Model Predictive Control: an

Introductory Survey

Eduardo F. Camacho
Universidad de Sevilla

Paris'2017 Eduardo F. Camacho MPC:An Introductory Survey 1

MPC successful in
industry.
 Many and very diverse and successful
applications:
 Refining, petrochemical, polymers,
 Semiconductor production scheduling,

 Air traffic control

 Clinical anesthesia,

 ….

 Life Extending of Boiler-Turbine Systems via

Model Predictive Methods, Li et al (2004)
 Many MPC vendors.
Eduardo F. Camacho MPC:An Introductory Survey 2
 MPC successful in
Academia

 Many MPC sessions in control

conferences and control journals, MPC
workshops.
 4/8 finalist papers for the CEP best paper
award were MPC papers (2/3 finally
awarded were MPC papers)

Eduardo F. Camacho MPC:An Introductory Survey 3

Takatsu et al. Report (1998) for the

Society of Instrumentation and Control
Engineering

Main Control problems found in the process industry

 Application state of the different advanced control
techniques
 Users Satisfaction level
 Expectations
…

Eduardo F. Camacho MPC:An Introductory Survey 4

 Main Control problems

Eduardo F. Camacho MPC:An Introductory Survey 5

Satisfaction degree
Comp. Retardo 72
89
Borroso 67
83
MPC 76
94
Gain-scheduling 78
87
PID avanzado 77
89
Autoajuste 60
65
Desacoplo 64
66
1989
Basado Reglas 43
61
1995
F. Kalman 70
66
Neuronal 0
69
LQ 79
70
Observador 67
62
Adaptativo 50
56
Hoo 0
50

0 10 20 30 40 50 60 70 80 90 100
Eduardo F. Camacho MPC:An Introductory Survey 6
 IFAC Pilot Industry Committee

Chaired by Tariq Samad (Honeywell), 28 total: 15 industry, 12 academia, 1 gov’t;

Members asked to assess impact of several advanced control technologies:

 Q1 Responses [23 responses]
• PID control: 23 High-impact
• Model-predictive control: 18 High-impact; 2 No/Lo impact
• System identification: 14 High-impact; 2 No/Lo impact
• Process data analytics: 14 High-impact; 4 No/Lo impact
• Soft sensing: 12 High-impact; 5 No/Lo impact
• Fault detection and identification [22]: 11 High-impact; 4 No/Lo impact
• Decentralized and/or coordinated control: 11 High-impact; 7 No/Lo impact
• Intelligent control: 8 High-impact; 7 No/Lo impact
• Discrete-event systems [22]: 5 High-impact; 7 No/Lo impact
• Nonlinear control: 5 High-impact; 8 No/Lo impact
• Adaptive control: 4 High-impact; 10 No/Lo impact
• Hybrid dynamical systems: 3 High-impact; 10 No/Lo impact
• Robust control: 3 High-impact; 10 No/Lo impact
7

Why is MPC so successful ?

 MPC is Most general way of posing the

control problem in the time domain:
 Optimal control
 Stochastic control

 Known references

 Measurable disturbances

 Multivariable

 Dead time

 Constraints

 Uncertainties

Eduardo F. Camacho MPC:An Introductory Survey 8

 Real reason of success: Economics
 MPC can be used to optimize operating points (economic
objectives). Optimum usually at the intersection of a set of
constraints.
 Obtaining smaller variance and taking constraints into account
allow to operate closer to constraints (and optimum).
 Repsol reported 2-6 months payback periods for new MPC
applications.

Tmax

P1 P2

Eduardo F. Camacho MPC:An Introductory Survey 9

ESQUEMA GENERAL CIRCUITO DE GASES

Flash

Contacto 1

Lavado
Línea 1
Contacto 3

Línea 2

Contacto 2

Eduardo F. Camacho MPC:An Introductory Survey 10

Paris'2013 Eduardo F. Camacho MPC:An Introductory Survey 11

0
1 100 199 298 397 496 595 694 793 892
-1
Serie1
-2

-3

-4

-5

Eduardo F. Camacho MPC:An Introductory Survey 12

 Eduardo F. Camacho MPC:An Introductory Survey 13

Benefits

 Yearly saving of more that 1900 MWh

 Standard deviation of the mixing chamber
pressure reduced from 0.94 to 0.66 mm
water column.
 Operator’s supervisory effort: percentage
of time operating in auto mode raised
from 27% to 84%.

Eduardo F. Camacho MPC:An Introductory Survey 14

 Outline
 A little bit of history
 Model Predictive Control concepts
 Linear MPC
 Multivariable
 Constraints
 Nonlinear MPC
 Stability and robustness
 Hybrid systems
 Implementation
 Some applications
 Conclusions
Eduardo F. Camacho MPC:An Introductory Survey 15

Bibliografía

 R. Bitmead, Gevers M. and Werts W. “Adaptive Optimal Control:

The Thinking man GPC”, Prentice Hall, 1990
 R. Soeteboek, Predictive Control: A Unified Approach, Prentice
Hall, 1991.
 E.F. Camacho and C. Bordons, “Model Predictive Control in the
Process Industry”, Springer-Verlag, 1995.
 J. Maciejowski, “Predictive Control with constraints”, Prentice Hall,
2002.
 A. Rossiter, “Model-Based Predictive Control: A Practical Approach”,
CRC Press, 2003
 E.F. Camacho and C. Bordons, “Model Predictive Control”, Springer-
Verlag, 1999 (Second edition 2004)

Eduardo F. Camacho MPC:An Introductory Survey 16

 A little bit of history: the
beginning
 Kalman, LQG (1960)
 Propoi, “Use of LP methods ...” (1963)
 Richalet et al, Model Predictive Heuristic Control (MPHC)
IDCOM (1976, 1978) (150.000 $/year benefits because of
increased flowrate in the fractionator application)
 Cutler & Ramaker, DMC (1979,1980)
 Cutler et al QDMC (QP+DMC) (1983)
 Clarke et al GPC (1987)
 First book: Bitmead et al, (1990)

Eduardo F. Camacho MPC:An Introductory Survey 17

The impulse of the 90s. A renewed

interest from Academia (stability)

 Stability was difficult to prove because of the finite horizon and the
presence of constraints (non linear controller, no explicit solution, …)
 A breakthrough produced in the field. As pointed out by Morari: ”the
recent work has removed this technical and to some extent
psychological barrier (people did not even try) and started wide
spread efforts to tackle extensions of this basic problem with the
new tools”. (Rawlings & Muske, 1993)
 Many contributions to stability and robustness of MPC: Allgower,
Campo, Chen, Jaddbabaie, Kothare, Limon, Magni, Mayne, Michalska,
Morari, Mosca, de Nicolao, de Olivera, Scattolini, Scokaert…

Eduardo F. Camacho MPC:An Introductory Survey 18

 The new millenium

 Linear MPC is a mature discipline. More than 4500 industrial

applications (not counting licensed technology companies app.) Qin
and Badgwell’03.
 The number of applications seems to duplicate every 4 years.
 Some vendors have NMPC products: Adersa (PFC), Aspen Tech
(Aspen Target), Continental Control (MVC), DOT Products (NOVA-
NLC), Pavilon Tech. (Process Perfecter)
 Efforts to developed MPC for more difficult situations:
 Multiple and logical objectives (Morari, Floudas)
 Hybrid processes (Morari, Bemporad, Borrelli, De Schutter, van den
Boom …)
 Nonlinear (Alamir, Alamo, Allgower, Biegler, Bock, Bravo, Chen, De
Nicolao, Findeisen, Jadbadbadie, Limon, Magni, …)

Eduardo F. Camacho MPC:An Introductory Survey 19

MPC Objetive
Compute at each time instant the sequence
of future control moves that will make the
future predicted controlled variables to best
follow the reference over a finite horizon and
taking into account the control effort.

Only the first element of the sequence is

used and the computation is done again at
the next sampling time.

Eduardo F. Camacho MPC:An Introductory Survey 20

 MPC basic concepts

Common ideas:
 Explicit use of a model to predict output.

 Compute the control moves minimizing an objective fuction.

 Receding horizon strategy. Estrategia deslizante (el horizonte

se desplaza hacia el futuro).
Thealgorithms mainly differ in the type of
model and objective function used.

Eduardo F. Camacho MPC:An Introductory Survey 21

MPC strategy

 At sampling time t the future

control sequence is compute so
that the future sequence of
Acciones de control
predicted output y(t+k/t) along a
horizon N follows the future
references as best as possible.
Setpoint

 The first control signal is used

and the rest disregarded.
t t+1 t+2 t+N
The process is repeated at the
next sampling instant t+1

Eduardo F. Camacho MPC:An Introductory Survey 22

 Errors minimized over
a finite horizon
t+1 t
t+2
Constraints
taken into Model of
account process used
for predicting

u(t)

Only the first

control move is
t+N applied t t+1 t+2 …….. t+N

Eduardo F. Camacho MPC:An Introductory Survey 23

t+2 t+1

u(t)

t+N t t+1 t+2 …….. t+N t+N+1

Only the first

t+N+1
control move is
applied again

Eduardo F. Camacho MPC:An Introductory Survey 24

 MPC vs. PID

PID: u(t)=u(t-1)+g0 e(t) + g1 e(t-1) + g2 e(t-2)

Eduardo F. Camacho MPC:An Introductory Survey 25

MPC basic structure

Future Reference
Past inpus
output trajectory
and outputs
Model
-

Future Future
controls errors
Optimizer

Cost function Constraints

Eduardo F. Camacho MPC:An Introductory Survey 26

 GPC

 CARIMA model

   
A q 1 yt  B q 1 ut 1 
 
C q 1
t

Cost function

J u, t    y t  j  rt  j     u t  j 1 
N2 Nu
2 2

j  N1 j 1

Eduardo F. Camacho MPC:An Introductory Survey 27

GPC

Control moves u  (G T G  I ) 1 G T (r  f )

with
 g0 0  0 
 g g0  0 
 1

     
G 
 g Nu 1 g Nu 2  g0 
    
 
 g N 2 1 g N 2 2  g N 2  Nu 

Eduardo F. Camacho MPC:An Introductory Survey 28

 Outline
 A little bit of history
 Model Predictive Control concepts
 Linear MPC
 Multivariable
 Constraints
 Nonlinear MPC
 Stability and robustness
 Hybrid systems
 Implementation
 Some applications
 Conclusions
Eduardo F. Camacho MPC:An Introductory Survey 29

Y1
U1
Y2
.
U2 .
.
.
Proceso .
. Yn
Un

Eduardo F. Camacho MPC:An Introductory Survey 30

 Y

R1 +
Gc1
- U
R2 + G
Gc2 Gd
-
(Proc.)
Rn +
Gcn
-

Eduardo F. Camacho MPC:An Introductory Survey 31

Multivariable MPC

 Direct extension
 But …
 Dead times, control horizon, ...
 Unstable transmission zeros

Eduardo F. Camacho MPC:An Introductory Survey 32

 Multivariable MPC

1/(1+0.1s) 5/(1+s)

1/(1+0.5s) 2/(1+0.4s)

Eduardo F. Camacho MPC:An Introductory Survey 33

Multivariable MPC

Paris'2013 Eduardo F. Camacho MPC:An Introductory Survey 34

 Outline
 A little bit of history
 Model Predictive Control concepts
 Linear MPC
 Multivariable
 Constraints
 Nonlinear MPC
 Stability and robustness
 Hybrid systems
 Implementation
 Some applications
 Conclusions
Eduardo F. Camacho MPC:An Introductory Survey 35

Constraints in process control

 All process are constrained

 Actuators have a limited range and slew
rate
 Safety limits: maximun pressure or
temperature
 Tecnological or quality requirements
 Enviromental legislation

Eduardo F. Camacho MPC:An Introductory Survey 36

 Real reason of success: Economics
 MPC can be used to optimize operating points (economic
objectives). Optimum usually at the intersection of a set of
constraints.
 Obtaining smaller variance and taking constraints into account
allow to operate closer to constraints (and optimum).
 Repsol reported 2-6 months payback periods for new MPC
applications.

Tmax

P1 P2

Eduardo F. Camacho MPC:An Introductory Survey 37

Work close to the optimal

but not violating it
Fine
400 Euros
120
3 points

Eduardo F. Camacho MPC:An Introductory Survey 38

 MPC: constraints IV

umin<u(t)< umax
Umin<U(t)< Umax
ymin (t)<y(t)< ymax(t)

Minu J(u,x(t))

s.t.: R u < r +S x(t)

Eduardo F. Camacho MPC:An Introductory Survey 39

Usual way of dealing with constraints

J (curvas
de nivel)

u(t+1)

u(t+1)
Señal de control
sin saturar
Señal
óptima
Solución sin u(t)
Señal restricciones
óptima
u(t)
Señal de control saturada

Eduardo F. Camacho MPC:An Introductory Survey 40

 MPC strategy
Consider a nonlinear invariant discrete time system:
x+=f(x,u), x  Rn, u  Rm

The system is subject to hard constraints

x  X, u  U

Let u={u(0),...,u(N-1) } be a sequence of N control inputs

applied at x(0)=x,

the predicted state at i is

x(i)= (i;x, u)=f(x(i-1), u(i-1) )

Eduardo F. Camacho MPC:An Introductory Survey 41

MPC strategy

1. Optimization problem PN(x,):

u*= arg minu (i=0,...,N-1) l(x(i),u(i)) + F(x(N))

 Operating constaints .
x(i)  X, u(i)  U, i=0,...,N-1

 Terminal constraint (stability): x(N)  

2. Apply the receding horizon control law:

KN(x)=u*(0).
Eduardo F. Camacho MPC:An Introductory Survey 42
 Linear MPC

 f(x,u) is an affine function (model)

 X,U, are polyhedra (constraints)
 l and F are quadratic functions (or 1-norm
or -norm functions)

 QP or LP

Eduardo F. Camacho MPC:An Introductory Survey 43

Control predictivo lineal

MODEL COST FUNCTION CONSTRAINTS SOLUTION

Linear Quadratic None Explicit

Linear Quadratic Linear QP

Linear Norm-1 Linear LP

Eduardo F. Camacho MPC:An Introductory Survey 44

 Otherwise

 If f(x,u) is not an affine function

 Or any of X,U, are not polyhedra
 Or any of l and F are not quadratic functions
(or 1-norm or -norm functions)

 Non linear MPC (NMPC)
 Non linear (non necessarily convex) optimization
problem much more difficult to solve.

Eduardo F. Camacho MPC:An Introductory Survey 45

MPC and nonlinear

processes
 Most processes are non-linear,
 Linear approximations works for small perturbations
around the operating point (well in most cases)
 There are processes with
 continuous transitions (startups, shutdowns, etc.) and spend a
great deal of time away from a steady-state operating region or
 never in steady-state operation (i.e. batch processes, solar
plants), where the whole operation is carried out in transient
mode.
 severe nonlinearities (even in the vicinity of steady states)
 hybrid

Eduardo F. Camacho MPC:An Introductory Survey 46

 NMPC vs. LMPC

 Consider the nonlinear system

y(t+1)= 0.9 y(t) + u(t)1/4
with 0<u(t)< 1.

 Linear model:
y(t+1)= 0.9 y(t) + u(t)

 Applying a L-MPC and a NL-

MPC with N=10, =0

Eduardo F. Camacho MPC:An Introductory Survey 47

NMPC vs. LMPC

 Better predictions should be obtained from

more accurate models.
 Better predictive control should be obtain
with better predictions.
 Is that so ?

Eduardo F. Camacho MPC:An Introductory Survey 48

 MPC turns out to be a linear
controller with a feedback of the
prediction of y(t+D)
u(t) y(t)
H C Process

^
y(t+D)
R Predictor

Eduardo F. Camacho MPC:An Introductory Survey 49

Falatious congeture

 An optimal predictor plus an optimal controller is going to

produce the “best” closed loop behaviour.
 J. Normey showed that Smith predictor (and other DTC
structures) produce “better” (more robust) controllers than
optimal predictor.
 Optimal state estimator + optimal controller (LQG/LTR)
 Optimal identifier + optimal controller does not produce
the optimal adaptive control (identification for control)
 A fundamental issue: The best model is the one that
produces the “best” close loop control not necessarily best
predictions..
Eduardo F. Camacho MPC:An Introductory Survey 50
 Nonlinear Identification is
more difficult

 Lack of a superposition principle

 A high number of plant tests required
 tests with many different size steps
 Multivariable processes the difference in the number
of tests required is even greater.
 Optimization problem for parameter
estimation is more difficult (offline)

Eduardo F. Camacho MPC:An Introductory Survey 51

Modeling:
Empirical Models
 Fixed structure, parameters determined from data
 State space x(t+1)=f(x(t),u(t)), y(t)=g(x(t))
 Input-output (NARMAX)
y(t+)=  (y(t), ..., y(t-ny), u(t),...,u(t-nu),e(t),...,e(t-ne+1))
 Volterra (FIR, bilineal)
y(t+1)= y0+ {i=0..N}h1(i) u(k-i)+  {i=0..M}  {i=0..M} h2(i,j)u(t-i)u(t-j)
 Hammerstein
 Wiener
 NN

Eduardo F. Camacho MPC:An Introductory Survey 52

 Neural Networks
 Multilayer Perceptron: a nonlinear function with good approximation
properties and “backpropagation”.
 Input-output with NN: y(t)= NN(y(t-1), ..., y(t-ny), u(t-1),...,u(t-nu))

 State space with NN x(t+1)=NNf(x(t),u(t)), y(t)=NNg(x(t))

z-1
Eduardo F. Camacho MPC:An Introductory Survey 53

Local Model Networks

 A set of models to accommodate local operating regimes

 The output of each submodel is passed through a
processing function that generates a window of validity.
y(t+1)=F((t),  (t)) =  {i=1.M}fi((t), (t)), i ( (t))

 local models are usually linear multiplied by basis

functions i ( (t)) chosen to have a value close to 1 in
regimes where is a good approximation and a value close
to 0 in other cases.

 Piece Wise Affine (PWA) systems

Eduardo F. Camacho MPC:An Introductory Survey 54

 PWA. (Sontag, 1981)

Nonlinear System: x(t+1)=f(x(t),u(t))

y(t)=g(x(t))

PWA Approx.

Some hybrid processes can be modeled by PWA system

Eduardo F. Camacho MPC:An Introductory Survey 55

Function

Eduardo F. Camacho MPC:An Introductory Survey 56

 PWA Approximation

Eduardo F. Camacho MPC:An Introductory Survey 57

Approximation error

Eduardo F. Camacho MPC:An Introductory Survey 58

 Outline
 A little bit of history
 Model Predictive Control concepts
 Linear MPC
 Multivariable
 Constraints
 Nonlinear MPC
 Stability and robustness
 Hybrid systems
 Implementation
 Some applications
 Conclusions
Eduardo F. Camacho MPC:An Introductory Survey 59

MPC: constraints II

 Difference in input and output constraints:

 manipulated variables can always be kept in
bound by the controller by clipping the control
action or by the actuator.
 Output constraints are mainly due to safety
reasons, and must be controlled in advance
because output variables are affected by
process dynamics.

Eduardo F. Camacho MPC:An Introductory Survey 60

 MPC: constraints III

 Not considering constraints on manipulated

variables may result in higher values of the
objective function. But this is not the main
problem.
 Violating the limits on the controlled variables
may be more costly and dangerous as it could
cause damage to equipment and losses in
production.
 Not considering input contraints may lead to
unstability
Eduardo F. Camacho MPC:An Introductory Survey 61

Stability
 Optimal controllers with infinite horizon
guaranty stability.
 Optimal finite horizon and the presence of
constraints make it very difficult to prove stability
(non linear controller, no explicit solution)
 A breakthrough has been made in the last few years
in this field. As pointed out by Morari , ”the recent
work has removed this technical and to some
extent psychological barrier (people did not even
try) and started wide spread efforts to tackle
extensions of thisbasic problem with the new
tools”.

Eduardo F. Camacho MPC:An Introductory Survey 62

Eduardo F. Camacho MPC:An Introductory Survey 63

Eduardo F. Camacho MPC:An Introductory Survey 64

 Stability and constraints
y(t+1)=1.2 y(t)+0.2 u(t-2) with -4 < u(t) < 4, N=5

Eduardo F. Camacho MPC:An Introductory Survey 65

MPC stability
 Infinite horizon, the objective function can be considered a Lyapunov
function, providing nominal stability. Cannot be implemented: an infinite set of
decision variables.
 Terminal cost. Bitmead et al’90 (linear uncostrained), Rawling & Muske’93
(linear contrained).
 Terminal state equality constraint. Kwon & Pearson’77 (LQR constraints),
Keerthi and Gilbert’88, x(k+N)= xS

x(t+2)
x(t+1)
x(t+N) xS

x(t)

Eduardo F. Camacho MPC:An Introductory Survey 66

 MPC stability: terminal region

 Dual control. Michalska and Mayne x(t+2)

x(t+1)
(1993) x(N)   x(t+N)
Once the state enters  the 
x(t)
controller switches to a previously
computed stable linear strategy.

Quasi-infinite horizon. Chen and Allgower (1998). Terminal

region and stabilizing control, but only for the computation of the
terminal cost. The control action is determined by solving a finite
horizon problem without switching to the linear controller even
inside the terminal region. The term (|| x(t+N)||P)2 added to the cost
function and approximates the infinite- horizon cost to go.

Eduardo F. Camacho MPC:An Introductory Survey 67

MPC and sliding mode control

 Robust controllers
 Impose surface S as
terminal constraints

Eduardo F. Camacho MPC:An Introductory Survey 68

 NMPC stability:
all ingredients
 Asymptotic stability theorem (Mayne 2001)
 The terminal set  is a control invariant set.
 The terminal cost F(x) is an associated Control
Lyapunov function such that
min{u  U} {F(f(x,u))-F(x) + l(x,u) | f(x,u)} ≤0  x
 Then the closed loop system is asymptotically
stable in XN( )

Eduardo F. Camacho MPC:An Introductory Survey 69

Removing the terminal constraint

maintaining stability

The optimization problem is simplified (especially when the state is not constrained)
 u*= arg minu (i=0,...,N-1) l(x(i),u(i)) + F(x(N))
s.t. x(i)  X, u(i)  U, i=0,...,N-1 and to the terminal constraint: x(N)  

There is a strong interplay between infinite horizon-terminal cost and terminal regions :
 A prediction horizon N and a quadratic terminal cost stabilizes the system in a
neighborhood of the origin. (Parisini and Zoppoli’95)

 The unconstrained (no terminal constraints) MPC satisfies the terminal constraint in
a neighborhood of the origin. (Jadbabaie et al’01)

 Given a terminal cost F(x) that is a CLF in  , define Fs(x)= F(x) if x   and Fs(x)= 
if x   where  = {x  Rn: F(x) ≤ }. The MPC with Fs(x) as terminal cost is
stabilizing for all initial state in the region where the optimal solution to PN(x,X)
reaches the terminal region. (Hu and Linnemann’02)

 Procedure for removing the terminal constraint while maintaining asymptotic stability
and computing the domain of attraction. (Limon et al’03). Suboptimality

Eduardo F. Camacho MPC:An Introductory Survey 70

 Robustness
Nonlinear uncertain system: x+=f(x,u,  ), x  Rn, u  Rm   Rp

With bounded uncertainties  and subject to hard constraints x  X, u  U

The uncertain evolution sets:

X(i)=(i;x, u)= {z  Rn |   , y  X(i-1), z=f(y, u(i-1), )}
and X(0)=x

y(t)

u(t)

t t+1 t+2 … t+N

Eduardo F. Camacho MPC:An Introductory Survey 71

Robustness (2)
 The stability conditions has to be
satisfied for all possible values of
the uncertainties.

 The terminal set  is a robust

control invariant set. (i.e.  x, 
  u  U | f(x,u, ))

 The terminal cost F(x) is an

associated Control Lyapunov
function such that
min{u  U} {F(f(x,u,))-F(x) + l(x,u) |
f(x,u,)} ≤0  x,  

Eduardo F. Camacho MPC:An Introductory Survey 72

 Computation of regions in robust NMPC
 Invariant regions,
 domain of attraction,
 uncertain evolution sets
 bounding sets (state estimation)
 bounding sets (identification) …

•Relatively easy if regions are

polyhedron and linear transformations
(f(x,u, ) ), Kerrigan’00

•NMPC more complex and

approximations are normally used.

Continuous stirred tank reactor (CSTR)

Eduardo F. Camacho MPC:An Introductory Survey 73

Example: continuos stirred tank

reactor (CSTR) Zonotopes: (Bravo et
al’03)

Sampling period 0.03.

A normalized additive uncertainty is

added. It is bounded by w1=(-
6.5*10-3, 6.5*10-3) and w2=(-
1.2*10-3, 1.2*10-3).

A terminal robust positively invariant

set is calculated.

A prediction horizon of N = 11 is
 E 
   C Af  C A   k0  exp  
dC A q
used.   CA
dt V  RT 
H  k0  E  UA
  T f  T  
dT q
 exp     CA   Tc  T 
dt V   Cp  RT  V    Cp

Eduardo F. Camacho MPC:An Introductory Survey 74

 Example

Some of these methods

are too conservative: Pure
interval arithmetic: boxes

Eduardo F. Camacho MPC:An Introductory Survey 75

State estimation

 The observer error must be small to guarantee

stability of the closed-loop and in general little
can be said about the necessary degree of
smallness.
 No general valid separation principle for
nonlinear systems exists.
 Nevertheless observers are applied successfully
in many NMPC applications.
 Extended Kalman Filter (EKF),
 High gain estimators.

Eduardo F. Camacho MPC:An Introductory Survey 76

 State estimation (2)
 Moving Horizon Estimation (MHE).
 Moving window looking backward
 A dual problem to MPC: Control moves: known, process state:
unknown, horizon: backwards. (Kwon et al’83), (Zimmer’94),
(Mishalska and Mayne’95) ..
 IDCOM (Richalet’76) was developed as the dual to identification !!!

 Set-membership estimation
 Compute a set of all states consistent with the measured output
and the given noise parameters.
 Ellipsoidal bounding (Schweppe’68), (Kurzhanski and Valyi’96)
 Polyhedron bounding (Kuntsevich and Lychak’85)
 Interval arithmetics, zonotopes (Kieffer et al’01), (Alamo et al’03)

Eduardo F. Camacho MPC:An Introductory Survey 77

Outline
 A little bit of history
 Model Predictive Control concepts
 Linear MPC
 Multivariable
 Constraints
 Nonlinear MPC
 Stability and robustness
 Hybrid systems
 Implementation
 Some applications
 Conclusions
Eduardo F. Camacho MPC:An Introductory Survey 78
 MPC and polytopes
 MPC with linear constraints is a
MultiParametric-QP o LP (Bemporad y
Morari, 2000)
 The solution of an mp_QP o mp_Lp is a
PWA function of state.
 Min-max quadratic is PWA (Ramirez y
Camacho, 2001).

Eduardo F. Camacho MPC:An Introductory Survey 79

Min-Max MPC with bounded global

uncertainties

Process:
y(t+1)=g(y(t),u(t),...)
Model:
z(t+1)=f(y(t),u(t),..., )
si   | z(t+1)=y(t+1)

e(t )
A( z 1 ) y(t )  z d B( z 1 )u(t  1)  C ( z 1 )

~
A( z 1 ) y(t )  z  d B( z 1 )u(t  1)   (t )  (t ) 

Eduardo F. Camacho MPC:An Introductory Survey 80

 Properties of J and J*

•The set of j-ahead optimal predictions for j=1,...,N2 can be written as:

y  Guu  G  f
J  , u   u T M uuu   T M   2 T M u u  2 M ufT u  2 MTf  f T f
where:
M uu  GuT Gu   I , M   GT G , M  u  GT Gu , M uf  GuT f , M  f  GT f

•M  is positive definite.

• Also, Muu is positive definite for positive values of 

Eduardo F. Camacho MPC:An Introductory Survey 81

Properties of J and J*


J * (u)  max  T M  M' (u)  M ' (u)
 

• Due to convexity of J and compactness of , the
maximum will be reached at one of the vertexes of the
polytope  . J* can be expressed as a convex function of
u (if Muu is positive definite) :
J * (u )  u T M uuu  M u* ( p )u  M * ( p )

• Due to its convexity, J* has a unique minimizer, thus

the min-max problem has a unique solution.

Eduardo F. Camacho MPC:An Introductory Survey 82

 Piecewise Linearity of the Control Law

• This situation can be generalized to N quadratic functions, thus it can be

stated that the solution will be attained on either a minimizer of one of them or
on an intersection point of two or more of them.

Eduardo F. Camacho MPC:An Introductory Survey 83

An illustrative example

• Consider the following first order integrated uncertainties prediction

model: yk 1  1.9 yk  0.9 yk 1  0.5  uk   k 1

• Let Nu=2, N2=2, =1.0 and -0.1  k  0.1 then the following
output predictions can be formulated:

 yk 1    uk   k 1   yk 1 
 y   Gu  u   G    Fy  y 
 k 2   k 1   k 2   k 
with :
 0.5 0  1 0   0.9 1.9 
Gu    G    Fy   
0.95 0.5 1.9 1  1.71 2.71

Eduardo F. Camacho MPC:An Introductory Survey 84

 An illustrative example

• The control law has been calculated numerically for yk , yk 1   0.5,0.5

(only uk shown):

• The lower region is due to the minimizer of J1.The upper region is due to the
minimizer of J4. Between them there is another region due to the intersection
of J1 and J4.
Eduardo F. Camacho MPC:An Introductory Survey 85

An illustrative example
• Boundary regions can be computed:
• J4 - J41: • J41 - J1:
y  y 
 0.2156  0.9873  1.6242  k 1   0  0.2156   0.9873 1.6242  k 1   0
 yk   yk 
y  y 
0.0747   0.7098 1.2209  k 1   0  0.0111  0.2775  0.4032  k 1   0
 yk   yk 
y  y 
 0.0111   0.2775 0.4032  k 1   0 0.0747  0.7098  1.2209  k 1   0
 yk   yk 

Eduardo F. Camacho MPC:An Introductory Survey 86

 An illustrative example

• An explicit form of the control law can be found:

y  R4 uk  0.1371  0.8872 yk 1  1.5262 yk
y  R42 uk  1.451 yk 1  2.4685 yk
other uk   0.1371  0.8872 yk 1  1.5262 yk
where y   yk 1 yk  and :
T

R4  y  0.2156   0.9873  1.6242  y  0 

 y  0.2156   0.9873  1.6242  y  0 
R42   
AND  0.2156    0.9873  1.6242  y  0 

• The output under the MMMPC

control law versus an
unconstrained GPC:

Eduardo F. Camacho MPC:An Introductory Survey 87

Min-max MPC: open-close

loop predictions

 Minu(k),u(k+1),.. Max w(k),w(k+1) … J(u,w)

 Minu(k) Max w(k) Minu(k+1) Max w(k+1) … J(u,w)

 Minu(k)c(x(k+1)),c(x(k+2)) .. Max w(k) w(k+1) … J(u,w)
 Control laws

Eduardo F. Camacho MPC:An Introductory Survey 88

 Min-max MPC: open-close
loop predictions (2)

x(k+1)=x(k)+u(k)+w(k) with
 -2 < x(k) < 2
 -2 < u(k) < 2
 -1 < w(k) < 1
 Qith horizon 3 there is no u(1),u(2), u(3) which
keeps -2 < x(k) < 2 dor all w(k)
 If u(k)= - x(k) then -2 < x(k) < 2 for all w(k)
 Game theory: available information by each
player

Eduardo F. Camacho MPC:An Introductory Survey 89

NMPC implementation

 Solving a Nonlinear (non QP), possibly

nonconvex.
 Real time and no convexity >>> suboptimal
solutions
 Sequential Quadratic Programming (SQP)
 Simultaneous approach (Findeisen and Allgower’02)
 Using a sequential approach with successive
linearization around the previous trajectory.
 PWA >>>> Mixed Integer Programming Problem.

Eduardo F. Camacho MPC:An Introductory Survey 90

 Hybrid System

Computer Control
Science Theory

X={1,2,3,4,5}
Discrete Dynamical x  Rn
U={A,B,C} Events u  Rm
systems
yRp
A 2
B C
1 B
B 3
x(k  1)  f (x(k ), u(k ))
C
C y (k )  g (x(k ), u(k ))
5
4 A
Eduardo F. Camacho MPC:An Introductory Survey 91
C

Hybrid System 1

Computer Control
Science Theory

Discrete Events Dynamical Systems

- Physical processes
- Collisions in control system
- Decision logic e.g. robots, aircraft, ...
- Discrete
communication

Eduardo F. Camacho MPC:An Introductory Survey 92

 Hybrid System 2

Computer Control
Science Theory

Model: Model:
Automata, Petri nets, - Differential equations
statecharts, etc. - Algebraic equations
A
2 - Invariant constraints
B C
1
x(k  1)  f (x(k ), u(k ))
B
3
B
C
5 C
4 A
y (k )  g (x(k ), u(k ))
C

Eduardo F. Camacho MPC:An Introductory Survey 93

Hybrid System 3

Computer Control
Science Theory

Analytical Tools: Analytical Tools:

Boolean algebra, Lyapunov functions,

formal logics, eigenvalue analysis,
recursion, etc. etc.

Eduardo F. Camacho MPC:An Introductory Survey 94

 Hybrid System 4

Computer Control
Science Theory

Software Tools: Software Tools:

Statemate, MATLAB,
Design CPN, MatrixX,
Slam II, VisSim, etc.,
SMV, etc.

Eduardo F. Camacho MPC:An Introductory Survey 95

PWA systems
x(k+1) = A1 x(k) + B1 u(t) + f1
x(k+1) = A3 x(k) + B3 u(t) + f3
u

x(k+1) = A2 x(k) + B2 u(t) + f2

x
Eduardo F. Camacho MPC:An Introductory Survey 96
 PWA approximations

Mixed
xk+1=Ak x k+Bk uk+f k
x =A x
xk+2=Ak+1xk+1+Bk+1uk+1+fk+1
+B u +f
k+3
k+2
k+2
k+2
k+2

? ? k+2

Integer-Real
X
X
y =C x +g
k+2
k+2 k+2
k+1 k+2 k+2
y =C x +g k+1 k+1
X k
Optimization
y =C
k
k
x +g k
k
Problem
k+1 k+1

The resulting optimization problem

U = {u(k), u(k+1), u(k+2), …,u (k+N-1)} real

I = {I(k), I(k+1), I(k+2),…, I(k+N-1)} Integer
Eduardo F. Camacho MPC:An Introductory Survey 97

NMPC implementation (2)

 Although there are many clever tricks to alleviate

the situation, these algorithms take time. This is
a major obstacle. Only slow or small processes.
 Approximations and simplifications
 Using short horizons
 Precomputation of solution over a grid in the state
space (only small systems)

Eduardo F. Camacho MPC:An Introductory Survey 98

 Outline
 A little bit of history
 Model Predictive Control concepts
 Linear MPC
 Multivariable
 Constraints
 Nonlinear MPC
 Stability and robustness
 Hybrid systems
 Implementation
 Some applications
 Conclusions
Eduardo F. Camacho MPC:An Introductory Survey 99

Paris'2013 Eduardo F. Camacho MPC:An Introductory Survey 100

 Application of NMPC to a mobile
robot: Path tracking in an
unstructured environment

 Problem
 Future trajectory known (computed by planner)
 Unexpected obstacles
 Control signals and state are constrained
 System model is highly nonlinear
 The objective function: position error, the acceleration, robot
angular velocity and the proximity between the robot and the
obstacles (detected with an ultrasound proximity system)
 Unexpected obstacles makes the objective function more
complex.

(Gómez & Camacho, 1994)

Eduardo F. Camacho MPC:An Introductory Survey 101

Mobile robot avoiding

obstacles

Eduardo F. Camacho MPC:An Introductory Survey 102

Mobile robot NMPC

Eduardo F. Camacho MPC:An Introductory Survey 103

NMPC at a Solar plant

Almeria

Eduardo F. Camacho MPC:An Introductory Survey 104

NMPC application at the
Solar plant in Almeria (PSA)

Eduardo F. Camacho MPC:An Introductory Survey 105

Distributed collectors

Eduardo F. Camacho MPC:An Introductory Survey 106

 Plant diagram

Eduardo F. Camacho MPC:An Introductory Survey 107

Nonlinear disturbances
prediction: Solar radiation

Eduardo F. Camacho MPC:An Introductory Survey 108

 Plant results: a clear day

Very fast setpoint tracking, little overshoot

Eduardo F. Camacho MPC:An Introductory Survey 109

Plant results: a cloudy day

Eduardo F. Camacho MPC:An Introductory Survey 110

 Conclusions
 Well established in industry and
academia
 Great expectations for MPC
 Many contribution from the research
community but …
 Many open issues
 Good hunting ground for PhD students.

Eduardo F. Camacho MPC:An Introductory Survey 111

Eduardo F. Camacho MPC:An Introductory Survey 112