Lecture 12

LECTURE TWELVE 054414 PROCESS CONTROL SYSTEM DESIGN
054414 Process Control System Design
LECTURE 12:
MODEL PREDICTIVE CONTROL
Daniel R. Lewin
Department of Chemical Engineering
Technion, Haifa, Israel
12-1 PROCESS CONTROL SYSTEM DESIGN - (c) Daniel R. Lewin Model Predictive Control
Objectives
On completing this section, you should:
n Be able to explain to a layman how model predictive
control (MPC) works.
o Understand why MPC can improve on the performance
that can be expected from decentralized controllers.
p Be able to correctly tune MPC to obtain a good
response.
q Be able to implement MPC using MATLAB.
Refs: 1) Seborg, D. E., Edgar, T. F. and Mellichamp, D. A., Process

Dynamics and Control, 2
1 Daniel R. Lewin, Technion

Introduction
Thus far, digital controllers have been designed using relatively
simple models to predict the future values of the process
outputs. We shall now consider a more general form of model
predictive control in which:
n The set points are computed based on a constrained
optimization of a steady-state process model.
o A model-based predictive algorithm ensures that the process
outputs track the setpoints over a prespecified horizon.
d
e u +
+
Opt. control process y
ys -
y model - +
constraints
Constrained Optimization
The most common approach is to perform constrained
optimization using a static process model at regular
intervals. This ensures that the process is maintained at
the optimal operating level when confronted with such
things as:
Changes in product demands.
Disturbances such as feedstock changes
Changes in equipment availability
Despite the fact that processes can be highly nonlinear,
the most commonly-used method is LP. This has been
covered in other courses (054374, 054402), so it will not
be discussed further here. This leaves MPC to cover…

MPC made simple…

MPC is a model-based control strategy using models in 2 ways:
Uses a reliable model to predict effect of past control
moves on P future outputs, assuming no future moves.
Uses the same model to compute the optimal M controller
moves. Implement first move and repeat procedure.
past Prediction horizon, P v
target YCL
v
yOL
yk
mk
k k+M k+P
Control horizon, M
Discrete Convolution Models

Evidently, a discrete model describing the effect of inputs on
outputs is required. Commonly, discrete impulse response, or
convolution, models are used.
Consider the response of a typical process to a unit step input:
ai are the step response coefficients
hi are the impulse response coefficients
hi = ai − ai −1 , i = 1, …,H - model horizon.
h0 = 0 (what does this assume?)
H, the model horizon, is selected to
capture 99% of the transient.
H
A discrete convolution model is: yk +1 = ∑ hi mk +1−i

i =1
Predicted Manipulated
value of y variable at k + 1 - i


H
Discrete convolution model: yk +1 = ∑ hi mk +1−i (11.1)
i =1
yk +1 = h1mk + h2mk −1 + … + hH mk +1−H
(11.2)
where mk +1−H = 0 when k + 1 − H < 0
Define ∆mk = mk − mk −1 and substituting in (11.2):

yk +1 = h1mk + h2mk −1 + … + hH mk +1−H
= ( a1 − a0 ) mk + ( a2 − a1 ) mk −1 + … + ( aH − aH −1 ) mk +1−H
= a0mk + a1 ( mk − mk −1 ) + … + aH −1 ( mk +1−H − mk +1−H ) + aH mk +1−H
H
Hence: yk+1 = ∑ ai ∆mk +1−i (11.3)
i =1
Note that this assumes that ∆mk +1−H = mk +1−H − mk −H = mn +1−H

Applying Eq. (11.3), a step change in m at k = 0 ⇒ ∆m0 = m0 – m-1,
with ∆mk = 0, k > 0:
H
k = 0: y1 = ∑ ai ∆m1−i
i =1 0
= a1 ∆m0 + a2 ∆m−1 + … = a1 ∆m0
H
k = 1: y2 = ∑ ai ∆m2−i
i =1 0 0
= a1 ∆m1 + a2 ∆m0 + a3 ∆m−1 + … = a2 ∆m0
In general, yk = ak ∆m0 − predicted values are step response

coefficients multiplied by ∆m0.

ZT-Convolution Model Equivalence

We shall now see how these convolution models are equivalent
to transfer function models studied so far. As an example,
consider the process: p ( s ) = s +1 with a hold and T = 0.2.
1
Then HP ( z ) =
(1 − e ) z
−0.2 −1
=
0.1813z −1
1 − e −0.2z −1 1 − 0.8187z −1
By long division: HP ( z ) = 0.1813z −1 + 0.1484z −2 + 0.1215z −3
+0.0995z −4 + 0.0815z −5 + … (11.4)
To obtain unit step response: Yˆ ( z ) = HP ( z ) M ( z ) = HP ( z ) ⋅ 1 −1

1 −z
y(T) y(2T) y(3T)
By long division: Yˆ ( z ) = 0.1813z −1 + 0.3297z −2 + 0.4512z −3
+0.5507z −4 + 0.6321z −5 + … (11.5)
This is identical to the response: y (kT ) = 1 − e
− kT
ZT-Convolution Model Equivalence

H
Note that Eq. (11.3) is yk+1 = ∑ ai ∆mk +1−i
i =1
Taking Z-transforms of this equation gives:
H
ˆ ( z ) z k +1 = a ∆Mˆ ( z ) z k +1−i Ŷ ( z ) H
Y ∑i = ∑ ai z −i (11.6)
i =1 ∆Mˆ ( z ) i =1
Hence, when ∆mo = 1 ( ∆Mˆ ( z )), Eq. (11.6) is equivalent to Eq.
(11.5), so coefficients a1 = 0.1813, a2 = 0.3297, etc., can be
obtained by measuring the unit-step response.
H
Furthermore, we can write Eq. (11.1): yk +1 = ∑ hi mk +1−i
i =1
Ŷ ( z ) H
in the Z-domain: = HP (z ) = ∑ hi z −i (11.7)
Mˆ ( z ) i =1
Hence, Eq. (11.7) is equivalent to Eq. (11.4), whose coefficients
are h1, h2, etc. and we can check that hi = ai – ai-1.

Matrix Form for Predictive Control

In general, at sample k, we will compute:
M future input changes - ∆mk , ∆mk +1 , …, ∆mk +M −1
input horizon
and P predicted future outputs - yk +1 , yk +2 , …, yk +P

prediction horizon
Generalizing Eq. (11.4):

H H
yk+1 = ∑ ai ∆mk +1−i = a1 ∆mk + ∑ ai ∆mk +1−i
i =1 i =2
future
past
H 2 H
yk+2 = ∑ ai ∆mk +2−i = ∑ ai ∆mk +2−i + ∑ ai ∆mk +2−i
i =1 i =1 i =3
future past

M H
yk +M = ∑ ai ∆mk +M −i + ∑ ai ∆mk +M −i
i =1 i M
= +1
future past
P H
y k +P = ∑
i =P −M +1
ai ∆mk +P −i + ∑ ai ∆mk
i =P +1
+P −i
future past
In summary, at sample interval n + j :

j H
yk + j = ∑
i
ai ∆mk
=1
+ j −i + ∑ ai ∆mk
i j
= +1
+ j −i j ≤M
future adjustments past adjustments (11.8)
j H
yk + j = ∑
i = j −M +1
ai ∆mk + j −i + ∑ ai ∆mk
i = j +1
+ j −i M < j ≤P
future adjustments past adjustments


Thus, in matrix form, the predicted trajectory is:
⎡y k0+1 ⎤
⎡yk +1 ⎤ ⎡a1 0 0 0 ⎤ ⎢ 0 ⎥ ⎡w k ⎤
⎢yk +2 ⎥ ⎢a2 a1 0 0 ⎥ ⎡ ∆mk ⎤ ⎢y k +2 ⎥ ⎢w k +1 ⎥
⎢ ⎥ ⎢ ⎥⎢ ⎥ ⎢ 0 ⎥ ⎢ ⎥
⎢yk +3 ⎥ ⎢a3 a2 a1 0 ⎥ ⎢ ∆mk +1 ⎥ ⎢y k +3 ⎥ ⎢w k +2 ⎥
⎢ ⎥=⎢ ⎥ ⎢ ∆mk +2 ⎥ + ⎢ ⎥+⎢ ⎥
⎢yk +M ⎥ ⎢aM aM −1 aM −2 a1 ⎥ ⎢ ⎥ ⎢y ⎥ ⎢w k +M −1 ⎥
0
⎢ ⎥ ⎢ ⎥⎢ ⎥ ⎢ k +M ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎣ ∆mk +M −1 ⎦ ⎢ ⎥ ⎢ ⎥
⎢⎣yk +P ⎥⎦ ⎢⎣aP aP −1 aP −2 aP −M +1 ⎥⎦ ⎢y 0 ⎥ ⎢⎣w k +P −1 ⎥⎦
⎣ k +P ⎦
y A ∆ m k
y
0
wk
k k
In the above, y k + j is the output predicted without further

0
H
control action: y k + j =
0
∑ ai ∆mk
i j= +1
+ j −i and wk is disturbance effect.
Dynamic Matrix Control

That, is, in matrix form: y k = A ⋅ ∆m k + y k + w k
0
(11.9)
In DMC, the effect of disturbances on the output at k + 1 is
considered to be modeling error, and is estimated by using the
actual value of y measured at instant k :
wˆk = yk − yk (11.10)
Now, our objective is to obtain a desired trajectory, y k

*
Thus, we need to minimize the closed loop prediction error:

e *k = y k* − A ⋅ ∆m k − y k0 − wˆk
Rearranging: e *k (
= −A ⋅ ∆m k + y − y − wˆk
*
k
0
k ) (11.11)
Closed-loop Vector of Open-loop
prediction future control
error prediction
moves error


Clearly, we would like to minimize the closed-loop prediction
error, e k . This can be formally stated as follows:
*
T
min φ = ⎡⎣e k0 − A ⋅ ∆m k ⎤⎦ ⎡⎣e k0 − A ⋅ ∆m k ⎤⎦ (11.12)
∆mk
where e k = y k − y k − wˆk is the open-loop prediction error.

0 * 0
Differentiating Eq. (11.12) with respect to ∆mk, and equating

to zero gives:
∂φ
= −AT ⎣⎡e k0 − A ⋅ ∆m k ⎦⎤ = 0
∂∆m k
( )
−1
Solving gives: ∆m k = A A AT e k0
T
(11.13)
A MxP matrix
of controller
gains

The generic DMC algorithm is implemented as follows:
n The process output, yk, is measured, and used to

estimate the process model error using Eq. (11.10).
o The open-loop prediction error, e k = y k − y k − wˆk , is
0 * 0
updated (accounting for changes in setpoint and

effect of previous controller moves)
( )
−1
p Solve Eq. (11.13): ∆m k = AT A AT e k0
q ∆mk (first step only) is implemented.
r Counter is updated: k = k + 1.

MPC in Action
n The process output, yk, is measured, and used to
estimate the process model error using Eq. (11.10).
past
yk
mk
MPC in Action
o The open-loop prediction error, e k = y k − y k − wˆk , is
0 * 0
updated (accounting for changes in setpoint and

effect of previous controller moves)
past Prediction horizon, P

Target y k*
0
y
yk k
mk
k k+P

MPC in Action
( )
−1
p Solve Eq. (11.13): ∆m k = AT A AT e k0
q ∆mk (first step only) is implemented.

Target y k*
0
y
yk k
mk
k k+M k+P
Control horizon, M
MPC in Action
r Counter is updated: k = k + 1.

Target y k*
yk
mk
k k+M k+P
Control horizon, M
Problem. Control law can result in ringing in ∆mk. This occurs
when A A is ill-conditioned or singular.
T


Solution. Modify the objective function, to penalize for
excessive control moves:
T
φMS = ⎡⎣e k0 − A ⋅ ∆m k ⎤⎦ ⎡⎣e k0 − A ⋅ ∆m k ⎤⎦ + ∆mTk W 2 ∆m k (11.14)
The inclusion of the positive definite matrix, W 2 , in the

objective function is refereed to as “move suppression.”
( )
−1
Minimizing with respect to ∆mk gives: ∆m k = A A +W 2 AT e k0
T
Example. Design MPC/DMC for the process: p ( s ) = e −s

1
(10 s +1)( 5s +1)
Use T = 2 min, and H = 40, and test: (a) M = 1, P = 3, and (b) M =

2, P = 4. Compare results with PID tuned by (c) ZN rule, and (d)
IMC-PID method. Test the performance of the designs for unit
step change in setpoint for nominal process model and for the
case with 50% gain and delay uncertainty.
Comparing MPC with PID (SISO)

Nominal servo responses.
P=3,M=1
P=4,M=2
P=3,M=1
P=4,M=2
ZN tuning is: Kc = 4.2,τI = 6, τD = 4. IMC-PID tuning (with

λ = 2) is: Kc = 3.75,τI = 15, τD = 3.33. Note that IMC-PID
gives similar performance to MPC, superior to ZN (Is this
surprising?)

Comparing MPC with PID (SISO)

Servo responses with 50% uncertainty in gain and delay.
P=3,M=1
P=4,M=2
P=3,M=1
P=4,M=2
Increasing M and P leads to more aggressive control

action and less robustness.
MIMO MPC
The results obtained so far are easily extended to cover MIMO
systems. Consider for example the 2x2 system described by
the following discrete convolution model:
H H
y1,k+1 = ∑ a11,i ∆m1,k +1−i + ∑ a12,i ∆m2,k +1−i
i =1 i =1
H H
y2,k+1 = ∑ a21,i ∆m1,k +1−i + ∑ a22,i ∆m2,k +1−i
i =1 i =1
This model can be transformed into the standard dynamic form:
* 0
E k = −A ⋅ ∆M k + E k
* 0
where E k and E k are vectors of length 2P and ∆Mk is a vector
of length 2M: * T
E k = ⎡⎣e1,*k +1 , e1,*k +2 , …, e1,*k +P , e2,*k +1 , e2,*k +2 , …, e2,*k +P ⎤⎦
T
∆M k = ⎡⎣ ∆m1,k , ∆m1,k +1 , …, ∆m1,k +M −1 , ∆m2,k , ∆m2,k +1 , …, ∆m2,k +M −1 ⎤⎦

MIMO MPC
In this 2x2 case, the matrix A has the following structure:
⎡A A 12 ⎤
A
A = ⎢ 11 ⎥ where each sub-matrix ij has the structure:
A
⎣ 21 A 22 ⎦
⎡aij ,1 0 0 0 ⎤
⎢a aij ,1 0 0 ⎥
ij ,2
⎢ ⎥
Aij = ⎢aij ,3 aij ,2 aij ,1 0 ⎥
⎢ ⎥
⎢ ⎥
⎣⎢aij ,P aij ,P −1 aij ,P −2 aij ,P −M +1 ⎦⎥
Even more general objective function is:

*T * T
φ = E k W 1 E k + ∆M k W 2 ∆M k
( )
−1 0
Minimizing gives: ∆M k = A W 1 A +W 2 AT W 1 E k
T
Comparing MPC with PID (MIMO)

Example. Consider the Wood and Berry distillation column:
⎡ 12.8 −18.9 −3s ⎤
e −s e
⎡xD ⎤ ⎢ 16.7s + 1 21s + 1 ⎥ ⎡L ⎤
= ⎢ ⎥⎢ ⎥
⎣⎢xB ⎦⎥ ⎢ 6.6 e −7 s −19.4 −3s ⎣V ⎦
e ⎥
⎣⎢ 10.9s + 1 14.4s + 1 ⎥⎦
We will compare two MIMO control systems for the process:

(a) an MPC design for H=30, P=6, M = 3, T=3, and W 2 = 0;
(b) a diagonal IMC-PID control system, tuned with λ = [6, 2.5].
Results. The following slides indicate that: (a) Nominally, MPC

provides superior performance to decentralized PID, since it
decouples the process interactions; (b) It is more sensitive to
model uncertainties (50% in gain and delay) than PID.


Example: Wood and Berry Column (Nominal MPC)
P=6, M=3

Example: Wood and Berry Column (Nominal PID)


Example: Wood and Berry Column (MPC - Model Error)
P=6, M=3

Example: Wood and Berry Column (PID - Model Error)

Tuning the MPC Controller

The following recommendations are often considered when
selecting the tunable parameters in MPC:
n H - model horizon.
HT should be selected so that it is greater than the open-loop settling
time (defined as 99 or 95% of response). Commonly 20 ≤ H ≤ 70.
o T – sampling period.
Should be selected to capture important dynamic information.
p P and M - prediction and control horizons.
Generally, P > M. Increasing M increases controller aggressiveness and
decreases robustness. A suitable first guess P ≈ t60/T (number of
samples to reach 60% of open-loop response).
q W1 and W2 – weighting matrices.
W2, generally defined as a diagonal matrix, permits selectively
penalizing particular manipulated variables for MIMO systems (“move
suppression”). W1 allows output scaling to be handled, and is referred
to as “equal concern error” in DMC.
Process Constraints
In practice, the manipulated variables are constrained to
physical, or allowable limits. Furthermore, the output variables
may also be subject to constraints. Thus, in practice, a
constrained optimization problem is solved:
*T * T
min φ = E k W 1 E k + ∆M k W 2 ∆M k (11.15)
∆M k
subject to: ∆m ≤ ∆m k ≤ ∆m (move limits)

L H
mL ≤ mk ≤ mH (manipulated variable bounds)

y ≤ yk ≤ y
L H
(output variable bounds)
This type of problem can be solved using quadratic

programming. Nonlinear versions are solved using successive
quadratic programming (SQP).

Why Model Predictive Control ?

Real processes are multivariable and often involve
significant interaction.
interaction Here’s a good analogy to convince
your future boss to implement MPC.
Driving a car down Haifa’s, Freud Street
When driving a car, we make coordinated use of:
Steering wheel
Brake pedal
Acceleration pedal
Gear stick
Consider how well you could maneuver a car at speed
down Freud Street using only the steering wheel, with
all other controls fixed? Equivalent to SISO control…
Summary
On completing this section, you should:
n Be able to explain to a layman how model predictive
control (MPC) works.
Recall that MPC uses a model in two ways: (a) to predict
the effects of past moves; (b) to optimize future moves.
o Understand why MPC can improve on the performance
that can be expected from decentralized controllers.
Recalling that MPC: (a) implicitly handles constraints and
thus maximizes DOF; (b) effectively decouples process
interactions.
p Be able to correctly tune MPC to obtain a good
response.
q Be able to implement MPC using MATLAB.

Lecture 12

Transféré par

Informations du document

Description originale:

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

Lecture 12

Transféré par

Droits d'auteur :

Formats disponibles

LECTURE TWELVE 054414 PROCESS CONTROL SYSTEM DESIGN

054414 Process Control System Design

Refs: 1) Seborg, D. E., Edgar, T. F. and Mellichamp, D. A., Process

1 Daniel R. Lewin, Technion

2 Daniel R. Lewin, Technion

MPC made simple…

Discrete Convolution Models

A discrete convolution model is: yk +1 = ∑ hi mk +1−i

3 Daniel R. Lewin, Technion

Discrete Convolution Models

Define ∆mk = mk − mk −1 and substituting in (11.2):

Note that this assumes that ∆mk +1−H = mk +1−H − mk −H = mn +1−H

Discrete Convolution Models

In general, yk = ak ∆m0 − predicted values are step response

4 Daniel R. Lewin, Technion

ZT-Convolution Model Equivalence

To obtain unit step response: Yˆ ( z ) = HP ( z ) M ( z ) = HP ( z ) ⋅ 1 −1

ZT-Convolution Model Equivalence

5 Daniel R. Lewin, Technion

Matrix Form for Predictive Control

and P predicted future outputs - yk +1 , yk +2 , …, yk +P

Generalizing Eq. (11.4):

Matrix Form for Predictive Control

In summary, at sample interval n + j :

6 Daniel R. Lewin, Technion

Matrix Form for Predictive Control

In the above, y k + j is the output predicted without further

Dynamic Matrix Control

Now, our objective is to obtain a desired trajectory, y k

Thus, we need to minimize the closed loop prediction error:

7 Daniel R. Lewin, Technion

Dynamic Matrix Control

where e k = y k − y k − wˆk is the open-loop prediction error.

Differentiating Eq. (11.12) with respect to ∆mk, and equating

Dynamic Matrix Control

n The process output, yk, is measured, and used to

updated (accounting for changes in setpoint and

8 Daniel R. Lewin, Technion

updated (accounting for changes in setpoint and

past Prediction horizon, P

9 Daniel R. Lewin, Technion

past Prediction horizon, P

past Prediction horizon, P

10 Daniel R. Lewin, Technion

Dynamic Matrix Control

The inclusion of the positive definite matrix, W 2 , in the

Example. Design MPC/DMC for the process: p ( s ) = e −s

Use T = 2 min, and H = 40, and test: (a) M = 1, P = 3, and (b) M =

Comparing MPC with PID (SISO)

ZN tuning is: Kc = 4.2,τI = 6, τD = 4. IMC-PID tuning (with

11 Daniel R. Lewin, Technion

Comparing MPC with PID (SISO)

Increasing M and P leads to more aggressive control

12 Daniel R. Lewin, Technion

Even more general objective function is:

Comparing MPC with PID (MIMO)

We will compare two MIMO control systems for the process:

Results. The following slides indicate that: (a) Nominally, MPC

13 Daniel R. Lewin, Technion

Comparing MPC with PID (MIMO)

Comparing MPC with PID (MIMO)

14 Daniel R. Lewin, Technion

Comparing MPC with PID (MIMO)