Professional Documents
Culture Documents
Takao Sato
University of Hyogo
2167 Shosha, Himeji
Hyogo
671-2201 Japan
Akira Inoue
Okayama University
3-1-1, Tsushima-naka
Okayama
700-8530 Japan
Toru Yamamoto
Hiroshima University
1-1-1 Kagamiyama,
Higashi-Hiroshima
739-8524 Japan
Abstract
This paper discusses the design methods of Generalized
Mininum Variance Control (GMVC)-based PID controllers
in a two-degree-of-freedom (2DOF) system. The PID pa-
rameters of PID controllers are designed based on GMVC,
and the PID controllers are designed as self-tuning con-
trollers. In design of a 2DOF PID controller based on
GMVC with a feed-forward compensator (abbreviated as
FF-type GMVC), because the order of a feed-forward com-
pensator depends on a dead-time, the longer the dead-time
is, the worse approximation error is. Therefore, in this
paper, to obtain better approximation, GMVC with a pre-
compensator (PC-type GMVC) is approximated by a 2DOF
PID controller. Further, to improve control performance,
a 2DOF PID controller having time-varying proportional
gain is designed based on strongly stable GMVCwith a pre-
compensator.
1. Introduction
This paper discusses the way to realize the performance
of advanced control by using a PID controller. Hence, in
this paper, the PID parameters of a PID controller are de-
signed based on Generalized Minimum Variance Control
(GMVC) [1, 7]. In design of GMVC, a cost function in
which control error and control input are evaluated using
weighting factors is minimized, and a control lawis derived.
So as to obtain the control performance of GMVC by using
a PID controller, the design methods of a GMVC-based PID
controller have been proposed [7, 4, 3].
At rst a GMVC-based PID controller has been de-
signed in a one-degree-of-freedom (1DOF) system [7, 4],
next it has been extended into a two-degree-of-freedom
(2DOF) system [3]. However, a GMVC-based PID con-
troller has been designed using approximation because gen-
erally the order of the coefcient polynomials of GMVC
is higher than that of a PID controller. The conventional
2DOF GMVC-based PID controller has been designed us-
ing a feed-forward compensator [3], and the order of a feed-
forward compensator of 2DOF GMVC [6] depends on a
dead-time. Hence, in design of the conventional 2DOF
GMVC-based PID controller, 2DOF GMVC cannot be ap-
proximated precisely by a 2DOF PID controller when a
dead-time is long. There are several notation for designing
a 2DOF system, and two design methods in a 2DOF system
have been proposed so far in design of GMVC. The one
utilizes a feed-forward compensator (FF-type GMVC) [6],
and the other is designed using a pre-compensator (PC-type
GMVC) [8]. Hence, even if the PC-type and the FF-type
2DOF GMVC are designed equivalently, designed PC-type
and FF-type 2DOF GMVC-based PIDcontrollers are not al-
ways equivalent beacuse in a 2DOF system the number of a
compensator approximated by a PID compensator is larger
than that of a 1DOF system.
In this paper, to improve the performance of a 2DOF
GMVC-based PID controller and approximation error, a
2DOF PID controller is designed newly based on PC-
type 2DOF GMVC [8]. Because the order of the pre-
compensator of PC-type GMVCdoes not depend on a dead-
time, 2DOF GMVCis approximated by the proposed 2DOF
GMVC-based PID controller better than the conventional
2DOF GMVC-based PID controller in case of a long dead-
time. Furthermore, because the proportional gain of the
proposed PID controller is time-varying, the control perfor-
mance is higher than that of xed proportional gain. To
obtain stable time-varying proportional gain, the proposed
PID controller is designed based on strongly stable GMVC
with a pre-compensator.
2. Controlled Plant and PID Controllers
Consider a plant model described by the following.
A[z
1
]y(t) = z
km
B[z
1
]u(t 1) +
(t)
(1)
0-7695-2882-1/07 $25.00 2007 IEEE
where, y(t), u(t), (t) and k
m
are the plant output, the con-
trol input, white Gaussian noise and the dead-time, respec-
tively. A[z
1
] and B[z
1
] are a second-order monic poly-
nomial and an mth-order polynomial, respectively. z
1
de-
notes the backward shift operator.
A 1DOF PID controller is given by
u(t) = k
c
(t)C[z
1
](w(t) y(t)) (2)
C[z
1
] =
+
T
s
T
I
+
T
D
T
s
. (3)
w(t) is the reference input. k
c
(t), T
I
and T
D
are time-
varying proportional gain, integral time and derivative time,
respectively. T
s
denotes a sampling time.
Using a pre-compensator C
p
[z
1
], the 1DOF PID con-
troller is extended to a PC-type 2DOF PID controller given
by the following.
u(t) = k
c
(t)C[z
1
](k
cp
(t)C
p
[z
1
]w(t) y(t)) (4)
C
p
[z
1
] =
1 +
T
Dp
T
s
(5)
where, k
cp
(t) and T
Dp
are time-varying proportional gain
and derivative time, respectively.
3. 1DOF GMVC-Based PID Controller
GMVC derives a control law which minimizes a cost
function expressed as follows.
J = E[(t +k
m
+ 1)
2
] (6)
(t +k
m
+ 1) = P[z
1
]y(t +k
m
+ 1)
+Q[z
1
]u(t) R[z
1
]w(t) (7)
where, P[z
1
], Q[z
1
] and R[z
1
] are the design polyno-
mials of GMVC. To calculate predictive output y(t+k
m
+1)
in Eq. (7), the following Diophantine equation is solved.
P[z
1
] = A[z
1
]E[z
1
] +z
(km+1)
F[z
1
] (8)
E[z
1
] = e
0
+e
1
z
1
+ +e
km
z
km
(9)
F[z
1
] = f
0
+f
1
z
1
+f
2
z
2
(10)
Then, a GMVC law is derived as follows.
u(t) =
R[z
1
]w(t) F[z
1
]y(t)
G[z
1
]
(11)
G[z
1
] = E[z
1
]B[z
1
] +Q[z
1
] (12)
The use of the derived GMVC law gives a closed-loop sys-
tem described by the following.
y(t) =
z
(km+1)
B[z
1
]R[z
1
]
T[z
1
]
w(t) +
G[z
1
]
T[z
1
]
(t)
(13)
T[z
1
] = P[z
1
]B[z
1
] + Q[z
1
]A[z
1
] (14)
In design of a PID controller, to obtain stable time-
varying proportional gain, the controller of a control law to
be compared with a PID controller has to be stable. Hence,
the derived control law is extended as follows [2].
G
e
[z
1
]u(t) = R
e
[z
1
]w(t) F
e
[z
1
]y(t) (15)
G
e
[z
1
] = U
d
[z
1
]G[z
1
] U
n
[z
1
]z
(km+1)
B[z
1
]
(16)
R
e
[z
1
] = U
d
[z
1
]R[z
1
] (17)
F
e
[z
1
] = U
d
[z
1
]F[z
1
] U
n
[z
1
]A[z
1
] (18)
where, U
d
[z
1
] and U
n
[z
1
] are new design polynomials.
Substituting the extended control law into the plant model,
the following closed-loop system is obtained.
y(t) =
z
(km+1)
B[z
1
]R[z
1
]
T[z
1
]
w(t) +
G
e
[z
1
]
T[z
1
]U
d
[z
1
]
(t)
(19)
It follows fromEq. (19) that the transfer function fromw(t)
to y(t) is not changed by introducing the new design poly-
nomials U
d
[z
1
] and U
n
[z
1
]. Consequently, the controller
can be redesigned independently to the closed-loop system,
and a strongly stable controller can be obtained.
To improve a reference response independently to a dis-
turbance response, a control law is extended using a pre-
compensator. The use of a pre-compensator S(z
1
) extends
the strongly stable GMVC described as follows.
G
e
[z
1
]u(t) = z
(km+1)
B[z
1
]R
e
[z
1
]S(z
1
)w(t)
F
e
[z
1
]y(t) (20)
Then, a closed-loop system is given by the following.
y(t) =
z
(km+1)
B[z
1
]R[z
1
]
T[z
1
]
S(z
1
)w(t)
+
G
e
[z
1
]
T[z
1
]U
d
[z
1
]
(t) (21)
Because the pre-compensator S(z
1
) is designed to be sta-
ble, the stability of the controller does not changed.
4. 2DOF GMVC-Based PID Controller
4.1. Design of 1DOF PID Controller
To design a 1DOF GMVC-Based PID controller, from
Eqs. (2) and (11) the following relations have to be satis-
ed.
R[z
1
] = F[z
1
] (22)
k
c
C[z
1
] = F[z
1
] (23)
k
c
(t) =
k
c
G[z
1
]
(24)
From the relations,
k
c
, T
I
and T
D
are obtained as follows.
k
c
= (f
1
+ 2f
2
) (25)
T
I
=
f
1
+ 2f
2
f
0
+f
1
+f
2
T
s
(26)
T
D
=
f
2
f
1
+ 2f
2
T
s
(27)
The proportional gain at step t is decided by the following.
k
c
(t) =
1
g
0
(
k
c
g
1
k
c
(t 1) g
ng
k
c
(t n
g
))
(28)
where,
G[z
1
] = g
0
+g
1
z
1
+ +g
ng
z
ng
. (29)
However, in order to obtain stable time-varying propor-
tional gain, G[z
1
] has to be stable.
4.2. Design of 2DOF GMVC PID Con-
trollers
To obtain stable time-varying proportional gain, the con-
troller to be compared with a PID controller has to be sta-
ble. Hence, in design of our proposed controller, the PID
controller (4) is designed based on strongly stable (20).
Further, to design a PC-type 2DOF GMVC-based PID con-
troller, the following relation has to be satised.
k
cp
(t)C
p
[z
1
] = S(z
1
) (30)
However, generally the order of the compensator S(z
1
) is
higher than that of C
p
[z
1
]. Further, S(z
1
) is a rational
function, but C
p
[z
1
] is a polynomial. Hence, S(z
1
) is
approximated by a rst-order polynomial, and C
p
[z
1
] is
designed using the approximated polynomial.
S(z
1
) =
S
n
[z
1
]
S
d
[z
1
]
S
n
[z
1
]
S
d
[z
1
]
(31)
To obtain an approximated polynomial
S
n
[z
1
], the roots
of S
n
[z
1
] = 0 are calculated, and
max
is selected such
that the absolute value of
max
is the largest among the cal-
culated values. Then the approximated polynomial
S
n
[z
1
]
is obtained as
S
n
[z
1
] = S
n
[1]
max
z
1
max
1
(32)
= s
n0
+ s
n1
z
1
. (33)
From Eqs. (30) and (31), the following equations are
derived.
k
cp
C
p
[z
1
] =
S
n
[z
1
] (34)
k
cp
(t) =
k
cp
S
d
[z
1
]
(35)
Solving Eq. (34),
k
cp
and T
Dp
are given as follows.
k
cp
= s
n0
+ s
n1
(36)
T
Dp
=
s
n1
s
n0
+ s
n1
T
s
(37)
Using Eq. (35), the time-varying proportional gain k
cp
(t)
is decided by the following.
k
cp
(t) =
1
s
d0
(
k
cp
s
d1
k
cp
(t 1)
s
dn
sd
k
cp
(t n
sd
)) (38)
where,
S
d
[z
1
] = s
d0
+s
d1
z
1
+ +s
dn
sd
z
n
sd
. (39)
To obtain stable time-varying proportional gain, S
d
[z
1
]
has to be stable.
5. Numerical Example
A controlled plant is given by transfer function G(s) =
1
(s+2)(s+3)
e
s
. Using the sampling time T
s
= 1[s], a con-
trolled system in discrete-time is given as (1 0.19z
1
+
0.0067z
2
)y(t) = z
1
(0.12 + 0.021z
1
)u(t 1) +
(t)