TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.6, Jun
e
201
4, pp. 4840 ~ 4
8
4
8
DOI: 10.115
9
1
/telkomni
ka.
v
12i6.584
3
4840
Re
cei
v
ed
Jan
uary 15, 201
4
;
Revi
sed Ma
rch 1
8
, 2014;
Acce
pted Ma
rch 3
0
, 2014
Passenger Flow Forecasting using Support Vector
Regression for Rail Transit
Bin Xia*
1
,
Fan
y
u
Kong
2
,
Song
y
u
an Xie
1
1
Logistic En
gin
eeri
ng Un
iversi
t
y
, Chi
na.
2
Chon
gqi
ng T
r
ansp
o
rt Plan
ni
ng Institute, Ch
ina
*Corres
p
o
ndi
n
g
author, em
ail
:
xia
b
i
n12
6@
12
6.com
A
b
st
r
a
ct
Supp
ort vector
regr
essio
n
is
a pr
o
m
isi
ng
method
for th
e f
o
recast
of p
a
s
s
eng
er flow
be
cause
it
uses a risk f
u
nction c
onsisti
ng of the
e
m
pirica
l error
a
nd a r
egu
lari
zed ter
m
w
h
ich
is bas
ed o
n
the
structural risk mi
ni
mi
z
a
t
i
o
n
princip
l
e. In this pap
er,
the pred
iction
mod
e
l of urba
n
rail trans
it passen
ger flo
w
is construct
ed.
It is to b
u
il
d
an ur
ba
n ra
il t
r
ansit p
a
sse
ng
er flow
foreca
st mo
del
an
d
select th
e o
p
ti
ma
l
para
m
eters fro
m
th
e su
pp
ort vector regr
essi
on thr
oug
h the
varia
b
le
metri
c
metho
d
to o
b
tain t
he
mi
ni
ma
l
valu
e fro
m
the
LOO error b
oun
ds. T
he
p
a
ssen
ger fl
ow
is forec
a
st b
y
means
of b
o
th su
pport v
e
cto
r
regressi
on
met
hod
an
d
BP ne
ural netw
o
rk method, an
d
the
results sh
ow
th
at the s
upp
ort
vector regr
essi
on
mo
de
l has s
u
ch theor
etical
super
iori
ty as
mi
ni
mi
z
e
d stru
ctural risk,
thu
s
havin
g a h
i
gher for
e
casti
n
g
accuracy
under sm
all sample conditio
ns f
o
r short-ter
m
r
a
il tra
n
sit pas
seng
er flow
, w
h
ich pre
d
icts
the
pro
m
isi
ng fore
casting p
e
rfor
ma
nce that the met
hod h
a
s.
Ke
y
w
ords
: rail
transit, passen
ger flow
, suppo
rt vector regres
sion, le
ave-o
n
e
-
bou
nd
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
Rail t
r
an
sit p
a
ssen
ger flo
w
fo
re
ca
st is
the fo
undati
on a
nd
qua
n
t
itative basi
s
for th
e
planni
ng, co
nstru
c
tion, o
peratio
n and
managem
en
t
of rail transpo
rtation, a
nd the sci
en
tific
accuracy of the re
sult
s is
dire
ctly relate
d to
the pre
p
a
ration
of rail
transportatio
n
planni
ng a
nd
approval im
plementatio
n
of the project, it
det
ermin
e
s th
e
developm
e
n
t mode of
rail
transpo
rtation
,
road network scale, line a
lignment
s,
hu
b setting
s an
d layout of interio
r
sp
ace.
Curre
n
tly, the rail tra
n
sit fo
recast m
e
tho
d
s
can
be div
i
ded into th
re
e cate
gori
e
s:
4-sta
g
e
forecast met
hod ba
se
d o
n
traffic dem
and an
alysi
s
, disag
g
re
gat
e model fore
ca
st method,
and
forecast meth
od not ba
sed
on pre
s
e
n
t pa
sseng
er flow
distrib
u
tion.
The first fore
ca
st method i
s
a conve
n
ti
o
nal method commonly use
d
both at home and
abro
ad, which is a
c
hi
eved
by colle
cting
or u
s
ing
re
si
d
ent travel
survey data,
to split the tran
sp
ort
mode
s in ord
e
r to forecast
the urban rai
l
transit
passenge
r flow o
n
the basi
s
of foreca
sting
the
total dema
n
d
of urb
an
pa
sseng
er tran
spo
r
t. It can
affect the fo
reca
sting
accura
cy to a
ce
rtain
degree
due t
o
heavy
workload of i
n
vest
igation, lo
w d
a
ta utilizatio
n
,
failure to ta
ke into
a
c
cou
n
t
the rea
c
tion
of traffic on land u
s
e. In view of
the disa
dvantag
e
s
of the 4-st
age metho
d
, th
e
disa
ggregate
model wa
s then rai
s
ed, whi
c
h is cl
ass I model in the unit of individual
s that
actually g
ene
rate tran
spo
r
t activi
ties
, to forec
a
s
t
the pers
o
nal travel ac
tivities
s
e
parately, and
make
statistics a
s
pe
r travel di
strib
u
tion
, tran
spo
r
tation m
ode
s
an
d tra
n
sit li
ne
s re
sp
ectively, in
orde
r to get the total amount of traffic dema
nd. L
i
terature [1] use
s
the disaggregate m
odel
based on the
4-sta
ge met
hod to predict the passen
ger volum
e
o
f
urban rail transit. The a
b
o
ve
two meth
od
s are fo
cu
se
d
on the
mid-/
l
ong-te
rm fo
reca
st of p
a
ssen
ger flow,
but they can
not
obtain effecti
v
e foreca
st re
sults in the
ca
se of
dynami
c
ch
ang
es in
the recent pa
sseng
er flow.
The third foreca
st metho
d
doe
s not t
a
ke
i
n
to a
c
count the p
r
e
s
ent p
a
sse
n
ger flo
w
distrib
u
tion, a
nd usually, the fore
ca
st method i
s
to transfe
r the
pre
s
ent p
a
ssenge
r flow of
the
relevant bu
s lines an
d bike traffic to th
e rail li
nes, so as to get a virtual base
-
year rail tran
sit
passe
nge
r flo
w
; and it d
e
te
rmine
s
the
growth
rate
of
passe
nge
r rai
l
tran
sit pa
ssenge
r flow,
a
n
d
cal
c
ulate
th
e
l
ong-te
rm rail transit pa
sse
nger
flow
based up
on the
h
i
story d
a
ta an
d growth l
a
w
of
relevant
bu
s line
s
. Lite
rat
u
re
[1] ap
pli
e
s
gray theo
ry to m
a
ke a
time
se
rie
s
forecast
on
the
annu
al urban
rail pa
sseng
er flow.
Currently, mo
st o
f
these meth
ods
unde
r re
sea
r
ch are to
forecast an
d analyze the chang
es in p
a
s
seng
er
flow
from the long
-term p
e
rspe
ctive.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Passeng
er Fl
ow Fo
re
ca
sting usi
ng Sup
port Ve
cto
r
Regre
s
sion fo
r Rail Tran
sit (Bin Xia)
4841
The f
o
re
ca
st
result
s of
sho
r
t
-
t
e
rm p
a
s
s
en
ger flo
w
to some
extent determine the
prep
aration a
nd adju
s
tme
n
t
of transp
o
rt
orga
niza
tion
plan
s and
co
ntingen
cy pla
n
s for
rail tran
sit.
If, in the
c
a
se of
dramatic
c
h
anges in pas
s
eng
er flow
during holidays and
major event
s, a
relatively accurate fo
re
ca
st can h
e
lp p
r
ovide effectiv
e de
cisi
on su
pport o
n
the
adju
s
tment of
th
e
above t
w
o
pl
ans.
Literature [2] ap
plie
s the F
u
zzy B
P
neu
ral
net
work mod
e
l t
o
ma
ke
a
da
ta
mining
predi
ction on
the
scale of
rail
way
pa
sseng
er
flow, b
u
t be
ca
use
of the
th
eoreti
c
al
defe
c
ts
the ne
ural n
e
t
work
ha
s, su
ch
metho
d
s h
a
ve thei
r final
sol
u
tion
too
much
d
epen
d
ed u
pon
initia
l
value an
d ov
er le
arni
ng, while havin
g lo
cal mi
nimality duri
ng the t
r
aining
pro
c
e
s
s, rel
a
tively low
rate of
conve
r
gen
ce, difficult cho
o
se of
hidde
n net
wo
rk
units, et
c., so the fo
re
ca
sting
results
are
not so ea
sy to be pro
m
ote
d
.
As a n
o
n
-
lin
ear m
odel fo
recastin
g me
thod, su
ppo
rt
vector
re
gre
ssi
on
(SVR)
has th
e
followin
g
adv
antage
s
com
pare
d
with t
he ne
ural
ne
tworks, in
clu
d
ing glo
bal
optimum al
ways
available
during the train
i
ng p
r
o
c
e
ss,
high
gene
ralizatio
n cap
ability, soluti
on spa
c
e
wi
th
spa
r
e
s
ity, hig
h
rate
of co
n
v
ergen
ce
and
good fo
re
ca
sting p
e
rfo
r
m
ance un
der the small
sam
p
le
con
d
ition. Short-te
rm pa
ssen
ger flo
w
forecast is it
self a com
p
l
e
x rando
m and no
n-lin
e
a
r
pro
c
e
ss,
a
nd its
variation h
a
s
a high un
certainty.
As a
new
algo
rith
m ba
sed u
p
o
n
the st
ru
ctural
risk minimi
zation prin
cipl
e, SVR has a high accura
cy and short time in
foreca
sting
the
passe
nge
r flow un
der
small sam
p
le
con
d
ition
s
. This pa
per i
s
based on th
e sup
port ve
ctor
reg
r
e
ssi
on al
gorithm, and
build
s
foreca
sting mod
e
ls
for
s
h
ort-term rail tran
sit
pa
sseng
er flow b
y
analyzi
ng a
n
d
mining
hi
story data
and
laws of
p
a
ssen
ger flow,
and it p
r
ovid
es a
ne
w
wa
y of
thinkin
g
in
ca
rrying
out tra
n
sp
ort o
r
ga
ni
zation,
a
d
ju
sting op
erationa
l pro
g
ra
ms
a
nd p
r
ep
arin
g
the
contin
gen
cy plan in a scie
ntific and re
a
s
on
able
way.
2. Support V
ector
Regr
es
sion Algorithm
Brief de
scripti
on of
sup
p
o
r
t vector re
gre
ssi
on al
go
rith
m [3] and
set of the give
n
training
sampl
e
as:
1,
1
k
k
{
(
x
y
),..
.,(
x
,
y
)
}
,
and the opti
m
al pro
b
lem i
s
:
,,
,
mi
n
ii
wb
22
11
1
()
22
2
kk
T
ii
ii
CC
ww
(1.a)
.
.
t
s
()
T
ii
i
i
by
wx
(1.b)
i
,0
i
,
1
,
...
,
ik
(1.c)
In Form
ula
(1
), the first item is ai
med to
maximize
th
e cla
s
sificatio
n
interval,
sm
oothing
the fun
c
tion;
the seco
nd
a
nd third item
s a
r
e
error
lo
ss fun
c
tion
s, for the
purpo
se
of re
du
cin
g
er
rors
, cons
tant
C
>
0, whi
c
h i
s
the d
e
g
r
ee of pe
nalty
beyond th
e e
rro
r
sampl
e
,
*
,,
ii
i
is the
introdu
ce
d
slack vari
able
s
. The
sampl
e
input p
o
ints are
mapp
ed
by the fun
c
tio
n
into a
high
d
i
me
ns
io
na
l s
p
ac
e
for
line
a
r
re
gr
es
s
i
on
is the in
sensitive lo
ss
function, an
d
the error lo
ss
function u
s
e
s
the squa
re
s and form
s of the minimum
squ
a
re fun
c
ti
on.
Usi
ng the d
uality princi
pl
e, the Lagra
nge optimi
z
a
t
ion method
conve
r
ts the
above
optimizatio
n probl
em into i
t
s dual proble
m
:
,
mi
n
ii
11
1
()
()
(
)
(
)
2
kk
T
ii
i
i
i
ii
y
K
,
(2.a)
..
s
t
1
()
0
k
ii
i
,
0,
ii
,
1
,
...,
.
ik
(2.b)
Whe
r
e,
ii
is the introd
uced
Lagrang
e m
u
ltiplier,
i
or
i
in the form
ul
a doe
s n
o
t
equal to zero,
and the co
rresp
ondi
ng sa
mple data is
sup
port vecto
r
.
Kernel Fu
nction
()
(
)
(
)
T
ij
i
j
xx
x
x
K
,
C
K
KI
in whi
c
h
I
k dime
nsi
onal unit
matrix and th
e basi
c
kern
e
l
function
s include the follo
wing fou
r
types:
i) linear k
e
rnel func
tion:
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4840 – 4
848
4842
()
T
ij
i
j
xx
x
x
K
,
(3)
ii) Polynomial
Kernel Fun
c
t
i
on:
()
(
)
,
0
Td
ij
i
j
r
xx
x
x
K
,
(4)
iii) Radi
al basis function (RBF):
2
2
()
e
x
p
(
)
ij
i
j
xx
x
x
K
,
(5)
iv) Sigmoid kernel fun
c
tion
:
()
t
a
n
h
(
)
i
T
ij
j
r
xx
x
x
K
,
(6)
Literatu
re [4]
study
sho
w
s:
Linea
r
ke
rne
l
function
can
not deal
with
the no
nline
a
r
inp
u
t
values;
RBF
ke
rn
el fun
c
t
i
on h
a
s le
ss high
-d
im
en
sional ke
rnel
para
m
eters
than polynom
ial
kernel fu
nctio
n
s, a
nd i
n
the
SVM traini
ng
pro
c
e
s
s,
the
use
of p
o
lyno
mial kernel fu
nction
re
qui
re
s
much mo
re training time than that for the use of
RBF
kern
el functi
on; upon u
s
in
g Sigmoid ke
rnel
function, cert
ain pa
ramete
rs h
a
ve erro
r values,
so this pa
pe
r ad
opts RBF
as the input ke
rnel
function.
w,b
Whi
c
h can be
calculated
wi
th the followin
g
formula:
1
()
(
)
k
ii
i
i
w
x
,
(7)
()
l
l
by
K
,
whe
r
e
0
l
(8)
()
l
l
by
K
,
whe
r
0
l
,
(9)
Obtain the re
gre
ssi
on fun
c
tion
()
(
)
ii
f
wb
xx
.
3. Stud
y
of the Selection
of Model Pa
ramete
rs
Mean
while, t
he fore
ca
stin
g perfo
rma
n
ce of sup
port
vector
reg
r
e
s
sion i
s
sen
s
itive to the
sele
ction
of p
a
ram
e
ters, b
u
t the ch
oice
of model p
a
rameters a
r
e
mostly mad
e
throug
h a
sim
p
le
cro
s
s-vali
dati
on method (CV) or geneti
c
algorith
m
(G
A). Literature
[4-6] Studies
sho
w
that these
two method
s have the sh
ortco
m
ing
s
of
overly l
ong trainin
g
time durin
g the se
lection p
r
o
c
e
ss,
CV can o
n
ly
use
so
me of
the sample
s for the
ca
li
b
r
ation
of pa
rameters d
u
ri
ng the
sel
e
ct
ion,
while GA, in
the sele
ction pro
c
e
ss, cannot
obtain
the impact of a single p
a
ram
e
ter on
the
forecastin
g p
e
rform
a
n
c
e
b
y
the optimi
z
i
ng p
r
o
c
e
ss,
but in the
u
s
e of Le
ave-o
ne-o
u
t (LOO
) for
para
m
eter
se
lection, the o
p
timal para
m
eters
c
an b
e
obtaine
d thro
ugh se
eki
ng the minimization
value from th
e least u
ppe
r bound
of the gene
rali
za
ti
on erro
r of the su
ppo
rt vector
ma
chin
es.
Comp
ared wi
th the above
two metho
d
s,
it has many
advantag
es
such a
s
small
e
r time cost a
nd
many appli
c
a
t
ion para
m
ete
r
s.
The suppo
rt vector regressi
on al
gorith
m
relie
s upo
n the choi
ce
of model pa
rameters,
and the a
d
ju
stable m
odel
para
m
eters of SVR are
error p
enalt
y
factor
C
, k
e
rnel
func
tion
para
m
eter
2
and insensitive loss function
, whe
r
e
C a
nd
2
are the i
ndep
ende
nt variable
s
for the abo
ve-mentio
ned
kern
el function
C
K
KI
, and all
param
eters are unifo
rmly
recorded a
s
.
3.1. LOO Error Bound o
f
Support Vec
t
or Re
gres
sion
In this
cha
p
ter, the mi
nim
u
m e
rro
r b
o
u
nd meth
od
of LOO
(L
eave
-
one
-o
ut) i
s
u
s
ed
for
para
m
eter se
lection
and
solution, an
d
LOO
error is
a qu
antitative crite
r
ia
used
to characte
ri
ze
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Passeng
er Fl
ow Fo
re
ca
sting usi
ng Sup
port Ve
cto
r
Regre
s
sion fo
r Rail Tran
sit (Bin Xia)
4843
the de
gre
e
of
excell
en
ce
o
f
the supp
ort
vector
ma
chi
ne al
gorith
m
,
whi
c
h i
s
defin
ed a
s
:
sele
cti
n
g
a sam
p
le in t
u
rn from
k
training
sam
p
l
e
s a
s
a te
st sampl
e
, and
obtain the re
gre
ssi
on fun
c
tion
with the rem
a
ining
k
-1 tra
i
ning sample
s as a trai
nin
g
set, whi
c
h
are sub
s
tituted into the inp
u
t
vector
of the
test sample
to obtain
t
he
predi
cted
val
ue of the
sa
mple. Repe
at
k
times
to
obtain
the predictive
value of e
a
ch trainin
g
sa
mple, and
fin
a
lly to obtain
the leave-one
-out e
r
ror val
ue,
with its expre
ssi
on a
s
follo
ws:
1
LOO
(
)
k
tt
t
f
xy
(10)
In this
formula,
()
t
f
x
and
t
y
are
re
spe
c
tively th
e p
r
edi
cted
value
and
the
true
value.
Whe
n
cal
c
ul
ating the
LO
O e
rro
r f
r
om
the trainin
g
set, it i
s
n
e
cessary
to u
s
e the
algo
rith
m
k
times for the training
set that contain
s
k
-1 sampl
e
s, to
obtain
k
reg
r
ession fun
c
tion values, so
it
is a
heavy workl
oad. T
h
e
r
efore, it i
s
ne
ce
ssary
to
give up the
accurate
cal
c
ul
ation of the
L
O
O
error valu
e, a
nd to use its
uppe
r bo
und
value that
is
easily calcula
t
ed instea
d. Mean
while,
such
a pro
perty ca
n be u
s
ed th
at the LOO u
pper
boun
d
i
s
the integ
r
al
function of the pa
ramete
r,
,
and th
e p
a
ra
meter val
ue
can
be
obtai
ned
by u
s
ing
the va
riabl
e
metri
c
m
e
th
od fo
r the
L
OO
uppe
r bou
nd,
so the para
m
eter selecti
on can b
e
attributed to the optimizatio
n of seeki
ng
the
LOO up
per b
ound
s.
Comp
ared wi
th the other LOO bou
nd
s, the radi
al interval upper bo
u
nd of LOO ha
s su
ch
advantag
es
as
simpl
e
op
eration
an
d l
e
ss p
r
omotio
n erro
rs, et
c.
, so the
re
se
arch a
dopt
s the
LOO radial in
terval uppe
r b
ound [3], with
its expre
ssio
n
as:
2*
4(
)
T
Rk
e
(11)
Whe
r
e,
T
e
is th
e
vecto
r
with t
he valu
e of
1
;
R
is the
minim
u
m hype
r-sp
here
radiu
s
t
hat
the input ma
pping fu
nctio
n
is
contai
ne
d in the hi
gh
-dime
n
si
onal spa
c
e, wi
th its expressio
n
as
follows
:
m
i
n
(
)
,
1
,
...
,
i
RR
R
i
k
x
c
c
(12)
Her
e
,
c
is th
e
cente
r
of the hypersph
e
re.The im
p
r
oved
input mappi
n
g
function i
s
defined
as:
()
(
)
T
i
ii
e
C
xx
(13)
Whe
r
e,
i
e
is the unit vector,
()
i
x
is the input m
appin
g
functi
on in Form
ula
(1).
3.2. Gradient Calculation
Whe
n
o
b
taini
ng the
optim
al pa
ram
e
ter of the m
o
d
e
l by u
s
ing
the
DFP met
hod, the
gradi
ent of
2
R
and
*
in Formul
a (11
)
.
2
R
is the optimal
value [13] of
the followin
g
issue
s
:
ma
x
11
1
(,
)
(
,
)
kk
k
ii
i
i
j
i
j
ii
j
KK
xx
xx
(13a)
1
1
k
i
i
,
0
i
,
1
,
...
,
ik
(13b)
So the gradi
e
n
t of
2
R
ca
n be e
x
presse
d as:
2
11
1
(,
)
(,
)
kk
k
ij
ii
ii
j
ii
j
K
K
R
xx
xx
(14)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4840 – 4
848
4844
The follo
wing
is the
deriva
t
ion of
*
. For the optimal
solution that
meets th
e
dual p
r
o
b
lem
(2), F
o
rm
ula
(9) a
nd F
o
rm
u
l
a (1
1)
are
e
s
tablish
ed a
n
d
can
be
written in the
matrix
form:
ˆ
0
b
p
G
(15)
Whe
r
e
,
T
e
G
0
e
K
,
ˆ
,
ˆ
0
ˆ
,0
ii
i
ii
y
p
y
,
. It is easy t
o
kno
w
that
the
followin
g
formula is e
s
tabl
ishe
d.
*
ˆ
()
ii
i
i
(16)
Whe
r
e
ˆ
0
1,
ˆ
0
1,
i
i
i
, for the parameter
except
2
,
C
, by differentiati
ng from
Formul
a (1
5),
the following
formula i
s
est
ablished:
ˆ
ˆ
0
T
b
b
G
G
(17)
The above fo
rmula i
s
deformed, and the
followin
g
formula ca
n be
obtaine
d:
1
ˆ
ˆ
K
G
(18)
Whe
n
the parameter
is
, there is the foll
owin
g equati
on that hold
s
:
1
ˆ
p
G
(19)
3.3. DFP Alg
o
rithm
Let
f
be the fun
c
tion of the d
e
sired p
a
ra
m
e
ter
k
, and here, the BFGS variable m
e
tri
c
algorith
m
is use
d
to solve the param
eter. Let
k
be the param
ete
r
of the k-th iteration, and
()
k
f
is the obje
c
tive function, the algo
rithm i
s
as follo
ws:
a.
Cal
c
ulate the
sea
r
ch directi
o
n
,
()
k
k
pH
f
.
b. M
a
ke o
n
e
-dim
en
siona
l se
arch
in th
e p
-
directio
n, and
dete
r
mi
ne the
optim
um
step
siz
e
based o
n
the followin
g
formula:
mi
n
(
)
k
f
p
(20)
From the a
b
o
v
e formula
,
the next para
m
eter poi
nt
1
kk
p
can be obtai
ne
d.
c. Calcul
atio
n
1
0
,
TT
T
T
k
TT
T
k
k
sy
y
s
ss
IH
I
i
f
y
s
ys
ys
ys
H
H
(21)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Passeng
er Fl
ow Fo
re
ca
sting usi
ng Sup
port Ve
cto
r
Regre
s
sion fo
r Rail Tran
sit (Bin Xia)
4845
Whe
r
e
,
1
kk
s
,
1
()
(
)
kk
yf
f
,
k
H
is the unit matrix tak
e
n from the firs
t
iteration.
d. The iteration termin
atio
n con
d
ition is
as follo
ws:
1
1
()
(
)
()
kk
k
ff
f
(22)
In the iteratio
n pro
c
e
s
s, seek the
opt
imal sol
u
tion
from the log
a
rithm value
of the
para
m
eter
2
ln
,
l
n
,
l
n
C
,
and in the iteration algo
rith
m, the gradie
n
t is
ln
f
f
.
4. Instanc
es
of Urb
a
n Rail Transit Pas
senger Flo
w
Foreca
st
4.1. Model Constr
uction
The p
a
sse
n
g
e
r flo
w
forecast of
sho
r
t-t
e
rm
urb
an
ra
il tran
sit is
b
a
se
d on
the
dynamic
cha
nge
s in hi
stori
c
al pa
sse
nger flo
w
dat
a, to det
ermi
ne the histo
r
i
c
al data that i
m
pact the fut
u
re
passe
nge
r fl
ow
as the in
put dime
nsi
o
n of the
mod
e
l, and
the p
r
edi
ctive valu
e a
s
the
out
put
dimen
s
ion
of
the mod
e
l through
analy
s
i
s
an
d a
rra
ng
ement of the
time se
rie
s
d
a
ta of pa
ssen
ger
flow. If there
are
n
histo
r
i
c
al data
for th
e sele
cted
m
odel
,
then the model is
c
o
ns
truc
ted to
mak
e
reg
r
e
ssi
on fo
reca
sts on
the
n+1 di
men
s
i
onal
hyper
pla
ne. Th
e
spe
c
i
f
ic forecastin
g mo
del
ca
n
be
expre
s
sed a
s
:
12
(
1
)
(
)
(
1)
.
.
.
(
1)
ii
i
n
i
vt
b
v
t
b
vt
b
v
t
n
(23)
Whe
r
e
,
j
b
is the weightin
g coefficient,
()
1
i
vt
is the rail tran
si
t passenge
r flow for the
t+1
perio
d. The
p
a
ram
e
ters (
2
,,
C
) i
n
the supp
ort
vector
reg
r
e
ssi
on m
odel
has
a si
gnificant
impact
on the
fore
ca
sting p
e
rform
a
n
c
e o
f
the model
,
and the
sel
e
ction of pa
ram
e
ters is
usual
ly
made
with th
e Cro
s
s Valid
ation meth
od,
that is, to
di
vide the fo
re
ca
sting
sam
p
le value
s
into
m
grou
ps,
and
sele
ct a
grou
p of p
a
ra
met
e
rs an
d tr
ain
the m-1 g
r
ou
p of d
a
ta fro
m
it, and th
e
rest
one serve
s
as the
che
cksum valu
e of forecast
in
g perfo
rman
ce
unde
r that p
a
ram
e
ter. After
several traini
ngs a
nd verifi
cation
s, an o
p
timal
set of
para
m
eters
can be d
e
term
ined to fore
cast
the future value.
4.2. Forecas
ting Results and An
aly
s
is
This paper is
foc
u
s
e
d on the
s
u
pport ve
c
t
or
regress
i
on algorithm, us
ing
LIBSVM
softwa
r
e
pa
ckag
e to fore
ca
st the hi
storical
data
o
f
passen
ger
flow on
a certain
route
of
Shangh
ai su
bway on
a d
a
ily basi
s
, where th
e train
i
ng sa
mple
set is the hi
storical data fo
r the
firs
t five week
s
,
the forecas
t
ing
s
a
mple s
e
t is
the 7
passen
ger v
a
lue
s
for the
sixth wee
k
s, and
the nu
mbe
r
o
f
cross-vali
da
tion group
s,
m, is
5,
at th
e same
time, ch
oo
se th
e
3-laye
r BP n
eura
l
netwo
rk m
e
th
od for control
l
ed trial, and
use the av
e
r
age rel
a
tive erro
r, root me
an sq
uare error,
maximum rel
a
tive erro
r, and minimum
relative error a
s
the evaluati
on indexe
s
.
With the
cro
s
s-validatio
n
pro
c
e
ss, th
e
sup
p
o
r
t vector regressio
n
pa
ramete
rs can
be
ultimately de
termine
d
; BP neural n
e
t
work pa
ram
e
ters
ch
oo
se them as:
input layer 7
,
interme
d
iate l
a
yer 13, o
u
tp
ut layer 7; Le
arnin
g
Netwo
r
k O
p
e
r
ators
as 0.8, da
mp
ing co
efficien
t
as 0.1, error adjustme
n
t factor a
s
0.2
,
error
o
b
je
ctive function as 0.05, and
the regre
s
si
on
forecastin
g re
sults i
s
sh
own in Table 1
2
(
,
,
)
(7
1.
1
,
5
.
4
,
0.
06
3
)
C
.
Table 1. Fo
re
ca
sting Samp
le Re
sult Values
No.
Actual value/person
BP forecast va
lue/person
SVR forecast value/person
36 599956
521313
555779
37 549094
525005
535026
38 557770
535809
536650
39 545399
543833
533642
40 604255
640660
601646
41 594073
640129
563103
42 564897
556048
544003
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4840 – 4
848
4846
Figure 1
sho
w
s th
e rel
a
tive error a
b
solute value f
unctio
n
curv
e for the fo
reca
sting
reg
r
e
ssi
on re
sults i
n
both
ways. As
se
e
n
from t
he fi
gure, ex
cept
for a few
poi
nts, the rel
a
tive
error ab
sol
u
te values for t
he su
ppo
rt vector
re
g
r
e
ssi
on (SVR) are
mostly small
e
r than the e
r
ror
values of the
BP neu
ral
n
e
twork. T
able
2
com
pares
t
he st
at
i
s
t
i
cal
re
sult
s f
r
om
f
o
rec
a
st
in
g t
h
e
relative e
r
ror ab
solute val
ues und
er th
e two
meth
ods, and the
statistical results in thi
s
table
sho
w
that except for the
minimum rela
tive errors,
the rest of the statistical indi
cators a
r
e lo
wer
than the latter, and thei
r forecasting
result
s are
re
l
a
tively stable
,
with less fl
uctuatio
n, wh
ich
indicates that
the forecasti
ng perfo
rma
n
c
e of SVR is
better than th
e BP neural n
e
twork.
Table 2. Statistics of Relati
ve Error F
o
re
ca
sting
Indicators BP
SVR
Average relative error
(%
)
5.2957
2.8052
Root mean squa
r
e
erro
r(
%)
6.9827
2.0292
Maximum relative erro
r (
%
)
7.75
5.3633
Minimum relative erro
r (
%
)
0.229
0.4318
Figure 1. The Abs
o
lute Relative Errors
of
Forecas
t
Samples
as
a Func
tion of t
4.3. Parameter Discu
ssio
n
Figure 2. The
Relation
ship
betwe
en the
Average
Rela
tive Error for
Fore
ca
sting
Sample
s and
the Paramete
r
C
0
2
0
4
0
6
0
8
0
1
00
12
0
140
0
5
10
15
20
25
C
0.15
36
37
38
39
40
41
42
0.00
0.03
0.06
0.09
0.12
t
/d
SV
R
BP
The Absolute rel
a
tive error(%
)
The averag
e rela
tive error (
%
)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Passeng
er Fl
ow Fo
re
ca
sting usi
ng Sup
port Ve
cto
r
Regre
s
sion fo
r Rail Tran
sit (Bin Xia)
4847
Based
upo
n t
he pa
ram
e
ters obtai
ned f
r
om t
he o
p
timization
LOO
u
pper bou
nd p
r
ocess,
make
an a
n
a
lysis
of the
impact of t
he chan
ge
s in a si
ngle
para
m
eter
o
n
the fore
ca
sting
perfo
rman
ce.
Ch
oo
se th
e
optimal
pa
rameter
2
ln
,
l
n
,
l
n
C
=
(3.8, 0.92,
-2
.13), that i
s
,
2
(,
,
)
C
= (44.
7, 2.50, 0.12) serve
s
as the fixed c
ondition for th
e study of sin
g
le paramete
r
s.
As ca
n be se
en from Fig
u
re 2, whe
n
2
and
are fixed,
the averag
e relative erro
r of
the fore
ca
stin
g sam
p
le de
crea
se
s with t
he incre
a
se o
f
C
, and it co
mes to a
sta
b
le statu
s
wh
e
n
increa
sing to
the vicinity of the optimal
p
a
ram
e
ter; a
s
see
n
from Fig
u
re 3, when
incr
ea
se
s,
t
he
averag
e rel
a
tive erro
r of the fore
ca
stin
g sam
p
le
ha
s a rel
a
tively stable chan
ge wh
en it is less
than 0.154, b
u
t it begins t
o
increa
se
ra
pidly whe
n
it is greater th
a
n
0.154;
in Fi
gure
4, with the
incr
ea
se of
2
,the average relative error
of the foreca
st
ing sampl
e
is in a rel
a
tively stable
cha
nge. As
can be
see
n
from the a
c
curacy chang
e curve of the fo
recastin
g sa
mple, the three
para
m
eters
a
r
e in
the vi
cin
i
ty of the opti
m
um valu
e,
with a
large
a
llowa
ble
ran
g
e
an
d a
rel
a
tively
small chan
ge
in accura
cy, whi
c
h is
simil
a
r
to the res
e
arch res
u
lt
s
in Literature [3].
Figure 3. The
Relation
ship
with the Average Relati
ve Erro
r of the Fore
ca
sting Sa
mple with the
Parameter
Figure 4. The
Relation
ship
with the Average Relati
ve Erro
r of the Fore
ca
sting Sa
mple with the
Paramete
r
0.
00
0
.
0
5
0
.
1
0
0.
15
0
.
2
0
0
.
25
0.
30
0.
35
2
4
6
8
10
12
14
16
18
20
02468
1
0
1
2
0
2
4
6
8
10
12
14
16
18
20
2
The averag
e rela
tive error (
%
)
The averag
e rela
tive error (
%
)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4840 – 4
848
4848
5. Conclusio
n
This pap
er a
nalyze
s
a
nd discu
s
ses
se
veral
m
odel
s of the u
r
b
a
n
rail tran
sit p
a
ssen
ger
flow fore
ca
sti
ng, com
p
a
r
e
s
the ran
ge of
appli
c
atio
n
s
a
nd re
se
arch
status of vario
u
s mo
del
s, a
n
d
prop
oses the
nece
s
sity of foreca
sting t
he sh
or
t-term rail trans
i
t pass
e
nger flow. It forec
a
s
t
s
sho
r
t-te
rm pa
sseng
er flo
w
by mean
s of
both
supp
ort vector reg
r
e
s
sion
m
e
thod and
BP
ne
ural
netwo
rk met
hod, an
d the
re
sults sho
w
that th
e
suppo
rt vecto
r
reg
r
e
ssi
on
model
ha
s such
a
theoreti
c
al su
perio
rity
a
s
minimi
zed structural risk,
t
hus
havin
g
a
high
er fo
re
casting
a
c
cura
cy
unde
r sm
all sample
conditi
ons fo
r sh
ort-term ur
ban ra
il transit pa
ssenge
r flow, which p
r
edi
cts
a
promi
s
in
g foreca
sting p
e
rf
orma
nce the method ha
s.
Referen
ces
[1]
W
u
Qian
g. T
he A
ppl
icati
o
n
of Gre
y
F
o
r
e
castin
g Met
h
od
in
urb
an r
a
il tra
n
sit
pas
seng
er fl
o
w
forecasting.
Study on Ur
ba
n Rail T
r
ans
it.
20
04; 7(3): 52-5
5
.
[2]
W
ang Ya
nh
ui
. Rai
l
w
a
y
Pa
sseng
er T
r
affic Vol
u
me
Da
ta Min
i
ng
F
o
r
e
castin
g M
e
th
od
an
d Its
Appl
icatio
n.
Jo
urna
l of Railw
a
y
.
2004; 26(
5): 1-7.
[3]
Ming-W
e
i Ch
a
ng,
Ch
ih-Je
n
Lin.
L
eav
e-on
e-out
Bo
un
ds
for Sup
port
Vector Re
gres
sion M
ode
l
Selecti
on.
Ne
u
r
al Co
mp
utatio
n
. 2005; 1
7
(5): 118
8-12
22.
[4]
K
y
o
u
n
g
ja
e Ki
m. F
i
nancia
l time series fore
casting us
in
g supp
ort vector machin
es.
Ne
uroco
m
putin
g
.
200
3; 55 (3): 3
07-3
19.
[5]
F
r
iedrichs F
,
Christia
n I. Evol
ution
a
r
y
tu
nin
g
of multiple SV
M parameters.
N
e
u
r
o
c
om
pu
ti
ng
. 2005; 6
4
:
107-
117
[6]
Cha
pel
le
O, Vapn
ik V.
Cho
o
s
ing
multi
p
le
p
a
rameters
for
supp
ort vector
machi
nes.
Mac
h
in
e l
ear
nin
g
.
200
2; 46: 131-
159.
Evaluation Warning : The document was created with Spire.PDF for Python.