TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.5, May 2014, pp
. 3348 ~ 33
5
6
DOI: http://dx.doi.org/10.11591/telkomni
ka.v12i5.4936
3348
Re
cei
v
ed O
c
t
ober 1
7
, 201
3; Revi
se
d Decem
b
e
r
1, 2013; Accepte
d
De
cem
ber
19, 2013
Forecasting Spatial Migration Tendency with FGM(1,1)
and Hidden Markov Model
Chan
g Jiang
*
1
, Jun Wang
1
, Yunsong Shi
2
1
Geograp
hic In
formation D
e
p
a
rtment, Nang
ji
ng Un
iversi
t
y
o
f
Posts and Co
mmunicati
ons,
Nanj
in
g, Chin
a
2
English D
epa
rtment, Nanji
n
g
Universit
y
of C
h
in
ese Med
i
ci
n
e
, Nanj
ing, Ch
i
n
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: jian
g
c@n
j
upt
.edu.cn
A
b
st
r
a
ct
Popu
latio
n
spa
t
ial migrati
on tend
ency forec
a
sting
is very importa
nt for the rese
arch of
spatia
l
de
mo
grap
hy.
T
r
aditio
nal
ap
p
r
oach
e
s ar
e to
o co
mp
lex to
be us
ed f
o
r ti
me
seri
es pr
e
d
ictio
n
. T
h
is p
aper
prese
n
ts a me
thod co
m
b
in
ati
ng Hi
dde
n Ma
rkov M
ode
l (H
MM) and Four
ier Seri
es Grey Model (FGM)
base
d
on Grey
Model (GM) to pred
ict the trend of Ji
angs
u
Province
’
s
mi
gratio
n in C
h
in
a. T
here are th
ree
parts of foreca
st. T
he first one is to b
u
il
d GM from a
s
e
rie
s
of coord
i
nate
data,
the sec
o
nd us
es the F
o
uri
e
r
series t
o
refi
n
e
the
resi
du
al
s pro
duce
d
by
the
mentio
ne
d
mo
del
a
nd t
he th
ird
uses
HMM to refi
ne
the
resid
uals
of F
G
M .It is evide
n
t that the pro
pose
d
ap
pro
a
c
h
gets the b
e
tter result
p
e
rfor
ma
nce i
n
stud
ying
the p
o
p
u
lati
on
migrati
on. S
a
tisfactory resu
l
t
s
have
b
een
obtai
ne
d, w
h
ic
h i
m
prove
HM
M-F
GM reach
ed
w
hen only GM w
a
s used for the pop
ul
ation sp
atial
mi
gratio
n tend
ency forec
a
sting.
Ke
y
w
ords
:
spatia
l de
mo
gr
aphy, grey
mo
del, fouri
e
r gre
y
mo
d
e
l, hid
d
e
n
Markov mod
e
l, forecast error,
gravity center
mo
de
l
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
Demo
graphy
is an inhe
rently spatial
scie
n
ce
whi
c
h involves t
he study of compl
e
x
pattern
s of interrel
a
ted so
cial, behavio
ral, ec
ono
mic,
and environ
mental phe
n
o
mena [1]. Thus,
schola
r
s hav
e in
crea
singl
y arg
ued
tha
t
spatial
an
swers to d
e
m
ogra
phi
c q
u
e
s
tion
s in
clu
d
i
ng
spatial a
naly
s
is of d
e
mog
r
aphi
c p
r
o
c
e
s
se
s and
o
u
tcomes. A gre
a
t deal of attention ha
s b
een
given to the p
henom
ena of
migration
an
d popul
ati
on
migratio
n pre
d
iction. Conti
nued hi
gh lev
e
ls
of migration
to advance
d
citie
s
will
lead to un
pre
c
ed
ented
cha
nge in
demog
ra
phic an
d
eco
nomi
c
co
mpositio
ns o
f
regi
onal
p
opulatio
ns
,
e
s
pe
cially i
n
Chin
a. As a
co
nsequ
en
ce,
Chinese prov
inces
will expe
rience a
shif
t towards
more uneven di
stributio
n,
which concerns to
most part
s
of the
Chine
s
e
publi
c
.
Althou
gh
d
ebate
fo
cuse
d m
a
inly
on the
p
r
ovin
ce
scale,
the
city
transfo
rmatio
ns
will be th
e most p
r
ofo
und. John
so
n et al. used
both glo
bal
and lo
cal
sp
atial
statistics to look for spatiot
e
mpo
r
al patte
rns in
mig
r
ati
on of the American South
w
e
s
t [2]. Bor
goni
et al. analysis the immigra
n
t resid
ential distrib
u
tion spatially with particul
a
r reference to densi
t
y
and
diversity-based m
e
tho
d
s [3]. Altho
u
gh the
s
e
ap
p
r
oa
che
s
allo
w the
inve
stig
ators to exa
m
in
e
dynamic
migratory pattern
s of spatial a
nd tempo
r
al
clu
s
terin
g
, they coul
dn’t b
e
used to m
a
ke
accurate pred
iction
s ba
sed
on the time serie
s
for pre
d
icting the te
nden
cy of migration.
Statistical a
n
d
artifici
al int
e
lligen
ce
app
roa
c
he
s
are
the two mai
n
techni
que
s fo
r time
seri
es predi
ction seen i
n
t
he lit
erature
[4-6], whi
c
h i
n
clu
de
simpl
e
moving
averag
e forca
s
t
i
ng
(SMAF), aut
oreg
re
ssive (AR), auto
r
eg
ressiv
e m
o
ving average
(ARMA) a
nd
neural net
wo
rk.
Ho
wever, the
y
are too
co
mplex to be u
s
ed i
n
predi
cting future val
ues
of a time
seri
es
and
h
a
ve
not pre
r
equi
sites for time seri
es n
o
rm
a
lity or
error
calib
ration [7
]. Grey system theory is an
interdi
sci
plinary scientific area
that
wa
s
first int
r
odu
ce
d in
ea
rly 19
80s by
De
ng.
Since the
n
, t
h
e
theory
h
a
s b
e
com
e
q
u
ite popul
ar with its
ability
to
deal
with the
system
s th
a
t
have pa
rtial
l
y
unkno
wn pa
rameters. As a supe
rio
r
ity to conv
entio
nal statisti
cal
models, Grey Model (G
M)
requi
re o
n
ly a limited amo
unt of data to
esti
mate the
behavio
r of
unkno
wn sy
st
ems [8]. Hid
d
en
Markov Mo
de
l (HM
M
) i
s
a
widely tool
to
analyse a
nd
predi
ct time
serie
s
p
hen
om
ena. HMM h
a
s
been used su
ccessfully
to analyse
vari
o
u
s type
s of ti
me serie
s
in
cluding fin
a
n
c
i
a
l time serie
s
[9,
10], spee
ch
signal re
co
gnit
i
on [11], and DNA sequ
en
ce an
alysi
s
[12] etc.
In this pape
r,
we propo
se
d HMM
-
FGM
combi
n
ing
HMM with
G
M
and Fou
r
i
e
r serie
s
refining the resid
ual
s to achieve bette
r foreca
sts. In
our model,
GM wa
s co
n
s
tru
c
ted to d
o
the
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fore
ca
sting
Spatial Migrat
ion Tend
en
cy with
FGM(1,1) and
Hidd
e
n
Marko
v
… (Cha
ng Ji
ang
)
3349
cal
c
ulatio
n of
the
co
ordi
nat
e of
pop
ulatio
n g
r
avity
ce
nter,and
the
re
sidu
al e
r
ror of
the m
odel
was
corre
c
ted by
Fouri
e
r seri
es. HMM wa
s u
s
ed to imp
r
ov
e forecastin
g accuracy.
2.Proposed Metho
d
2.1. GM(1,1
)
GM predi
ct the future val
ues
of a time
seri
es
ba
se
d on a
set of
the most
re
cent data,
and the
sam
p
ling freq
uen
cy of the time seri
es i
s
fixe
d
.
The main ta
sk
of grey
system theo
ry is to
extract re
alist
i
c governing l
a
ws of
the system usi
ng a
v
ailable data.
This pro
c
e
ss is kno
w
n a
s
the
gene
ration of
the grey se
q
uen
ce. In gre
y
syst
em the
o
ry, GM(n,m
) denote
s
a grey model, wh
ere
is the
order o
f
the differe
nce eq
uation
an
d is t
he n
u
mb
er of va
riabl
e
s
. Althoug
h v
a
riou
s type
s
of
grey m
odel
s
can
be
me
ntioned,
mo
st o
f
the p
r
ev
iou
s
re
sea
r
che
r
s h
a
ve fo
cu
sed thei
r
atten
t
ion
an GM
(1,1
)
model
s in th
eir p
r
edi
ction
s
be
ca
use
of
its co
mputat
ional effici
en
cy. It should
be
noted that i
n
real time
ap
p
lication
s
, the
comp
utationa
l burden i
s
th
e mo
st impo
rtant paramet
er
after the
perf
o
rma
n
ce. G
M
(1,1) type
of grey m
o
d
e
l is th
e mo
st widely u
s
e
d
in the lite
r
at
ure,
pron
oun
ce
d
as
“Grey Mo
del First O
r
d
e
r O
ne Va
ria
b
le”. Thi
s
m
o
del is tim
e
serie
s
forecasting
model. The
differential e
quation
s
of the GM(1
,1) model
have time-varying coeffici
ents. The
GM(1,1
)mo
d
e
l can
only b
e
use
d
in po
sitive data seq
uen
ce
s . In this pap
er,
since all the primi
t
ive
data a
r
e
po
si
tive, grey mo
dels can
be
use
d
to fo
re
cast the
future
value
s
of th
e pri
m
itive d
a
ta
points.
In orde
r to smooth the ra
ndomn
e
ss, th
e primit
ive da
ta obtained f
r
om the sy
ste
m
to form
the GM
(1,1
)i
s
subj
ecte
d t
o
an
op
erator, name
d
A
c
cumulating
G
e
neratio
n
Ope
r
ator (AG
O
).
The
differential
e
quation
(i.e.
GM(1,1
))i
s solved to
obta
i
n the
n-step
ahead
predi
cted value of
the
system. Fi
na
lly, using th
e predi
cted
value, t
he In
verse
Accum
u
lating
Gene
ration
Ope
r
at
o
r
(IAGO) i
s
app
lied to find the predi
cted v
a
lue
s
of origi
nal data.
Con
s
id
er
a
time sequ
ence
0
X
that denote
s
x-coordi
nat
e or
y-co
ordinat
e
of
demog
ra
phy gravity cente
r
.
00
0
0
1,
2
,
,
,
4
,
Xx
x
x
n
n
(1)
W
h
er
e
0
X
is a
n
on-n
egative
seque
nce a
nd
n is the
samp
le si
ze
of th
e
data.
When
t
h
is
seq
uen
ce i
s
subj
ecte
d t
o
the Accu
mulating
Ge
neratio
n O
p
eration
(AG
O
), the follo
wing
seq
uen
ce
1
X
is monotoni
call
y increa
sing.
11
1
1
12
n
n
4
Xx
x
x
,,
,
,
,
(2)
10
1
,1
,
2
,
3
,
,
.
k
i
x
kx
i
k
n
(3)
The gen
erate
d
mean sequ
ence
1
Z
of
1
X
is def
ined a
s
:
11
1
1
1,
2
,
,
,
Z
zz
z
n
(4)
Whe
r
e
1
zk
is the mean value o
f
adjace
n
t data, i.e.:
11
1
0.5
0
.5
1
,
2
,
3
,
,
zk
x
k
x
k
k
n
(5)
The le
ast
sq
uare
e
s
timat
e
sequ
en
ce
of t
he
g
r
ey differen
c
e eq
uation of
G
M
(1,1)
is
defined a
s
fol
l
ows:
01
.
x
ka
z
k
b
(6)
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: 3348 – 33
56
3350
The white
n
in
g equatio
n is
therefo
r
e, as
follows:
1
1
.
dx
t
ax
t
b
dt
(7)
In above,
,
T
ab
is a
seq
uen
ce of
para
m
eters that can b
e
found a
s
follows:
1
,,
T
TT
ab
B
B
B
Y
(8)
Whe
r
e:
00
0
2,
3
,
,
,
T
Yx
x
x
n
(9)
1
1
1
21
31
1
z
z
B
zn
(10)
Acco
rdi
ng to equatio
n (6
),the sol
u
tion of
1
x
t
at time
k
:
10
11
.
ak
p
bb
xk
x
e
aa
(11)
2.2. Grav
it
y
Model
Gravity cente
r
wa
s int
r
od
u
c
ed to th
e re
sea
r
ch
, nam
ely, with the
balan
ce
point
that the
popul
ation sp
atial distri
buti
on re
ac
hed
spatial torq
ue i
n
the re
se
arch are
a
du
ring
a ce
rtain tim
e
.
We can analy
z
e this a
r
ea'
s evolution in the popul
at
ion
migration an
d disclose the
characte
ri
stics
and fore
ca
st the tende
ncy
of the res
earch area
's p
opu
lation migration.
x
and
y
are written a
s
:
11
11
,
nn
ii
i
i
ii
nn
ii
ii
px
p
y
xy
pp
(12)
Whe
r
e,
n
mean
s the
numb
e
r of the admi
n
istrative u
n
it;
,
ii
x
y
are th
e ge
og
raphi
c
gravity cente
r
of each
ba
si
c unit;
i
p
mea
n
s the popul
ation num
ber;
,
x
y
are the pop
ulat
ion gravity center in
Jian
gsu.
2.3. Hidden
Markov
Model
Hidd
en Markov
Mod
e
l (HMM)
i
s
co
m
posed of
a
five-tuple:
,
,,,
SO
A
B
, where
12
,,
,
n
SS
S
S
is a
set of di
stinct
states
and
n
is the
n
u
mbe
r
of stat
es;
12
,,
,
m
OO
O
O
is
an o
b
served
se
que
nce
and
m
is the
num
ber of
ob
se
rvation
se
que
nce;
1,
ij
ij
N
Aa
is
transition pro
babilities and
1
|
ij
t
j
t
i
aS
S
is the probability of
a transition from
state
i
to state
j
;
1
1
jM
iN
ij
Bb
is emissi
on
probalilities and
|
ij
t
k
t
j
bO
v
S
is the
probability of
state
j
emitting
k
v
;
1
i
iN
is a vector of initial probalilities.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fore
ca
sting
Spatial Migrat
ion Tend
en
cy with
FGM(1,1) and
Hidd
e
n
Marko
v
… (Cha
ng Ji
ang
)
3351
Fore
ca
sting
the
popul
atio
n gravity center
po
sition
erro
r is a
n
HMM
proce
ss. Th
e
predi
ction
e
r
ror
is th
e
ob
servation
se
q
uen
ce
,
and t
he p
r
edi
ction
ability of th
e mod
e
l
is t
he
seq
uen
ce st
ate.
Acco
rdin
g to the
historical fore
ca
sting erro
rs
di
st
rib
u
t
i
on,
f
o
re
ca
st
err
o
r
wer
e
devided into
the N cla
s
se
s, analysi
z
in
g the hidden
state
seq
u
e
n
ce mo
del. T
he state
tran
sfer
matrix and observation
probab
ility
matrix were
estimated to
forecast to the forecast
errors. Finally
, prediction e
rro
r wa
s
corrected a
c
cordi
ng to the pre
d
ictive value.
2.4. FGM(1,1
)
In orde
r to im
prove the m
o
deling a
c
cu
ra
cy of
GM(1,1), Fourie
r seri
es a
nd g
r
avity model
wa
s used to
modify the grey m
odel
s. The fouri
e
r
series can ma
ke
the forecast
ed re
sult
s m
o
re
pre
c
ise. The
forecasting
algorith
m
ba
sed o
n
t
he combine
d
FG
M(1,1) m
odel
is described
as
follows
:
1)
Comp
uting th
e gravity cent
er of sp
atial data.
2)
Testing
the
q
uasi
-
expo
nen
tial and
the
q
uas
i
-
smooth
n
e
ss of th
e
series. If it is qu
asi-
expone
ntial and qu
asi
-
ex
poential, the
n
goto
3;else do the sm
ooth pro
c
e
ss or IAGO
p
r
oc
es
s
.
3)
Con
s
id
erin
g the origi
nal da
ta seri
e
s
p
r
o
c
essed by ste
p
1 and ste
p
2,
00
0
0
1,
2
,
,
,
4
.
Xx
x
x
n
n
Comp
uting th
e 1-AGO of serial
0
X
,
11
1
1
12
n
n
4
Xx
x
x
,,
,
,
,
whe
r
e
10
1
,1
,
2
,
3
,
,
.
k
i
x
kx
i
k
n
4) Con
s
tru
c
ting
the
matrix
1
1
1
21
31
1
z
z
B
zn
Whe
r
e
11
1
0.
5
0
.5
1
,
2
,
3
,
,
z
k
xk
xk
k
n
.
5) Acco
rdi
ng
to
1
,
T
TT
ab
B
B
B
Y
,the estimat
ed value
a
an
d
b
we
re cal
c
ulated,
whe
r
e
00
0
2,
3
,
,
.
T
Yx
x
x
n
6)
Cal
c
ulating th
e simulate
d value
1
x
t
acco
rdi
ng to the equ
ation (10
)
.
7)
Cal
c
ulating th
e first-o
r
de
r resid
ual erro
r seri
es
00
0
0
1,
2
,
T
En
,
Whe
r
ein,
00
0
ˆ
kx
k
x
k
.
8)
Modelin
g the
resi
dual
serie
s
b
a
sed o
n
t
he Fo
urie
r
se
ries a
c
cordi
n
g to the foll
o
w
ing
formula:
0
1
12
2
()
c
o
s
s
i
n
,
2
,
3
,
,
.
2
N
ii
i
kk
Ek
a
a
t
b
t
f
o
r
k
n
TT
(13)
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: 3348 – 33
56
3352
Whe
r
e
T
indica
tes the len
g
th
(pe
r
iod
)
of th
e re
sidu
al series, which is
equal to
(1
)
n
;
k
a
and
k
b
are co
efficient
s to be determin
ed by the least sq
u
a
re meth
ods;
0
a
is the averag
e
value of the f
unctio
n
in th
e
use
d
rang
e;
N is
the n
u
m
ber
of ha
rmo
n
ics of the
se
ries; t i
s
the order
nu
mber
given i
n
the seri
es.
The lea
s
t
sq
uare
metho
d
wa
s u
s
ed to
cal
c
ulate
the coeffici
en
ts
0
a
,
k
a
and
k
b
expre
s
sed a
s
follo
ws:
1
0
TT
CP
P
P
E
(14)
Whe
r
e,
01
1
2
2
,,
,
,
,
,
,
T
NN
Ca
a
b
a
b
a
b
, and
21
21
2
2
1
2
co
s
2
sin
2
co
s
2
sin
2
21
21
2
2
1
2
c
o
s
3
sin
3
cos
3
sin
3
21
21
2
2
2
1
2
co
s
s
in
co
s
s
in
NN
TT
T
T
NN
P
TT
T
T
N
nn
n
n
TT
T
T
9)
Cal
c
ulating th
e forecaste
d
value of seri
e
s
as the follo
wing formula:
00
0
ˆ
ˆˆ
(
1
)
1
()
()
(
)
2,
3
,
,
X
X
a
n
d
Xk
Xk
E
k
fo
r
k
n
(15)
2.5. HMM-F
G
M(1,1)
HMM
we
re
e
m
ployed to
i
m
prove
the
modelin
g a
ccura
cy of F
G
M(1,1).
To
a
pply the
HMMS, the
n
u
mbe
r
of
stat
es, the
types of mod
e
ls a
nd the
pa
ram
e
ters to b
e
m
odele
d
mu
st
be
deci
ded. Th
e
fore
ca
sting a
l
gorithm
ba
se
d on t
he
co
m
b
ined
HMM
-
F
G
M(1,1
)
mo
d
e
l is d
e
scri
be
d
as
follows
.
1)
Processin
g
the fore
castin
g
error
seri
es o
f
FGM(1,1) in
the discrete
way. The stat
es
and ob
se
rvation se
que
nce can b
e
de
cid
ed acco
rdin
g to the thresho
l
ds.
2)
Estimating th
e transitio
n probabilitie
s an
d the emissio
n
prob
alilities.
3)
Get the p
r
ed
iction of fo
re
ca
sting e
r
ror
se
ries
acco
rding to th
e
curre
n
t state
as
follows
:
ii
i
eS
T
H
(16
)
Whe
r
e
i
S
i
s
the cu
rre
nt stat
e;
i
T
is the probability of a
transition from state
i
to the other
state;
H
is the thre
sho
l
d discretin
g
fore
ca
sting error
se
ri
es. In this pape
r, the threshold is denoted a
s
0.2
0
.4
0.5
0
.6
0.8
,,
,
,
T
H
P
PP
PP
.
4)
Cal
c
ulating th
e forecaste
d
value of seri
e
s
as the follo
wing foum
ula
:
'
ii
i
X
Xe
(17)
Whe
r
e
i
X
is x-co
ordinate
or y-coo
r
din
a
te
of the p
opulatio
n gra
v
ity center;
'
i
X
is
forecastin
g x-co
ordinate o
r
y-co
ordi
nate
of the popula
t
ion gravity center.
3.Applica
t
ion of the Pro
posed Me
th
od
3.1. Data Se
t
Acco
rdi
ng to
algorith
m
d
e
scrib
ed a
b
o
v
e, the com
b
ined m
odel
s were ba
se
d on th
e
popul
ation of
Jian
gsu province i
n
China
(199
1-201
0) colle
cted f
r
o
m
from the
Ji
ang
su Statisti
cs
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fore
ca
sting
Spatial Migrat
ion Tend
en
cy with
FGM(1,1) and
Hidd
e
n
Marko
v
… (Cha
ng Ji
ang
)
3353
Almana
c, whi
c
h i
s
shown i
n
Fig
u
re
1. A
c
cordi
n
g
to
Ji
ang
su Stati
s
tics Alman
a
c i
n
20
08
,
Jian
gsu
has
13 p
r
efe
c
ture-l
evel citi
es
,
5
4
urban
distri
cts, 27
county-level
cities a
nd 2
5
co
unties. Su
zho
u
,
Wuxi, Ch
ang
zho
u
, Na
njin
g and
Zhe
n
ji
ang al
ong
wit
h
the di
stri
cts and
cou
n
tie
s
subo
rdi
nate
to
them con
s
i
s
t of Southern
Jiang
su; Ya
ngzhou,
Tai
z
hou an
d Na
ntong with t
he distri
cts
and
cou
n
ties
su
b
o
rdin
ate to th
em con
s
ist
of Cent
ral
Jian
gsu; Xu
zh
ou,
Huai’
an, Su
qian, Yan
c
h
e
ng
and
Lianyu
n
gang
with
t
heir bo
rou
g
h
s
a
n
d
count
ies--North
ern
Jia
n
g
s
u.An
d the
mod
e
l
i
ng
,
forecastin
g a
nd analy
s
is p
r
ocess a
r
e a
s
follows:
Figure 1. Eco
nomic S
k
etch
Map of Jiang
su in 20
08
3.2. Experimental Me
tho
dolog
y
The g
r
avity center
co
ordi
n
a
tes of
popul
ati
on di
stribut
ion were
co
mputed
acco
rding
to
the demo
g
ra
phy data.The
popul
ation
gravity center coordi
nate wa
s achieved b
a
se
d on eq
ua
tion
(12
)
, which di
scl
osed the spatial distri
bu
tion and
tend
ency of popul
ation migratio
n. GM(1,1)
was
con
s
tru
c
ted
b
a
se
d o
n
the
popul
ation
gravity cente
r
coordi
nate
s
. A
s
sho
w
n
in
Fi
gure
2,
altho
ugh
the preci
s
io
n
of applyin
g
GM(1,1
) to f
o
re
ca
st
the
coordi
nation
seems to b
e
accepta
b
le, the
predi
ction performance could still be improved. In or
der to improv
ing the preci
s
ion. In order to
improve the
accuracy, Fo
urie
r se
rie
s
wa
s used
to modeling th
e resi
dual
se
ries of GM
(1
,1)
according to
Equation
(1
3), and
the
coeffici
ents
0
a
,
k
a
and
k
b
were
calcul
ated a
ccordin
g to
Equation (1
4
)
. To apply HMM to the re
sidu
al se
rie
s
of FGM(1,1
)
, the re
sidual
serie
s
mu
st be
transfo
rme
d
into the distin
ct states
an
d observation
seque
nce. The
number
n
of states
S
was 3
according to
cla
ssifi
cation.
The numb
e
r
m
of obse
r
vati
on se
que
nce
O
wa
s 5 acco
rding to the
percentile
s of
foreca
sting e
rro
r as Eq
uati
on(1
8
).
0.
2
0.2
0
.4
0.4
0
.6
0.6
0
.8
0.8
1
2
3
4
5
eP
Pe
P
OP
e
P
Pe
P
eP
(18)
The state
transition
probabilities matri
x
T
and emi
s
si
on p
r
ob
abilities of
a HM
M for
seq
uen
ce
O
with kno
w
n
states
S
were cal
c
ulate
d
. The
predi
ction of
forecastin
g
error
se
ries
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: 3348 – 33
56
3354
were
cal
c
ulat
ed a
c
co
rding
to the
cu
rrent
state
a
s
Eq
u
a
tion
(16
)
.
T
he fo
re
ca
sted
value
of
HM
M-
FGM is
sh
o
w
n in
Table
1. The results of the
f
o
re
ca
st
ing a
c
cur
a
cy
f
o
r
co
ordinate of g
r
a
v
ity
cente
r
by u
s
ing GM
(1,1),
FGM(1,1) a
n
d
HMM
-
FG
M
(
1,1)
are
sho
w
n in Ta
ble
1. As sh
own
in
Figure 2
and
Figure 3,
G
M
(1,1) cann
o
t
predi
ct t
he
acute
pa
rts o
f
the co
ordi
n
a
te of po
pula
t
ion
gravity in Ji
a
ngsu satisfa
c
torily. Ho
wev
e
r, t
he fo
re
casting
erro
rs are
obviou
s
ly redu
ce
d by
FGM(1,1) an
d HMM-FGM
(
1,1). The HM
M-FGM
(
1,1
)
is more accu
rate than FGM
(
1,1) a
pproa
ch.
Table 1. Fo
re
ca
sting Resul
t
s of Different
Models
Ye
a
r
Actual Value
GM(1,
1
)
FGM
(
1,1)
HMM-F
G
M(
1,1)
X (m)
Y (m
)
X (m)
Y (m
)
X (m)
Y (m
)
X (m)
Y (m
)
1991
40455200
40455308
40455200
3637940
40455308
3637921
40455075
3637923
1992
40455000
40455008
40454905
3638267
40455008
3638096
40454994
3638106
1993
40454800
40454892
40454619
3638562
40454892
3638090
40454701
3638085
1994
40454600
40454606
40454342
3638828
40454606
3638310
40454594
3638376
1995
40454300
40454383
40454075
3639069
40454383
3638554
40454201
3638616
1996
40454200
40454210
40453816
3639286
40454210
3639832
40454194
3640328
1997
40453700
40453675
40453565
3639482
40453675
3641061
40453704
3640908
1998
40453400
40453478
40453322
3639659
40453478
3641017
40453301
3641198
1999
40453200
40453150
40453088
3639818
40453150
3641507
40453204
3641506
2000
40452600
40452622
40452861
3639962
40452622
3642003
40452529
3642023
2001
40452500
40452554
40452641
3640093
40452554
3641087
40452429
3640768
2002
40452300
40452370
40452428
3640210
40452370
3640539
40452229
3640603
2003
40452200
40452202
40452222
3640316
40452202
3640622
40452194
3640641
2004
40451900
40451953
40452023
3640412
40451953
3640372
40451829
3640426
2005
40451600
40451546
40451831
3640498
40451546
3640321
40451604
3640276
2006
40451400
40451554
40451644
3640576
40451554
3640208
40451353
3640233
2007
40451200
40450994
40451464
3640646
40450994
3640188
40451204
3640196
2008
40450300
40450395
40451289
3640710
40450395
3640669
40450175
3640808
2009
40450700
40450896
40451120
3640767
40450896
3640627
40450653
3640468
2010
40451000
40451088
40450957
3640819
40451088
3640353
40450901
3640296
(a) X-co
ordi
n
a
te of gravity cente
r
(b) Y-co
ordi
n
a
te of gravity cente
r
Figure 2. Coo
r
dinate F
o
re
casted
Re
sults
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fore
ca
sting
Spatial Migrat
ion Tend
en
cy with
FGM(1,1) and
Hidd
e
n
Marko
v
… (Cha
ng Ji
ang
)
3355
Figure 3. Co
mpari
s
o
n
of GM(1,1
), FG
M(1,1)
and HHM-FGM
(
1,1
)
from 19
91 to 2010
3.3. Results Accu
racy
The utimate goal of
any
fore
ca
sting
en
deavor
i
s
to
provide
an
a
c
curate
a
nd
unbia
s
e
d
forecast. F
o
reca
st e
rro
r i
s
the differe
nce
betwe
en a
c
t
ual qu
antity and the fo
re
ca
sted. Thi
s
stu
d
y
make
s a
co
mpari
s
o
n
of t
he
re
sults fro
m
19
91 to
2
010 to
a
s
sess the
forecast perfo
rma
n
ce o
f
GM(1,1
),
FG
M(1,1) and HHM
-F
GM(1,1)
by way o
f
the mean
absolute e
rro
r
M
AE
, mean
absolute pe
rcentage e
r
ror
M
APE
and ro
ot mea
n
squ
r
e erro
r
RMSE
.
M
AE
is a quantity use
d
t
o
mean
su
re
how
clo
s
e f
o
re
ca
st
s o
r
pr
edi
ction
s
a
r
e to the
eventual o
u
tco
m
es.
M
APE
is a
measure of
a
c
cra
c
y in
a fit
t
ed time
se
ri
es val
ue i
n
statistics,whi
ch
us
ually
ex
p
r
es
se
s a
c
cur
a
cy
as a
pe
rce
n
tage.
RMSE
a way
to quantify the differen
c
e
betwe
en a
n
estimato
r an
d the tru
e
value of the quantity being
estimated. Th
e indicators were exp
r
e
s
se
d as follo
w:
1
1
n
tt
t
M
AE
A
F
n
(19)
1
1
n
tt
t
t
A
F
MAPE
nA
(20)
2
1
1
n
tt
t
RMSE
A
F
n
(21)
Whe
r
e
t
A
is act
ual value for
perio
d t , and
t
F
is
forec
a
s
t
value for period t.
From
Tabl
e
2, GM
(1,1) p
r
ovided
x-co
rrdinate
p
r
edi
ction
accu
ra
cy to 30
0m in
RM
SE
whi
c
h F
G
M(1
,
1) and
HM
M
-
FGM
(
1,1
)
a
c
hieved a
c
curacy to 93
m a
nd 69m
in
RM
SE
, res
p
ec
tivly.
As a
bove, th
e FG
M(1,1
)
u
s
ed
an
integ
r
al ap
pro
a
ch t
o
imp
r
ove th
e
fore
ca
sting
value
of GM
(1,
1
)
further an
d
was i
m
proved
by Fou
r
ier
se
ries.
Th
e HMM-
F
G
M(
1
,
1)
w
a
s
impr
o
v
ed
b
y
H
MM.
HMM-
FGM(1,1) i
s
better in com
pari
s
on to all
studi
ed
app
roache
s re
gardless of
the
adaptin
g inde
x of
M
AE
,
M
APE
or
RMSE
.
Table 2. Simulation Results of Coo
r
din
a
te Fore
ca
sted
Name
GM(1,
1
)
FGM
(
1,1)
HMM-F
G
M(
1,1)
MAE[m] MAPE
RMSE[m]
MAE[m]
MAPE
RMSE[m]
MAE[m]
MAPE
RMSE[m]
X-coordinate
217
0.00054
300
73
0.00018
93
53
0.00013
69
Y
-
coo
r
dinate
639
0.01756
856
92
0.00361
191
58
0.00160
65
4. Conclusio
n
GM i
s
very
common
tech
nique
s
used
for time
se
rie
s
fo
re
castin
g. Ho
weve
r, th
e HM
M-
FGM which combi
ned
use of HMM a
nd Fou
r
ie
r se
rie
s
ba
se
d
on GM in
po
pulation mi
gration
forecastin
g is a novel app
roac
h, whi
c
h h
a
s be
en p
r
ov
ed to provid
e an ade
quate
perfo
rman
ce.
In
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: 3348 – 33
56
3356
this p
ape
r, th
e HM
M-F
G
M
is p
r
e
s
ente
d
t
o
u
s
e the
gre
y
model to
ro
ughly p
r
edi
ct
the next datu
m
from
a
set
of
the
mo
st re
ce
nt
data. The
model
u
s
e
HMM an
d th
e
Fouri
e
r serie
s
to fit
the
resi
dual
errors p
r
o
d
u
c
ed by th
e G
M
. It is evid
e
n
t t
hat the
propo
sed
ap
proach
HMM-F
G
M h
a
s a
hi
gher
forecastin
g a
c
cura
cy than
GM in popul
a
t
ion
spatial mi
gration ten
d
e
n
cy fore
ca
sting.
Ackn
o
w
l
e
dg
ements
This re
se
arch is
supp
orted in
pa
rt by
th
e Na
n
jin
g Un
iver
s
i
ty o
f
Po
sts
a
nd
Comm
uni
cati
ons for L
abo
ratory Con
s
truction a
nd Equipme
n
t Ma
nagem
ent Rese
arch Proj
ect
unde
r g
r
a
n
t
no. 20
12XSG
16; the
Qingl
an Proje
c
t
un
der grant n
o
. NY20
803
9; t
he Key Bid
d
i
ng
Proje
c
t in T
eaching
Ref
o
rm
und
er
g
r
ant n
o
. JG
0321
2JX0
2;
the National
Natu
ral S
c
i
ence
Found
ation of
China u
nde
r grant no. 61
2
7108
2.
Referen
ces
[1]
PR Voss. De
mogra
p
h
y
as
a Spati
a
lSoc
ial
Science.
Po
p
u
lati
on R
e
sear
ch an
d Policy
Review
.
20
07
;
26(5-
6): 457-4
76.
[2]
KM Johns
on,
PR Voss, RB
Hammer, GV Fuguitt,
S McNiven. T
e
mporal
and
S
patial V
a
riation in Age-
Specific N
e
t Migratio
n
in the U
n
ited States.
D
e
mogr
aphy.
2
0
05; 42(4): 7
91-
812.
[3]
G
Borruso.
G
eogr
aph
ica
l
A
nalysis
of F
o
r
e
ig
n I
m
migr
ati
on
an
d Sp
atia
l Patterns
in
Urba
n Are
a
s:
Density Esti
mation, Sp
atia
l
Segr
e
gati
on a
nd D
i
versity A
nalysis.
Pr
oce
edi
ng s
o
f the
intern
atio
na
l
confere
n
ce o
n
Comp
utation
a
l
Scienc
e an
d Its App
lic
ations (
I
CCSA). Perug
i
a. 200
8; 507
2: 459-4
74.
[4]
YH Li
n, CC
C
h
iu, PC
Le
e,
YJ Lin. A
ppl
yi
ng F
u
zzy
Grey Modification
Mode
l on Inflow
F
o
recasting.
Engi
neer
in
g Applic
atio
ns of Artificial Intel
lig
e
n
ce.
201
2; 25(
4): 734-7
43.
[5]
CT
Lin, IF
Lee. Artificial Inte
lli
genc
e Di
ag
nos
is
Alg
o
rithm for
Exp
a
n
d
in
g a P
r
ecisio
n E
x
p
e
rt F
o
recastin
g
Sy
s
t
e
m
.
Expert
Systems w
i
th Appl
icatio
ns
. 2
009; 36(
4): 838
5-83
90.
[6]
Yan B
a
i, Ke
M
a
, Qing
han
g R
en.
A N
N
-GM(1,1) Mod
e
l-B
a
s
ed A
nalys
is of
Netw
ork T
r
affic F
o
recastin
g.
Procee
din
g
s of
3
rd
Intern
ation
a
l C
onfer
ence
on Bro
a
d
ban
d
Ne
t
w
ork an
d Multimed
ia
T
e
chno
log
y
(IC
-
BNMT
). Shang
hai. 20
10; 3: 1
91-1
94.
[7]
HL W
o
ng, JM
Shiu.
Comp
ari
s
ons
of F
u
zz
y T
i
me Series
and
H
y
br
id
Gre
y
Mo
del
for
Non-stati
onar
y
Data Forecasting.
Appl
ie
d Ma
thematics & Informatio
n
Scien
c
es
. 2012; 6(
2
S
): 409S-41
6
S
.
[8]
E Ka
yacan, B Ulutas, O Ka
yn
ak. Gre
y
S
y
ste
m
T
heor
y
-
B
a
s
ed Mod
e
ls in T
i
me series Pre
d
i
ction.
Expert
Systems w
i
th Appl
icatio
ns
. 2
010; 37(
2): 178
4-17
89.
[9]
Md Rafi
ul
Has
s
an, Ba
ikunt
h
Nath.
Stock M
a
rket F
o
rec
a
sti
ng Us
in
g H
i
dd
en M
a
rkov M
o
del: A
New
Appro
a
ch
. Pro
c
eed
ings
of the 20
05 5th Int
e
rnati
ona
l
Co
n
f
erence
on Int
e
lli
ge
nt S
y
ste
m
s Desig
n
a
n
d
Appl
icatio
n (ISDA’0
5
), W
r
oclo
w
.
2
005: 1
92-1
96.
[10]
Md Rafiu
l
Has
s
an, Baiku
n
th
Nath, Micha
e
l
Kirle
y
. A Fusio
n
Mode
l of HMM, ANN and
GA for Stock
Market Forecasting.
Expert Systems w
i
th Ap
plicati
ons
. 2
0
0
7
; 33(2): 17
1-1
80.
[11]
Bhup
ind
e
r Si
n
gh, Ne
ha K
a
p
u
r, Pun
eet Ka
ur.
Spe
e
ch R
e
cog
n
itio
n
w
i
t
h
Hi
dde
n Mar
k
ov Mod
e
l:
A
Revie
w
.
Inter
n
ation
a
l Jour
nal
of Advanced
Rese
arch
in C
o
mputer Sci
e
n
c
e and Softw
are Engi
neer
in
g.
201
2; 2(3): 400
-403.
[12]
Ale
x
a
nder V, L
u
kash
in, Mark
Borod
o
vsk
y
.
G
eneM
ark.hmm: Ne
w
So
luti
ons
for Gene F
i
n
d
i
ng.
N
u
cl
e
i
c
Acids Res
earc
h
.199
8
; 26(4):
110
7-11
15.
Evaluation Warning : The document was created with Spire.PDF for Python.