Professional Documents
Culture Documents
FY Eng PDF
FY Eng PDF
Multivariable control
LIC LT
PT FT
Pipe pressure to be kept constant, while desired flow decided by tank level
controller
State-space form:
dx
= Ax + Bu
dt
y = C x + Du
0.018e−2s −0.02e−1.4s
5 6 5 6
ypres 0.42s + 1 0.9s + 1 upump
=
yflow 4.1e−2.64s 12e−4s uvalve
3s + 1 3s + 1
Multivariable system descriptions
0.02
0.01
To: Out(1)
-0.01
-0.02
Amplitude
-0.03
14
12
10
8
To: Out(2)
6
4
2
0
-2
0 5 10 15 20 0 5 10 15 20
Time (seconds)
Multivariable system descriptions
u1 y1
G11 Σ
G21
G12
u2 y2
G22 Σ
Block diagram computations for multivariable systems
Serial connection:
U (s) Y (s)
G1 (s) G2 (s) Y (s) = G2 (s)G1 (s)U (s)
Feedback:
U (s) Y (s)
Σ G1 (s)
−G2 (s)
! "
Y (s) = G1 (s) U (s) − G2 (s) Y (s)
! "
I + G1 (s)G2 (s) Y (s) = G1 (s)U (s)
! " −1
Y (s) = I + G1 (s)G2 (s) G1 (s)U (s)
Each feedback loop by itself might be stable when no other loops closed, but
full system might become unstable when more (stable) loops are closed
Example:
−1
Process
r1 u1 y1
Σ Gc1 G11 Σ
G21
G12
r2 y2
Σ Gc2 u2 G22 Σ
−1
Stability for multivariable systems
Only first loop closed:
−1
r1 u1 y1
Σ Gc1
u2 y2
1 + G11 Gc1 = 0
r1 u1 y1
Σ Gc1
r2 u2 y2
Σ Gc2
−1
G11 Gc1 + Gc1 Gc2 (G11 G22 − G12 G21 ) G12 Gc2
Y1 = R1 + R2
A A
G21 Gc1 G22 Gc2 + Gc1 Gc2 (G11 G22 − G12 G21 )
Y2 = R1 + R2
A A
Stability decided by characteristic equation
A(s) = (1 + G11 Gc1 )(1 + G22 Gc2 ) − G12 G21 Gc1 Gc2 = 0
Multivariable stability – General case
r y
Σ Gc (s) Gp (s)
−1
Characteristic equation:
1 5
0.1s+1 s+1 K1 0
Gp (s) =
1 2 ,
Gr (s) =
0.5s+1 0.4s+1
0 K2
∆ yi
k i j = Gi j (0) =
∆u j
u1 y1
k11 Σ
k21
k12
u2 y2
k22 Σ
y1 = k 11 u1
RGA for 2 $ 2-system
u1 y1
k11 Σ
k21
k12
u2 y2
k22 Σ
Closed loop gain from u1 to y1 when all outputs controlled to zero except y1 :
y2 = k 21 u1 + k 22 u2 = 0 (perfect control)
k 21
u2 = − u1
k 22
3 4
k 12 k 21
y1 = k 11 − u1
k 22
| {z }
=l 11
RGA for 2 $ 2-system
Relative gain:
k 11 k 11
λ11 = =
l 11 k 11 − k12k22k21
K
Note that l 11 = det
k 22
"T
Λ = K . ∗ K −1
!
For Λ , we have
n
X n
X
λi j = λi j = 1
i=1 j =1
Rule of thumb: Inputs and outputs should be connected such that relative
gain is positive and as close to one as possible.
Example 1:
5 Λ 6 Connectivity Interaction
1 0 u1 − y1
None
0 1 u2 − y2
5 6
0 1 u1 − y2
None
1 0 u2 − y1
5 6
0.85 0.15 u1 − y1
Weak
0.15 0.85 u2 − y2
5 6
2 −1 u1 − y1
Difficult
−1 2 u2 − y2
Connecting inputs and outputs
Example 2: 3 $ 3-system
2 1 1
s+1 3 s+3 s+1
1 1
G(s) =
s+1
s+1 1
s+2
1 2
s+1 s+1 s+2
RGA-computation:
2 1
1 −2 0 3
3
3
1
G(0) =
1 2 ,
1
Λ=
4 − 1 − 2
1 1 1 −1 2 0
Connectivity suggestion: u1 − y2 , u2 − y3 , u3 − y1
u1 y1
Gc2
u2 y2
Gc3 Process
u3 y3
Gc1
Decoupling
Idea: Introduce new inputs m and decoupling filter D(s) to make system
easier to control
D
m1 u1 y1
Gc1 Σ G11 Σ
G21
G12
Gc2 m2 Σ G22 Σ
u2 y2
Decoupling
G12
1 −
D(s) =
G11
G21
− 1
G22
G21
G11 − G12 0
T (s) =
G22
G12
0 G22 − G21
G11
This can be controlled using two ordinary controllers
Interpretation as two feedforwards:
m1 u1 y1
Gc1 1 Σ G11 Σ
G21
−G22
G21
G12
−G11
G12
Gc2 m2 1 Σ G22 Σ
u2 y2
Implementing the decoupling
G12
Decoupling elements − G and − G21
G cannot be implemented if they contain
11 22
◮ pure derivatives
Solution options:
Process:
0.018e−2s −0.02e−1.4s
0.42s + 1 0.9s + 1
G(s) =
4.1e−2.64s
12e−4s
3s + 1 3s + 1
RGA: 5 6
0.73 0.27
Λ=
0.27 0.73
Conclusion: Control pressure with pump and flow with valve, some interaction
between loops
Simulation with decoupling (solid) and without (dashed):
Flow Pressure
85 35
80
30
75
70 25
65
20
60
55 15
0 100 200 300 0 100 200 300
Valve Pump
50 70
45 65
40 60
35 55
30 50
0 100 200 300 0 100 200 300
Model predictive control (MPC)
Basics:
y1
Outputs
y2
t
k k+H
u2
Inputs
u1
t
k k+H
Optimization over receding horizon
y1
Outputs
y2
t
k+1 k+H+1
u2
Inputs
u1
t
k+1 k+H+1
Optimization problem example
Linear model:
dx
= Ax + Bu
dt
y = C x + Du
+H
kX k+
X H −1
2 2
ρ 1 ∆u21 + ρ 2 ∆u22
! " ! "
J= w1 (r1 − y1 ) + w2 (r2 − y2 ) +
t= k+1 t= k
◮ Unidirected influence from control loop to another does not affect stability
and the effect can be eliminated by feedforward
◮ Multivariable interaction can affect both stability and performance to the
worse. Interaction can be reduced with decoupling filter
◮ Analysis tool RGA describes the difficulty of the interaction and what
outputs that should be fed back to what inputs
◮ Difficult interaction or high performance demands might require a
multivariable controller, e.g., MPC