Model Predictive Control

Model Predictive Control

(MPC)

Overview of Model Predictive Control

Impulse/Step Response Model Identification

Predictions for SISO and MIMO Models

Model Predictive Control Calculations

Selection of Design and Tuning Parameters

1
 MPC - Motivation
Practical Problems
Model Predictive Control

multivariable
difficult dynamic behavior
nonlinear
constraints (input, output)

Overall Objectives of MPC

Prevent violations of input and output constraints.
Drive some output variables to their optimal set points,
while maintaining other outputs within specified ranges.
Prevent excessive movement of the input variables.

2
 MPC - Basic Concepts
1. Future values of output variables are predicted using a
Model Predictive Control

dynamic model of the process and current measurements.

Unlike time delay compensation methods, the predictions are
made for more than one time delay ahead.
2. The control calculations are based on both future predictions
and current measurements.
3. The manipulated variables, u(k), at the k-th sampling instant
are calculated so that they minimize an objective function, J.
Example: Minimize the sum of the squares of the deviations
between predicted future outputs and specific reference
trajectory.
4. Inequality & equality constraints, and measured disturbances
are included in the control calculations.
5. The calculated manipulated variables are implemented as set
point for lower level control loops. (cf. cascade control).

3
Model Predictive Control
MPC Block Diagram

Basic Elements of MPC

 Reference Trajectory Specification
 Process Output Prediction (using Model)
 Control Action Sequence Computation (programming problem)
 Error Prediction Update (feedback)

4
 Model Predictive Control
Control Hierarchy

5
 MPC - Calculations
1. At the k-th sampling instant, the values of the manipulated
variables, u, at the next M sampling instants, {u(k), u(k+1), ,
Model Predictive Control

u(k+M -1)} are calculated.

This set of M control moves is calculated so as to minimize
the predicted deviations from the reference trajectory over the
next P sampling instants while satisfying the constraints.
Typically, an LP or QP problem is solved at each sampling
instant.
Terminology: M = control horizon, P = prediction horizon
2. Then the first control move, u(k), is implemented.
3. At the next sampling instant, k+1, the M-step control policy is
re-calculated for the next M sampling instants, k+1 to k+M, and
implement the first control move, u(k+1).
4. Then Steps 1 and 2 are repeated for subsequent sampling
instants.
Note: This approach is an example of a receding horizon
approach.
6
Model Predictive Control

Figure 20.2 Basic concept for Model Predictive Control

7
Model Predictive Control
Moving Horizon Concept of MPC

8
 When Should MPC be Used?

1. Processes are difficult to control with standard PID

Model Predictive Control

algorithm (e.g., large time constants, substantial time

delays, inverse response, etc.)
2. There is significant process interactions between u
and y.
i.e., more than one manipulated variable has a
significant effect on an important process variable.

3. Constraints (limits) on process variables and

manipulated variables are important for normal control.

Terminology:

y CV, u MV, d DV

9
 MPC History
Model Algorithmic Control (MAC) (1978)
Model Predictive Control

Finite impulse response model

Dynamic Matrix Control (DMC) (1980)
Step response model
Control calculation by least-squares method (no constraints)
Quadratic Dynamic Matrix Control (QDMC) (1984)
Step response model
Control calculation by quadratic programming (with constraints)
Generalized Predictive Control (GPC) (1987)
Transfer function model
Nonlinear Model Predictive Control (NMPC)

Over 5000 applications of MPC since 1980

(Qin and Badgwell, 2003)

10
 Dynamic Models for MPC
Could be either:
Model Predictive Control

1. Physical or empirical (but usually empirical)

2. Linear or nonlinear (but usually linear)
3. Parametric or non-parametric

Typical linear models used in MPC:

1. Impulse response models
2. Step response models
3. Transfer function models
4. State-space models

Note: Can convert one type of linear model to the

other types
Discrete-time models are more convenient for
prediction
11
 Discrete Impulse Response Models
Consider a single input, single output process:
Model Predictive Control

uu (k) Process yy (k)

where u and y are deviation variables, u(k) and y(k) are

their measurements at k-th sampling instant.

Definition: impulse response is the response of a

relaxed process to a unit pulse (impulse) excitation at t = 0

{h i } i = 0, 1, 2 , 3,
Process input-output relationship

y ( k ) = hi u ( k i ) (convolution summation)
i =0
12
 For a stable process: h i 0 for i > N
 Finite Impulse response (FIR)
N
Model Predictive Control

y ( k ) = hi u ( k i ) (h0 = 0)
i =0

For a multivariable process with r inputs and m outputs,

the representation becomes a impulse response
matrix
h11,k h1r ,k

Hk = k = 0, 1, 2, 3,
h h
m1 ,k mr ,k
hij,k : the FIR between the j-th input and the i-th output
N
y (k ) = H i u (k i)
i =0
13
 FIR Identification
Identification problem: given input u(k) and output
measurement y(k)  find the FIR, hi
Model Predictive Control

e(k) (noise)

u(k) Process w(k) + y(k)

(FIR, hi)

Assumptions:
The input u(k) is a continuing driving function of the process.
We observe u(k) for 0 k n + N where n > N.
The output sequence y(k) for N k n + N is also observed.
The noise e(k) is a random sequence with zero mean and is
uncorrelated with u(k).

14
 Using the input-output relationship, we have n+1 eqs.

k = N, y ( N ) = h0 u ( N ) + h1 u ( N 1) + + hN u ( 0 ) + e ( N )
Model Predictive Control

k = N + 1, y ( N + 1) = h0 u ( N + 1) + h1 u ( N ) + + hN u (1) + e ( N + 1)

k = N + n, y ( N + n ) = h0 u ( N + n ) + h1 u ( N + n 1) + + hN u ( n ) + e ( N + n )

Written in vector form y = Uh + e

y(N ) e( N ) h0 u(N) u ( N 1) u ( 0)

y ( N + 1) e ( N + 1) h1
u ( N + 1) u(N) u (1)
y= ; e= ; h = ; U =


y ( N + n ) e ( N + n ) h N u ( N + n ) u ( N + n 1) u ( n )

n+1 equations, N+1 unknowns (n > N)

 No exact solution
15
 Least-Squares Estimation
Estimate the unknown parameter vector h by the
method of least-squares
Model Predictive Control

(minimize the error criterion J = eTe with respect to h

where e = y Uh )

( )
1
= UT U
Solution : h UT y (Least-squares estimator)

Accuracy of h
( )
=h
h is an unbiased estimate of h, i.e. E h
h is a consistent estimate of h if n

Potential problems
The selection of settling time parameter N (unknown a priori)
Usually needs to repeat the computation with progressively
increasing N values
A large N value may cause computational difficulty
16
 Multivariable FIR Identification
Input-output relationship for the i-th output
r N
yi ( k ) = hij , u j ( k )
Model Predictive Control

j =1 =0

y i = [ U1 U 2 U r ] hi + ei = Uhi + ei

yi ( N ) ei ( N ) uj (N ) u j ( N 1) u j ( 0)

y ( N + 1) e ( N + 1) u j ( N + 1) uj (N) u j (1)
yi = i ; ei = i ; Uj = ;


yi ( N + n ) ei ( N + n ) u j ( N + n ) u j ( N + n 1)
u j ( n )
hij ,0

h
ij ,1 T T
hij = ; h = h T
h T
h
i1
i i2 ir

hij ,N

( )
1
Least-squares estimation = UT U
h UT y i
i
17
 Discrete Step Response Models
Model Predictive Control

The step response Si is related to the impulse response hi

For unit step input, u(k)=1

N
S k = y ( k ) = hi u ( k i )
i =0
k
S k = hi (S0 = 0)
i =0

or hi = Si Si 1

18
 Prediction for SISO Step Response Models
The step response model of a stable SISO process
N N
( )
y ( k + 1) = hi u ( k i + 1) = S i S i 1 u ( k i + 1)
Model Predictive Control

i =1 i =1
N 1
y ( k +1) = Si u ( k i + 1) + S N u ( k N + 1)
i =1

Si = the i-th step response coefficient

N = an integer (the model horizon)
u(k ) = u(k ) u(k 1)
= change in the input from one sampling instant to the next

Let k denote the current sampling instant and y ( k +1)

denote the prediction of y(k+1) at time k
One-step-ahead prediction
N 1
y ( k +1) = Si u ( k i + 1) + S N u ( k N + 1)
i =1 19
 Rearrange as
N 1
y (k + 1) = S1u (k )
 
+ Si u(k i + 1) + S N u(k N + 1)
i =2

Effect of current
Model Predictive Control

control action Effect of past control actions

Two-step-ahead prediction (k = k+1)

N 1
y (k + 2) = S1u (k + 1) +

S2 u (k )
 
+ Si u(k i + 2) + S N u(k N + 2)
i =3

Effect of future Effect of current
control action control action Effect of past control actions

j -step ahead prediction

j N 1
y (k + j ) = Si u (k + j i) + Si u (k + j i) + S N u (k + j N )
i =1
 
 i = j +1
 
Effects of current and Effects of past
future control actions control actions
j
y ( k + j ) = i
S u ( k + j i ) +
y o
(k + j )
i =1
Predicted unforced response (no current or future
control actions, u ( k + i ) = u ( k 1) for i 0 )
20
 Example : Derive a predictive control law based on the
following concept. A single control move, u (k ) , is calculated so
that the J-step-ahead prediction is equal to the set-point.
Assume u (k + i ) = 0 for i > 0.
Model Predictive Control

Solution:
A single prediction for J step ahead y ( k + J ) = S J u (k ) + y ( k + J )
o

Setting y ( k + J ) = ysp and rearranging gives the desired predictive

controller
ysp y o ( k + J )
u ( k ) =
SJ

Apply the predictive

control law to the process
Y (s) 1
=
U ( s) ( 5s + 1)5

Large J  sluggish response

21
 Vector Notation for Multiple Predictions
Define vectors:
Y ( k +1) : predicted response for the next P sampling instants
Model Predictive Control

Y o ( k +1) : predicted unforced response

U ( k ) : control actions for the next M sampling instants

y ( k+1) y o ( k+1)

y ( )
k+ 2 y o ( k+2 )
Y ( k +1) Y ( k +1)
o



y ( k+P )
y ( k+P )
o

u ( k )

u ( )
k+1
U ( k )


u ( k+M -1)
22
 Dynamic Matrix Model
The MPC control calculations are based on calculating
U ( k ) so that the predicted outputs move optimally to the
Model Predictive Control

new set-point
The model predictions can be written as
Y ( k +1) = S U ( k ) + Y o ( k +1)

where S is the P M dynamic matrix

S1 0 0
S2 S1 0

0

S SM S M -1 S1
S M +1 SM S2


SP S P-1 S P-M +1

23
 Output feedback: Bias Correction
Sources of inaccurate prediction:
model inaccuracy and unmeasured disturbance
d(k) (disturbance)
Model Predictive Control

u y ( k ) + y(k)
Process

The model predictions can be corrected by utilizing the

latest measurement, y(k).
The corrected prediction is defined to be:
y ( k + j ) y ( k + j ) + d ( k + j ) d ( k + j ) = y ( k ) y ( k )
Estimated disturbance
y ( k + j ) + y ( k ) y ( k ) (Residual)
Adding this bias correction to each prediction gives
Y ( k +1) = S U ( k ) + Y o ( k +1) + 1 y ( k ) y ( k )
(Assume process disturbance is constant for j=1,2,,P)
24
 EXAMPLE
The benefits of using corrected predictions will be illustrated by a simple
example, the first-order plus-time-delay model:
Y(s) 5e-2s
Model Predictive Control

=
U(s) 15s + 1
Assume that the disturbance transfer function is identical to the
process transfer function, Gd(s)=Gp(s). A unit step change in u
occurs at time t=2 min and a step disturbance, d=0.15, occurs at
t=8 min. The sampling period is t = 1 min.
(a) Compare the process response y(k) with the predictions that
were made 15 steps earlier based on a step response model with
N=80.
(b) Repeat part (a) for the situation where the step response coefficients
are calculated using an incorrect model:
Y(s) 4e-2 s
=
U(s) 20s + 1

25
 y ( k +15 ) y ( k +15 ) + d ( k +15 )
y ( k +15 ) + y ( k ) y ( k )
Model Predictive Control

Bias correction begins Bias correction begins

after t = 25 (k = 10) after t = 19 (k = 4)

(a) Without model error. (b) With model error.

26
 Prediction for MIMO Step Response Models
By the Principle of Superposition
N 1 N 1
y1 ( k + j ) = S11,i u1 ( k + j i ) + S11,N u1 ( k + j N ) + S12,i u2 ( k + j i ) + S12,N u2 ( k + j N )
Model Predictive Control

i =1 i =1
N 1 N 1
y2 ( k + j ) = S21,i u1 ( k + j i ) + S21,N u1 ( k + j N ) + S22,i u2 ( k + j i ) + S22,N u2 ( k + j N )
i =1 i =1

Let the output vector be y = [ y1 y2 ym ] and the input

T

vector be u = [u1 u2 ur ]
T

The MIMO model for the corrected predictions

Y ( k +1) = S U ( k ) + Y o ( k +1) + y ( k ) y
( k )

y ( k+1) y o ( k+1) u ( k )

y ( k+ 2 ) y o ( k+2 ) u ( k+1)
Y ( k +1) Y ( k +1)
o
U ( k )


y ( k+P ) mP 1

y ( k+P ) mP 1
o u ( k+M -1) rM 1
27
 The mP m matrix is defined as
= [Im Im Im ]
T

  
P times

where Im denotes the m m identity matrix

Model Predictive Control

The dynamic matrix is defined as

S1 0 0
S2 S1 0

0

S SM S M -1 S1
S M +1 SM S2


SP S P-1 S P-M +1
mP rM

where Si is the m r matrix of step response coefficients

for the i-th time step S S
11,i 1r ,i

Si =
S
m1,i S mr ,i 28
 Example: Individual step-response models for a distillation column
with three inputs and four outputs. Each model represents the step
response for 120 minutes. Reference: Hokanson and Gerstle (1992).
Model Predictive Control

29
 MPC Calculations
The control objective is to calculate a set of control moves
(input changes) that make the corrected predictions as close to
Model Predictive Control

a reference trajectory as possible.

Reference Trajectory
A reference trajectory can be used to make a gradual
transition to the desired set point.
Let the reference trajectory over the prediction horizon P
be denoted as: y ( k+1) r

y ( k+ 2 )
Yr ( k +1) r


r (
y k+P ) mP 1
Exponential Trajectory from y(k) to ysp(k)
A reasonable approach for the i-th output is to use:
yi ,r ( k + j ) = ( i ) yi ( k ) + 1 ( i ) yi , sp ( k )
j

j
( (i ) = 0 y r = y sp
j
)
for i=1,2,, m and j=1, 2, , P.
30
 Unconstrained MPC
The control calculations are based on minimizing the predicted
deviations between the reference trajectory.
Model Predictive Control

The predicted error vector is defined as:

( k +1) Y ( k +1) Y ( k +1)
E r

Similarly, the predicted unforced error, E o ( k +1) , is defined as

o ( k +1) Y ( k +1) Y o ( k +1)
E r
where the corrected prediction for unforced case is defined as
Y o ( k +1) Y o ( k +1) + y ( k ) y
( k )
( k +1) = E
E o ( k +1) S U ( k )

Note that all of the above vectors are of dimension, mP.

The objective of the control calculations is to determine the

control moves, U ( k ) , for the next M time intervals.
The rM-dimensional vector U ( k ) is calculated so that an
objective function (also called a performance index) is minimized.
31
 MPC Performance Index
The rM-dimensional vector U(k) is calculated so as to minimize:
a. The predicted errors over the prediction horizon, P.
Model Predictive Control

b. The size of the control move over the control horizon, M.

Example: Consider a quadratic performance index:

( k + 1)T Q E
min J = E ( k + 1) + U ( k )T R U ( k )
U ( k )

where Q and R are weighting matrices used to weight the

most important outputs and inputs.

Both Q and R are usually diagonal matrices with positive

diagonal elements.

32
 MPC Control Law: Unconstrained Case
The MPC control law that minimizes the quadratic objective
function can be calculated analytically J
Model Predictive Control

=0
( ) U ( k )
-1
U ( k ) = S Q S + R
T T
S QE o
( k +1)
where S is the dynamic matrix.
This control law can be written in a more compact form
U ( k ) = K c E
o ( k +1)
where controller gain matrix Kc is defined to be:

( )
-1
Kc S Q S + R
T
ST Q Dimension: rM mP

(a multivariable, proportional control law based on the predicted error)

Note that Kc can be evaluated off-line, rather than on-line,

provided that the dynamic matrix S and weighting matrices, Q and
R, are constant.
33
 MPC Control Law:
Receding Horizon Approach
MPC control law: U ( k ) = K c E
o ( k +1)
Model Predictive Control

where: u ( k )

u ( k+1)
U ( k )


(
u k+M -1) rM 1
Note that the controller gain matrix, Kc, is an rM mP matrix.
.

In the receding horizon control approach, only the first step of

the M-step control policy, u ( k ) , is implemented.
u ( k ) = K E
o ( k +1)
c1

where matrix Kc1 is defined to be the first r rows of Kc.

Thus, Kc1 has dimensions of r mP .
Advantage: new information in the form of the most recent
measurement y(k) is utilized immediately
34
 Unconstrained MPC (DMC) Algorithm
The process output, y(k), is measured, and used to estimate
Model Predictive Control

the disturbance. d ( k + j ) = y ( k ) y ( k )

The predicted unforced error, E o ( k +1) , is updated

(accounting for changes in set-point and effect of previous
controller moves).
o ( k +1) = Y ( k +1) Y o ( k +1) y ( k ) y
E ( k )
r

Solve for control move.

( )
-1
U ( k ) = S Q S + R
T o ( k +1)
ST Q E

u ( k ) (first step only) is implemented.

Counter is updated: k = k + 1.

35
 MPC with Inequality Constraints
Inequality constraints on input and output variables are
important characteristics for MPC applications
Model Predictive Control

Input constraint: physical limitations on plant equipments

such as limits on input value and rate-of-change
Output constraint: related to plant operating strategy such
as constraint on product quality

MPC inequality constraints

u ( k ) u ( k + j ) u+ ( k ) j = 0,1, , M 1
u ( k ) u ( k + j ) u + ( k ) j = 0,1, , M 1
y ( k + j ) y ( k + j ) y + ( k + j ) j = 1, 2, , P

The introduction of inequality constraints results in a

constrained optimization problem
Can be solved numerically using linear or quadratic
programming techniques
36
 Selection of Design Parameters
Model predictive control techniques include a number
of design parameters:
Model Predictive Control

N: model horizon
t : sampling period
P: prediction horizon (number of predictions)
M: control horizon (number of control moves)
Q: weighting matrix for predicted errors (Q > 0)
R: weighting matrix for control moves (R > 0)

37
 Selection of Design Parameters
1. N and t
These parameters should be selected so that N t > open-loop
settling time. Typical values of N:
Model Predictive Control

30 < N < 120

2. Prediction Horizon, P
Increasing P results in less aggressive control action
Set P = N + M
3. Control Horizon, M
Increasing M makes the controller more aggressive and increases
computational effort, typically:
5 < M < 20 or N/3 < M < N/2
4. Weighting matrices Q and R
Diagonal matrices with largest elements corresponding to most
important variables
output weighting matrix Q : the most important variables having the
largest weights
input weighting matrix (move suppression matrix) R : increasing the
values of weights tends to make the MPC controller more conservative
by reducing the magnitudes of the input moves
38
 Example: SISO system
e-s
G( s ) = Assume N = 70, t = 1
(10 s + 1)( 5s + 1)
Model Predictive Control

The controller gain matrix, Kc, for two cases (Qii = 1, Rii = 0):

Set-point response [MPC vs. PID (ITAE set-point tuning)]

MPC: small settling time

39
 Disturbance response [MPC vs. PID (ITAE disturbance tuning)]
Model Predictive Control

MPC: small maximum deviations and non-oscillatory

40
 MIMO Example:
Effects of MPC design parameters (M, P, Q, R)
Wood and Berry process
Model Predictive Control

t = 1 min
Saturation limits of 0.15 were imposed on each input
+1% set-point change in XB at t=0, followed by two feed flow rate
disturbances: +30% increase at t=50 and a return to the original
value at t=100 (Fs = 2.45)
Compare a variety of MPC controllers and a multi-loop control
system (Lee et al., 1998)
Control loop Kc I
XD R 0.85 7.21
XB S -0.089 8.86
41
 Unconstrained MPC vs. Multi-loop Control
Model Predictive Control

MPC is superior to the multi-loop control system (faster responses, less oscillation).
42
 Effect of input weighting matrix R
Model Predictive Control

When Rii are increased, the MPC inputs become smoother and the output
responses have larger deviations and longer settling times. 43
 Effect of output weighting matrix Q
Model Predictive Control

Control of the more heavily weighted output improves at the expense of

the other output. 44
 Effect of control horizon M
Model Predictive Control

The y responses are similar, but the u responses are smoother for M = 5.
45
 MATLAB Tools for MPC
Unconstrained MPC (DMC)
Kmpc = mpccon(model,ywt,uwt,M,P)
Model Predictive Control

Calculate MPC controller gains for unconstrained case.

Inputs:
model : Step response coefficient matrix of model.
ywt,uwt : output and input weighting matrices.
M : control horizons.
P : prediction horizon.
Output:
Kmpc : Controller gain matrix

model = tfd2step(tfinal,delt,nout,g1,...,g25)
Determines the step response model of a transfer function model.
Inputs:
tfinal: truncation time for step response model.
delt: sampling interval for step response model.
nout: number of outputs, ny.
g1, g2,...: SISO transfer function ordered to be read in columnwise
(by input). The number of transfer functions required is
ny*nu. (nu=number of inputs). Limited to ny*nu <= 25.
Output:
model: step response coefficient matrix in MPC step format. 46
 g = poly2tfd(num,den,delt,delay)
Create transfer functions for MPC toolbox
Inputs:
Model Predictive Control

num : Coefficients of the transfer function numerator.

den : Coefficients of the transfer function denominator.
delt : Sampling time. Can be 0 (for continuous-time system)
or > 0 (for discrete-time system).
delay : Pure time delay (dead time). Can be >= 0.
Output:
g: transfer function.

47
 Unconstrained MPC (DMC)
[yp,u,ym] = mpcsim(plant,model,Kmpc,tend,r,usat,...
tfilter,dplant,dmodel,dstep)
Simulation of the unconstrained Model Predictive Controller.
Model Predictive Control

REQUIRED INPUTS:
plant(model): the step response coefficient matrix of the plant
(model) generated by the function tfd2step.
Kmpc: the constant matrix computed by the function mpccon.
tend: final time of simulation.
r: reference trajectory (set-point).
OPTIONAL INPUTS:
usat: the matrix of manipulated variable constraints.It is a vector
of the lower limits (Ulow), upper limits (Uhigh) and rate of
change limits (DelU) on the manipulated variables.
tfilter: time constants for noise filter and unmeasured
disturbance lags.
dplant: step response coefficient matrix for the disturbance
generated by the function tfd2step.
dmodel: step response coefficient matrix for the measured
disturbance effect on the model output generated by
the function tfd2step.
dstep: matrix of disturbances to the plant.
OUTPUT:
y (system response), u (manipulated variable) and ym (model
response) 48
 Constrained MPC (QDMC)
[yp,u,ym] = cmpc(plant,model,ywt,uwt,M,P,tend,r,ulim,ylim...
tfilter,dplant,dmodel,dstep)
Model Predictive Control

Simulation of the constrained Model Predictive Controller.

REQUIRED INPUTS:
plant, model, ywt, uwt, M, P, tend, r

OPTIONAL INPUTS:
ulim: matrix of manipulated variable constraints. It is a trajectory
of lower limits (Ulow), upper limits (Uhigh) and rate of change
(DelU).
ylim: matrix of output variable constraints. It is a trajectory of
lower (Ylow) and upper limits (Yhigh) on the output variables.
tfilter, dplant, dmodel, dstep

OUTPUT:
y (system response), u (manipulated variable) and ym (model
response)

49