TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.7, July 201
4, pp
. 5244 ~ 52
5
0
DOI: 10.115
9
1
/telkomni
ka.
v
12i7.469
6
5244
Re
cei
v
ed O
c
t
ober 1
0
, 201
3; Revi
se
d Febru
a
ry 21, 2
014; Accepte
d
March 12, 2
014
Delay S
e
parated Neural Network Inver
se Control in
Main-Steam Temperature System
Lingfang Su
n*, Yihang Li, Dan Li
Schoo
l of Auto
mation En
gi
ne
erin
g/Northe
as
t Dianli U
n
iv
ers
i
t
y
NO.169, Cha
n
g
chu
n
Ro
ad, Ji
lin Cit
y, J
ili
n Provinc
e
, Chin
a, +
86-04
32-6
4
8
0
620
1
*Corres
p
o
n
id
n
g
author, e-ma
i
l
: dr_sun
lf@16
3
.com
A
b
st
r
a
ct
In ord
e
r to
i
m
p
r
ove th
e c
ontr
o
l
effect of the
mai
n
stea
m temper
ature w
i
t
h
l
a
rge
ti
me
d
e
lay, th
i
s
pap
er pro
pose
d
a de
lay se
p
a
rated
neur
al
netw
o
rk in
vers
e (DSNNI) con
t
rol sche
m
e. T
he de
lay ti
me
a
n
d
the pos
itive
mo
del w
i
tho
u
t del
ay w
e
re giv
en
by usi
ng a
d
a
p
ti
ve lin
ear
ele
m
ent an
d BP net
w
o
rk. T
he neur
a
l
netw
o
rk inv
e
rse
mod
e
l
of th
e pos
itive
mo
del w
i
tho
u
t d
e
l
ay w
a
s b
u
ilt
on th
at bas
is. An a
ppro
p
ri
a
t
e
referenc
e
mod
e
l w
a
s s
e
l
e
cte
d
to
make
the
invers
e
mo
de
l
’
s outp
u
t s
m
o
o
thin
g. It is
an
o
pen-
loo
p
c
ontr
o
l
system
w
h
en the
m
o
del
is cascaded in
t
o
ori
g
inal system
. It w
ill avoid th
e i
n
stabi
lity caus
ed by the cl
os
ed-
loop control system
s. Off-line
identifi
cation and on-line
identification
ar
e com
b
ined t
o
get
t
he inverse model
in ord
e
r to reduce the ste
a
d
y-st
ate error and
mak
e
the
system hav
e
fine ada
ptive
capacity. Det
a
il
simulati
on test
s are carri
ed
out on th
e giv
en 3
00MW
po
w
e
r unit. T
e
sts show
that the ne
ural
netw
o
r
k
invers
e co
ntrol
w
i
th del
ay ti
me
separ
atio
n ca
n get r
api
d
and
smooth
outp
u
t for t
he
main
stea
m te
mp
eratu
r
e
system. It is able to overco
me
the adverse ef
fects c
aused b
y
the time de
la
y and the par
a
m
eters ch
an
ge
s.
Com
p
ared with the casc
ade P
I
D cont
roller, t
he adjustment
tim
e
of DSNNI reduc
es from
600s t
o380s and
show
s faster respons
e, better robustn
ess an
d anti-i
n
terfere
n
ce perfor
m
an
ce.
Ke
y
w
ords
:
m
a
in steam tem
p
erature syst
em,
large tim
e
delay, neural net
work
, inverse
dynam
i
cs, adaptiv
e
line
a
r el
e
m
ent
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
The m
a
in
ste
a
m temp
erat
ure
of a
coal
-fired
po
we
r
gene
rating
u
n
it is
one
of
the key
para
m
eters required to be
contro
lled
stri
ctly to ensu
r
e
boiler’
s safet
y
and eco
n
o
m
y [1].
Due
to the l
o
ng pip
e
s of th
e supe
rhe
a
te
r resulting the
therm
a
l ine
r
ti
a and
lag
larg
er, the
conve
n
tional
feedba
ck
con
t
rol i
s
difficult to a
c
hieve
g
ood
effect. S
o
me n
e
w co
ntrol
strate
gi
es,
su
ch a
s
pre
d
i
ctive co
ntrol,
fuzzy PID contro
l
and
ge
netic al
go
rith
m have b
een
applie
d to t
he
main
steam
tempe
r
ature
system with
th
e devel
opme
n
t of intellig
e
n
t co
ntrol te
chnolo
g
y [2-4].
But
most re
se
arch is ba
sed
on
feedba
ck
co
ntrol an
d it
s control effe
ct is not goo
d fo
r the larg
e de
lay
sy
st
em,
su
ch
as t
h
e mai
n
st
eam
sy
st
e
m
.
S
o
we
introdu
ce the
co
nce
p
t of inve
rse
dynami
cs of
the thermal
system in this paper. Th
e basi
c
ide
a
of
inverse
cont
rol is to drive
the plant with a
sign
al fro
m
a
co
ntrolle
r
wh
ose
tra
n
sfe
r
f
unctio
n
i
s
the
inverse
of th
e pla
n
t itself [
5
]. It’s an
op
en
loop control
system whe
n
the inverse m
odel is
ca
sca
ded into the
origin
al syste
m
[6]. Feedback
is u
s
ed
only for the
adju
s
t
m
ent of the
controlle
r p
a
ra
meters in trai
ning. Th
e de
sign
prin
cipl
e
of
this co
ntrol m
e
thod is
simpl
e
.
Delay
sep
a
rated n
eural
netwo
rk inve
rse
(D
SNNI) co
ntrol i
s
p
r
opo
se
d to
solve the
delay’s effe
ct on the main
steam temp
eratu
r
e
sy
ste
m
. In this pa
per, sectio
n
2 pre
s
e
n
ts t
he
neural n
e
two
r
k i
n
verse
m
odelin
g p
r
o
c
e
s
s an
d the
m
e
thod
of dela
y
time identifi
c
ation. T
hen
a
control
strate
gy of delay
separ
ated
neu
ral n
e
two
r
k i
n
verse
control
wa
s p
r
op
ose
d
at the e
nd
of
se
ct
ion 2.
I
n
se
ct
ion 3
，
we
ca
rri
ed
out
d
e
tail si
mulati
on te
sts on
a
given
300
M
W
p
o
wer unit
at
rand
om to verify the validity of the control sy
ste
m
prop
osed in this pa
per. Section 4 give
s a
con
c
lu
sio
n
to the whole
pape
r that the DSNNI
control sy
ste
m
has faste
r
resp
on
se b
e
tter
robu
stne
ss a
nd anti-inte
rfe
r
en
ce pe
rformance
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Dela
y Sepa
ra
ted Neu
r
al Network In
verse Control
in Main-Steam
Tem
peratu
r
e
…
(Lingfa
ng
Sun)
5245
2. Neural Netw
o
r
k Iden
tifi
cation o
f
Inv
e
rse Dy
namics Proces
s
2.1. Neural Net
w
o
r
k Inv
e
r
se Modeling
In
ve
r
s
e
s
y
s
t
em is a s
y
s
t
em th
a
t
imp
l
eme
n
t
s
a map
p
ing relatio
n
ship
of
a syst
em
fro
m
the outp
u
t to t
he in
put. That
is, if the
de
si
red
output
()
d
y
t
is
the inp
u
t of in
verse
mo
del,
then the
output of inv
e
rse mo
del i
s
the a
m
ou
nt of cont
rol
()
ut
which
drive
s
th
e plant g
e
ts t
he de
sired
output
()
d
y
t
. The diagra
m
is sho
w
n in Figu
re
1 [7].
Figure 1. Inverse Syste
m
a
nd its Com
p
o
und System
Inverse
syste
m
can
be
di
vided into
in
verse
system
on th
e
right
and
the l
e
ft inverse
system. The l
e
ft inverse sy
stem
can
be
called func
tional observabi
lity. T
he syst
em will get two
different o
u
tp
uts if the
inp
u
ts a
r
e
different at
the
sa
me initial
stat
e. The
syste
m
’s in
put
can
be
resto
r
e
d
thro
ugh the l
e
ft inverse
sy
stem
; the invers
e
system
on th
e right
can
be
calle
d fun
c
tio
nal
rep
r
od
uci
b
ility or fu
nctio
n
a
l
co
ntrol
-
abilit
y. In the b
r
oa
d sen
s
e, it
re
fers to the
tra
cki
ng
ca
pabili
ty
of the sy
ste
m
to a given
referen
c
e
si
gnal. Fo
r a
n
y
desi
r
ed
out
put for the
gi
ven syste
m
, the
inverse mode
l’s output is the cont
rol a
m
ount
()
ut
which will make the original system to track
the de
si
red
o
u
tput. The
r
ef
ore
we n
eed
to buil
d
the
ri
g
h
t inverse
mo
del a
nd th
e
specifi
c
m
odeli
ng
method
is sh
own
in
Figu
re 2. T
he i
n
verse m
odel
ca
scade
s th
e
ori
g
inal
sy
stem
so th
at t
h
e
pse
udo
-line
a
rization
syste
m
ca
n b
e
ob
tained[8]. Th
e erro
r b
e
tween
y
(
k
) a
nd
y
d
(
k
) is
used to
train the inverse mod
e
l.
Figure 2. Inverse System
s on
the Right
Modelin
g Structure
2.2. Dela
y
Ti
me Parameter Identification
It is
imposs
ible to
get the co
mplete
inverse
mod
e
l for
the la
rge
time del
ay sy
ste
m
. Thi
s
is be
cau
s
e sy
stem’s o
u
tput
is zero in del
ay time, and the inverse m
odel’
s
output sho
u
ld be ten
d
to infinite to o
ffset the del
a
y
portion
at th
e same
time.
There is no
so mu
ch
ene
rg
y in the inve
rse
model. So
we excl
ude
the
delay
pa
rt a
nd g
e
t the
in
verse
mo
del f
o
r th
e p
o
rtion
witho
u
t del
a
y
of
the system [9
]. This mean
s we nee
d to know the d
e
la
y paramete
r
exactly.
Adaptive line
a
r
eleme
n
t (A
daline
)
i
s
pro
posed
by Dr.
W
idrow fro
m
Stanford Univ
ersity in
1961
[10]. It is a
continu
o
u
s
-time
line
a
r
netwo
rk a
s
Fi
gure
3.
()
Z
nT
Is the
input ve
ctor,
()
Wn
T
is
the weig
ht vector. The n
e
twork’
s output
is:
T
ˆ
()
()
()
yn
T
W
n
T
Z
n
T
(1)
The wei
ght training
s use least mea
n
square
learnin
g
algorith
m
(LMS). The o
b
jective
func
tion is
J
,
whic
h form is
:
2
ˆ
[(
)
(
)
]
J
Ey
n
T
y
n
T
(2)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5244 – 52
50
5246
Figure 3. Adaline Structu
r
al
Model
The d
e
lay tim
e
is e
s
timate
d u
s
ing th
e e
r
ror fun
c
tion m
i
nimizatio
n
-
()
F
i
. We
su
ppo
se t
he
trained weig
h
t
s
are
i
,
(1
,
2
,
.
.
.
,
)
ip
and combine
i
with the error function.
*
()
i
F
iW
w
(3)
*
W
Is the sum of
weights
coef
ficients. Whe
n
()
F
i
is minimu
m, we get the delay time
estimated val
u
e
ˆ
d
:
mi
n
ˆ
dd
i
(4)
The metho
d
to sep
a
rate th
e delay time from po
sitive model is
sho
w
n in Figu
re
4.
Figure 4. Del
a
y Separate
d
Neural Net
w
ork M
odeli
ng
Diag
ram
Usi
ng pa
rall
el
netwo
rk
stru
cture to i
denti
f
y the positive model
and
delay time ca
n avoid
th
e
sh
or
ta
g
e
o
f
us
ing
a
s
i
n
g
l
e
ne
tw
ork
to id
en
tify.
T
h
e tw
o type
s
o
f
ne
tw
ork
s
us
e d
i
ffe
re
n
t
algorith
m
s to
train respe
c
tively. The positiv
e mod
e
l
without d
e
la
y is identifie
d by BP
（
Bac
k
-
Propa
gation
）
neural net
work. And
bef
ore that
we
should
kn
ow t
he exa
c
t dela
y
time. The pl
ant’s
output is
()
d
yk
an
d the network’s
output is
()
y
k
. Usi
ng
e
to t
r
ain the
po
sitive model wit
hout
delay in form
ula (5
) whi
c
h
ˆ
d
is the estimat
ed value of d
e
lay time identified by Adaline.
ˆ
()
(
)
d
ey
k
y
k
d
2
1
2
Ee
(5)
We u
s
e BP algorithm
with momentum to
train the net
work.
(1
)
(
)
[
(
1
)
(
)
(
1
)
]
l
l
Dl
Dl
(6)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Dela
y Sepa
ra
ted Neu
r
al Network In
verse Control
in Main-Steam
Tem
peratu
r
e
…
(Lingfa
ng
Sun)
5247
2.3. Dela
y
Se
parated Neu
r
al Net
w
o
r
k I
n
v
e
rse Control Sy
stem
Figure 5
sh
ows the whol
e co
ntrol
syst
em’s
stru
cture. We got the
delay time a
n
d
'
()
P
z
—the model
without delay
by using ada
ptive li
near el
ement and B
P
network, a
nd then we m
ade
the ne
ural n
e
twork inverse mo
del
of
'
()
P
z
. It can
avoi
d t
he limit
ca
used by
delay
for inve
rse
modeling. Considered the
stabilit
y and
robustness of
the whole inv
e
rse
cont
rol
system, we add
a refe
ren
c
e
model to m
a
ke a
mod
e
l-referen
c
e
inverse. The
whole
cont
rol system accu
racy
depe
nd
s o
n
t
he a
c
cu
ra
cy
of the m
odel
identification
.
In o
r
de
r to
minimize the
sy
stem’s sta
t
ic
error, we
com
b
ine off-line i
dentificatio
n and on
-li
ne id
entification to
get the final inverse mod
e
l
.
Figure 5. Del
a
y Separate
d
Neural
Net
w
ork Inve
rse Control System
3. Contr
o
l Simulation Te
sts
In orde
r to v
e
rify the qual
ity and rob
u
stnes
s of the
control sy
ste
m
, we take a
300M
W
boiler
unit as the plant. Two cases
are
includ
ed in
t
h
is si
mulat
i
o
n
t
e
st
s:
o
r
igin
al ca
sc
ade
d
P
I
D
control a
nd t
he d
e
lay sep
a
rated
ne
ural
network
inve
rse
control. T
he
system
ca
n be
de
scrib
e
d
as;
Inert zo
ne:
53
01
8
()
18
9
s
Ws
e
s
(7)
Leadi
ng area:
6
02
1.25
()
13
3
s
Ws
e
s
(8)
In inert zo
ne,
=
5
3s
, T
=
89
s,
/T
0.6>0.5. It is a typical l
a
rge time
del
ay system. F
o
r
the
cascaded
PID control, we us
e PI
controller for the
main ci
rcuit
and P
c
ontroller for auxiliar
y
circuit. PID param
eters are obtai
ne
d by decay curve
method.
Auxiliary cont
roller:
2
=2.
5
；
Main controll
er:
1
3.9
2
,
9
8
i
Ts
While fo
r DS
NNI
control system, we u
s
e the main
st
eam temp
era
t
ure a
s
contro
l amount
instea
d of d
e
su
perheate
r
outlet temp
eratu
r
e,
so
we n
eed n
o
inert zone.
We u
s
e
off-line
identificatio
n
to get a
con
v
ergent
controller
whi
c
h
will meet th
e
control
re
quire
ments in
so
me
degree. But error
still exists and t
he
whole sy
stem
can
not ada
pt to the interferen
ce a
nd ti
me
-
varying. The
r
efore, we a
dd on-li
ne id
entificat
ion to improve the adaptive
cap
a
city of the
controlle
r. Th
e identificatio
n of controlle
r is ide
n
tificat
i
on of the inverse mod
e
l. We u
s
e a three
layer BP network wh
ose
input layer, hidden lay
e
r and o
u
tp
ut layer’s n
ode
s is 7,1
0
,1
,
respe
c
tively. We
ado
pt out
put value
s
of the
refe
ren
c
e mo
del i
n
th
e p
r
eviou
s
m
o
ments a
nd t
h
e
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5244 – 52
50
5248
disturban
ce a
s
the network’s input to en
sure the
dyna
mic ch
aracte
ristic an
d the anti-inte
rfere
n
ce
ability. The input vector of the network which
will be trained i
s
:
ˆ
(
)
[
(
1
)
,(
2
)
,(
3
)
,(
4
)
,(
5
)
,
(
6
)
,
(
)
]
mm
m
m
m
m
y
k
N
N
yk
yk
yk
yk
yk
y
k
d
k
(9)
Usi
ng a
si
ne
wave a
s
th
e
excitation
sig
nal, after trai
ning, the
co
m
pari
s
on
of th
e inverse
system’
s
outp
u
t and the ref
e
ren
c
e
cu
rve and t
he traini
ng error a
r
e
shown in Figu
re 6.
(a) T
r
aini
ng si
gnal
s co
ntra
st curve
(b) Er
ro
r cu
rv
e
Figure 6. Neu
r
al Net
w
o
r
k I
n
verse Syste
m
Trainin
g
Curve
The outp
u
t of the system i
s
un
stabl
e in
sh
o
r
t time an
d then follo
w
the refe
ren
c
e
curve
stability. A step signal i
s
gi
ven
to the sy
stem, the compari
s
on of
the inverse
sy
stem’s
output
and
the refe
ren
c
e
cu
rve an
d th
e amou
nt of
control
a
r
e shown
in
Fig
u
r
e
7. We can
see
th
e
inve
rse
model’
s
o
u
tp
ut is exe
c
tly
the co
ntrol
a
m
ount to
di
rv
e the o
r
igin
al
system
tra
c
e the referen
c
e
curv
e.
(a) F
o
llow
cu
r
v
e
(b) Amo
unt of control
Figure 7. DNSSI Output
Figure 8. Con
t
rol System Step Re
spo
n
se Curve
Com
pari
s
on
0
10
0
20
0
30
0
-1.
5
-1
-0.
5
0
0.
5
1
1.
5
2
DS
NN
I
o
u
t
p
u
t
ti
m
e
(
s
)
r
e
f
e
r
enc
e c
u
r
v
e
DS
N
N
I
ou
t
p
ut
0
10
0
200
30
0
-0
.
5
0
0.
5
1
1.
5
er
r
o
r
ti
m
e
(
s
)
0
50
100
150
200
250
30
0
-0.
2
0
0.
2
0.
4
0.
6
0.
8
1
1.
2
out
p
u
t
c
u
rv
e c
o
m
p
a
r
i
s
on
ti
m
e
(
s
)
r
e
f
e
r
enc
e c
u
r
v
e
D
S
N
N
I
out
pu
t
0
20
0
40
0
600
80
0
100
0
-0
.
6
-0
.
4
-0
.
2
0
0.
2
0.
4
am
o
unt
of
c
ont
r
o
l
ti
m
e
(
s
)
0
200
400
600
800
1000
-0
.
2
0
0.
2
0.
4
0.
6
0.
8
1
1.
2
ti
m
e
(
s
)
out
p
u
t
c
u
r
v
e c
o
m
p
a
r
i
s
on
c
a
s
c
aded P
I
D
c
ont
r
o
l
D
S
NN
I
o
u
t
put
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Dela
y Sepa
ra
ted Neu
r
al Network In
verse Control
in Main-Steam
Tem
peratu
r
e
…
(Lingfa
ng
Sun)
5249
We
ma
ke a unit
step perturbation exp
e
rime
nt for these two
co
ntrol sy
stem
and the
comp
ari
s
o
n
of the two
systems i
s
sh
own i
n
Fig
u
re 8. From t
he dia
g
ram
we
can
see
the
ca
scade
d
PID control system
ex
i
s
t
o
v
ersh
oot
a
n
d
sh
oc
k
se
rio
u
sly
.
W
h
ile
DS
N
N
I
c
ont
r
o
l
system’
s
re
spon
se is fa
st
er an
d more stable.
Th
e a
d
justme
nt time of DSNNI control is 3
8
0
s
while
the
ca
scad
ed PI
D
control
is almo
st 60
0s.
An
d
the T
r
a
c
king
cu
rve
of DSNNI co
ntrol
with
no ove
r
shoot
is m
o
re
smoo
th than
ca
sca
ded PID
c
ont
rol o
b
viou
sly. The
r
efo
r
e, al
l sig
n
s sugg
e
s
t
that DNSSI control ha
s bet
te
r dynami
c
chara
c
te
risti
c
s.
The m
a
in
ste
a
m sy
stem i
s
a time
-varyi
ng
sy
stem
at the
same
time an
d the
cha
nge
s
often ap
pea
rs in
ine
r
t zon
e
. So we
cha
nged
the T
in
inert
zo
ne f
r
om 89
s to
20
0s, the
outp
u
ts of
the two co
ntrol system i
s
shown in Figu
re 9.
Figure 9. Curve Compa
r
i
s
on after Co
nstant Time Ch
ange
s
Whe
n
T
cha
n
ged g
r
e
a
tly, the adj
ustm
en
t time of ca
scaded PI
D con
t
rol be
com
e
s
longe
r.
While
DSNNI control
still shows goo
d dynamic characteri
stics, and
the adjustm
ent time is much
sho
r
ter tha
n
cascad
ed PID control’s.
In orde
r to verify the anti-interferen
ce
ability of the system,
we a
dded
a pe
rtu
r
bation
()
dk
=0.1 at
80
0s.
Figu
re 10. shows
th
e ste
p
res
pon
se
with di
stu
r
ba
nce
of th
ese
two
control
ca
se. DS
NNI’
s
di
sturb
a
n
c
e
amplitude i
s
a little larg
e
r
t
han PID
cont
rol’s
whi
c
h i
s
1
.
12, the time to
eliminate the
disturban
ce
is mu
ch
sh
ort
e
r tha
n
PID control and re
duces
abo
ut 120
s. The
r
ef
ore,
DSNNI has a
good a
n
ti-inte
r
fere
nce abilit
y
relatively.
Figure 10. Step Re
sp
on
se
with Distu
r
ba
nce
4. Conclusio
n
A delay sep
a
rated neu
ral n
e
twork inve
rse (D
S
NNI
) co
ntrol sy
stem
wa
s propo
se
d for the
main ste
a
m
system i
n
thi
s
pa
per. It a
v
oids t
he lim
it cau
s
ed
by delay for i
n
verse
mod
e
li
ng.
Adaline a
nd
BP neural
ne
twork a
r
e u
s
e
d
here to
get
the delay tim
e
and th
e inverse mo
del. It is
an o
pen
-loo
p
co
ntrol
syste
m
wh
en
putting the
train
e
d
inverse
mo
del in
front
of the pl
ant
whi
c
h
will avoid the instability caused
by
the closed-loop control syst
em
s. The whole system’
s
out
pu
t
can
tra
c
k the
pred
etermi
ne
d traj
ecto
ry a
c
curately
.
Wh
en a
pplying
this
syste
m
to
the m
a
in
ste
a
m
0
200
400
600
800
1000
-0
.
2
0
0.
2
0.
4
0.
6
0.
8
1
1.
2
ti
m
e
(
s
)
st
e
p
r
e
s
p
o
n
se
c
a
s
c
aded P
I
D
c
ont
rol
D
S
N
N
I
out
put
0
50
0
100
0
1
500
-0
.
2
0
0.
2
0.
4
0.
6
0.
8
1
1.
2
ti
m
e
(
s
)
S
t
ep
res
p
o
n
s
e
w
i
t
h
di
s
t
u
r
ba
nc
e
c
a
s
c
ade
d P
I
D
c
o
n
t
r
o
l
DS
NN
I
o
u
t
p
ut
di
s
t
ur
b
a
n
c
e
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5244 – 52
50
5250
temperature
system, it sh
ows good
dy
namic cha
r
a
c
teristics.
The re
spon
se i
s
faster a
nd sta
b
le.
It is able to o
v
erco
me the
external p
e
rt
urbatio
n qui
ckly. Whe
n
th
e obje
c
t’s
parameter
ch
an
ged,
the whol
e sy
stem can get
stabl
e outp
u
t
without larg
e oversho
o
t and sho
ck. S
o
DSNNI co
n
t
rol
system p
r
op
ose
d
in this pape
r is si
gn
ificantly
improved com
pared with the ca
se of origi
nal
ca
scade PID
control. Thou
gh this control schem
e
sh
ows better p
e
rform
a
n
c
e t
han the cascade
PID c
o
ntrol
，
some
curve
shocks a
ppe
ared
whe
n
tra
c
king
the
given outp
u
t curv
e inevitably.
So,
we will try to find out a feasible method t
o
redu
ce o
r
el
iminate the tracking
sho
cks in future.
Referen
ces
[1]
Lia
n
g
y
u M
a
, Z
hen
xing S
h
i, K
w
a
n
g
Y. Lee.
Super
he
ater S
t
eam T
e
mper
a
t
ure Co
ntrol B
a
sed
on th
e
Expan
de
d-Stru
cture Ne
ural
N
e
tw
ork Inverse
Mode
ls.
ISECS Internati
ona
l
Coll
oq
uium
o
n
Com
putin
g,
Commun
i
cati
o
n
, Control, an
d
Manag
eme
n
t. Yangz
ho
u. 201
0; 131-1
34.
[2]
Lich
uan
Yu
an,
Yanj
un
Di
ng,
Don
g
h
a
i
Li.
F
i
eld
ap
plic
ati
on of
multi
p
le
mode
l pr
ed
ictive co
ntrol f
o
r
super
heat te
mperatur
e in b
o
il
ers.
J
T
s
inghu
a Univ (Sci&T
ech). 2010; 5
0
(8
): 1258-1
2
6
2
.
[3]
Haib
o Lu
o. De
sign a
nd Simu
l
a
tion o
n
PID Control
l
er.
Co
mputer Si
mu
latio
n
. 2012; 3
45-3
48.
[4]
Shih
e Ch
en,
Xin Li, C
h
u
n
le
i Cui, Ya
nju
n
F
a
ng.
Res
earch
on Improv
ed G
A
Boiler
of the
Main Steam
T
e
mperature C
ontrol
l
er Para
meters Optimization.
Automation & Instrumentation
. 20
12; (7
): 6-10.
[5]
Shou
Din
g, Qi
ngh
ui W
u
.
R
e
search
on
Rub
u
stenss of BP
Neur
al
netw
o
rk Based
Inver
s
e Mod
e
l fo
r
Inductio
n
Moto
r Drives.
Internation
a
l Co
nfer
ence on El
ectroni
cs a
nd Optoelectro
n
ics (ICEOE 2011)
.
Dali
an. 20
11; 1
27-1
31.
[6]
Ping W
a
n
g
, Baoh
ua C
hen,
W
encha
o Xin
g
, Hui Di
ng.
T
he Direct Inv
e
rse-
mo
del C
o
ntrol Base
d on
Neur
al
Netw
orks for Inverts
.
Internati
o
n
a
l Confer
ece on
Measur
i
n
g
T
e
chno
log
y
a
nd Mechatro
nics
Automatio
n
. Chan
gsh
a
. 201
0
;
855-85
8.
[7]
Widro
w
B. Adapti
ve Inverse C
ontrol.
US
:
IEEE Press Series
on Pow
e
r Engi
neer
ing
. 2
007.
[8]
Xi
e Pe
izh
ang,
Z
hou
Xi
ng
pe
n
g
. T
i
me Dela
y MI
MO Decou
p
lin
g C
ontro
l B
a
sed
on
DOB
SVM Inverse
Sy
s
t
e
m
.
T
e
lko
m
n
i
ka Ind
o
n
e
si
an Jour
nal
of Electrical E
ngi
ne
erin
g
. 201
3; 11
(12): 752
5-7
5
3
2
.
[9]
Xi
e
P
e
izh
a
n
g
,
Z
hou
Xi
ngp
e
ng.
Sa
nitiz
e
r Dosin
g
Dec
o
u
p
lin
g Co
ntrol base
d
on
IM
C-NN
Inv
e
rs
e
Sy
s
t
e
m
.
T
e
lko
m
n
i
ka Ind
o
n
e
si
an Jour
nal
of Electrical E
ngi
ne
erin
g
. 201
3; 11
(9): 5048-
50
54
.
[10]
Luo P
e
n
g
, Li
Ship
ing, Z
h
ou
Yun. R
e
searc
h
on
per
io
dic
error se
parat
e
base
d
o
n
a
d
aptive
lin
ea
r
elem
ent.
F
o
rei
gn
Electro
n
ic Measur
e
m
ent T
e
chno
logy
. 2
011; 30(
6): 127
-131.
Evaluation Warning : The document was created with Spire.PDF for Python.