TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.5, May 2014, pp
. 3483 ~ 34
9
0
DOI: http://dx.doi.org/10.11591/telkomni
ka.v12i5.3238
3483
Re
cei
v
ed Ma
y 22, 201
3; Revi
sed
De
ce
m
ber 9, 2013
; Accepte
d
Decem
b
e
r
29, 2013
Demand Forecasting Model of Port Critical Spare Parts
Zhijie Song
1
, Zan Fu*
1
, Han Wang
1
, Guibin Hou
2
1
Departme
n
t of Economics a
n
d
Mana
gem
ent
, Y
anshan U
n
iv
ersit
y
, He
be
i, 0660
04, Ch
ina
2
Qinhua
ng
dao
Port Co., Ltd.,
Heb
e
i, 06
600
2, Chin
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: 5311
54
09@
q
q
.com
A
b
st
r
a
ct
De
ma
nd forec
a
sting for
port
critical s
pare
p
a
rts (CSP)
is
notori
ously
diffi
cult as it is
ex
pens
ive,
lu
mpy
an
d i
n
te
rmittent w
i
th
hi
gh v
a
ria
b
il
ity. In this
pa
per, s
o
me i
n
flu
enti
a
l
factors w
h
ich
have
an
effect
on
CSP cons
u
m
pt
ion w
e
re
pro
p
o
s
ed acc
o
rdi
ng
to port CSP
c
h
aracteristics an
d
histor
ical dat
a.
And ana
lytic
hier
archy proc
ess (AHP) is used to siev
e out the more
i
n
flue
ntial facto
r
s. Comb
in
ed
w
i
th the influe
ntia
l
factors, a le
a
s
t squares
s
upp
ort vector
mach
i
nes (
L
S-SVM) mo
de
l opti
m
i
z
e
d
b
y
particl
e sw
arm
opti
m
i
z
at
ion
(PSO) w
a
s de
velo
ped
to for
e
cast the
de
ma
nd. A
nd th
e effective
nes
s of the
mo
d
e
l i
s
de
mo
nstrated
throu
g
h
a r
eal
case stu
d
y, w
h
ich s
how
s that
the
pro
pose
d
mo
de
l ca
n for
e
cast the
de
ma
nd
of port CSP more accur
a
tely,
and effective
l
y
reduce i
n
ve
ntory backl
og.
Ke
y
w
ords
: sp
are p
a
rts, de
ma
nd for
e
casti
ng, an
alytic h
i
erarchy pr
oce
ss (AHP), leas
t squares s
u
p
port
vector mac
h
i
n
e
s
(LS-SVM), particle sw
arm
o
p
timi
z
a
ti
on (PS
O
)
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
Port enterpri
s
es play a very important ro
le
in eco
nomi
c
develo
p
me
nt, and they also pla
y
a sig
n
ifica
n
t role in the
wh
ole logi
stic
ch
ain. In
order t
o
avoid d
o
wn
time of the e
quipme
n
t due
to
spa
r
e pa
rts
shortag
e
, port
enterp
r
i
s
e
s
tend to st
ore la
rge am
ount
s of spa
r
e pa
rts, whi
c
h take
up
a lot of inventory ca
pital, but even so, spare
part
s
sh
ortage
phe
no
menon
still o
c
curs freq
uen
tly.
Therefore,
predictin
g spa
r
e pa
rt
s
dema
nd ha
s
be
co
me the
key t
o
solve the
s
e
pro
b
lem
s
[1]
.
In
this pape
r, we focus o
n
the critical spa
r
e par
ts, whi
c
h are more importa
nt to the equip
m
ent
and
occupi
ed mo
re inventory capital. But CSP have the
cha
r
a
c
t
e
ri
st
ic
s of
mult
iple i
n
f
l
uen
ce f
a
ct
ors
,
non-li
nea
rity and hig
h
vari
ability, which
bring
s
difficult
ies to pre
d
ict
the deman
d.
Re
sea
r
che
r
s
have develo
p
ed many fo
re
ca
sting te
chn
i
que
s in t
he l
a
st
de
ca
de
s,
su
ch a
s
time
se
rie
s
p
r
edi
ction met
hod, Cro
s
ton’
s
meth
od,
Bo
otstrap
p
ing,
neural n
e
two
r
k [2,
3] an
d
so
on. Li and Kuo appli
ed e
nhan
ce
d fuzzy neural net
work (EF
N
N) to foreca
st the dema
nd for
automobil
e
spare
part
s
in
a cent
ral warehou
se,
u
s
in
g analytic hi
e
r
archy p
r
o
c
e
s
s (A
HP) meth
od
to determi
ne
factor’
s
wei
ght. The exp
e
rime
ntal re
sults sho
w
th
at EFNN
out
perfo
rms
oth
e
r
model
s in fill
rate an
d sto
ck
co
st mea
s
ures
[4]. Hu
a and Z
han
g
prop
osed a
n
appli
c
ation
of
sup
port ve
cto
r
ma
chi
n
e
s
(SVM) re
gression
meth
o
d
for forecastin
g spare p
a
rt
s dema
nd,
wh
ich
aiming fi
rstly
to fore
ca
st th
e o
c
curren
ce
of no
nzero
deman
ds,
an
d then
to e
s
ti
mate lea
d
-ti
m
e
deman
d. Th
eir test u
s
in
g real d
a
ta sets of 30
kinds of spa
r
e part
s
from
a petro
che
m
ical
enterp
r
i
s
e in
Chin
a su
gge
sted thi
s
met
hod pe
rf
orm
better than
Crosto
n’s, b
o
o
t
strappi
ng a
n
d
other methods
[5].
Based
on the
above literat
ure
s
, there
are not
many i
n
vestigatio
ns focu
sed o
n
the CSP
requi
rem
ent predi
ction. In
vestigation
s
on port in
dus
tries a
r
e eve
n
fewe
r. In general, there
is no
approp
riate fore
ca
sting m
odel for pred
icting t
he re
quire
ment of
port CSP. Furthe
r more, no
method
s hav
e been
previously p
r
op
osed that ma
ke
full use of
real data b
a
sed on p
o
rt
CSP
relative facto
r
s, as
we do i
n
this stu
d
y. In th
is arti
cle,
Lea
st Square
s
Supp
ort Ve
ctor M
a
chi
n
e
s
(LS-SVM
)
[6] reg
r
e
ssi
on,
a semi
-pa
r
am
etric modeli
ng
t
e
ch
niqu
e, is
use
d
to p
r
edi
ct the p
o
rt
CSP
deman
d. Firstly, we p
r
o
posed
some
infl
uential f
a
ctors
whi
c
h
have an
e
ffect on
CSP
con
s
um
ption
after analyzi
n
g port CSP chara
c
te
ri
sti
c
s and histo
r
ical data. And applie
d analy
t
ic
hiera
r
chy pro
c
e
ss (A
HP) t
o
sieve out the more
influential facto
r
s as the in
p
u
ts of LS-SVM
model. Seco
ndly, aiming at the param
eter optimiz
ation pro
b
lem i
n
LS-SVM, the particl
e swarm
optimizatio
n (PSO) algo
rith
m wa
s ad
opt
ed to opt
imi
z
e the pa
ram
e
ter and i
m
pro
v
e the learni
ng
performance and
genera
lization ability of LS-SVM model. The
proposed PSO-LSSVM m
odel
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3483 – 34
90
3484
take
s into accou
n
t influent
ial factors of port C
SP, and the real
case
study wi
th real data i
n
Qinhu
ang
dao
port illustrate
s the effectivene
ss of the
model.
2. Rese
arch
Metho
d
2.1. Analy
t
ic
Hierar
ch
y
Pr
ocess
Analytic hie
r
a
r
chy
process
(AHP) was propo
se
d
by T
homa
s
L.Sa
a
t
y in 198
0 [7]
,
whi
c
h
is one of wi
dely used mul
t
i-criteria deci
sion-m
a
k
in
g
method
s. AHP involves th
e prin
cipl
es
of
decompo
sitio
n
, pair-wise comp
ari
s
o
n
s,
and prio
ri
ty vector gene
ration an
d synthesi
s
. In this
study, the AHP method wa
s u
s
ed to
sel
e
ct influe
nce
factors of the
port CSP a
n
d
determine t
he
relative impo
rtance.
The thre
e ma
in step
s of AHP are
sho
w
n as follo
ws:
Step 1: Const
r
uctio
n
of hierarchical structure.
Step 2: Calcu
l
ation of weig
hts between f
a
ctors at ea
ch hiera
r
chi
c
al
level.
Step 3: Calcu
l
ation of the ov
erall hie
r
archical
weight
s.
Ask
evaluato
r
s to m
a
ke p
a
ir-wi
s
e
com
pari
s
on
s of t
he rel
a
tive importa
nce of
variable
s
usin
g the
scale. Based
o
n
the
re
sults of the
que
stionnai
re, a
p
a
ir-wi
s
e
com
pari
s
on
matri
x
is
con
s
tru
c
ted
t
o
calculate t
he
cha
r
a
c
teristic val
u
e
s
and th
e
cha
r
acte
ri
stic ve
ctors, the
r
eb
y
examining th
e con
s
i
s
ten
cy of the matrix to derive a con
s
i
s
te
ncy index (
C.I
). For each
alternative, th
e con
s
iste
ncy
ratio
(
C.R
) is me
asu
r
ed
by the
ratio
of the
consi
s
ten
c
y in
dex to the
rand
om ind
e
x (
RI
). The eq
uat
ions a
r
e a
s
follows:
ma
x
λ
-n
CI
C.
I
=
,
C
.
R
=
.
n-
1
R
I
(1)
Gene
rally, the value of
C
.
R
sho
u
ld be le
ss tha
n
0.1 to guara
n
tee
con
s
iste
ncy.
If
con
s
i
s
ten
c
y doe
s not co
mply with the requi
rem
e
nt, it means that judgm
ents ma
de
are
inco
nsi
s
tent. And the re
se
arche
r
sh
all e
x
plain t
he pro
b
lem of every
pair-wi
se
co
mpari
s
o
n
. After
cal
c
ulatin
g weights of eve
r
y facto
r
, we
can
obtai
n th
e mo
re i
n
flue
ntial facto
r
s
as th
e in
puts of
LS-
SVM model.
2.2. Least Sq
uares Supp
o
r
t Vector Ma
chines
Lea
st sq
uare
s
supp
ort ve
ctor m
a
chine
s
(LS
-
SVM) i
s
a mo
dificat
i
on of the st
anda
rd
sup
port ve
cto
r
machine
(SVM) and
wa
s develop by
S
u
yken
s [6]. LS-SVM is use
d
for the opti
m
al
contr
o
l of non
-linea
r sy
ste
m
s for cl
as
sifi
cation a
s
w
e
ll
as reg
r
e
s
sio
n
.
Given the sa
mple of
kk
D
=
x
,
y
,
k
=
1
,
2
,
,
N
, with input vectors
n
k
xR
and outpu
t
values
k
yR
. The goal is to e
s
timate a model
of the form:
T
y(
x)
=
ω
(x
)
+
b
(2)
Whe
r
e
()
is the mappi
ng to a high dimensi
onal fe
ature spa
c
e.
Combin
e the functional
compl
e
xity and fitting error,
the optimizat
ion pro
b
lem o
f
LS-SVM is given as:
N
2
2
k
ω
,b
,
e
k=
1
1
mi
n
Q
ω
,b
,
e
=
ω
+e
22
>0
s.t
.
T
kk
k
y=
ω
(x
)
+
b
+
e
k=
1
,
2
,
,
N
(3)
This form
ula
t
ion con
s
ist
s
of eq
uality inste
ad
of ineq
uality
con
s
trai
nts.
Con
s
tru
c
ting
the
Lagrangi
an:
N
T
kk
k
k
k=
1
L
ω
,b
,
e
,
a
=
Q
ω
,b
,
e
-
a
ω
(x
)
+
b
+
e
-
y
(4)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Dem
and Fo
re
ca
sting Mod
e
l
of Port Cr
itical Spare Pa
rts (Zhiji
e Song
)
3485
Whe
r
e
k
aR
are
the Lan
grang
e multiplie
rs.
Se
tting
the partial deriva
t
ives
of
L
ω
,b
,
e
,
a
respec
t to
ω
, b
,
e
and
a
to be 0,
we get:
T
N
-1
NN
01
0
b
=
1
Ω
+
Ι
y
α
(5)
With
N
1=
1
,
,
1
,
12
N
α
=
α
,
α
,
α
,
12
N
y=
y
,
y
,
y
,
and
Mercer’
s
con
d
ition is appli
ed within
the
Ω
matrix:
T
ij
i
j
Ω
=(
x
)
(
x
)
i
,
j
=
1,
2,
,
N
(6)
The output of
LS-SVM reg
r
ession i
s
:
N
ki
j
k=
1
y(
x)
=
a
K
x
,
x
+
b
(7)
Whe
r
e
k
a
and
b
are the sol
u
tions to the lin
ear sy
stem
. Note that in the ca
se of RBF Kernel,
one ha
s only
two addition
al tuning pa
rameter
whi
c
h
is
and
,
stand
s for the
weig
ht at
whi
c
h th
e te
st
ing e
r
rors
will
be
treate
d
in
rel
a
tion to
th
e sepa
ration
margi
n
, an
d
stand
s
fo
r
th
e
width of the kernel fun
c
tion
. The RBF Kernel is d
e
fine
d as:
2
2
ij
i
j
Kx
,
x
=
e
x
p
-
x
-
x
/
σ
(8)
2.3. Particle S
w
arm Opti
miz
a
tion
As s
h
o
w
n in
(6)
and
(9
),
and
are i
m
po
rtant pa
ram
e
ters in LS
-SVM model
whi
c
h
determi
ne th
e accu
ra
cy and g
ene
rali
zation
ability of LS-SVM model. Ho
wever, the
s
e
two
para
m
eters a
r
e given a
r
bit
r
arily in g
ene
ral LS-
SVM
model. In this section, a
new evol
utio
nary
comp
utation
call
ed p
a
rti
c
le
swarm
optimizatio
n
(PSO) [8] i
s
appli
ed to
obtain
opti
m
al
para
m
eters
of LS-SVM. PSO is an
evoluti
ona
ry comp
utation
techni
que b
a
s
ed o
n
swa
r
m
intelligen
ce. It has many a
d
vantage
s ov
er othe
r heu
ri
stic techniq
u
e
s. PSO algo
rithm ca
n exp
l
oit
the di
stribut
ed and
parallel
co
mputing capabilities,
to
escape local
opt
ima and quick
conve
r
ge
nce.
In PSO, individual
s are ca
lled pa
rticle
s
and the
pop
u
l
ation is
call
e
d
a swa
r
m. I
n
a n-
dimen
s
ion
a
l compl
e
x
se
ar
ch spa
c
e,
t
h
e
i
th
parti
cle update
s
its p
o
sition and speed with
th
e
rule
s as give
n belo
w
:
k+
1
k
k
k
k
k
id
i
i
d
1
1
i
d
i
d
2
2
g
d
i
d
V=
ω
V
+
c
r
(P
-
X
)
+
c
r
(P
-
X
)
(9)
k+
1
k
k+
1
id
id
id
X
=
X
+
V
,
i
=
1,
2
,
n
,
d
=
1,
2,
D
(10
)
Whe
r
e
T
ii
1
i
2
i
D
V=
(
V
,
V
,
V
)
is the
spe
ed vecto
r
of the
i
th particle,
T
ii
1
i
2
i
D
X=
(
x
,
x
,
x
)
is the
positio
n ve
ctor of
the
i
th particle,
i
ω
is the inertia weight,
1
c
and
2
c
are l
e
arnin
g
p
a
ra
m
e
ters,
1
r
and
2
r
are
ran
dom value be
tween 0 a
nd
1.
3. Modeling Based on AHP-PSO-LSSVM
for Port
CSP Demand Forcasting
3.1. Analy
s
is
of Port CSP Influential F
actor
s Ba
se
d on AHP
As the
po
rt e
n
terp
rises lo
cate in
coa
s
tal
are
a
s, th
e
surroun
ding
s
of the e
quip
m
ent a
r
e
very bad, an
d
the equi
pme
n
t must u
nde
rgo the b
ad
e
n
vironm
ent,
such as high
-l
ow
temp
eratu
r
e,
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3483 – 34
90
3486
vibration a
n
d
so o
n
. Besid
e
s, p
o
rt
take
s l
ong
-term continu
ous
ope
ratio
n
with a hi
gher
prod
uctivity, and the m
a
n
ageme
n
t leve
l of the
equip
m
ent mainte
n
ance pe
rsonn
el also
plays
an
importa
nt part. Therefore, the re
ason
s for CSP re
pla
c
ement are infl
uen
ced by ta
sks, equi
pme
n
t,
environ
ment, human, a
cci
d
ents an
d ma
n
y
other com
p
l
i
cated fa
ctors.
In ord
e
r to q
uantify these
factors fo
r th
e mod
e
l'
s in
p
u
t, we
analy
z
ed the
hi
stori
c
al
data
and
referred
to the opi
ni
ons
of the e
x
perts. Be
sid
e
s,
the auth
o
rs
have discu
s
sed with the
experts a
nd the que
stionn
aire ba
se
d o
n
AHP
wa
s distrib
u
ted to
30 manag
ers and
staffs, 28
effective
que
stionn
aire
s were coll
ecte
d
.
After
que
stionnai
re inve
stigation, d
a
ta analy
s
is a
n
d
weig
ht calcul
ation a
c
co
rdi
ng to
AHP
m
e
thod, th
e
d
e
s
cription
s an
d weight
s
of
each influ
enti
a
l
factor a
r
e list
ed as Ta
ble
1. In order t
o
elim
inate irrelevant noi
se and de
rive
the foreca
sti
n
g
result more a
c
curate, the
author
will first assume
five factors of the front as th
e input varia
b
le.
They are eq
u
i
pment wo
rking hou
rs, eq
uipment ha
nd
ling volume, failure time, m
a
intena
nce time
and fail
ure
ra
te. In the
nex
t experim
ent
se
ction, the
r
eal d
a
ta
of fo
ur fa
ctors, fiv
e
facto
r
s a
n
d
six
factors will b
e
tested to see whi
c
h d
a
taset ha
s the
higher
accu
racy, and the
n
the numb
e
r
of
input variabl
e
s
is dete
r
min
ed.
Table 1. The
Weig
hts of ea
ch Influential
Facto
r
s
The influential factors
Weights
Equipment w
o
rki
ng
hours
0.1525
Equipment handl
ing volume
0.1311
Failure time
0.1249
Maintenance time
0.1130
Failure rate
0.1022
CSP lifetime
0.0925
The historical req
u
irement at the s
a
me month
0.0879
Quantit
y of CSP i
n
one equipment
0.0757
Environmental factors
0.0720
Accidental factors
0.0482
3.2. The PSO-LSSVM Model for Port CSP
In orde
r to b
u
ild the LS-SVM model
for Po
rt CSP
,
we ma
ke the influential
factors
prop
osed ab
ove as the in
put ve
ctors of LS-SVM, and make
CSP deman
d as t
he output values
[9]. The LS-SVM model for port CSP de
mand is
sho
w
n in Figure 1.
Figure 1. LS-SVM Model for Port CSP
Dema
nd
The pa
ramet
e
rs
and
are
optimize
d
by PSO [10] with flowch
art sh
own in Fig
u
re
2.
w
o
rking hours
Handling volume
Failure time
Maintenance time
1
K(x
,
x
)
2
K(
x
,
x
)
M
K(
x
,
x
)
. . .
CSP demand
Input vectors
Kernel
Output values
Fai
l
u
re rate
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Dem
and Fo
re
ca
sting Mod
e
l
of Port Cr
itical Spare Pa
rts (Zhiji
e Song
)
3487
Figure 2. PSO Optimizatio
n
Flowcha
r
t o
f
LS-SVM
4. Experiments and
Disc
ussion
4.1. Data Pre
p
aratio
n
In this stu
d
y, the time serie
s
data
of po
rt
CSP co
nsum
ption alo
ng
wi
th the five influential
factors
cove
red from
Jan
u
a
ry, 20
08 to
De
cemb
er
,
2
012,
whi
c
h i
s
obtain
ed f
r
o
m
Qin
huan
gd
ao
Port. From the dataset, there a
r
e 60 sa
mples a
s
ea
ch month is a sampl
e
, and
the first 80% is
use
d
for train
i
ng whil
e the
balan
ce of 2
0
% is for te
sting. The mo
d
e
l is si
mulate
d with the LS
-
SVMlab1.8 toolbox in M
A
TLAB environment. Prio
r to training,
all input and output were
norm
a
lized
u
s
ing
fun
c
tion
scalefo
r
SVM. The
obje
c
tive is to i
ndep
ende
ntly no
rmali
z
e
e
a
ch
feature comp
onent
to
the spe
c
ified ran
g
e
0,
1
, as norm
a
lizatio
n tech
nique m
a
y improve th
e
predi
ction a
c
curacy an
d d
a
ta mining al
gorithm [11].
4.2. Parameters Op
timiza
tion
The
origi
nal
values of th
e
pa
ramete
rs i
n
PSO
are
shown in
Ta
bl
e 2. An
d
eval
uate the
desi
r
ed
optim
ization
fitness functio
n
fo
r
each p
a
rt
i
c
le as
th
e
M
ean
Square
Erro
r (MSE) over
the
data
set. Th
rough
the o
p
timization
ste
p
s
in
Figu
re
1
,
the optimal
para
m
eters
a
r
e o
b
taine
d
a
s
=
26.
8024
,
σ
=
1
.7686
.
Table 2. PSO
Paramete
rs
Parameter
value
Sw
ar
m size
30
Evolution genera
t
ions
300
Learning pa
rame
ters
1
c
1.5
Learning pa
rame
ters
2
c
1.7
numeric area of
[0.01,1000]
numeric area of
[0.01,100]
4.3. Ev
aluation Metric
s
For the
purp
o
se
of evalu
a
ting the p
r
o
pos
ed te
chni
que, two q
u
antitative evaluation
metrics are utilized, namel
y Mean Abso
lute Pe
rce
n
ta
ge Erro
r (MA
PE), and Mean Square Error
(MSE), whi
c
h
are define
d
a
s
follows:
set initial parameters of PSO
determine searc
h
space of
and
train model of LS
-SVM
compute values of fitness
select optimal value of ever
y partic
l
e
select optimal value of particle swarm
if meet the condition of ending?
update curr
ent s
peed and positio
n
No
Ye
s
obtain optimal
a
nd
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3483 – 34
90
3488
ˆ
n
ii
i=
1
i
y-
y
1
MA
P
E
=
1
0
0
ny
(11
)
ˆ
n
2
ii
i=
1
1
MS
E
=
y
-
y
n
(12
)
Whe
r
e
i
y
s
t
an
ds
fo
r ac
tu
a
l
va
lu
es
,
ˆ
i
y
stand
s for predi
cte
d
value
s
, an
d
n
sta
n
d
s
for th
e num
ber
of samples. It is generally cons
idered that the smaller the va
lue of
MAPE and MSE, the better
accuracy of the pre
d
ictio
n
.
4.4. Results and Disc
uss
i
ons
As
shown in
Figure 3, the
proposed model
PSO-LSSVM c
a
n obtain relatively ac
curate
predi
ction
s
fo
r CSP dema
nd, as 75% of the sa
mpl
e
s are pre
d
icted absol
utel
y correct and
the
error
of the rest is at mo
st 2. By using
the
same tes
t
samples
,
comp
a
r
ed
with
LS-SVM mo
de
l
using cross-v
a
lidation opti
m
izat
ion met
hod. Com
puti
ng the MAPE and MSE shown in Tabl
e 3, it
is obvious that approximat
e accuracy of PSO-L
SSVM is much better than LS-SVM with cross-
validation opti
m
ization.
Figure 3. The Compari
s
on
Chart of Predicted
Value and Actual Val
ue of PSO-LSSVM Model
Table 3. Tabl
e of Error
Co
mpari
s
o
n
model MSE
MAPE
PSO-LSSVM 0.4500
5.2944
LS-SVM
0.8913
8.6571
In the above
experim
ent, five influential factors
are ta
ken a
s
the in
put variable
s
.
In orde
r
to see if it is t
he best
choi
ce,
we
will
compare with the models
of
four factors
and
six factors of
the front in t
he re
sult of
AHP method.
The
co
mp
arison
ch
arts o
f
predi
cted v
a
lue a
nd a
c
t
ual
value of fou
r
factors m
odel
and
six facto
r
s
model
are
sho
w
n i
n
Fig
u
re
4 an
d Fig
u
re
5. And th
e
MSE comp
ari
s
on
of the th
ree m
odel i
s
sho
w
n i
n
Ta
ble 4. From t
hese fiure
s
a
nd table
s
, three
promi
nent fin
d
ing
s
ca
n be
observed.
First, it i
s
o
b
vious that LS
-SVM model
d
i
d
a g
r
e
a
t job
in po
rt
CSP
deman
d fo
re
ca
sting.
And the perfo
rman
ce of PSO is better th
an normal pa
ramete
r opti
m
ization m
e
thod.
Secon
d
, the
model
of six i
nput vari
able
s
h
a
s th
e b
e
st MSE, but Figure
3
and
5
illustrat
e
that five and six factors m
odel
s have th
e sam
e
predi
ction result. It mean
s that the sixth facto
r
CSP lifetime doe
s not pla
y
an importa
nt role in the
model. So, the dem
and f
o
re
ca
sting m
odel
still cho
o
se the first five factors in the result of AHP method.
Last, but not
least. Using t
he propo
sed
model, we ca
n pre
d
ict the
mean d
e
man
d
of the
spa
r
e p
a
rt
s
next month, and
set the
rang
e of sa
fety inventory, which will
effectively re
duce
inventory ba
cklog.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Dem
and Fo
re
ca
sting Mod
e
l
of Port Cr
itical Spare Pa
rts (Zhiji
e Song
)
3489
Figure 4. The
Compa
r
i
s
on
Cha
r
t of Pred
icted
Value and A
c
tual Value of Four F
a
cto
r
s
Model
Figure 5. The
Compa
r
i
s
on
Cha
r
t of Pred
icted
Value and A
c
tual Value of Six Factors Model
Table 4. Tabl
e of MSE Compari
s
o
n
Model w
i
th diffe
r
ent number
of fa
ctors
Four facr
ors
Five facrors
Six factors
MSE
0.6528
0.4500
0.4405
5. Conclusio
n
Spare
pa
rts
manag
eme
n
t ha
s al
way
s
been
a ve
ry i
m
porta
nt pa
rt
in fa
ctori
e
s,
esp
e
ci
ally
in port industry. Excess
ive spare part
s
will
cause backl
og
of th
e i
n
ventory
and insufficiency
will
cause termination of eq
ui
pment operation, leadi
ng t
o
loss.
In this paper, a PS
O-LSSVM m
odel
for port CSP i
s
pro
p
o
s
ed.
Unli
ke previo
us spar
e pa
rts dema
nd research, so
me i
n
fluential fact
ors
are
in
clude
d
as i
nput
s in
t
he m
odel,
wh
ich
ma
ke
s th
e mo
del
more comfort
to t
he
cha
r
a
c
teri
stics
of the port CSP. In additio
n
, a new PSO algorithm
i
s
introd
uced
to optimize th
e param
eters of
LS-SVM mod
e
l, and the ex
perim
ents
with real d
a
ta
in
Qinhua
ngd
a
o
port sho
w
that the fore
cast
accuracy of PSO-LSSVM method is be
tter than norm
al LS-SVM method.
The fo
re
ca
sting mo
del
of
this p
ape
r
ca
n be
provide
d
a
s
a
refe
re
nce
of
critica
l
sp
are
parts
man
a
g
e
ment in p
o
rt enterp
r
ises
to make
plan
ning a
nd red
u
ce
risks an
d co
sts. And
to
other type
of
comp
anie
s
which
also h
a
ve sp
are pa
rts pro
b
lem
s
, ch
ange
so
me o
f
the influenti
a
l
factors in the
model may
also a
d
ju
st to their
si
tuati
on. It is note
d
that the pa
per give
s a n
e
w
forecastin
g m
e
thod for
spa
r
e pa
rts dem
and, and the
method is
clo
s
er to the e
n
terp
rise pra
c
ti
ce.
Ackn
o
w
l
e
dg
ements
This work wa
s finan
cially suppo
rted by the
Education
a
l Commi
ssio
n of Hebei Province
of Chi
na
(Z
H2012
021
) a
n
d
Th
e
Natu
ral Sci
e
n
c
e F
ound
ation
of He
bei P
r
ovi
n
ce
Youth
F
und
(G20
112
031
9
5
).
Referen
ces
[1]
F
e
iLo
ng C
h
e
n
, YunC
hin
Ch
e
n
, JunYu
an K
u
o. Appl
yi
n
g
mo
ving
back-pr
op
agati
on
neur
al
net
w
o
rk
an
d
movin
g
fuzz
y
neur
on
net
w
o
r
k
to pr
ed
ict th
e re
quir
e
me
nt
of critica
l
sp
ar
e p
a
rts
. Expert
System
s with
Appl
icatio
ns
. 2
010; (37): 4
358
–43
67.
[2]
Somnath Muk
h
opa
dh
ya
y
,
Adr
i
ano O.Solis, R
a
fae S Guti
err
e
z.
T
he Accura
c
y
of Non-tra
d
i
t
iona
l versus
T
r
aditional Met
hods of F
o
rec
a
sting Lum
p
y
D
e
man
d
.
Journ
a
l
of F
o
recastin
g.
2012; (3
1): 721-7
35.
[3]
Andre
a
Bacc
h
e
ttin, Nico
l
a
S
a
ccan
i
. Spar
e
parts
cl
assifi
cation
an
d d
e
m
and f
o
recast
ing for
stock
control: Investi
gatin
g the ga
p bet
w
e
en res
e
a
r
ch and pr
actic
e
.
Omeg
a.
201
2; (40): 722
–73
7.
[4]
SG Li, X Kuo.
T
he inventor
y
mana
geme
n
t system for auto
m
obil
e
s
par
e p
a
rts in a centra
l
w
a
re
hous
e.
Expert Systems w
i
th Applicati
ons.
200
8; (34)
: 1144–
11
53.
[5]
Hua H, Z
h
a
ng
B. A h
y
br
id su
pport vector m
a
chi
nes
a
nd l
o
gistic regr
essio
n
ap
proac
h for
forecastin
g
intermittent de
mand of spar
e
parts.
Appli
ed
Mathe
m
atics a
nd Co
mputati
o
n.
2006; (1
81): 103
5–
48.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3483 – 34
90
3490
[6]
Su
y
k
ens JAK,
Vand
e
w
a
l
l
e
J. Least Sq
uar
es
Supp
or
t Vector Machi
ne
Cl
assifiers. Ne
ur
al Proc
essin
g
Letters. 199
9; 9(3): 293-
30
0.
[7]
Saat
y
T
L
.
T
he ana
l
y
t
i
c hier
arc
h
y
process. Ne
w
York: McGra
w
-
H
il
l. 198
0.
[8]
Kenn
ed
y J, Eb
erhart R.
Partic
le sw
arm
opti
m
i
z
at
io
n.
Proceedings of IEEE Inte
rnational Conferenc
e on
Neur
al Net
w
o
r
ks. Perth: I
EEE Press. 1995: 1
942-
194
8.
[9]
Peng B
a
i, Xibi
n Z
hang, Bi
n Z
hang et a
l
.
Supp
ort Vector Machi
ne a
nd it
s Applic
atio
n i
n
Mixed Ga
s
Infrared Spectr
um A
nalys
is
. Xi'
an Electro
n
ic
Sienc
e &T
echnolo
g
y
Univ
ersit
y
Press. 20
08: 69-7
6
.
[10]
Qian
w
e
n
Xi
an
g, Yukun
Sun,
Xi
aofu J
i
.
Mo
deli
ng In
ducta
nce for Be
ari
n
gless Sw
itche
d
Re
luctanc
e
Motor based on PSO-LSSVM
. Chines
e Cont
rol an
d Decis
i
o
n
Confer
enc
e (CCDC), 20
11:
800-
803.
[11]
LA Sha
l
ab
i, Z
Shaa
ba
n, B Kasasb
eh.
Dat
a
minin
g
: a pre
p
rocess
ing
en
gin
e
.
J Co
mp
u
t
Sci.
2006; 2:
735-
739.
Evaluation Warning : The document was created with Spire.PDF for Python.