POD-based

POD based MPC controllers

Oscar Mauricio Agudelo Mañozca

SCD Research Division

Course :
ESAT- K. U. Leuven
Computergestuurde regeltechniek
December 6th, 2012

K. U. LEUVEN
 Introduction
Many physical systems

Heat Exchangers
spatial
C C 
E
discretization
modeled  v  k0Ce RT
t z
Tubular reactors
by T T 
E
 v  Gr Ce RT  H r (Tw  T )
t z

Partial Differential Equations (PDEs)

… High-dimensional
Ri
River systems
t models
Reduced-order
models x (t )  F  x(t ), u(t ) 
Allow the design of
Large number of states
a (t )  ΦTn F  Φ na(t ),
) u(t ) 
Few number of states What to do?

POD-based MPC control schemes

Proper Orthogonal
Decomposition (POD)
+
PLC
• Galerkin p
projection
j
PC Physical system • System Identification

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 2

 p Orthogonal
Introduction to Proper g p
Decomposition ((POD))

Depending of the field of application, POD is also known by other names such as:

 Karhunen-Loève Decomposition (KLD) or Expansion

 Principal Component Analysis (PCA)

 Singular Value Decomposition (SVD)

 other names ….

A little bit of History…

Histo

POD has been developed by several people:

 We can trace
t ace the idea of POD back to independent investigations
in estigations by
b Kosambi (1943),
(1943)
Loève (1945), Karhunen (1946), Pougachev (1953) and Obukhov (1954).

 However if we consider the history of PCA and SVD, we can not forget the work of Pearson
who introduced PCA in 1901,
1901

 and we have to mention the contributions of Beltrami (1873), Jordan (1874), Sylvester
(1889), Schmidt (1907) and Weyl (1912), who were responsible for establishing the
existence of SVD and developing
p g its theory.
y

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 3

 p about POD
Basic concepts
x(t )   N   x1 (t ), x2 (t ), , xN 
T
Let be the state vector of a given dynamical system, and let
X   N  Nd with N d  N be the so-called snapshot matrix

 x1 (t1 ) x1 (t2 )  x1 (t Nd ) 
 
 x2 (t1 ) x2 (t2 )  x2 (t Nd ) 
X   x(t1 ),
) x(t2 )), , x(t Nd )   
    
 
 xN (t1 ) xN (t2 )  xN (t Nd ) 

snapshots

containing a finite number of samples or “snapshots” of the evolution of x(t) at t  t1 , t2 , , t Nd .

In POD we start by observing that each snapshot can be written as a linear combination of a set
of ordered orthonormal basis vectors (POD basis vectors)

N  j   N  basis vector
x(ti )   a j (ti ) j (1)
j 1 a j (ti )  x(ti ), i   Tj x(ti )  coordinate of x(ti) with respect to  j
(POD coefficients)

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 4

 p about POD
Basic concepts

Since the first n most relevant basis vectors capture most of the energy contained in the data,

we const ct an “nth order

e can construct o de approximation”
app o imation” of the snapshots by
b means of the following
follo ing
truncated sequence

n
x n (ti )   a j (ti ) j , i  1, 2, , N d , n  N .
j 1

In POD, the basis vectors are computed in such a way that the reconstruction of the snapshots
using the first n most relevant basis vectors is optimal in the sense that the Sum Squared Error
(SSE) between x(ti ) and x n (ti )

Nd
En (ti )   x(ti )  x n (ti )
2
2
i 1

is minimized.
minimized Herein  2
denotes the L2-norm
norm or Euclidean norm.
norm

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 5

 p about POD
Basic concepts
In other words, the POD basis vectors are the ones that solve the following constrained
optimization problem:
x n (ti )
2
Nd n
min
1 ,,n
 x(t )  
i 1
i
j 1
x(ti ),  j  j
2

subject to orthonormality
1, if i  j condition
iT  j  
0,
0 otherwise.
th i

The orthonormal basis vectors that solve this problem can be found by computing the Singular
Value Decomposition (SVD) of the snapshot matrix X

 1 0  0 Φ   N  N  1 , 2 , ,  N 
Σ   N  Nd       ΦT Φ  ΦΦT  I N
X  ΦΣΨT  N 0  0 
Ψ   Nd  Nd   1 , 2 , , Nd 
POD basis 1   2   3     N  0 ΨΨT  ΨT Ψ  I Nd
vectors

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 6

 p about POD
Basic concepts
If we write equation (1) using a matrix formulation
Γ   N  Nd

X Φ   N  N  a1 (t1 ) a1 (t2 )  a1 (t Nd ) 
 
a ( t ) a ( t )  a (t )
 x(t1 ), , x(t Nd )   1 , ,  N   2 1 2 2 2 Nd  Evolution of the
     
POD coefficients
 
 aN (t1 ) aN (t2 )  aN (t Nd ) 

X  ΦΓ ΦT Φ  I N

Then we obtain the proper orthogonal decomposition of X.

How to choose n (the number of basis vectors) ?

It can be done by using the so-called energy criterion.


n

j 1 j
2 Modeled energy The closer Pn to 1, or similarly the closer
Pn  of 1-Pn to 0, the better the approximation

N

j 1 j
2 Total energy
contained in X of X will be.

determines the truncation degree of Ad-hoc rule: n has to be determined for Pn = 0.99
Pn  the selected POD basis vectors

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 7

 Model Reduction
Model reduction aims to approximate a given high-dimensional system

x (t )  f  x(t ), u(t )  x(t )   N  state vector

(2)
y (t )  g  x(t ), u(t ) 

by a lower complexity model (a model with less number of states and state equations).

Model reduction by means of POD

We can basically distinguish two steps:

 The derivation and selection of the n most relevant basis vectors from an ensemble of
simulation or experimental data of the process described by (2).
(2)

 SVD of the snapshot matrix + Energy criterion (n « N)

 The derivation of the dynamical model for the POD coefficients a1(t), …, an(t) associated to
the selected basis vectors. These coefficients are the states of the reduced-order model.

 Galerkin projection

 System identification

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 8

 p
Examples

 Controll off the

h temperature profile
f l off a one-
dimensional bar

 Control of a non-isothermal tubular chemical

reactor

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 9

 p
Control of the temperature p
profile of a one-dimensional bar
Problem Formulation
Given a perfectly insulated heated bar,
bar
Actuators (Process Inputs)
u1 (t ) u2 (t ) u3 (t ) u1 (t ), u2 (t ), u3 (t ) are the temperatures at
1
x  0, x  L, and x  L
2
Measured Outputs (Process Outputs)
y1 (t ), y2 (t ) are the temperatures at
y1 (t ) y2 (t ) 1 3
x
x  L, and x  L
0 4 4
L

where the dynamics of the temperature profile T(x, t) is modeled by the following PDE (only
temperature variations in the x-direction
direction are considered) :

T ( x, t )  2T ( x, t ) Initial conditions: T ( x, 0)  h( x)
G (3)
t x 2
Boundary conditions : T (0, t )  u1 (t )
  Thermal Conductivity [J  s-1  m -1  K -1 ] (Dirichlet Type) 1
 T ( L, t )  u2 (t )
G   Density [kg  m -3 ] 2
 cp
cp = Heat
H tC it [J  Kg
Capacity K -11  K -11 ] T ( L, t )  u3 (t )

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 10

 p
Control of the temperature p
profile of a one-dimensional bar

Design a control scheme that allows the system to reach a desired temperature
distribution
d st but o in steady state as fast
ast as poss
possible,
b e, while
e sat
satisfying
sy gp process
ocess co
constraints.
st a ts

Desired temperature
Desired profile
Temperature Profile
70
Parameters of the problem
60

 L  0.1 m
[°C ]

50

 G  105 m 2  s -1
C)
Temperature (C

40
emperature

30
 Initial temperature distribution :
T ( x, t  0)  h( x)  0
20
Te

 The process inputs have to satisfy

10 the following constraints:

0
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
0  u1 (t )  150C
x
x [m] 0  u2 (t )  150C
0  u3 (t )  150C
This profile corresponds to the steady state
temperature distribution when the bar is heated
f
from 0°C byb constantt t temperatures
t t ( ) = 30°C,
u1(t)
u2(t) = 60°C and u3(t) =10°C.

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 11

 p
Control of the temperature p
profile of a one-dimensional bar
Discretization of the PDE
For design and simulation purposes, the PDE (3) is discretized by means of “The Implicit
Backward Euler method” (a finite difference method).

u1 (k ) u2 ( k ) u3 (k )
x

••• •••
x0 x1 x2 x3 xP 1 xP xP 1 xN  2 xN 1 xN

T i ( k )  T i ( k  1)  T ( k )  2 T i ( k )  T i 1 ( k ) 
 G  i 1  (4) Ti (k )  T ( xi , tk )
t  x2 
N = Number of sections in
Backward difference Approx. Central difference approx. which the bar is divided
P = Grid-point where u2(k)
for i  1, 2, , P  1, P  1, , N  1 is located.

M = Number of time steps

for k  1, 2, , M
with: T0 (k )  u1 (k ), TP (k )  u2 (k ), TN (k )  u3 (k )

Remark: The implicit backward Euler method is unconditionally stable

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 12

 p
Control of the temperature p
profile of a one-dimensional bar
Discretization of the PDE
If we define T(k)
T(k)=[[ T1(k),…,TP-1 T
P 1(k),TP+1(k),…, TN-1
N 1(k)] as the vector containing the temperatures
of the grid-points x1,…,xP-1, xP+1,… , xN-1 at each time step k, Equation (4) can be cast as follows:

1  2r r 0  0   T1 (k  1))   T1 (k )   r 0 0 t
  r 1  2r      T (k  1)   T (k )  0 0 0 
r G

0  2   2   x 2
 0 r  r 0         
      
    1  2r r  TP  2 (k  1)  TP  2 (k )  0 0 0
 u1 (k ) 
 0  0  r 1  2r   TP 1 (k  1)   TP 1 ( k )  0 r 0  
     u2 ( k ) 
 1  2 r  r 0  0 T
  P 1 ( k  1) T ( k
  P 1  0) r 0
u3 (k ) 
  r 1  2r     TP  2 ( k  1)  TP  2 (k )  0 0 0 
      
 0 r  r 0         
0
    1  2r  r  TN  2 (k  1)  TN 2 (k )  0 0 0
      
 0  0 r 1  2r   TN 1 (k  1)   TN 1 (k )  0
   0 r 

T(k  1) T(k ) B u( k )
A

AT(k  1)  T(k )  Bu(k ) N = 400 T(k )  

398

(5)
1 1
T(k  1)  A T(k )  A Bu(k ) The system has 398 states !!!

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 13

 p
Control of the temperature p
profile of a one-dimensional bar
Reduced-order model of the bar by using POD
g the reduced-order model the subsequent
For deriving q p were followed:
steps

1. Generation of the Snapshot matrix

Pseudo Random Binary Noise signals (PRBNS) were applied to the inputs u1(k), u2(k), and u3(k)
t excite
to it the
th discrete
di t model
d l off the
th bar.
b Th resulting
The lti d t (system
data ( t response)) were used d to
t
construct the snapshot matrix Tsnap.
150
100 T(k  1) T(k  2)  T( k  M )
u1u1(k)
(k)

50
0
 T ( x1 , t1 ) T ( x1 , t2 )  T ( x1 , tM ) 
-50
0 50 100 150 200 250 300 350 400 450 500  T (x , t ) T ( x2 , t2 )  T ( x2 , tM ) 
150  2 1 
100      
 
2(kk)
uu2(k)

50 T(k  1)  A 1T(k )  A 1Bu(k ) Tsnap  T ( xP 1 , t1 ) T ( xP 1 , t2 )  T ( xP 1 , tM ) 

0
-50
T ( xP 1 , t1 ) T ( xP 1 , t2 )  T ( xP 1 , tM ) 
0 50 100 150 200 250 300 350 400 450 500  
150 Discrete model of      
100
tthe
e ba
bar T ( x , t ) T ( x , t )  T ( xN 1 , tM ) 
u3((k)

 N 1 1 N 1 2
u3(k)

50
0
-50
0 50 100 150 200 250 300 350 400 450 500
Tsnap  398500
k
k - samples
p g time  Ts = 1 s
Sampling
Switching probability Number of time steps  M = 500
for the PBRNS = 2%

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 14

 p
Control of the temperature p
profile of a one-dimensional bar
2. Derivation of the POD basis vectors
The POD basis vectors are found by calculating the SVD of the snapshot matrix Tsnap

Tsnap  Φ ΣΨT The columns of  are the POD basis vectors.

Φ   398398  1 ,  2 , , 398 
Spatial Temporal
dynamics dynamics Σ   398500 , Ψ   500500

3. Selection of the n most relevant POD basis vectors

It was done by using the energy criterion presented before.
0
10

-2
10
For this problem, the first n = 10
basis vectors were selected.
-4
10

-6
10 1  Pn  2.454 10-5
-8
10
1-Pn

10

-10
Tn (k )   a j (k ) j  Φ n a(k )
10
j 1

1 , 2 ,10 
-12
10
with: Φ n 

a(k )   a1 (k ),
) a2 (k )), , a10 (k ) 
-14
10 T

-16
10
0 5 10 15 20 25 30 35 40 45 50
n

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 15

 p
Control of the temperature p
profile of a one-dimensional bar
1 2 3
-0.03 0.15 0.05

-0.05
0 05 0 025
0.025 -0.075
0 075

-0.07 -0.1 -0.2

0 0.05 0.1 0 0.05 0.1 0 0.05 0.1

4 5 6
0.15 0.15 0.15

0.025 -0.025 -0.025

-0.1 -0.2 -0.2

0 0.05 0.1 0 0.05 0.1 0 0.05 0.1

7 8 9
0.15 0.1 0.2

-0.05 -0.075 0

-0.25 -0.25 -0.2

0 0.05 0.1 0 0.05 0.1 0 0.05 0.1
x [m] x [m]
 10
0.25

0.025
“Selected POD basis vectors”

-0.2
0 0.05 0.1
x [m]

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 16

 p
Control of the temperature p
profile of a one-dimensional bar
4. Construction of the model for the first n =10 POD coefficients

4.a. Galerkin Projection

If we define a residual function R (T(k )) for Equation (5) as follows:

R(T(k ))  T(k  1)  A 1T(k )  A 1Bu(k )

and we replace T(k) by its nth order approximation Tn(k), the Galerkin projection states that
the projection of R(Tn(k)) on the space spanned by the basis vectors n vanishes. That is,

R(Tn (k )),  j  0; j  1, , n  ,   Euclidean inner product

Replacing T(k) by its nth order approximation Tn(k)= n a(k) in the discrete model of the system
(5),
Φ na(k  1)  A 1Φ n a(k )  A 1Bu(k )

and applying the inner product criterion (Galerkin projection) to the resulting equation, we have
that
Φ na(k  1),  j  A 1Φ na(k )  A 1Bu(k ),  j

j  1, 2, , n  10

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 17

 p
Control of the temperature p
profile of a one-dimensional bar
4. Construction of the model for the first n =10 POD coefficients

4.a. Galerkin Projection

By evaluating the inner product in the previous expression,

ΦTn Φ na(k  1)  ΦTn A 1Φ n a(k )  ΦTn A 1Bu(k ) a, b  aT b  bT a

1 1 ΦTn Φ n  I n
a(k  1)  Φ A Φ n a(k )  Φ A Bu(k )
T
n
T
n

we obtain the model for the first n POD coefficients. The reduced order model of the bar with
only 10 states is then given by

 (k )  Bu
a(k  1)  Aa  (k )
Tn (k )  Φ na(k )

  1010  ΦT A 1Φ , and B  103  ΦT A 1B.

where A n n n

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 18

 p
Control of the temperature p
profile of a one-dimensional bar
4.b. System Identification
Subspace identification techniques can be used to derive the dynamical model between the
process inputs u1(k),
(k) u2(k) and u3(k) and the first n POD coefficients a1(k),a
(k) 2(k),…,a
(k) 10(k).

Tsnap  Φ ΣΨT The first n = 10 rows of ΣΨT corresponds to the evolution

Spatial Temporal
p along time of the first 10 POD coefficients.
coefficients
dynamics dynamics

150 2000 1000

100
Subspace
p identification a1 0 a2 0
(k)k)
u1u1(k

50
algorithm -2000
0 100 200 300 400 500
-1000
0 100 200 300 400 500
0 500 500
k k
-50
0 50 100 150 200 250 300 350 400 450 500 a3 0 a4 0

150
-500 -500
0 100 200 300 400 500 0 100 200 300 400 500
100 500 200
k k
k)
k)

a5 a6
uu2(k

50
2(k

0 0
0
-50
z ( k  1)  Az (k )  Bu(k ) -500
0 100 200 300 400 500
-200
0 100 200 300 400 500
0 50 100 150 200 250 300 350 400 450 500 200 100
k k
150 a(k )  Cz (k )  Du(k ) a7 0 a8 0
100
-200 -100
3((k)

Tn ( k )  Φ n a ( k )
k)

0 100 200 300 400 500 0 100 200 300 400 500
uu3(k

50 50 50
k k
0
a9 0
a10 0
-50
0 50 100 150 200 250 300 350 400 450 500 -50 -50
0 100 200 300 400 500 0 100 200 300 400 500
k Selected model order: 10
k - samples k k
k - samples k - samples
It offers the best trade-
off between accuracy and
complexity Sampling time Ts = 1 s

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 19

 p
Control of the temperature p
profile of a one-dimensional bar
Validation of the reduced-order models
For validating the POD models, constant inputs u1((k)=0°C
) , u2((k)) =100°C and u3((k)=50°C
) were
applied to the full-order model and to the reduced-order models, afterwards their outputs were
compared (Number of time steps = 100).

 POD model derived by

y using
g Galerkin p
projection
j
Blue - Full Order Model Red - Reduced Order Model Average
Average of theAbsolute
of the absolute error
Error
100 100 0.18
Temp. at Time step k = 25
e step k = 1

1 100
ture [°C]

k=1 k = 25
ET   T(k )  T10 (k )
[°C]
922108
error: 0.838564

0.16
50 50 100 k 1
ure
Temperat
Temp. at Time

error: 0.09
Temperatu

0.14
0 0
0.12

[°C]
ure (C)
-50 -50
0.1

Temperatture
0 0.05 0.1 0 0.05 0.1

Temperatu
x x[m] x
x [m]
0.08
100 100
e step k = 80
e step k = 50

k = 50 k = 80
[°C]

ture [°C]

80 80 0.06
280382

042363
error: 0.0ture

60 60
0 04
0.04
Temp. at Time
error: 0.0
Temp. at Time
Temperat

Temperat

40 40
0.02
20 20

0 0 0
0 0.05 0.1 0 0.05 0.1 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
x x[m] x x[m] x x[m]
POD model Full-order model

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 20

 p
Control of the temperature p
profile of a one-dimensional bar
Validation of the reduced-order models

 POD model
d l derived
d i d byb means off subspace
b identification
id tifi ti techniques
t h i

Blue - Full order Model Red - Reduced Order Model Average

Average of theAbsolute
of the absolute error
Error
100 100
Temp. at Time step k = 25
0.3
mp. at Time step k = 1
C]

k=1 k = 25
[°C

1 100
[°C]]

ET   T(k )  T10 (k )
error: 0.844423

error: 0.100039

50 50
Temperature

100 k 1
Temperature

0.25

0 0
0.2

[°C]
Tem

Temperature (C)
-50 -50
0 0.05 0.1 0 0.05 0.1

Temperature
x
x [m] x
x [m] 0.15
100 100
mp. at Time step k = 50

mp. at Time step k = 80

k = 50 k = 80
5C]

2C]
[°C

[°C

80 80

T
error: 0.0555075

error: 0.0850122

0.1
Temperature

Temperature

60 60

40 40 0.05
20 20
Tem

Tem

0 0 0
0 0.05 0.1 0 0.05 0.1 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
x
x [m] xx[m] x x[m]

POD model Full order model

Full-order

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 21

 p
Control of the temperature p
profile of a one-dimensional bar
POD-based MPC control scheme (without a disturbance model)

Control goal: to allow the bar to reach a desired temperature distribution in steady state
as fast as possible, while satisfying the process constraints

The control of the temperature profile of the bar is achieved indirectly by controlling the
POD coefficients. The reference for the POD coefficients are calculated as follows:

aref  ΦTn Tref

Temperature
profile
a ref ( k ) MPC u(k ) Full order Model T( k )
Controller of the Bar y (k )
Measurement
Points

aˆ (k ) Kalman
Filter

Remark:k Theh POD model

d l used
d by
b the
h MPC and
d the
h Kalman
l Filter
l is the
h one derived
d d via
Galerkin projection.

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 22

 p
Control of the temperature p
profile of a one-dimensional bar

Kalman Filter
Observer equations:

Cs  Selection matrix which selects

aˆ (k  1)  Aa  (k )  L  y (k )  yˆ (k ) 
 ˆ (k )  Bu the measured temperatures from
Tn(k)
y (k )  Measured outputs y1(k) and y2(k)
yˆ ( k )  C s Tˆ n ( k )  C s Φ n aˆ ( k )
L  Estimator Gain (Kalman gain)

Th Kalman
The K l gain
i is
i the
th one that
th t minimizes
i i i th steady-state
the t d t t covariance
i lim
k 

li   x(k )  xˆ (k )  x(k )  xˆ (k ) 
T

of the estimation error x(k )  xˆ (k ) .

For
o calculating
ca cu at g the
t e Kalman
a a gagain tthe
e following
o o g design
des g matrices e e used:
at ces were

106 0 
RV   6  R W  I10 x10
 0 10 

Covariance matrix of the measurement Covariance matrix of the process Noise. It tries to
Noise. Its diagonal contains the noise explain unknown disturbances or imperfections in
variance
a a ce oof eac
each p
process
ocess Output ((For
o tthiss the model of the p plant. There is not an specific
p
example such values were assumed). way to select its parameters. (trial and error).

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 23

 p
Control of the temperature p
profile of a one-dimensional bar

The MPC formulation is as follows:

Np N c 1

min  a ref (k  i )  a(k  i ) Q   u(k  i )

2 2
R
a Np , u Nc i 1 i 0

subject to:
 (k  i )  B
a(k  i  1)  Aa
A  (k  i ),
Bu ) i  0,
0 , Np 1
u(k  i )  u(k  i  1)  u(k  i ), i  0, , N c  1
u(k  i )  u(k  i  1), i  N c , , N p  1
u min  u(k  i )  u max , i  0,
0 , Nc  1

with a N P  a( k  1); a(k  2); ; a( k  N p )  u Nc   u(k ); u(k  1); ; u(k  N c  1) 

The control horizon Nc was set to 7 and the prediction horizon Np was selected according the
following criterion:

Prediction
d h
horizon (Np) = Controll horizon
h (Nc) + Largest settling
l time = 80 samples
l

Parameters:
u min   0 0 0 , u max  150 150 150 , Q  I10 , R  I 3
T T

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 24

 p
Control of the temperature p
profile of a one-dimensional bar

Temperature profile in steady-state

and absolute error Evolution of the Temperature
profile
Temperature profile in steady state
60
ylabelre [°C]

40 150
Temperatur

mperature [°C]
20 100
Reference

Temp(C)
Closed-loop
0

Tem
0 0 01
0.01 0 02
0.02 0 03
0.03 0 04
0.04 0 05
0.05 0 06
0.06 0 07
0.07 0 08
0.08 0 09
0.09 01
0.1 50

T
x [m]
Absolute error
0.06 0
ylabel re [°C]

0
0.05
Temperatur

50
0.04
100
0.03 0.1
tk 0.08
150 0.06
0.02 0.04
0 0 01
0.01 0 02
0.02 0 03
0.03 0 04
0.04 0 05
0.05 0 06
0.06 0 07
0.07 0 08
0.08 0 09
0.09 01
0.1 0 02
0.02 x
200 0
x [m] x [m]

The maximum deviation in the temperature

profile is 0.055 °C.

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 25

 p
Control of the temperature p
profile of a one-dimensional bar

Temperature at the measurement points and control actions

of the MPC controller
50 40
[°C]

[°C]
40
30
rature

y2labrature
30

el
el
y1labe
y1(k) - Temper

y2(k) - Temper
20
20
10
10

0 0
0 50 100 150 200 0 50 100 150 200
samples
samples- -kk samples
samples -- kk
130 150 80
[[°C]
[°C]

[°°C]
104
125 60
u2(k) - Temperature
Temperature

Temperature
78
u1label

u2label

u3label
100 40
52
u1(k) - T

u3(k) - T
75 20
26

0 50 0
0 100 200 0 100 200 0 100 200
samples
samples --kk samples
samples--kk samples
samples -- kk

Open loop response Closed loop response

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 26

 p
Control of the temperature p
profile of a one-dimensional bar
Let’s suppose that the left edge of the bar is exposed to the ambient temperature. Therefore
the temperature at the left edge of the bar would be given by :

Temperature profile in steady state

u1 (t ) u2 (t ) u3 (t ) 60

ylabel [°C]
20°C +
40

Temperature
20
y1 (t ) y2 (t ) Reference
Closed-loop
0
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
xlabel
x [m]
T ( x  0, t )  u1 (t )  20C Absolute error
15
ylabel [°C]

maximum
10 deviation =14.22 °C
mperature

Remark :
5
The poor performance was an
Tem

expected result since the control

0
system does not include a 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
mechanism for rejecting such kind xlabel
x [m]
of disturbances.
disturbances

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 27

 p
Control of the temperature p
profile of a one-dimensional bar
POD-based MPC control scheme (with a disturbance model)
Temperature
a reff (k )
MPC u(k ) Full order Model T(k ) profile
Controller of the Bar y (k )
Measurement
Points
dˆ (k ) d (k )
Kalman
aˆ (k )
Filter

The MPC formulation is as follows:

Np N c 1

min  a ref (k  i )  a(k  i ) Q   u(k  i )

2 2
R
a Np , u Nc i 1 i 0 d (k )  disturbance signal
g
subject to:
F  disturbance
a(k  i  1)  A a(k  i )  Bu
 (k  i )  Fd (k  i ), i  0, , N p  1 model matrix
disturbance
equation d (k  i  1)  d (k  i ), i  0, , N p  1
u(k  i )  u(k  i  1)  u(k  i ), i  0, , N c  1
u(k  i )  u(k  i  1), i  N c , , N p  1
u min  u(k  i )  u max , i  0, , N c  1

with a N P  a( k  1); a(k  2); ; a( k  N p )  u Nc   u(k ); u(k  1); ; u(k  N c  1)

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 28

 p
Control of the temperature p
profile of a one-dimensional bar

Since the disturbance d(k) and the state a(k) are unknown, they must be estimated by a new
Kalman filter:

 aˆ (k  1)   A F   aˆ (k )  B
  La 
ˆ 
 
0 1  ˆ     u ( k )  L   y (k )  yˆ (k ) 
 d (k  1)   0
  d (k )     d

yˆ ( k )  C s Tˆ n ( k )  C s Φ n aˆ ( k )

The number of disturbance terms that can be estimated without losing observability is
equal to the number of sensors.

Disturbance model
Since we are interested in rejecting the disturbance associated to the input u1(k), the matrix
F was chosen as follows:
F  B (:,1)  first column of the matrix B

Other parameters of the MPC controller

N c  7 samples, N p  80 samples, u min   0 0 0 , u min  150 150 150 , Q  I10 , R =I 3
T T

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 29

 p
Control of the temperature p
profile of a one-dimensional bar

Temperature profile of the bar Temperature profile of the bar

before the disturbance after the disturbance

60 60
150 80
kk201
= 201 kk210
= 210
k k1 k k20

l [°C]

l [°C]
=1 = 20 50
ylabel e [°C]

ylabel e [°C]
60 40

ure

ure
100 40
Temperature

Temperature

ylabel

ylabel
Temperatu

Temperatu
40 30
20
50
20 20

0 10
0 0 0 0.05 0.1 0 0.05 0.1
0 0.05
0 05 01
0.1 0 0.05
0 05 01
0.1
xlabel
x [m] xlabel
x [m]
xlabel
x [m] xxlabel
[m]
60 60
80 60 kk220
= 220 kk400
= 400
ylabel ure [°C]

k40 k200
k= 200

[°C]

el [°C]
k = 40 50 50
ylabel e [°C]

60

ure

ure
40 40
0 40
0
Temperatu

el
ylabe

ylabe
Temperatu

Temperatu
Temperatur

40
30 30
20
20 20 20

10 10
0 0 0 0.05 0.1 0 0.05 0.1
0 0.05
0 05 01
0.1 0 0.05
0 05 01
0.1
xlabel
x [m] xlabel
x [m]
xlabel
x [m] xlabel
x [m]

Reference Closed loop response

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 30

 p
Control of the temperature p
profile of a one-dimensional bar

Temperature profile in steady-state and

absolute error after the disturbance Evolution of the Temperature
profile
Temperature
Temperature Profile. profileBlue-Steady
Red -Reference, in steady state
State Closed Loop
60
°C]

50 150
Temperature [((C)

mperature [°C]
40
Temperature

30 100
Reference

emp(C)
20 Closed loop response

Tem
Te
10 50
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
x x[m]
SteadyAbsolute error Error
State Absolute
0.25 0
0
C]

0.2
C)
(C
Temperature [°

100
0.15
Temperature

0.1 200
0.1
0.05 kt 300 0.08
0.06
0 04
0.04
0 400 0.02
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 0 x x [m]
x x[m]

The maximum deviation in the temperature

profile is 0.231°C

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 31

 p
Control of the temperature p
profile of a one-dimensional bar

C] 50 40

C]
[°C

[°C
40
emperature

emperature
30
30
1label

2label
ddk 20
20
(k) ddk
(k)
y1

y2
y1(k) - Te

y2(k) - Te
10
10

0 0
0 100 200 300 400 0 100 200 300 400
xlabel - k
samples xlabel - k
samples
130 150 80
u1label ure [°C]

u2label ure [°C]

u3label ure [°C]

104
125 60
ddk
(k)
u1((k) - Temperatu

u2((k) - Temperatu

u3((k) - Temperatu
78 ddk
(k) ddk
(k)
100 40
52
75 20
26

0 50 0
0 200 400 0 200 400 0 200 400
xlabel - k
samples xlabel - k
samples xlabel - k
samples

POD-based MPC controllers Oscar Mauricio Agudelo Mañozca 32