EE4C04

MPC Exercise Session

Ton van den Boom ([email protected])

Delft Center for Systems and Control, TU Delft

MPC Exercise Session – EE4C04

 1. Computation of predicted output

Given a system
   
1 a 0
x(k + 1) = x(k) + u(k)
0 1 2
 
y(k) = 1 0 x(k)
 
1
with x(k) = and
1
(
1 for j = 20
u(k + j|k) =
0 elsewhere

1. Compute ŷ(k + 25|k).

Solution:
 
1 a
A=
0 1
 
1 pa
Ap =
0 1
Use the formula
for p(k + j) = y(k + 25):

ŷ(k + 25|k) = C A25 x(k) + C A4 B u(k + 20|k)

     
  1 25 a 1   1 4a 0
= 1 0 + 1 0
0 1 1 0 1 2

= 1 + 25a + 8a

= 1 + 33a

MPC Exercise Session – EE4C04 1

 2. Standard formulation

Consider the IO model

xo (k + 1) = Ao xo (k) + Ko eo (k) + Lo do (k) + Bo u(k)

y(k) = Co xo (k) + eo (k)

where Ao ∈ IRna ×na , Ko ∈ IRna ×1 , Lo ∈ IRna ×1 , Bo ∈ IRna ×1 and

Co ∈ IR1×na . The state is xo (k), do (k) is the known disturbance, u(k) is the input
signal and eo (k) is zero-mean white noise. Note that this is a SISO model.

A performance index J(k) is given by:

XN 
T T
J(u, k) = x̂o (k+j|k)Qx̂o (k+j|k) + u (k+j −1|k)Ru(k+j −1|k)
j=1

T
+ ŷ (k+j|k)S ŷ(k+j|k) (1)

where ˆ denotes prediction, and Q, R, S are positive definite matrices.

Consider the following constraint:

u(k + j) = u(k + j − 1) , for j ≥ Nc (2)

1. Describe the problem of finding the control-law that minimizes performance index (1)
subject to (2) as a standard predictive control problem, and give the corresponding
system matrices and the selection matrices Γ(j).

Solution:
Define the performance signal
 
S 1/2 y(k + 1)
 
z(k) =  Q1/2 x (k + 1) 
 o 
R 1/2 , u(k)

MPC Exercise Session – EE4C04 2

 and selection matrix

 
1 0 0
Γ(j) =  0 1 0 
0 0 1

Then it easy to see that

NX−1
J(v, k) = ẑ T (k + j)Γ(j)ẑ(k + j) =
j=0

XN 
= ŷ T (k+j|k)S ŷ(k+j|k) + x̂T
o (k+j|k)Qx̂o (k+j|k)
j=1

+uT (k+j −1|k)Ru(k+j −1|k)

MPC Exercise Session – EE4C04 3

 For z(k) we obtain:
 
S 1/2 y(k + 1)
 
z(k) =  1/2 x (k + 1) 
 Q o 
R 1/2 u(k)
   
S 1/2 Co Ao xo (k)+Co Ko eo (k)+Co Lo do (k)+Co Bo u(k)
   

= Q1/2 Ao xo (k)+Ko eo (k)+Lo do (k)+Bo u(k)


 
R1/2 , u(k)
   
S 1/2 Co Ao S 1/2 Co Ko
   
=  Q1/2 Ao  xo (k) +  Q1/2 Ko  eo (k)
0 0
   
1/2 S 1/2 Co Bo
S Co Lo  
 
+  Q1/2 Lo  do (k) +  1/2  u(k)
 Q Bo 
0 R 1/2

= C2 x(k) + D21 e(k) + D22 w(k) + D23 v(k)

for    
S 1/2 Co Ao S 1/2 Co Ko
   
C2 =  Q1/2 Ao  D21 =  Q1/2 Ko 
0 0
   
S 1/2 Co Lo S 1/2 Co Bo
 

D22 =  Q1/2 Lo

D23 =  Q 1/2 B 
  o 
0 R 1/2
x(k) = xo (k), e(k) = eo (k), w(k) = do (k), v(k) = u(k)
Finally by choosing

A = Ao , B1 = Ko , B2 = Lo , B3 = Bo

MPC Exercise Session – EE4C04 4

 we obtain the state equation
 
xo (k + 1)
x(k + 1) =
xu (k + 1)
 
Ao xo (k) + Ko eo (k) + Lo do (k) + Bo u(k)
=
Au xu (k) + Bu u(k)

= A x(k) + B1 e(k) + B2 w(k) + B3 v(k)

Define  
v(k|k)

 v(k + 1|k) 

ṽ(k) =  .. 
 . 
v(k + N − 1|k)
with v(k) = u(k) we obtain the constraint:
 
u(k + Nc ) = u(k + Nc − 1)
 u(k + Nc + 1) = u(k + Nc ) 
 
 .. 
 . 
u(k + N − 1) = u(k + N − 2)
 
0 ··· 0 1 −1 0 ... 0

 0 0 0 1 −1 0 

= .. .. .. ..  ṽ(k)
 . 0 . . . 
..
0 . 0 0 0 ··· 1 −1
| {z }
D̃33

= C̃3 x(k) + D̃31 e(k) + D̃32 w̃(k) + D̃33 ṽ(k)

where C̃3 = 0, D̃31 = 0, and D̃32 = 0. ✷

MPC Exercise Session – EE4C04 5

 3. Compute v(k)

Consider a LTI system given in the state space description

x(k + 1) = A x(k) + B3 v(k)

z(k) = C2 x(k) + D23 v(k)

   
3 1
where A = −2, B3 = 1, C2 = and D23 = . Note that there is no noise
0 1
(so e = 0) and there is no external input (w = 0). A performance index is defined as

NX−1
T
min J(ṽ, k) = min ẑ (k + j|k)ẑ(k + j|k) (3)
ṽ(k) ṽ(k) j=0

for N = 1 and x(0) = −10.

1. Compute v(k) for k = 0, . . . , 2.

Solution:    
1 3
Compute Ev = 1, Γ̄ = 1, D̃23 = D23 = , C̃2 = C2 = ,
1 0
 
1 0
and Γ̄ = . The optimal control law is given by
0 1

v(k) = −F x(k)
T −1 T
= −Ev (D̃23 Γ̄D̃23 ) D̃23 Γ̄C̃2 x(k)
   
  1 −1   3
= 1( 1 1 · I ) 1 1 I = −1.5 · x(k)
1 0

MPC Exercise Session – EE4C04 6

 We find

v(0) = −F x(0) = −1.5 · (−10) = 15

x(1) = A x(0) + B3 v(0) = (−2)(−10) + 15 = 35

v(1) = −F x(1) = −1.5 · (35) = −52.5

x(2) = A x(1) + B3 v(1) = −2 (35) − 52.5 = −122.5

v(2) = −F x(2) = −1.5 · (−122.5) = 183.75

x(3) = A x(2) + B3 v(2) = −2 (−122.5) + 183.75 = 428.75

MPC Exercise Session – EE4C04 7

 4. Solution of the SPCP

system

x(k + 1) = A x(k) + B1 e(k) + B2 w(k) + B3 v(k)

y(k) = C1 x(k) + D11 e(k) + D12 w(k)

z(k) = C2 x(k) + D21 e(k) + D22 w(k) + D23 v(k)

with v = ∆u (so we have an IIO system) and a steady-state

(vss , xss , wss , zss ) = (vss , xss , wss , 0)

at time k with state x(k), external signal w̃(k) and ZMWN signal e(k).

1. Define
J ∗ (k) = min J(k)
ṽ
Describe how the optimized performance index J ∗ (k) of the SPCP changes when
• the prediction horizon is increased (decreased),
• the minimum cost horizon is increased (decreased), (but Nm < N )
• the control horizon is increased (decreased), (but Nc < N ),
• an inequality constraint is introduced (removed).

2. Define the optimized performance indices

∗
Ji (k) = min Ji (k) , i = 1, . . . , 6
ṽ

• J1 is the performance index for the SPCP with infinite prediction horizon (N = ∞),
minimum cost horizon Nm = 4 and control horizon Nc = 20 subject to inequality
constraints ψ(k) ≤ Ψ(k).
• J2 is the performance index for the SPCP with prediction horizon N = 50,
minimum cost horizon Nm = 1 and control horizon Nc = 20.

MPC Exercise Session – EE4C04 8

 • J3 is the performance index for the unconstrained SPCP with infinite prediction
horizon (N = ∞), minimum cost horizon Nm = 4 and infinite control horizon
(Nc = ∞).
• J4 is the performance index for the SPCP with infinite prediction horizon (N = ∞),
minimum cost horizon Nm = 1 and control horizon Nc = 20 subject to inequality
constraints ψ(k) ≤ Ψ(k).
• J5 is the performance index for the SPCP with infinite prediction horizon (N = ∞),
minimum cost horizon Nm = 1 and infinite control horizon (Nc = ∞)
• J6 is the performance index for the SPCP with prediction horizon N = 50,
minimum cost horizon Nm = 1 and finite control horizon Nc = 20, subject to
inequality constraints ψ(k) ≤ Ψ(k).

Find a ranking in magnitude Ji∗ ≤ Jj∗ for all combinations i = 1, . . . , 6, j = 1, . . . , 6.

(For example, four fictive performance indices J7∗ , J8∗ , J9∗ and J10 ∗ give a ranking:
J7∗ ≤ J8∗ ≤ J9∗ and J7∗ ≤ J10 ∗ ≤ J ∗ if J ∗ is the smallest, J ∗ is the largest, but the
9 7 9
order in J8∗ and J10
∗ is indefinite). Comment on your choices.

Solution:

1. Define
J ∗ (k) = min J(k)
ṽ
• Increasing (Decreasing) the prediction horizon means that more (less)
terms are taken into account, so performance index J(k) will increase
(decrease).
• Increasing (Decreasing) the minimum cost horizon means that less
(more) terms are taken into account, so performance index J(k) will
decrease (increase).
• Increasing (Decreasing) the control horizon means that we obtain more
(less) degrees of freedom for the optimization, so performance index
J(k) will decrease (increase).
• Introducing (removing) an inequality constraint means that we obtain
less (more) degrees of freedom for the optimization, so performance

MPC Exercise Session – EE4C04 9

 index J(k) will increase (decrease).

2. Define the optimized performance indices

Ji∗ (k) = min Ji (k) , i = 1, . . . , 6

ṽ

PI N Nm Nc ψ
J1∗ ∞ 4 20 Yes
J2∗ 50 1 20 No
J3∗ ∞ 4 ∞ No
J4∗ ∞ 1 20 Yes
J5∗ ∞ 1 ∞ No
J6∗ 50 1 20 Yes
We find the following inequalities

J1∗ ≥ J3∗ J1∗ ≤ J4∗ J2∗ ≤ J4∗

J3∗ ≤ J4∗ J3∗ ≤ J5∗ J4∗ ≥ J5∗ J2∗ ≤ J6∗

∗ ∗
J4 ≥ J6

Based on the inequalities we obtain:

J3∗ ≤ J1∗ ≤ J4∗

J2∗ ≤ J6∗ ≤ J4∗

J3∗ ≤ J5∗ ≤ J4∗

MPC Exercise Session – EE4C04 10

 5.Computation of MPC controller

Consider a LTI system given in the state space description

x(k + 1) = A x(k) + B1 e(k) + B3 v(k)

y(k) = C1 x(k) + D11 e(k)

z(k) = C2 x(k) + D21 e(k) + D23 v(k)

where A = −2, B1 = 2, B3 = 1, C1 = −1, D11 = 1, C2 = 3, D21 = 1, and

D23 = 1. Note that there is no external input (w = 0). A performance index is defined as

NX−1
T
min J(ṽ, k) = min ẑ (k + j|k)ẑ(k + j|k) . (4)
ṽ(k) ṽ(k) j=0

There are no constraints.

1. Compute the poles of the closed loop for N = 1. Is the closed loop stable?
2. Using MATLAB, compute the poles of the closed loop for N = 3. Is the closed loop
stable?
3. Using MATLAB, compute the poles of the closed loop for N = 3 with a control horizon
Nc = 1. Is the closed loop stable?

Solution:

T =D
1. For N = 1 we obtain: Ev = 1, Γ̄ = 1, D̃23 23 = 1
and C̃2 = C2 = 3. The optimal control law is given by

T Γ̄D̃ )−1 D̃ T Γ̄C̃ = 3

F = Ev (D̃23 23 23 2
And so
A − B3 F = −2 − 1 · 3 = −5
−1
A − B1 D11 C1 = −2 − 2 · 1 · (−1) = 0

MPC Exercise Session – EE4C04 11

 There is a closed-loop pole in −5 (outside the unit circle), so system is
unstable.
 
2. For N = 3 we obtain: Ev = 1 0 0 , Γ̄ = I ,
   
C2 3
C̃2 =  C2 A  =  −6 
C2 A2 12

   
D21 1
D̃21 =  C2 B1  =  6 
C2 AB1 −12

 
D23 0 0
D̃23 =  C2 B3 D23 0 
C2 AB3 C2 B3 D23
 
1 0 0
=  3 1 0 
−6 3 1

The optimal control law is given by

T −1 T
F = Ev (D̃23 Γ̄D̃23 ) D̃23 Γ̄C̃2
 T  −1
  1 0 0 1 0 0

= 1 0 0   3 1 0   3 1 0 
−6 3 1 −6 3 1
 T  
1 0 0 3
×  3 1 0   −6 
−6 3 1 12

MPC Exercise Session – EE4C04 12

  −1  
  46 −15 −6 −87
= 1 0 0  −15 10 3   30 
−6 3 1 12
 
  3
= 1 0 0  −15 
75

=3

A − B3 F = −2 − 1 · 3 = −5

−1 C = −2 − 2 · 1 · (−1)0
A − B1 D11 1

3. For Nc = 1 we obtain the constraint

 
0 1 0
ṽ = D̃33 ṽ = 0
0 0 1

and so
 
0
C̃3 = D̃31 =
0

(In this example there is no D̃32 ). First we compute:

  

0 0 1
r r⊥
D̃33 =  1 0  , D̃33 =  0 
0 1 0

r = I and D̃ D̃ r⊥ = 0.)
(check that D̃33 D̃33 33 33

MPC Exercise Session – EE4C04 13

 Now

r⊥ )T D̃ T Γ̄D̃ D̃ r⊥ −1 (D̃ r⊥ )T D̃ T Γ̄
 
Ξ= (D̃33 23 23 33 33 23
    −1
  1 3 −6 1 0 0 1
= 1 0 0  0 1 3   3 1 0   0 
0 0 1 −6 3 1 0
 
  1 3 −6
× 1 0 0  0 1 3 
0 0 1

= (46)−1
 
1 3 −6
 
F = −Ev D̃33 r⊥ ΞD̃ D̃ r C̃ − D̃ r⊥ ΞC̃ − D̃ r C̃
23 33 3 33 2 33 3
r⊥ ΞC̃
= Ev D̃33 2
   
  1   3
=− 1 0 0  0  1/46 3/46 −6/46  −6 
0 12

= 3/46 − 18/46 − 71/46

= −85/46

= −1.8478

A − B3 F = −2 − 1 · (−1.8478) = −0.1522
The system is stable.
✷

MPC Exercise Session – EE4C04 14

 6. Tuning

We design a GPC controller for the following system:

y(k) − 0.6 y(k − 1) + 0.05 y(k − 2) = 3u(k) − 1.1 u(k − 1) + e(k) − 0.5 e(k − 1)

Find the initial setting for the tuning parameters N , Nm and Nc .

Solution:
The transfer function is given by

3 z 2 − 1.1z
G(z) =
(z − 0.6z + 0.05)

Step response:
−1 z
s(k) = Z (G(z) )
z−1

3z 3 − 1.1z 2
= Z −1 (
z 3 − 1.6z 2 + 0.65z − 0.05
38 −1 z 1 z 2 −1 z
= Z ( ) − Z −1 ( )− Z ( )
9 z−1 2 z−1 90 z−1

= 4.2222 − 0.5 · 0.5k − 0.0222 · 0.1k

The settling time ts is the smallest number k for which holds

|0.5 · 0.5k + 0.022 · 0.1k | ≤ 0.05 ∗ 4.222 .

We find ts = 2. This means that the initial setting for the prediction horizon
is: 3 ≤ N ≤ 4. There is no delay or inverse response, so with d = 0 we find
Nm = d + 1 = 1. Finally the IIO system will be a third-order system, so
Nc = 3.
✷

MPC Exercise Session – EE4C04 15