Te
knol
og
i M
ARA
Ked
a
h,
P.O. Bo
x 1
87, 084
00 Mer
bok, Ke
da
h
, Mal
a
ysi
a.
Em
a
il
: srafid
ah
192@ke
da
h.uit
m
.ed
u.
m
y
1.
INTROD
U
CTION
Fo
r
ecast
ing
R
WL
is
cru
ci
al
fo
r
reserv
oir
’
s
op
e
rato
r
in
m
aking
de
ci
sion
on
the
res
ervoir
wate
r
release
(R
WR
)
of
a
pa
rtic
ular
re
ser
vo
i
r.
It
is
a
chall
en
ging
a
nd
c
om
plex
ta
sk
,
esp
eci
al
ly
durin
g
flo
od
an
d
dro
ught
occ
ur
a
nces
due
to
un
pr
e
dicta
ble
infl
ow
s
uc
h
as
RF
[1]
.
Th
us
,
a
fe
w
resea
rch
es
ha
ve
f
ocu
se
d
on
non
-
structu
ral
ap
proach
e
s
pr
e
dicti
ng
rese
r
vo
ir
i
nf
l
ow
s
[
2]
.
H
ow
e
ve
r,
duri
ng
flo
od
or
dro
ught,
the
decis
ion
on
R
W
R
is
no
t
only
based
on
th
e
avail
abili
ty
o
f
water
infl
ows
,
bu
t
al
so
o
n
th
e
pr
e
vious
rele
ase,
dem
and
s,
tim
e,
et
c.
Be
sides
da
il
y
RF,
severa
l
researc
hes
al
so
co
ns
ide
re
d
cha
nges
in
t
he
R
W
L
(
W
C
)
as
an
i
nput
in
the
m
ul
ti
pu
r
po
se
r
eservoir
f
or
ec
ast
ing
m
od
el
[
2]
.
RF
(
hydro
log
ic
al
data)
a
nd
re
ser
vo
i
r
WC
a
re
fou
nd
to
be
correla
te
d
i
n
th
e floo
d pr
e
dicti
on
m
od
el
[
3]
.
Ma
ny
li
te
rature
cond
ucted
on
the
R
W
R
oper
at
ion
ha
ve
util
iz
ed
RF
data
an
d
R
W
L
as
in
puts
[4]
,
a
nd
hav
e
a
ppli
ed
di
ff
ere
nt
m
et
ho
ds
an
d
te
ch
ni
ques
of
Ar
ti
fici
al
In
te
ll
igence
and
m
achine
le
arn
i
ng
[
5
–
8]
.
On
ly
a
sm
a
ll
nu
m
ber
of
resea
rch
e
s
c
onduct
ed
on
R
WR
decisi
ons
highli
gh
te
d
on
the
tim
e
delay
betwee
n
the
R
F
an
d
the inc
rease
of
R
W
L.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
14
, N
o.
1
,
A
pr
il
2019
:
443
–
449
444
In
[
9]
discreti
z
ed
data
are
no
rm
alized
us
ing
Mi
n
-
Ma
x
te
chn
i
qu
e
.
I
n
this
stud
y,
the
r
es
ults
sh
owe
d
ei
gh
t
days’
ti
m
e
la
g
relat
in
g
to
upstream
RF
an
d
R
WL
with
a
n
ANN
m
od
el
of
24
-
15
-
3.
Lat
er,
t
he
m
od
e
l
reco
m
m
end
ed
five
da
ys’
tim
e
la
g
with
8
-
23
-
2
A
N
N
m
od
el
with
a
0.0
0708
5%
er
ror.
Ty
pe
2
SV
M
regre
ssion
has
bee
n
use
d
by
[2]
to
f
or
e
c
ast
the
daily
R
WL
of
the
Kla
ng
reserv
oir,
Ma
la
ysi
a.
The
stud
y
em
plo
yed
Z
-
Score te
ch
niqu
e for
data
nor
m
al
iz
ation
and
f
ound
ou
t t
hat the b
est
in
put var
ia
bles are
c
om
bin
at
ion
of
bo
t
h
RF
and
R
WL,
wh
i
ch
we
re
us
e
d
to
determ
ine
the
best
tim
e
la
g
wh
ic
h
a
re
tw
o
da
ys
of
RF
a
nd
with
1.64
%
error.
Au
t
or
e
gr
e
ssive
In
te
grat
ed
Mo
ving
A
ver
a
ge
(A
RIM
A)
m
odel
was
dev
el
op
ed
in
[
4]
for
predict
in
g
the
Kain
j
i
Dam
,
Niger
ia
d
ai
ly
water
le
ve
ls
us
in
g
a
te
n
-
ye
ar
recor
d.
T
he
stu
dy
res
ulted
in
a
m
od
el
w
it
h
a
relat
ive error
o
f
0.039%
ha
d
the
be
st
predic
ti
on
.
In
[
10
]
ANN
with
fee
dforwa
rd
bac
k
pro
pag
at
io
n
was
c
on
cl
ud
e
d
as
the
su
it
able
pr
e
dic
tor
for
real
-
ti
m
e
water
le
vel
f
or
ecast
in
g
of
t
he
S
ukhi
Re
ser
vo
i
r,
India.
T
he
inputs
a
re
t
he
daily
data
of
i
nf
lo
w
,
R
W
L,
a
nd
R
W
R
where
t
he
best
ti
m
e
l
ag
is
te
n
days
with
a
0.8
2%
error
.
NN
wa
s
al
so
e
m
plo
ye
d
in
[
11
]
to
predict
R
W
L
a
nd
co
nc
lud
e
d
a
5
-
25
-
1
NN
m
od
el
a
s
the
best
arc
hitec
ture.
The
stud
y
fou
nd
out
that
five
days’
obse
rv
at
io
ns
of
R
W
L
a
re
si
gn
ifi
cant
f
or
the
R
WR
de
ci
sion
with
a
0.0
3875
6%
e
rror.
A
N
N
arc
hitec
ture
of
4
-
17
-
1
in
forecast
in
g
the
cha
nge
of
RWL
sta
ge
was
pro
po
se
d
in
[3]
.
The
in
pu
t
pa
tt
er
ns
wer
e
t
he
c
hang
es
an
d
sta
ges
of
R
W
L
instea
d
of
the r
eal
va
lue o
f
R
WL.
T
he
resea
rch
sho
wed
that
the
ch
ang
e
s
in
the
sta
ges
of
R
W
L
wer
e
i
nf
l
uen
ce
d
by
the
tw
o
days
of
delay
.
H
ow
e
ve
r,
m
od
el
li
ng
us
in
g
N
N
te
ch
niques
can
be
est
a
blished
without
know
le
dge
of
the
m
at
he
m
at
ic
al
relat
ion
sh
ip
betwee
n
the
inputs
and
th
e
corres
pondin
g
ou
tp
uts.
Whereas
m
ult
iple
reg
re
s
sio
n
is
us
ed
to
ex
pl
or
e
the
relat
ion
s
hi
p
betwee
n
one
con
ti
nu
ous
de
pende
nt
va
riab
le
(DV)
an
d
a
num
ber
of
in
dep
e
ndent
va
riables
(
IVs)
or
pr
e
dicto
rs
(u
s
ually
con
ti
nu
ous).
It
can
determ
ine
how
well
a
set
of
va
riables
is
able
to
pr
e
dict
a
par
ti
c
ular
ou
t
com
e
[12
–
18]
.
This
stud
y
a
pp
li
ed
m
ul
ti
ple
reg
res
sion
i
n
or
der
t
o
identify
w
hich
IV
s (
sli
ces
of R
W
L
a
nd
RF
)
can
be
st
be
t
he
input
pr
e
dictor
s to
pr
edict
DV
(R
WL
t
).
Anothe
r
im
po
r
ta
nt
issue
to
be
co
ns
ide
red
w
hi
ch
is
relat
ed
t
o
f
oreca
sti
ng
is
duri
ng
t
he
preprocessi
ng
ph
a
se
w
he
re
m
os
t
no
n
-
pa
ra
m
et
ric
te
chn
iq
ues
norm
al
iz
e
data
into
disc
r
et
iz
ed
data.
D
at
a
norm
al
iz
a
tio
n
can
influ
e
nce
the re
su
lt
s
of
for
eca
sti
ng
. N
orm
al
i
zat
ion
can b
e pe
rfor
m
ed
at
the
le
vel
of
the
i
nput
fe
at
ur
e
s
or at
the
le
vel
of
t
he
ke
rn
el
[
19
]
.
I
n
m
any
ap
plica
tio
ns,
t
he
a
vaila
ble
feat
ur
es
are
c
on
ti
nu
ous
val
ues,
w
here
eac
h
featur
e
is
m
easur
e
d
in
a
di
ff
e
r
ent
scal
e
and
ha
s
a
diff
e
ren
t
r
ang
e
of
possib
le
values.
I
n
s
uc
h
cases,
it
is
of
te
n
ben
e
fici
al
to
s
cal
e
al
l
feature
s
to
a
com
m
on
ra
nge
by
sta
nd
a
r
dizing
th
e
data.
Pr
e
vious
st
ud
ie
s
m
entione
d
above,
ha
ve
no
t
repor
te
d
a
ny
com
par
at
ive
st
ud
y
done
on
t
he
norm
al
iz
a
t
ion
m
et
ho
d
us
e
d
in
thei
r
res
ear
ch.
In
[19
–
22]
,
nor
m
al
iz
ation
process
ha
s
inc
reased
the
cl
assifi
cat
ion
ac
cur
acy
w
hile
in
certai
n
da
ta
set
s,
norm
al
iz
a
ti
on
m
ay
n
ot d
em
onstrat
e si
gn
ific
ant adva
ntages
[
23
]
.
In
R
W
L
f
orec
ast
ing
,
the
dat
a
is
in
the
f
orm
of
te
m
po
ral
seq
ue
nces,
w
her
e
tim
e
(m
on
th,
day
or
hours)
is
crit
ic
al
[24]
.
T
he
c
hanges
in
t
he
patte
rn
s
of
t
he
data
can
infl
ue
nce
certai
n
de
ci
sion
-
m
akin
g.
T
he
Tem
po
ral
Data
Mi
ning
(TDM
)
te
ch
nique
i
s
re
qu
i
red
to
unco
ve
r
the
va
lues
of
the
at
tribu
te
s
in
vo
l
ved
from
tem
po
ral
seq
ue
nces
re
prese
nting
te
m
po
ra
l
inform
ation
relat
ed
to
c
ertai
n
decisi
ons
by
the
al
gorithm
form
ulati
on
.
The
sig
nificant
tim
e
delay
bet
ween
th
e
cause
of
e
ve
nt
an
d
the
act
ual
eve
nt
need
s
to
be
ca
ptur
e
d
accuratel
y. Se
ve
ral stu
dies r
e
porte
d on the
use
of tem
po
ral
da
ta
in
f
or
e
cast
ing
[
3],
[11],
[25
–
33]
.
This
pap
e
r
pr
esents
reserv
oi
r
water
le
vel
(R
WL)
f
or
ec
ast
ing
us
in
g
norm
al
iz
ation
a
nd
m
ulti
ple
regressio
n.
I
n
this
stud
y,
co
nt
inu
ous
data
of
RF
an
d
cha
ng
es
of
rese
rvoir
water
le
vel
(
WC)
a
re
no
rm
al
iz
ed
us
in
g
tw
o
dif
f
eren
t
norm
al
izati
on
m
et
ho
ds,
Mi
n
-
Ma
x
an
d
Z
-
Sc
or
e
te
ch
ni
qu
es
.
Its
c
omparati
ve
stu
dies
an
d
forecast
in
g
pro
cess
are
car
rie
d
ou
t
us
i
ng
m
ulti
ple
re
gr
es
sion.
T
hr
e
e
in
pu
t
scena
rios
f
or
m
ulti
ple
reg
re
ssion
wer
e
desig
ne
d
wh
ic
h
com
pr
is
e
of
te
m
po
ral
patte
rn
s
of
W
C
and
RF.
The
sli
din
g
window
te
ch
nique
ha
s
bee
n
us
e
d
to
capt
ure
the
delay
in
tem
po
ral
data.
The
ex
pe
rim
en
ta
l
resu
lt
s
sh
owed
t
hat
the
be
st
inp
ut
sce
na
rio
f
or
forecast
in
g
the
R
WL
t
em
plo
ys
bo
th
the
RF
and
the
W
C
,
in
wh
ic
h
the
best
pr
edict
ors
are
three
day’s
delay
of
WC
a
nd
tw
o
da
ys’
de
la
y
of
RF.
T
he
fin
dings
al
s
o
s
ugge
s
te
d
that
t
he
pe
rfor
m
ance
of
t
he
R
WL
f
or
ec
ast
ing
m
od
el
us
ing
m
ulti
ple
regressi
on
was
dep
e
nd
ent
on
the
nor
m
al
iz
ation
m
eth
ods.
Ro
ot
Me
an
S
qu
a
r
e
(R
MSE),
Me
an
Ab
s
olu
t
e
Er
ror
(MA
E)
a
nd
Me
an
A
bs
ol
ute
Pe
r
centage
Er
ror
(MA
PE)
ha
ve
bee
n
us
e
d
as
the
par
am
et
ers
to
m
easur
e the
fo
recast res
ults
ba
sed o
n
the
actual
d
at
a
analy
s
is.
2.
RESEA
R
CH MET
HO
D
Figure
1
de
pic
ts
the
app
r
oac
h
that
has
bee
n
us
e
d
in
co
nduct
in
g
the
re
search
.
The
re
servoir
data
wh
ic
h
c
on
sist
of
RF
an
d
R
W
L
from
19
97
unti
l
2006,
hav
e
been
colle
ct
ed
from
the
Departm
ent
of
Irri
gation
and
Dr
ai
na
ge
(DID),
w
hich
is
in
char
ge
of
m
on
it
or
i
ng
and
m
anag
in
g
the
Ti
m
ah
T
aso
h
reserv
oir.
This
reserv
oir
is
one
of
the
la
r
ges
t
m
ulti
pu
rpose
reservo
ir
s
sit
uated
in
the
nort
hern
Pe
nin
s
ular
of
Ma
la
ysi
a.Th
e
data
co
ns
ist
s
of
operati
onal
and
hydrol
og
ic
al
data.
The
operati
onal
data
has
the
daily
R
W
Ls
m
easur
ed
i
n
m
et
re
(m
)
un
it
w
hile
the
hydrol
og
ic
al
data
has
th
e
daily
R
F
r
ea
dings
m
easur
e
d
i
n
m
ili
m
et
re
(m
m
),
reco
r
de
d
from
f
ive g
a
ug
ing
sta
ti
on
s.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Reservoir w
ate
r leve
l f
or
ec
as
ti
ng
us
i
ng nor
m
alizati
on
and mu
lt
iple re
gr
es
sion
(
Siti
Ra
fi
dah M
-
D
awam
)
445
Figure
1. The
process
f
lo
w f
or R
W
L
f
or
eca
sti
ng
In
the
data
preparati
on
sta
ge
,
the
at
tribu
te
s
are
descr
i
be
d
an
d
rec
ords
with
m
issi
ng
values
we
re
interp
olate
d.
T
his stu
dy u
se
d t
he
R
W
L as t
he
o
ut
pu
t
wh
il
es
the ch
a
ng
e
s of
the r
ese
rvoir wat
er level
(
WC)
an
d
RF we
re
us
e
d as t
he
i
nput. T
hese
WC
will
b
e cal
culat
e
d u
sing eq
uatio
n
[
3]
(1):
1
t
t
t
R
W
L
R
W
L
WC
(1)
wh
e
re
WC
t
is
the
cha
nge
of
R
W
L
at
c
urre
nt
tim
e
t
,
RWL
t
is
the
R
WL
at
curre
nt
tim
e
t
and
R
W
L
t
-
1
is
the
R
W
L
at
on
e
previ
ou
s
day
t
-
1
.
The
RF
dat
a
are
aver
a
ged
by
the
nu
m
ber
of
sta
ti
ons
th
at
hav
e
RF
ba
sed
on
[30]
(
2)
:
r
a
i
n
w
i
t
h
s
t
a
t
i
o
n
s
of
n
u
m
b
e
r
r
a
i
n
t
o
t
a
l
RF
A
v
e
r
a
g
e
_
_
_
_
_
_
(2)
Nex
t,
the
c
ha
nge
-
point
detect
ion
te
c
hn
i
qu
e
is
ap
plied,
w
he
re
rec
ords
w
hich
c
on
sist
of
ga
te
op
e
ning
decisi
on
only
are
extracte
d
[
34]
w
hile
rec
ords
with
gate
cl
os
in
g
decisi
on
wer
e
rem
ov
ed
.
A
total
of
501
record
s
wer
e
de
te
ct
ed
f
ro
m
ten years
of
reservo
i
r op
e
rati
on
(19
97
–
2006).
The
RF an
d
W
C
data
use
d
i
n
this
stu
dy
is
te
m
po
ral
data
wi
th
the
ti
m
e
delay
ed
eve
nt.
T
he
cha
ng
es
in
R
W
L
a
re
the
i
m
pact
of
seve
r
al
sequences
e
ven
ts
of
RF.
I
n
order
to
ca
pt
ur
e
t
he
te
m
po
r
al
inf
or
m
at
ion
of
WC
and
RF,
sli
ding
window
te
c
hniq
ue
is
a
ppli
ed
[
34]
.
Fi
gure
2
s
hows
t
he
pse
udo
-
co
de
for
the
sli
ding
window
wh
e
re
n
is
the size
of
the
wi
ndow. I
n
this
st
ud
y,
n
is
ta
ken
as
the
value
of
sev
e
n
to
inv
e
sti
gat
e
on
the
e
ffec
t
of
seve
n pr
e
vious
ev
e
nt on
c
urre
nt R
W
L
[
35]
as sho
wed in
Ta
ble 1 an
d Tabl
e 2
.
______
______
______
______
______
______
______
for
ti
m
e
t
to end
of f
il
e
read data at
ti
m
e
t
get d
at
a
at
(t
-
1)…
(
t
-
n)
add into
w
i
ndow sli
ces set
__
next
______
___
______
______
______
______
_____
Figure
2. Steps
for
Sli
ding
W
i
ndow
Table
1
.
Sli
ced
Reser
vo
ir
W
C
Date
RWL
t
W
Ct
-
1
W
Ct
-
2
W
Ct
-
3
W
Ct
-
4
W
Ct
-
5
W
Ct
-
6
W
Ct
-
7
12
-
Feb
-
97
2
9
.27
5
0
.02
0
0
.03
5
0
.05
5
0
.03
5
0
.02
5
0
.15
0
0
.00
5
13
-
Feb
-
97
2
9
.33
5
0
.06
0
0
.02
0
0
.03
5
0
.05
5
0
.03
5
0
.02
5
0
.15
0
14
-
Feb
-
97
2
9
.33
5
0
.00
0
0
.06
0
0
.02
0
0
.03
5
0
.05
5
0
.03
5
0
.02
5
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
14
, N
o.
1
,
A
pr
il
2019
:
443
–
449
446
Table
2
.
Sli
ced
RF
Date
Av
erage_
RF
RFt
-
1
RFt
-
2
RFt
-
3
RFt
-
4
RFt
-
5
RFt
-
6
RFt
-
7
12
-
Feb
-
97
2
0
.25
0
7
.33
0
5
.38
0
1
3
.00
0
.00
0
4
6
.25
0
2
4
.50
0
1
0
.00
0
13
-
Feb
-
97
1
3
.87
5
2
0
.25
0
7
.33
0
5
.38
0
1
3
.00
0
0
.00
0
4
6
.25
0
2
4
.50
0
14
-
Feb
-
97
8
.25
0
1
3
.88
0
2
0
.25
0
7
.33
0
5
.38
0
1
3
.00
0
0
.00
0
4
6
.25
0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
In
th
e
ne
xt
sta
ge,
the r
eser
vo
ir W
C
an
d
RF are
norm
al
iz
ed,
w
her
e
the
at
tr
ibu
te
data
is
sc
al
ed
so
as
t
o
fall
within
a
sm
al
l
sp
eci
fied
range.
I
n
a
real
app
li
cat
ion
,
be
cause
of
the
di
ff
ere
nces
in
th
e
ran
ge
of
at
tribu
te
s
’
values
,
one
at
t
rib
ute
m
igh
t
ov
er
powe
r
the
oth
e
r.
Norm
al
i
zat
ion
pr
e
ve
nts
the
ou
t
weig
hi
ng
at
tri
bu
te
s
with
a
la
rg
e
ra
nge.
T
he
goal
is
to
equ
al
iz
e
the
siz
e
or
m
agn
it
ude
and
the
var
i
abili
ty
of
these
at
tribu
te
s.
T
her
e
ar
e
m
any
ty
pes
of
data
norm
al
iz
a
ti
on
,
ho
wev
e
r
on
ly
two
te
ch
ni
qu
es
are
us
ed
to
m
ake
a
co
m
par
is
on
in
this
stud
y;
Z
-
Sc
or
e
and M
in
-
Ma
x N
orm
a
li
zat
ion
.
In
Z
-
Sc
ore
nor
m
al
iz
ation
,
t
he
values
f
or
the
at
tribu
te
s
of
re
servoir
WC
a
nd
RF
a
re
norm
al
iz
ed
base
d
on the m
ean a
nd sta
nd
a
rd d
e
vi
at
ion
.
The
equ
at
ion
for
s
uc
h
t
ran
s
f
or
m
at
ion
is g
i
ven as
fo
ll
ows
(3):
SD
Z
Z
Z
n
e
w
(3)
wh
e
re
Z
is
the
m
ean
of
at
tri
bu
te
an
d
SD
is
the
sta
nd
a
r
d
dev
ia
ti
on
of
the
at
trib
ute.
T
his
m
et
ho
d
of
norm
al
iz
a
ti
on
is
us
ef
ul
if
the
act
ual
m
ini
m
u
m
and
m
ax
i
m
u
m
values
of
the
at
trib
ut
es
are
unkn
own
.
Th
e
adv
a
ntage
of
this
sta
ti
sti
cal
norm
is
that
it
reduces
t
he
e
ff
ect
s
of
ou
tl
ie
rs
in
the
data.
Tab
le
3
a
nd
T
able
4
sh
owe
d
t
he nor
m
al
iz
ed W
C a
nd RF
us
in
g Z
-
Score te
ch
niqu
e.
Table
3
.
Z
-
Sc
ore
of
Rese
r
vo
ir
W
C
Date
zRWL
t
zW
Ct
-
1
zW
Ct
-
2
zW
Ct
-
3
zW
Ct
-
4
zW
Ct
-
5
zW
Ct
-
6
zW
Ct
-
7
12
-
Feb
-
97
0
.69
4
0
.26
6
0
.29
2
0
.39
3
0
.14
8
0
.01
7
1
.31
0
-
0
.20
4
13
-
Feb
-
97
0
.90
8
0
.62
7
0
.15
6
0
.20
7
0
.33
7
0
.11
6
0
.00
3
1
.34
9
14
-
Feb
-
97
0
.90
8
0
.08
6
0
.51
9
0
.06
7
0
.14
8
0
.31
4
0
.10
8
0
.01
0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Table
4
.
Z
-
Sc
ore
of
R
F
Date
zRFt
zRFt
-
1
zRFt
-
2
zRFt
-
3
zRFt
-
4
zRFt
-
5
zRFt
-
6
zRFt
-
7
12
-
Feb
-
97
0
.43
3
-
0
.46
3
-
0
.61
7
-
0
.19
2
-
1
.03
9
1
.93
8
0
.55
6
-
0
.35
1
13
-
Feb
-
97
0
.02
2
0
.29
8
-
0
.50
3
-
0
.64
2
-
0
.19
1
-
1
.03
8
1
.97
9
0
.60
5
14
-
Feb
-
97
-
0
.34
0
-
0
.07
7
0
.25
4
-
0
.52
7
-
0
.68
8
-
0
.20
1
-
1
.04
5
2
.04
9
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
The
seco
nd
te
chn
i
qu
e
is
Mi
n
-
Ma
x
Norm
alizat
ion
.
This
m
et
ho
d
rescal
es
the
at
trib
utes
or
outp
uts
from
on
e
ran
ge
of
value
s
to
a
new
ra
ng
e
of
values.
T
he
at
tribu
te
s
are
re
s
cal
ed
to
li
e
wit
hin
a
range
of
0
to
1
or from
-
1
t
o 1. The
r
escal
in
g i
s accom
plished
by u
si
ng the
fo
ll
owin
g
e
qua
ti
on
(4):
m
i
n
m
a
x
m
i
n
M
M
M
M
M
n
e
w
(4)
wh
e
re
M
is
the
act
ual
val
ue
of
an
at
trib
ute.
T
his
m
eth
od
has
t
he
a
dv
a
ntage
of
preser
ving
e
xac
tl
y
al
l
relat
ion
s
hip
s
in
the
data. Ta
bl
e 5
a
nd Ta
ble
6
s
howe
d
t
he n
or
m
al
iz
ed W
C
and RF
us
in
g M
in
-
Ma
x t
ech
nique.
Table
5
.
Min
-
Ma
x of
Rese
r
voir
WC
Date
m
R
W
Lt
m
W
C
t
-
1
mWC
t
-
2
m
W
C
t
-
3
m
W
C
t
-
4
m
W
C
t
-
5
m
W
C
t
-
6
m
W
C
t
-
7
12
-
Feb
-
97
0
.58
3
8
0
.27
3
5
0
.28
6
3
0
.30
3
4
0
.29
4
7
0
.28
6
3
0
.39
1
8
0
.26
9
4
13
-
Feb
-
97
0
.61
8
5
0
.30
7
6
0
.27
3
5
0
.28
6
3
0
.31
1
6
0
.29
4
7
0
.28
6
3
0
.39
1
8
14
-
Feb
-
97
0
.61
8
5
0
.25
6
4
0
.30
7
6
0
.27
3
5
0
.29
4
7
0
.31
1
6
0
.29
4
7
0
.28
6
3
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Reservoir w
ate
r leve
l f
or
ec
as
ti
ng
us
i
ng nor
m
alizati
on
and mu
lt
iple re
gr
es
sion
(
Siti
Ra
fi
dah M
-
D
awam
)
447
Table
6
.
Min
-
Ma
x of
R
F
Date
m
R
Ft
m
R
Ft
-
1
m
R
Ft
-
2
m
R
Ft
-
3
m
R
Ft
-
4
m
R
Ft
-
5
m
R
Ft
-
6
m
R
Ft
-
7
12
-
Feb
-
97
0
.13
8
7
0
.05
0
2
0
.03
6
8
0
.08
9
0
0
.00
0
0
0
.38
4
6
0
.20
3
7
0
.08
3
1
13
-
Feb
-
97
0
.09
5
0
0
.13
8
7
0
.05
0
2
0
.03
6
8
0
.10
8
1
0
.00
0
0
0
.38
4
6
0
.20
3
7
14
-
Feb
-
97
0
.05
6
5
0
.09
5
1
0
.13
8
7
0
.05
0
2
0
.04
4
7
0
.10
8
1
0
.00
0
0
0
.38
4
6
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Mult
iple
re
gr
e
ssion
is
us
e
d
t
o
e
xp
l
or
e
the
r
el
at
ion
sh
i
p
between
one
co
nt
inuous
dep
e
nd
ent
va
riable
(DV)
a
nd
a
num
ber
of
ind
e
pe
nd
e
nt
va
riabl
es
(IVs)
or
pr
e
dictors
(u
s
ual
l
y
con
ti
nuous)
.
It
can
determ
i
ne
how
well
a
set
of
va
riables
is
a
ble
to
pr
e
dict
a
pa
rtic
ular
outc
om
e.
The
re
gr
es
sion
eq
uatio
n
(
5)
ta
kes
t
he
fo
l
lowing
form
:
n
n
X
B
X
B
X
B
A
Y
....
`
2
2
1
1
(5)
wh
e
re
Y
`
is
t
he
pr
e
dicte
d
val
ue
on
the
D
V,
A
is
the
inte
rcept
,
the
Xs
r
e
pr
es
ent
the
va
rio
us
IVs,
a
nd
the
Bs
are
the co
e
ff
ic
ie
nts assig
ned to ea
ch of
the
IVs
duri
ng r
e
gr
es
sio
n.
The
ou
pu
t
f
or
this
stud
y
is
th
e
R
WL
t
an
d
th
e
inputs
are
re
servoir
WC
an
d
RF.
T
his
stu
dy
desig
ne
d
three
dif
fer
e
nt
input sce
nar
io
s
f
or m
ulti
ple r
egr
essi
on in o
rder to
i
den
ti
fy
w
hich
i
nput sce
na
rios (IVs
)
ca
n best
be
the
in
put
pr
edict
or
s
t
o
f
oreca
st
RWL
t
(
D
V)
.
The
first
s
cenari
o
co
ns
id
ers
the
daily
RF
betwee
n
ti
m
e
(
t
-
1
)
and
(
t
-
7
)
as
t
he
so
le
input, w
hi
le
the
second s
cenari
o
co
ns
i
der
s both
the
R
F
(at
t
-
1
–
t
-
7
)
dan
r
ese
rvoir
WC
(at
t
-
1
–
t
-
7
)
as
i
nputs.
T
he
t
hir
d
sce
nar
io
us
e
s
the
rese
r
vo
ir
W
C
only
bet
ween
ti
m
e
(
t
-
1
)
an
d
(
t
-
7
)
as
inputs.
Eq
uations (
6),
(7)
a
nd (8) re
present the
f
irst
, se
co
nd and t
hir
d
sce
nar
i
os
,
r
es
pecti
vely
.
R
WL
t
=
f
RF
(t
-
i
)
i
=
{
-
1,
-
2,
-
3,
-
4,
-
5
,
-
6
,
-
7}
(6)
R
WL
t
=
f
(RF(t
-
i
),
W
C(t
-
j
))
i
=
{
-
1,
-
2
,
-
3
,
-
4
,
-
5
,
-
6,
-
7}
j
=
{
-
1,
-
2,
-
3,
-
4,
-
5,
-
6,
-
7}
(7)
R
WL
t
=
f
W
C(t
-
i
)
i
= {
-
1
,
-
2
,
-
3,
-
4
,
-
5
,
-
6
,
-
7}
(8)
3.
RESU
LT
S
AND A
N
ALYSIS
In
this
sect
io
n,
the
resu
lt
s
of
the
stu
dy
a
re
di
scusse
d
base
d
on
in
puts
scen
ario
an
d
data norm
al
iz
ation
te
chn
iq
ue.The
best
in
put
sce
nar
i
o
is
deter
m
ined
be
f
or
e
procee
ding
f
ur
ther
i
nto
t
he
f
or
ecast
in
g
c
al
culat
ion
.
B
ased
on
sta
ti
sti
cal
te
st
in
Table
7,
the
f
or
ecast
e
d
val
ue
s
ob
ta
i
ned
by
e
m
plo
yi
ng
s
econ
d
in
put
scenari
o
achieve
the
be
st
resu
lt
s
f
ro
m
oth
er
t
wo
sce
nar
i
os
.
T
he
sc
enar
i
o
em
plo
ys
m
or
e
in
pu
t
da
ta
,
thu
s
pro
vid
in
g
a
bette
r
f
oreca
sti
ng
est
i
m
at
ion
.
It
has
great
er
R
2
wh
ic
h
is
0.
319
as
com
pared
to
t
he
first
and
seco
nd
sc
enar
i
o
wh
ic
h
has
R
2
values
e
qu
a
l
to
0.1
93
a
nd
0.2
79
r
especti
ve
ly
.
The
seco
nd
in
put
scen
ario
al
so
ha
s
sm
a
ll
er
sta
nd
a
rd
e
rror
of
est
im
at
e
(
SEE)
for
both
norm
al
iz
a
ti
on
m
et
ho
ds
.
Th
e
SEE
f
or
Mi
n
-
Ma
x
Tech
ni
qu
e
i
s
0.135
88,
an
d
SEE
f
or
Z
-
Score
te
chn
i
qu
e
is
0.
83
3856.
T
he
refor
e
,
this
second
in
pu
t
sce
nar
i
o
will
be
use
d
as
the b
e
st i
nputs
for fu
rthe
r data
run
s
.
Table
7
.
Stat
ist
ic
al
Test for T
hr
ee
Input Sce
nar
i
os
Inp
u
t Scenario
R
R
2
SEE
(M
in
-
Max
Te
ch
n
iq
u
e)
SEE
(Z
-
Sco
re
T
ec
h
n
iq
u
e)
First
0
.44
0
0
.19
3
0
.14
6
7
3
0
.90
5
4
8
Seco
n
d
0
.56
5
0
.31
9
0
.13
5
8
8
0
.83
8
56
Third
0
.52
8
0
.27
9
0
.13
8
7
2
0
.85
6
0
7
The
sli
ding
wi
ndow
te
chn
i
que
has
been
s
uc
cessf
ully
app
li
ed
on
R
W
L
da
ta
to
extract
a
nd
se
gm
ent
the
tem
po
ral
da
ta
and
preser
ved
the
delay
.
The
stu
dy
us
e
d
m
ult
iple
reg
r
ession
to
fin
d
ou
t
that
the
be
st
tim
e
la
g
f
or f
ore
cast
ing
R
WL
t
is
th
ree d
ay
s’
delay
of r
eser
voir WC
an
d
t
wo
day
s
of
RF
.
Ba
se
d
on
t
his
fi
ndin
g,
tw
o
set
of
re
gressi
on
m
od
el
f
or
R
W
L
t
are
de
ve
lop
e
d
in
orde
r
to
in
vestigat
e
wh
ic
h
norm
al
iz
at
ion
te
chni
qu
es
pro
du
ces
le
ss
error.
T
he
firs
t
reg
r
essio
n
m
od
el
us
e
d
the
Min
-
Ma
x
w
hile
the
seco
nd
m
od
el
us
ed
Z
-
Score
norm
al
iz
a
ti
on
techn
i
qu
e
as s
how
n
i
n
e
qu
at
io
n (9) a
nd (1
0)
:
RWL
t
=
(0.1
7
5
)
+
(0.3
7
5
)
m
W
C
t
-
2
+
(
0
.22
8
)
m
W
C
t
-
3
+ (
0
.35
8
)
m
W
C
t
-
4
+
(0.
1
7
2
)
m
RF
t
-
1
+
(0.1
8
3
)
m
R
F
t
-
2
(9)
RWL
t
=
(
0
.00
)
+
(
0
.21
8
)
z
WC
t
-
2
+
(0.
1
2
9
)
z
WC
t
-
3
+
(0.
197)
z
WC
t
-
4
+ (
0
.12
3
)
z
RF
t
-
1
+
(0.1
3
2
)
z
RF
t
-
2
(10)
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
14
, N
o.
1
,
A
pr
il
2019
:
443
–
449
448
Tw
o
set
s
of
da
ta
based
on
two
diff
e
re
nt
data
no
rm
al
iz
ati
on
wer
e
te
ste
d
us
ing
the
tw
o
reg
res
si
on
m
od
el
dev
el
op
ed.
F
our
sta
ti
sti
cal
fo
rm
ula
a
re
sel
ect
ed
to
evaluate
the
f
or
ecast
in
g
ef
fi
ci
ency
in
this
stud
y,
nam
ely
Root
Mea
n
Square
(R
MS
E),
Mea
n
A
bsolute
Err
or
(
MA
E),
Mea
n
Abs
olut
e
Per
ce
nt
a
ge
Err
or
(MA
PE)
and
th
e
Corre
lation
Coef
fic
i
ent
(R).
The
com
par
ison
of
sta
ti
sti
cal
evaluati
on
on
two
norm
al
iz
a
ti
on
te
chn
i
qu
e
s
is
sh
ow
n
in
Table
8.
T
he
resu
lt
s
s
how
ed
that
the
obt
ai
ned
val
ues
of
RM
SE
,
MAPE
a
nd
M
A
E
by
us
i
ng
M
in
-
Ma
x
te
chn
iq
ue
are
0.141
25,
0.2
4191
a
nd
0.111
22
res
pecti
vely
.
Wh
il
e
us
in
g
the
Z
-
Sc
ore
te
chn
i
qu
e
the
res
ults
are
0.871
65, 6.
90884 an
d 0.6
8677
resp
ect
ively
.
All t
he
RM
SE
,
MAE a
nd MA
PE
values o
btained usi
ng Mi
n
-
M
ax
data
norm
alizat
ion
a
re
cl
os
er
to
0
t
han
usi
ng
Z
-
Sc
or
e
te
ch
nique,
in
dicat
ing
that
t
he
Mi
n
-
Ma
x
te
c
hn
i
ques
is
bette
r
than
Z
-
S
cor
e
.
Howe
ver,
the
Z
-
Sc
or
e
te
chn
i
qu
e
pro
vid
es
sli
gh
tl
y
great
er
correla
ti
on
coeffic
ie
nt
va
lue
s
(R
=
0.
48
858),
than
the
Mi
n
-
Ma
x
te
chn
i
que
(R
=
0.488
56).
I
n
overall
,
forecast
in
g
usi
ng
Mi
n
-
Ma
x
data
norm
al
iz
a
ti
on
te
chn
i
qu
e
s
yi
el
d
le
ss
er
ror
tha
n
us
i
ng
the
Z
-
Score
te
ch
niqu
e.
The
pr
e
dicte
d
outp
ut
usi
ng
Mi
n
-
Ma
x norm
alizat
ion
is m
or
e
reli
able tha
n
that
of the Z
-
Sc
or
e
norm
al
iz
a
ti
on
techn
i
qu
e
.
Table
8
.
C
om
par
iso
n of Sta
ti
sti
cal
Ev
al
uatio
n for
Norm
al
izati
on
Tec
hniq
ue
No
r
m
aliz
atio
n
T
ec
h
n
iq
u
e
RMSE
MAPE
MAE
R
Min
-
Max
0
.14
1
2
5
0
.24
1
9
1
0
.11
1
2
2
0
.48
8
5
6
Z
-
Sco
re
0
.87
1
6
5
6
.90
8
8
4
0
.68
6
7
7
0
.48
8
5
8
4.
CONCL
US
I
O
N
This
pa
per
ha
s
pr
ese
nted
rese
rvoir
water
le
v
el
(RWL)
f
or
e
cast
ing
us
i
ng
norm
al
iz
ation
and
m
ulti
ple
regressio
n.
T
he
resea
rch
on
t
he
c
om
par
ison
of
in
put
sce
na
rio
for
m
ulti
ple
re
gr
essi
on
co
nclu
des
t
hat
th
e
best
input
scena
rio
for
m
ulti
ple
re
gr
essi
on
is
the
seco
nd
in
put
scenari
o
w
hich
consi
sts
of
c
om
bin
a
ti
on
data
of
RF
and
W
C
.
The
sli
ding
wi
ndow
te
chn
i
que
has
been
s
uc
cessf
ully
app
li
ed
on
R
W
L
da
ta
to
extract
a
nd
se
gm
ent
the
tem
po
ral
da
ta
and
preser
ved
the
delay
.
The
stu
dy
us
e
d
m
ult
iple
reg
r
ession
to
fin
d
ou
t
that
the
be
st
tim
e
la
g
f
or foreca
sti
ng
R
WL
t
is
th
ree
days’ del
ay
of r
e
ser
vo
i
r W
C an
d
tw
o day
s of RF.
The
c
om
par
at
i
ve
stu
dies
on
t
he
tw
o
dif
fer
e
nt
norm
al
iz
ati
on
m
et
ho
ds
of
the
Ti
m
ah
Tasoh
rese
rvoi
r
data
us
in
g
m
ulti
ple
reg
ressi
on
sho
wed
t
hat
data
norm
al
i
zed
us
i
ng
Mi
n
-
Ma
x
te
ch
nique
can
e
nhanc
e
the
reli
ab
il
it
y
of
the
f
or
eca
sti
ng
m
od
el
fo
r
R
WL
t
.
F
oreca
sti
ng
us
in
g
Mi
n
-
M
ax
te
ch
niques
yi
el
d
le
ss
err
or
tha
n
us
in
g
the
Z
-
Sc
or
e
te
c
hn
i
qu
e
and
t
he
pre
dicte
d
outp
ut
is
m
or
e
reli
able.
T
he
ex
pe
rim
ent
al
resu
lt
s
sho
w
ed
that
the pre
dicti
on
of the
R
WL
t
usi
ng
MLR
w
as
dep
e
nde
nt
on t
he norm
al
iz
a
tio
n m
et
ho
ds
use
d.
In
the
f
utu
re
,
ot
her
i
nput
var
i
ables
s
uch
as
s
edim
ent,
vo
l
um
e
of
water
re
le
ase
an
d
s
patia
l
eff
ect
ca
n
be
ex
plored
t
o
i
m
pr
ove
th
e
fo
reca
sti
ng
m
od
el
of
R
WL
t.
The
c
om
par
iso
n
of
othe
r
va
rio
us
sta
ti
sti
ca
l
norm
al
iz
a
ti
on
m
et
ho
ds suc
h as m
edian,
sig
m
oid
an
d st
at
ist
ic
al
co
lum
n
norm
al
iz
ation
c
an
al
s
o be m
ea
su
re
d.
ACKN
OWLE
DGE
MENTS
The
aut
hors
wi
sh
to
tha
nk
t
he
Mi
nistry
of
H
igh
e
r
Ed
ucati
on
Ma
la
ysi
a
fo
r
fund
i
ng
this
s
tud
y
unde
r
the
Lo
ng
Te
r
m
Re
search
G
ran
t
Sc
hem
e
(
LRGS/b
-
u/2012/
U
UM/Te
knol
og
i
K
om
un
ika
si
dan
I
nfo
rm
a
si),
a
nd
the D
e
pa
rtm
en
t of I
rr
igati
on
a
nd Drai
na
ge
M
al
ay
sia
f
or sup
plyi
ng
hydr
ology an
d rese
rvo
ir operati
onal
da
ta
.
REFERE
NCE
S
[1]
T.
M.
Sa
tt
ar
i,
K.
Yurekl
i
,
and
M.
Pal,
“
Perf
orm
anc
e
ev
al
ua
ti
on
of
ar
ti
f
ic
i
a
l
neur
a
l
ne
twork
appr
oac
h
es
i
n
fore
ca
st
ing
r
ese
r
voir
inf
low,
”
Ap
pl.
Ma
th. Model.
,
vol
.
36
,
no
.
6
,
p
p.
2649
–
2657
,
J
un.
2012
.
[2]
H.
Afiq,
E.
Ahm
ed,
N.
Al
i,
K
.
Othm
an
Abdul,
H.
Aini
,
and
M
.
Muham
m
ad,
“
Daily
Fore
ca
stin
g
of
Dam
W
at
e
r
Le
ve
ls:
Com
par
ing
a
Support
Vec
tor
Mac
hin
e
(SV
M)
Model
W
it
h
Adapti
ve
Neuro
Fuzz
y
Infe
re
nc
e
S
y
ste
m
(AN
FIS
),
”
Water
Re
sour
.
Manag
.
,
vol. 27, no. 10, pp. 3803
–
3823,
Jun.
2013.
[3]
N.
As
haa
r
y
,
W
.
IShak,
and
K.
Ku
-
Maha
m
ud,
“
N
eur
al
Network
Applic
a
ti
on
in
t
he
cha
nge
of
re
s
erv
oir
wate
r
l
ev
el
stage
for
ecasti
ng
,
”
Ind
ian J. Sci.
Technol
.
,
vo
l. 8,
no.
13
,
pp
.
1
–
6
,
2015.
[4]
C.
C.
Nw
obi
-
Oko
y
e
and
A.
C.
Igboa
nugo,
“
P
re
dicting
W
at
er
Le
vel
s
at
Ka
i
nji
Dam
using
Artifi
c
ia
l
Neur
a
l
Networks,”
N
iger
.
J
.
Techno
l.
,
vo
l
.
32
,
no
.
1
,
pp
.
1
29
–
136,
2013
.
[5]
A.
Afs
har
and
A.
Sale
h
i,
“
Gate
d
Spill
wa
y
s
Ope
ra
ti
on
Rule
s
Consi
der
ing
W
at
e
r
Surfac
e
Eleva
t
ion
a
nd
Flood
Peak ;
Applic
a
ti
on
to
K
ark
heh
D
am,”
W
orld E
nv
iron.
W
ate
r R
esour.
Co
ngr.
2011
,
no.
2
000,
pp
.
3007
–
3
015,
2011
.
[6]
M.
H.
Afs
har
,
“L
arg
e
sca
le
re
ser
voir
oper
ation
by
Constrai
n
ed
Particle
Sw
arm O
pti
m
iz
ation
al
go
rit
hm
s,”
J.
Hydro
-
env
ironment
R
es.
,
vol. 6, no. 1, p
p.
75
–
87
,
Mar
.
2
012.
[7]
E.
T.
Al
emu,
R.
N.
Palmer,
A.
Polebi
tski,
and
B.
Mea
k
er,
“
De
ci
sion
Support
S
y
stem
for
Op
tim
iz
ing
Reser
vo
i
r
Opera
ti
ons
Us
ing
Ensemble
Strea
m
flow
Predic
tions
,
”
J.
Wat
er
Re
sour
.
Pl
an.
Ma
nag.
,
vol
.
137,
n
o.
Februa
r
y
,
pp
.
72
–
82,
2011
.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Reservoir w
ate
r leve
l f
or
ec
as
ti
ng
us
i
ng nor
m
alizati
on
and mu
lt
iple re
gr
es
sion
(
Siti
Ra
fi
dah M
-
D
awam
)
449
[8]
C.
-
C.
W
ei,
“
Discre
t
iz
ed
and
Co
nti
nuous
Ta
rg
et
Fiel
ds
for
the
Reser
voir
Re
lea
se
Rule
s
During
Floods
,
”
Wat
er
Res
our.
Manag
.
,
vol. 26, no. 12,
pp.
3457
–
3477
,
Jun.
2012.
[9]
W
.
H.
Ishak,
K
.
R.
Ku
-
Maha
m
ud,
and
N.
Norw
awi,
“
Neura
l
Network
Applicati
on
in
Reser
v
oir
W
at
er
Le
v
el
Forec
asti
ng
and
Rel
e
ase
De
ci
sio
n,
”
In
t. J.
N
ew
C
omput.
Arch
it.
T
hei
r A
pp
l.
,
vol
.
1
,
no
.
2
,
pp
.
2
65
–
274,
2011
.
[10]
S.
Rani
and
F.
P
are
kh,
“
Predic
t
in
g
Reser
voir
W
at
er
Le
ve
l
Us
ing
Artifi
c
ia
l
Neura
l
Network,
”
Int.
J
.
Innov
.
Re
s.
Sc
i.
Eng.
Te
chnol.
,
v
ol.
3
,
no
.
7
,
pp
.
1
4489
–
14496,
20
14.
[11]
S.
Mokhtar,
W
.
Ishak,
and
N
.
N
orwawi,
“
Model
li
ng
of
R
ese
rvoi
r
W
at
er
R
el
e
ase
Dec
ision
Us
ing
Neura
l
Ne
twork
and
T
emporal
Patt
ern
of
R
ese
rvoir
W
ater
L
e
vel
,
”
in
F
if
th
I
nte
rnational
Co
nfe
renc
e
on
Int
el
li
g
ent
Syste
ms
,
Mode
ll
ing
and
S
imulat
ion
,
2014
,
pp.
127
–
130.
[12]
I.
Gonzá
lez
-
T
ab
oada
,
B.
Gonzá
le
z
-
Font
eboa
,
F.
Martí
nez
-
Ab
e
lla,
and
J.
L.
Pér
ez
-
Ordóñez
,
“
Prediction
of
the
m
ec
hani
c
al
prop
ert
i
es
of
struct
u
ra
l
re
c
y
c
le
d
con
cre
t
e
using
m
ultivariable
re
gre
ss
ion
and
g
ene
t
ic
progra
m
m
ing,
”
Constr.
Build.
M
ate
r.
,
vol
.
106
,
p
p.
480
–
499
,
201
6.
[13]
M.
Krz
y
,
“
AN
AP
PLICATION
OF
FU
NCTIO
NA
L
MU
LT
IVA
RIATE
REGR
ESS
ION
,
”
vol.
18,
no.
3,
pp.
4
33
–
442,
2017
.
[14]
A.
Canc
e
ll
i
ere,
G.
Giuli
ano
,
A.
Anca
ra
ni
,
and
G.
Ross
i,
“
Deri
vat
ion
of
ope
ra
t
ion
rule
s
for
an
irri
gation
wat
er
suppl
y
b
y
m
ult
i
ple
li
n
ea
r
r
egr
es
sion
and
neur
al
net
works
,
”
in
Tools
for
Dr
ou
gh
t
Mit
iga
ti
on
i
n
Me
dit
erranea
n
Re
gions
,
Vol.
4
4.
,
Giuseppe
Ro
ss
i,
Antonino
C
anc
e
ll
i
ere,
Lui
s
S.
Pere
ira
,
Th
eib
Ow
ei
s,
Muha
m
m
ad
Shata
naw
i,
and
Abdel
aziz
Z
ai
ri
,
Eds.
Ne
the
r
la
nd:
Springer
Nethe
r
la
nds,
2003
,
pp
.
275
–
291
.
[15]
A.
M.
Ti
c
la
vi
lca
and
M.
McKee,
“
Mul
ti
var
iate
B
a
y
esia
n
R
egr
essi
on
Approac
h
to
Forec
ast
Re
le
ase
s
from
a
Sy
stem
of
Multi
pl
e
R
ese
rvoirs,
”
Wate
r
R
esour.
Manag.
,
v
ol.
25
,
no
.
2
,
pp
.
523
–
543,
Sep
.
2
010.
[16]
V.
P.
Oikonom
o
u,
A.
Maronidi
s,
G.
Li
aro
s,
S.
N
ikol
opoulos,
and
I.
Kom
pat
siari
s,
“
Sparse
Ba
y
esi
an
Le
arn
ing
for
subjec
t
inde
p
end
ent
c
la
ss
ifica
t
ion
with
app
li
c
at
io
n
to
SS
VEP
-
BCI,
”
In
t.
I
EEE/EMBS
Conf.
Neur
al
Eng.
NER
,
pp
.
600
–
604,
2017
.
[17]
N.
Basant,
S.
Gu
pta
,
and
K.
P
.
Si
ngh,
“
Predic
t
ing
hum
an
int
est
inal
absorpti
on
o
f
d
ive
rse
ch
emicals
using
en
sem
ble
le
arn
ing
bas
ed
Q
SA
R
m
odel
ing
a
pproa
che
s,
”
Co
mput.
B
iol.
Che
m.
,
2016
.
[18]
A.
Esm
al
i,
K.
G.
Bal
d
erl
ou,
M.
Alizade
h
,
N.
Ze
in
al
i
,
and
R.
H.
Gholizade
h
,
“
Predic
ti
on
of
copi
ng
st
y
l
es
an
d
happi
ness
base
d
on
the
m
al
ada
p
tive
sche
m
a
in
cli
ent
s
of
Aid
comm
it
t
ee
in
Urm
ia
Ira
n.
pdf
,
”
R
es.
J
.
Appl.
Sci.
,
vol
.
11,
no
.
4
,
pp
.
14
4
–
153,
2016
.
[19]
P.
S.
Patki
,
V.
W
est,
and
V.
V
Kelka
r,
“
Cla
ss
ifica
t
ion
using
Diffe
re
nt
Norm
al
izat
ion
Te
chn
ique
s
i
n
Support
Vec
tor
Mac
hine,” pp. 4
–
6,
2013
.
[20]
Y.
Li
and
J.
Rui
z
-
ca
st
il
lo
,
“
The
compari
son
of
norm
al
iz
a
ti
on
pro
ce
dure
s
bas
ed
o
n
diffe
re
n
t
c
la
ss
i
fic
a
ti
on
s
y
s
te
m
s,”
J.
In
formetr.
,
vol
.
7
,
no
.
4
,
pp
.
94
5
–
958,
2013
.
[21]
T.
Ja
y
alakshm
i
and
A.
Santha
kum
ara
n,
“
Stat
isti
c
al
Norm
al
iz
a
ti
on
and
Bac
k
Propaga
ti
on
for
Cla
ss
i
fic
a
ti
on,
”
vol.
3
,
n
o.
1
,
pp
.
1
–
5
,
2
011.
[22]
M.
Eft
ekh
ar
y
,
P.
Gholami,
S.
Safa
ri,
and
M.
Shojae
e
,
“
Ran
king
Norm
al
iz
a
t
ion
Methods
fo
r
Im
proving
the
Acc
ura
c
y
of
SV
M Algori
thm b
y
DEA Met
hod,
” vol.
6
,
no
.
10
,
20
12.
[23]
H.
Han
and
K.
Men,
“
How
does
norm
al
iz
ation
impact
RNA
-
seq
d
isea
se
di
agnosis
?
,
”
J.
Bi
omed.
Inform.
,
vol.
85
,
no.
Novem
ber
2
017,
pp
.
80
–
92
,
2018.
[24]
K.
R.
Ku
-
Mah
a
m
ud,
N.
Za
k
aria,
N.
Katuk
,
an
d
M.
Shbier
,
“
Flood
Patt
ern
Detect
ion
Us
ing
Slidi
ng
W
indow
Te
chn
ique
,
” in
2
009
Thir
d
Asia
I
nte
rnational
Co
nfe
renc
e
on
Mod
el
li
ng
&
Simulat
ion
,
2009
,
pp
.
45
–
50.
[25]
K.
Ruhana
Ku
-
Maha
m
ud
and
K.
Jia
Yun,
“
Forest
Fire
Patter
n
Ext
ra
ct
ion
an
d
Rule
Gene
r
at
i
on
using
Slidi
ng
W
indow
Te
chn
i
que.
”
[26]
S.
A.
Mokhtar,
W
.
H.
W
.
Ishak,
and
N.
M.
Nor
wawi,
“
Modell
i
ng
of
re
servoir
wate
r
re
l
ea
se
de
ci
sion
using
neu
ra
l
net
work
and
te
m
pora
l
pa
ttern
of
r
ese
rvoir
wa
te
r
level,”
Proc
.
-
Int.
Conf.
In
te
l
l.
Sys
t.
Mod
el
.
Simula
ti
on,
IS
MS
,
vo
l.
2015
–
Septe
,
pp.
127
–
130,
2015
.
[27]
W
.
H.
Ishak,
K.
R.
Ku
-
Maha
m
ud,
and
N.
Norw
awi,
“
Conce
p
tual
Model
of
In
te
l
li
gent
Dec
ision
Support
S
y
stem
Based
on
Natur
a
li
stic
Dec
ision
T
heor
y
for
Res
erv
oir
Opera
ti
on
du
ring
Emerge
n
c
y
Situa
ti
on
,
”
Int.
J
.
Civ.
Env
iron.
Eng.
,
vol
.
11(2)
,
no.
April
,
pp
.
6
–
11,
2011
.
[28]
M.
H.
Hass
in,
N.
Norw
awi,
and
A.
Ab
-
Aziz
,
“
Te
m
pora
l
Case
-
Based
Rea
so
ning
for
re
servoir
spill
wa
y
g
ate
oper
ation
r
ec
om
m
enda
ti
on,
”
200
6
Int. Conf.
Com
put.
In
formatic
s
,
pp.
1
–
4,
Jun.
20
06.
[29]
W
.
H.
W
an
Ishak,
K.
R
.
Ku
-
Maha
m
ud,
and
N
.
Md
Norw
awi,
“
Modell
ing
of
h
um
an
expe
rt
d
e
ci
sion
m
aki
ng
i
n
re
servoir
op
erati
on,
”
J. Tek
no
l.
,
vol.
77
,
no
.
22
,
p
p.
1
–
5
,
2015
.
[30]
N.
Norw
awi,
“
Com
puta
ti
ona
l
Re
cogni
ti
on
-
Prim
e
d
Dec
ision
Model
Based
On
Tem
pora
l
Data
Mi
ning
Approac
h
i
n
A Mult
ia
g
ent
En
vironment
For R
ese
rvoir
Flood
C
ontrol
De
ci
sion
,
”
2004.
[31]
S.
Mokhtar,
W
.
Ishak,
and
N.
Norw
awi,
“
In
ve
stiga
ti
ng
the
Spatial
R
el
a
ti
onshi
p
bet
we
en
th
e
u
pstrea
m
gaugi
ng
stat
ions
and
th
e
r
ese
rvoir,”
J. A
d
v
.
Manag
.
Sc
i.
,
vo
l.
4
,
no
.
6
,
pp
.
50
3
–
506,
2016
.
[32]
K.
Ruhana
Ku
-
Maha
m
ud,
N.
Za
kar
i
a,
N.
Katuk
,
and
M.
Shbier,
“
Flood
Patt
ern
Dete
c
ti
on
Us
ing
Slidi
ng
W
in
dow
Te
chn
ique
.
”
[33]
K.
R.
Ku
-
Maha
m
ud,
N.
Za
k
ari
a
,
N.
Katuk
,
and
M.
Shbier,
“
Flo
od
Predic
t
ive
M
odel
gen
erate
d
b
y
t
emporal
d
at
a
m
ini
ng
techniqu
e.
”
.
[34]
N.
Norw
awi,
K
.
R.
Ku
-
Maha
m
ud,
and
S.
Der
i
s,
“
Rec
ognition
Dec
ision
-
Mak
i
ng
Model
Us
in
g
Te
m
pora
l
Dat
a
Mini
ng
T
ec
hniq
ue,
”
J
.
ICT
,
vol
.
4,
pp
.
37
–
56
,
20
05.
[35]
M.
H.
Hass
in,
N.
Norw
awi,
and
A.
Ab
-
Aziz
,
“
Te
m
pora
l
Case
-
Based
Rea
soning
for
re
servoir
spill
wa
y
g
ate
oper
ation
r
ec
om
m
enda
ti
on,
”
Co
mput.
Informati
c
s,
…
,
pp
.
4
–
7
,
20
06.
Evaluation Warning : The document was created with Spire.PDF for Python.