TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.6, Jun
e
201
4, pp. 4624 ~ 4
6
3
0
DOI: 10.115
9
1
/telkomni
ka.
v
12i6.544
4
4624
Re
cei
v
ed
De
cem
ber 2
8
, 2013; Re
vi
sed
F
ebruary 28,
2014; Accept
ed March 1
6
, 2014
Forecast Model of Water Quantity Based on Back
Propagation Artificial Neural Network
Shihua Li
Schoo
l of Busi
ness, Jinl
i
ng In
stitute of
T
e
chnol
og
y
No. 99, Hon
g
ji
ng Aven
ue, Jia
ngn
ing D
i
strict, Nanji
ng 2
1
1
1
6
9
email: lis
hi
hua
88@
163.com
A
b
st
r
a
ct
Back Prop
ag
ation
(BP)
neur
a
l
netw
o
rk, W
i
d
e
ly a
dopt
ed a
n
d
util
i
z
e
d
in
au
tomatic c
ontro
l
,
ima
g
e
recog
n
itio
n, hydrol
ogic
a
l fore
casting a
nd w
a
ter qua
lity
ev
alu
a
tion, etc., as one of the
Artificial Neur
a
l
Netw
orks, has stronger fu
ncti
on a
nd pr
op
erty of ma
ppi
ng,
classificati
on, f
unctio
nal fittin
g
.
T
h
is article ta
kes
the w
a
ter flow
of Lan
z
h
ou s
e
ction
of Yell
o
w
river
in Chi
n
a as an
exa
m
ple by th
e w
a
y of BP mo
del
to
pred
ict the w
a
ter qua
ntity. It is
w
e
ll proved th
at BP netw
o
rk
mo
de
l can
rea
c
h the purp
o
se
s of early w
a
rnin
g
and forec
a
stin
g.
Ke
y
w
ords
: the
forecast mo
de
l, w
a
ter quantity,
BP, artificial neur
al netw
o
rk
, ANN
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
With
the so
cial
and econ
omic develo
p
m
ent,
La
nzhou se
ction of
Yellow Ri
ver
was
grad
ually poll
u
ted alon
g the river.
The
stru
ctural ch
a
r
acte
ri
stic of
indu
strial poll
u
tion is expli
c
it;
the p
r
opo
rtio
n of d
o
me
sti
c
p
o
llution
is increa
si
n
g
consta
ntly;
area-sou
r
ce po
llution
i
s
seri
ous.
These
conditi
ons
give ri
se
to a neg
ative impact
o
n
survival and dev
elopme
n
t of Gan
s
u Provin
ce
and the wh
ol
e of the Yellow River in ma
sses. In orde
r to decrea
s
e l
o
sse
s
cau
s
e
d
by the sudden
water poll
u
tion a
c
cide
nt in lo
wer reaches of
L
a
n
z
hou
se
ction
of Hu
ang
he
River, it i
s
v
e
ry
necessa
ry to establi
s
h p
r
e
d
iction m
odel
of "wat
er am
ount" and "water qu
ality" of Lanzh
ou re
ach
of Hua
ngh
e River
so a
s
to
respon
d to e
m
erg
e
n
c
ie
s,
ensure ecolo
g
ical se
cu
rity,
water se
cu
rit
y
.
It
is sig
n
ifica
n
t to built warni
ng and fo
re
casting
sy
ste
m
of Lan
zho
u
se
ction of
Yellow
River,
to
provide te
chn
i
cal supp
ort for eme
r
ge
nt investi
gatio
n and ha
ndling
of water poll
u
tion incid
ents.
At prese
n
t, the wate
r re
so
urces
and th
e qua
lity of the wate
r envi
r
onm
ent ha
s
become
one of the pri
m
ary goal
s of
the sustai
na
ble devel
o
p
m
ent of soci
ety and econo
m
y
. Therefore, in
the re
se
arch
proje
c
t of "Unified man
a
g
e
ment a
nd
sche
duling,
all
o
catio
n
of th
e Yellow Riv
e
r", it
is not only ab
solutely ne
ce
ssary, but also very
timing to cond
uct th
e re
sea
r
ch o
n
"early wa
rni
ng
and fo
re
ca
st
based
on th
e
wate
r q
uantit
y and
wate
r q
uality". Espe
cially, there i
s
theoreti
c
al
an
d
pra
c
tical
si
gni
fican
c
e in
the
integrated
water m
anag
e
m
ent an
d all
o
catio
n
, sche
duling
of Yell
ow
River. T
o
e
s
t
ablish
su
ch
a
syste
m
m
o
d
e
l, deb
ug
an
d succe
s
sfull
y
ope
rate
will
get
huge
so
cial
benefits,
ca
n
greatly stren
g
then th
e u
n
i
f
ied man
age
ment
a
nd pro
t
ection of
water re
sou
r
ces in
the Yello
w
Ri
ver b
a
si
n, an
d avoid
di
re
ct and
indi
r
e
c
t
e
c
o
n
o
mic los
s
or
r
e
du
c
e
w
a
te
r po
llu
tion
.
Mean
while it
will play an i
nestima
ble ro
le in
the soci
al se
cu
rity, the peopl
e'
s life and p
r
o
perty
se
curity an
d
the rapi
d de
velopment of
eco
nomi
c
constructio
n
. This p
a
rt onl
y introdu
ce
s
the
water q
uantit
y predictio
n.
2. The Forec
ast Mod
e
l of Wa
ter Qu
antit
y
on the Ba
sis of BP Ar
tificial Neural
Net
w
o
r
k
2.1. Brief intr
oduction o
f
Artifici
al Ne
ural Ne
t
w
o
r
k
—
ANN
Artificial Neu
r
al Network-ANN
is
an
emer
ging
int
e
rdi
sci
plina
r
y scien
c
e
rel
a
ted to
mathemati
cs,
physi
cs, brai
n scie
nce, p
sycholo
g
y,
co
gnitive scie
nce, com
pute
r
sci
en
ce, a
r
tificial
intelligen
ce, etc.. Study o
n
predi
ction
model ha
s g
r
adu
ally beco
m
e a very importa
nt con
t
ent
based
on
Ne
ural
Net
w
o
r
k
[1]. BP network-BP, a
s
a
feed-fo
rward
netwo
rk,
po
sse
s
ses
st
ron
ger
function of m
appin
g
, cla
ssi
fication an
d functio
n
fitting, which h
a
ve been
widely
applie
d in ma
ny
fields,
su
ch
a
s
a
u
tomatic control, im
age
re
cog
n
ition, hydrolo
g
ical
f
o
re
ca
sting an
d
water qualit
y
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fore
ca
st Mod
e
l of Water Q
uantity Based
on Back
Pro
pagatio
n Artificial Neural…
(Shihu
a Li)
4625
evaluation[2
-
6]. BP neural network, through
which
multiple varia
b
les
can be
effective anal
yed
without cre
a
ting a mathem
atical mo
del, is a kin
d
of sup
e
rvised m
u
ltilayer feed
-forwa
rd n
e
u
r
al
netwo
rk
com
posed
of inp
u
t layer,
hid
den l
a
yer an
d outp
u
t lay
e
r i
n
whi
c
h
each laye
r
h
a
s
a
plurality of ne
uron
s, a
s
sho
w
n in
Figu
re
1 [7-9
]. It is a
topology
stru
cture
of the t
h
ree l
a
yer fe
ed-
forwa
r
d
neu
ral network, where
the first
layer i
s
the in
put nod
es, th
e se
co
nd i
s
t
he hid
den l
a
yer
node
s, and th
e third layer i
s
the output n
ode.
Figure 1. BP
Neu
r
al Netwo
r
k Stru
ctural Dra
w
in
g
For the i
nput
sign
al, it will first p
r
op
agat
e
forward to the hid
den n
o
des, the
n
through th
e
action fun
c
tio
n
transport th
e output information
of the hidden no
de
to the output node. Finally,
the output re
sults
co
uld b
e
obtaine
d. Action fun
c
tion
of node
s u
s
ually sele
ct sigmoid fun
c
ti
on,
calle
d the S functio
n
(see
Figure 2):
1
()
1
x
fx
e
,
2
()
(1
)
x
x
e
fx
e
(
1
)
Figure 2. S Type Functio
n
Curve
2.2. BP Lear
ning Algorithm
Assu
me that
the neu
ron n
ode of the in
put la
yer, hid
den laye
r an
d output laye
r as
N1,
N2,
N3,
so th
e relation
ship
between
inp
u
t and
out
p
u
t of BP ne
ural
netwo
rk is a
highly n
on-li
n
ear
mappin
g
relat
i
onship
and
the n
e
two
r
k is a m
appin
g
from N1 di
men
s
ion
a
l Eu
clid
ean
sp
ace to
the
N3 dim
e
n
s
io
nal Eucli
dean
spa
c
e. Tran
sfer fun
c
tion
of the hidde
n
layer and
ou
tput layer ne
uron
is S function (2):
x
e
x
f
1
1
)
(
(
2
)
Suppo
se
that
there i
s
N
pa
ir of
sa
mple
（
n
I
,
n
T
, n=1, 2…
N. Among
that,
In
∈
R
N1
i
s
in
put
of Nth trainin
g
sampl
e
; T
n
∈
3
N
R
is output
of Nth trainin
g
samp
l
e
. Th
us the course that input
sign
al wa
s transporte
d fro
m
input laye
r to output
lay
e
r can b
e
ind
i
cated
by equ
ation from
(3
) to
(7):
0
0.
5
1
1.
5
-6
-5
-
4
-
3
-2
-1
0
1
2
3
4
5
6
x
f(
x)
……
……
……
P
1
P
2
P
N
O
1
O
2
O
M
u
11
u
12
u
1
N
u
21
u
22
u
2
N
u
N
1
u
N
2
u
NN
v
11
v
12
v
1
M
v
21
v
22
v
2
M
v
N
1
v
N
21
v
NM
First lay
e
r
Second lay
e
r
Third la
yer
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: 4624 – 4
630
4626
pi
i
I
net
(
3
)
pi
i
ni
I
net
O
(
4
)
1
1
N
i
j
ni
ji
j
O
W
net
(5)
)
(
j
nj
net
f
O
(6)
2
1
N
i
k
ni
kj
k
O
W
net
(7)
)
(
k
nk
net
f
O
(8)
Among abov
e
eq
uation:
1
,
2
,
3
,
...,
1
iN
;
1
,
2
,
3
,
...,
2
j
N
;
1
,
2
,
3
,
...,
3
kN
. The
i
ne
t
,
j
ne
t
,
k
ne
t
rep
r
e
s
ent resp
ectively a node
i
in the input layer, the node
j
in hidden laye
r
and node
k
in the o
u
tput lay
e
r;
j
i
W
and
kj
W
indi
cates
re
spe
c
ti
vely con
n
e
c
tion
weight
bet
wee
n
two no
de
s (n
ode
i
,
j
, and
k
);
j
and
k
indica
tes threshold
value of no
de
j
and n
ode
k
;
The
ni
O
,
nj
O
,
nk
O
represent the outpu
t produ
ce
d by node
i
,
j
, and
k
when i
nput
Nth traini
ng
sampl
e
.
Input vecto
r
n
I
of inp
u
t
sa
mples tra
n
sf
erred
by thre
e laye
rs of f
eed-fo
rward
netwo
rk
gene
rate
s o
u
t
put vector
nk
O
.
T
he sum of
sq
uare
s
0f
the error betwee
n
nk
O
and d
e
si
re
d outp
u
t
nk
T
(k
=1
，
2
，
…
N
3)
can b
e
e
x
presse
d by:
N
n
N
n
N
k
nk
nk
n
O
T
E
E
11
3
1
2
)
(
2
1
2
1
(9)
As for
situati
on that the
whole n
e
two
r
k
has
only on
e output n
o
d
e
, the N3
=1,
equatio
n
(9) cha
nge
s
to:
N
n
nk
nk
N
n
n
O
T
E
E
1
2
1
)
(
2
1
2
1
(10)
The obj
ectiv
e
of netwo
rk learni
ng is
to achieve t
he minimu
m
E through
adju
s
ting
con
n
e
c
tion
weights i
n
the
netwo
rk.
O
p
timization
o
f
the error f
unctio
n
is
a
n
un
con
s
trai
ned
nonlin
ear opt
imization
pro
b
lem by the
way of p
u
tting Equatio
n (3)
～
(7) i
n
to (8). By usi
ng
the
gradi
ent meth
od of optimization, as to
the hidde
n laye
r and outp
u
t layer exists:
k
nk
n
j
nn
kj
k
k
j
ne
t
EE
O
Wn
e
t
W
(
1
1
)
In the above formul
a:
1
nk
nk
kn
k
k
nk
nk
k
pk
pk
nk
nk
O
EE
ne
t
O
ne
t
TO
f
n
e
t
TO
O
O
(12
)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fore
ca
st Mod
e
l of Water Q
uantity Based
on Back
Pro
pagatio
n Artificial Neural…
(Shihu
a Li)
4627
Between the i
nput layer an
d hidde
n laye
r exists:
nj
j
nk
k
ji
n
k
k
n
j
j
ji
p
kp
k
k
k
j
j
n
j
nk
nj
nj
n
On
e
t
On
e
t
EE
W
O
net
O
net
W
TO
f
n
e
t
W
f
n
e
t
O
O
(
1
3
)
In the above formul
a:
1
n
nj
nk
nk
k
k
j
j
k
j
nk
k
j
j
nj
nj
nk
k
j
E
TO
f
n
e
t
W
f
n
e
t
net
Wf
n
e
t
OO
W
(
1
4
)
As for the de
rivation of the threshold val
ue exists:
nk
n
k
E
,
nj
n
j
E
(
1
5
)
The weight
s
and
th
re
sho
l
ds can be obtaine
d by
the gradient
des
ce
nt met
hod, the
formula i
s
:
)
(
)
(
)
1
(
t
W
E
t
W
t
W
(16
)
)
(
)
(
)
1
(
t
E
t
t
(17
)
Among
up
e
quation,
th
e
W
and
indicates th
e
weig
ht vecto
r
th
reshold ve
ctor;
is
calle
d le
arni
n
g
effect
(o
r
called
Lea
rnin
g efficie
n
cy
). The
put
eq
u
a
tion to
(1
0),
(1
2),
(14
)
,
(15)
into (16
)
, (17
)
, the formula of the weight
vector an
d a thre
shol
d vect
or ca
n be gott
en:
n
ny
nx
xy
xy
O
t
W
t
W
)
(
)
1
(
(18)
n
nx
x
x
t
t
)
(
)
1
(
(
1
9
)
Type
()
xy
Wt
rep
r
e
s
e
n
ts the iterati
on value of
the con
nectio
n
weig
hts between
x
and
y
of any adja
c
e
n
t two layer
o
f
feed-forwa
r
d network. Th
e
()
x
t
indi
cate
s the t-th iteration value
of
x
node'
s th
re
sh
old value in
hi
dden l
a
yer o
r
the output lay
e
r. Fo
r a n
o
d
e
x
of output la
yer, there
is:
)
1
(
)
(
)
(
)
(
nx
nx
nx
nx
x
px
nx
nx
O
O
O
T
net
f
O
T
(
2
0
)
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: 4624 – 4
630
4628
For a no
de
x
of hidden layer, we coul
d ab
ain:
()
(1
)
nx
x
n
x
x
x
x
nx
nx
px
x
x
x
fn
e
t
W
OO
W
(21)
What i
s
discussed
ab
ove
is
erro
r b
a
ck p
r
op
agatio
n alg
o
rithm
o
f
error of a
multilayer
feed-fo
rward
netwo
rk. The
learnin
g
function of
BP network is mat
e
riali
z
ed thro
ugh the iterat
ive
pro
c
e
ss,
so the above iterat
ive algorith
m
is calle
d the BP learning
algorithm.
3.
Forecas
t Model's Calc
ulation
and Prediction
Results
Becau
s
e
the
r
e are n
o
t en
o
ugh
outlet flo
w
d
a
ta a
bout
the oil
pipe,
East big
-
dit
c
h
,
We
st
big-dit
c
h
bet
wee
n
L
a
n
z
ho
u rea
c
h
and
An-nin
g fordi
ng. And
flow i
s
little, so
ne
glect th
e influ
ence
of water quantity, utilize the properties of infi
nite approxim
ation of BP algori
thm to seek
the
relation
shi
p
betwe
en flo
w
of Lan
zho
u
se
ction
a
n
d
An-nin
g fording. The
di
stan
ce b
e
tween
Lan
zho
u
se
ct
ions of the Ye
llow Rive
r to down-
stre
am'
s
An-ni
ng fording is a
bout
170
kms.
The ave
r
ag
e
annu
al rate o
f
flow is 1.63
m/s, wate
r flo
w
pa
ssing
fro
m
Lan
zh
ou
section to
the An-ning
fordin
g
will ta
ke
abo
ut 29
hours. So
we
ca
n a
pply th
e flow of th
e
pre
s
ent
pe
rio
d
of
time for predi
cting
usi
n
g
th
e traffic flow
predi
ct
ion
of
Lan
zho
u
se
ction of
the Y
e
llow
Rive
r d
o
w
n-
strea
m
'
s
An-n
ing se
ction'
s
corre
s
p
ondin
g
flow abo
ut 29 hou
rs late
r.
Assu
me that the numbe
r
of neuro
n
s i
s
10; l
earnin
g
rate (L
R) is
0.05; the momentum
con
s
tant
(MC) is 0.9; TE (target e
r
ror) i
s
0.00
1; ME
(maximum
n
u
mbe
r
of ite
r
ations) i
s
50
00.
The
sampl
e
s are traine
d
by Levenb
erg-ma
rqu
a
rdt method a
nd
predi
cted t
w
o
times; the t
w
o
simulatio
n
re
sults
sh
ould
be optimi
z
ed
so a
s
to g
e
t the final p
r
edi
ction. Acco
rdi
ng to the "n
orms"
of hydrol
ogi
cal fore
ca
sting
,
relative e
r
ro
r of
<20%
is
qualified, th
e
predi
ction
results a
r
e
sh
own in
Table 1 an
d Table 2.
Table 1. The
Predi
ction Re
sult of An-nin
g Fordi
ng fro
m
Aug.24 to Sept.2.18
Time13
An-ning
Fi
rs
t
output
Relative
e
r
r
o
r1
(%
)
Second
output
Relative
e
r
r
o
r2
(%
)
predict
Relative
e
r
r
o
r1
(%
)
Aug.23
975
965.17
-1.00821
976.11
0.113846
956.65
1.9886
Aug.24
994
845.81
-14.9085
851.46
-14.34
841.41
-14.229
Aug.25
863
857.03
-0.69177
855.83
-0.83082
857.96
2.3824
Aug.26
938
930.88
-0.75906
934.43
-0.3806
928.12
10.754
Aug.27
844
986.28
16.85782
998.14
18.26303
977.05
15.764
Aug.28
907
746.01
-17.7497
673.2
-25.7773
802.69
-12.656
Aug.29
1060
1060.3
0.028302
1056.3
-0.34906
1063.4
3.2441
Aug.30
832
869.34
4.487981
886.19
6.513221
856.22
0.73198
Aug.31
653
742.56
13.71516
677.09
3.689127
793.53
19.508
Sept.1
1360
1184.4
-12.9118
1049.7
-22.8162
1289.3
0.72402
Qualified rate
100%
90%
100%
Table 2. The
predi
ction
re
sult of An-n
ing
fording from
Aug.24 to Sept.2.23
Time18
An-ning
Fi
rs
t
output
Relative
e
r
r
o
r1
(%
)
Second
output
Relative
e
r
r
o
r2
(%
)
predict
Relative
e
r
r
o
r1
(%
)
Aug.23
1260
1161
-7.85714
1193.2
-5.30159
1043
-17.224
Aug.24
1050
1117.3
6.409524
1127.3
7.361905
1080.6
2.9189
Aug.25
981
831.3
-15.2599
825.56
-15.8451
852.34
-13.115
Aug.26
969
997.83
2.975232
991.46
2.317853
1021.2
5.3847
Aug.27
988
1123.5
13.71457
1137.4
15.12146
1072.6
8.5581
Aug.28
900
1123.5
24.83333
1137.4
26.37778
1072.6
19.173
Aug.29
1070
997.83
-6.74486
991.46
-7.34019
1021.2
-4.5629
Aug.30
950
1112.8
17.13684
1123.2
18.23158
1074.7
13.124
Aug.31
719
1192.4
65.84145
1233.7
71.58554
1041
44.788
Sept.1 1370
1173.2
-14.365
1210
-11.6788
1038.3
-24.21
Qualified rate
100%
90%
100%
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Fore
ca
st Mod
e
l of Water Q
uantity Based
on Back
Pro
pagatio
n Artificial Neural…
(Shihu
a Li)
4629
The
real
me
a
s
ured
value
is
com
pared
with the
p
r
edi
cted
re
sult
s a
s
i
s
sho
w
n
in
Figure 3
to Figure 4.
Figure 3. Co
mpari
s
o
n
of An-nin
g's Fl
o
w
betw
een th
e Real Me
asured a
nd Pre
d
iction fro
m
Aug.
24 to s
ept.2 .18
Figure 4. Co
mpari
s
o
n
of An-nin
g's Fl
o
w
betw
een th
e Real Me
asured a
nd Pre
d
iction fro
m
Aug.
24 to s
ept.2 .23
It is seen from Table
1 to Table 2 t
hat
the accu
racy of the
model
can
meet the
requi
rem
ents,
of predi
ction
:
the qualified
rate of
is m
o
re than
80%. Whe
n
usi
ng
this mod
e
l, we
can p
r
edi
ct the flow ba
se
d former pe
ri
od of fl
ow, and rea
c
h the
purp
o
ses of early wa
rnin
g
and
forecas
t
ing.
4. Conclusio
n
It is seen from Table
1 to Table 2 t
hat
the accu
racy of the
model
can
meet the
requi
rem
ents,
of predi
ction
:
the qualified
rate of
is m
o
re than
80%. Whe
n
usi
ng
this mod
e
l, we
can p
r
edi
ct the flow ba
se
d former pe
ri
od of fl
ow, and rea
c
h the
purp
o
ses of early wa
rnin
g
and
forecastin
g. T
he artifici
al n
eural
network has
so
me b
a
si
c c
har
act
e
rist
ic
s t
o
sim
u
lat
e
t
he h
u
m
an
brain,
su
ch a
s
self ada
ptation, self-orga
n
izatio
n, highl
y parallel i
n
tel
ligen
ce, ro
bu
stne
ss
and fa
ult
informatio
n p
r
ocessin
g
fun
c
tion
s, which
has very
im
p
o
rtant p
r
a
c
tical si
gnifican
c
e for the
co
rrect
descri
p
tion of
nonlin
ear
pro
b
lems. An
d it is go
od
at a
s
so
ciation, g
e
neral
i
z
atio
n, and a
nalogy
and
rea
s
oni
ng;
ca
n refin
e
statistical la
w from
a la
rge
am
o
unt
of statisti
cal data;
a
nd has
be
en wid
e
ly
applie
d and studied in ma
ny fields. Study on predi
cti
on model (B
P-ba
ck p
r
op
a
gation mod
e
l) o
f
Lan
zho
u
section of the Y
e
llow
River can p
r
ovide t
e
ch
nical sup
port fo
r wate
r qu
ality, water
pollution in
ci
dent eme
r
ge
ncy investig
a
t
ion. Espe
ci
al
ly, it has theoretical and
pra
c
tical valu
e in
the integrate
d
water ma
nag
ement, alloca
tion and sch
e
duling in the f
l
ow of Yellow
River.
0
500
1000
1500
123
456
789
1
0
the
real
m
easured
the
p
redicte
d
0
500
1000
1500
123
456
789
1
0
the
real
m
easured
the
p
redicte
d
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: 4624 – 4
630
4630
Referen
ces
[1]
Shui-j
in
g Z
h
u,
Xi
an-c
hen
Z
hu.
T
he princ
i
p
l
e
and
meth
od of pred
icting. Sha
ngh
ai:
S
han
gh
ai
Ji
ao
T
o
n
g
Univers
i
t
y
pres
s.
1991:5-7.
[2]
X
ia
o-do
ng
T
ang. Imag
e R
e
c
ogn
itio
n
B
a
s
ed on Ne
ura l Net
w
orks in
Matlab.
Co
mp
uter a
nd
Dig
ita
l
Engi
neer
in
g.
2007; 35(
8).
[3]
W
e
i-gu
o
Xi
ao,
Chi-
bin
g
H
u
. Investig
atio
n of
meth
o
d
for i
n
telli
ge
nt contro
l
base
d
on
neu
ral n
e
t
w
ork
.
Electrical Drive Automation
. 2
000; 22(
3): 19-
21.
[4]
K En, L
Xu,
H
Chi. Improv
ed
lear
nin
g
a
l
g
o
rithms
for mi
xtu
r
e of e
x
perts i
n
multic
lass classification.
Neur
al Netw
or
ks
. 1999; 12(
9)
: 1229~
12
52.
[5]
I Maqsoo
d, MR Khan, A Abr
aham. An e
n
s
e
mbl
e
of neur
al net
w
o
rks for
w
e
ather for
e
c
a
sting.
Ne
ura
l
Co
mp
uter & Applic
atio
n
. 200
4; 13: 112~
1
2
2
.
[6]
Z
Ahmad, J Z
han
g. Ba
yes
i
a
n
sel
e
ctive c
o
mbin
at
ion
of m
u
ltipl
e
n
eur
al n
e
t
w
o
r
ks for im
provi
ng l
o
n
g
-
rang
e pre
ddicti
ons in n
o
n
lin
ea
r process mod
e
lin
g.
Neur
al C
o
mputer & Ap
p
licatio
n
. 20
05; 14: 78~
87.
[7]
Ji-gu
ang
Ch
en
. T
he fuzz
y
a
r
tificial
neur
al
net
w
o
rk
an
al
ysis of D
a
m o
b
servati
on
dat
a.
Jour
nal o
f
Hydra
u
lic En
gi
neer
ing
. 2
000;
(1): 19~
22.
[8]
Shou-r
en HU. Introducti
on to
Neur
al Net
w
o
r
ks. C
hangs
ha:
Natio
nal D
e
fen
s
e Univ
ersit
y
P
r
ess. 1993.
[9]
Shun-ti
an Lo
u.
S
y
stem Anal
ysis and Des
i
g
n
Based o
n
Matlab. Xi'
an: Xi
'
an
Univers
i
t
y
of
Electroni
c
Scienc
e an
d T
e
chn
o
lo
g
y
Pre
ss. 1998.
[10]
W
en Xi
ng. ne
u
r
al net
w
o
rk ap
plicati
on a
nd d
e
si
g
n
bas
ed on
Matlab. Beij
in
g: Science Pre
ss. 2000.
Evaluation Warning : The document was created with Spire.PDF for Python.