You are on page 1of 34

Explicit Non-linear Optimal Control

Law for Continuous Time Systems


via Parametric Programming

Vassilis Sakizlis,

Vivek Dua, Stratos Pistikopoulos

Centre for Process Systems Engineering


Department of Chemical Engineering
Imperial College, London.
Brachistrone Problem
Find closed-loop trajectory γ(x,y) of a gravity driven ball such
that it will reach the opposite wall in minimum time
x

γ wall- target
g

x=l
plane-obstacle

y=xtanθ+h
y
Outline

• Introduction

• Multi-parametric Dynamic Optimization

• Explicit Control Law

• Results

• Concluding Remarks
Introduction
Model Predictive Control

Solve an optimization problem at each time interval

Accounts for Shortcomings


- Optimality -Demanding Computations
- Constraints -Applies to slow processes
- Logical Decisions -Uncertainty handling
Application - Parametric Controllers
(Parcos)
Optimization Parametric Solution
Problem v( x* )

Parametric Controller
v(t)=g(x*) x*

Control v Plant State x*

Input PLANT Process Outputs y


Disturbances
w
• Explicit Control law
• Eliminate expensive, on-line computations
Theory of PARCOS
What is Parametric Programming?

f ( x )  min f ( v, d , x ) v( x)
v ,d
s.t. g ( v, d , x )  0 d ( x)
x : parameters
v : continuous
x
δ : binary
Region CR1
Features
•Complete mapping of optimal conditions in parameter space
•Function fc(x),vc (x),dc (x)
•Critical regions CRc(x)0 c=1,Nc
Parametric Programming Developments
Theory, Algorithms and Software Tools
for Multi-parametric Optimization Problems

 Quadratic and convex nonlinear


 Mixed integer linear, quadratic and nonlinear
 Bilinear Applications

 Process synthesis and planning


 Design under Uncertainty
 Reactive scheduling / Bilevel Programming
Stochastic Programming
 Model based and hybrid control
Model – based Control via Parametric
Programming
N 1
φ(x )  min
*
N
x
v
Px T

N |t
N |t  ( 0. 0116 x 2
1,t  k
k 0
 0.0084 x 2
2 ,t  k
Objective
 0.0198 x1,t  k x2,t  k  10  4 vt2 k )
s.t. x1,t  k 1  1.269 x1,t  k  0.107 x2,t  k  0.16954vt  k
Discrete Model
x2,t  k 1  0.170 x1,t  k  0.992 x2,t  k  0.01221vt  k
xt  x* " " Current States
0  g ( xk |t , vk |t )  1.5 x1,t  k  x2,t  k Constraints
k  0,  N
N  10, t  0.15 vcN (x * ),φc(x * ),
Formulate mp-QP (mp-LP)
Obtain piecewise affine control law CR( x* )  0, X o
Pistikopoulos et al., (2002) c  1, N c
Bemporad et al.,(2002)
Parco / Explicit MPC Solution

Control law
v0  ac x*  bc
if CR1c x*  CRc2  0
c  1,...N c

• Complex
• Approximate
Multi-parametric Dynamic Optimization
mp-DO
tf

f ( x* )  min xTf Px f   [ x(t )T Qx(t )  v(t) T Rv(t)]dt


v
t*

x  A  x  B  v
g  D1  x(t )  D2  v(t )  b2  0,   G  x(t f )  0
xo  x*
t  [to , t f )

• Feasible Set X*
For each x*  X* there exists an optimizer v*(x*,t) such
that the constraints g(v*,x*) are satisfied.
• Value Function f(x*), x*  X*
• Optimizer, states v*(x*,t), x(x*,t), x*  X*
mp-DO Solution
Three methods

mp- (MI)DO (1)


Complete discretization •Lagrange Polynomials for
Parameterizing the Controls
Discrete state space model (Vassiliadis et al., 1994)
(Bemporad and Morari, 1999)
•semi-infinite program - two
mp-(MI)QP (LP) stage decomposition .(similar
(Dua et al., 2000,2001) to Grossmann et al., 1983)

mp- (MI)DO (2)


• Euler – Lagrange conditions of Optimality
• No state or control discretization
Multi-parametric Dynamic Optimization
mp-DO
Optimality Conditions - Unconstrained problem
(No inequality constraints)
Augmented Differenti al System of Equations
x  Ax  Bv
  Qx  AT 
v   R 1 B T 
t  [to , t f ]
Boundary Conditions
x(to )  x*
 (t f )  Px(t f )
Two point boundary value problem
Multi-parametric Dynamic Optimization
mp-DO
Optimality Conditions - Unconstrained problem

Constraint bound
g(x,v)

to tf
Multi-parametric Dynamic Optimization
mp-DO
Optimality Conditions - Constrained problem

Boundary constrained arc


g(x,v) - constraint

Unconstrained arc

to t1 t2 tf
Unknowns
Switching points
Multi-parametric Dynamic Optimization
mp-DO
Optimality Conditions - Constrained problem
Augmented Differenti al System of Equations
x  Ax  Bv
 g 
T

  Qx  A     
T

 x 
 g 
T

v  R (B      )
1 T

 x 
0  gi  i , i  0 Complementarily Conditions
t  [to , t f ]
Multi-parametric Dynamic Optimization
mp-DO
Optimality Conditions - Constrained problem
Boundary Conditions
x(to )  x*
 
x(t1 )  x(t1 ) States - Continuity
 
x(t 2 )  x(t 2 )

 (t f )  Px(t f )
 g 
T
Costates - Adjoints
 (t1 )   (t1 )    
 x 
 (t 2  )   (t 2  )

 
H (t1 )  H (t1 ) Hamiltonian – Switching points
 
H (t 2 )  H (t 2 )
H  x T   xT Qx  vT Rv  g T 
Multi-parametric Dynamic Optimization
mp-DO
Optimality Conditions - Constrained problem
• Solve analytically the dynamics, get time profiles of variables

• Substitute into Boundary Conditions  Eliminate time

M (t1, 2 )    S (t1, 2 )  x*
  ( N (t1, 2 )    F (t1, 2 ) x*  b(t1, 2 ))  0,  (t )  0 Linear in 

where :   x f  o 1 1 ,   1 1 
Non Linear in t1,2

• Solve for ξ (sole unknown) and back-substitute into dynamics

• Get profiles of x(t,x*), v(t,x*), λ(t,x*), μ(t,x*)


Solution of mp-DO
1. Fix a point in x-space

2. Solve DO and determine active constraints and boundary


arcs

3. Determine optimal profiles for μ(t,x*),λ(t,x*),v(t,x*),t1(x*),t2(x*)

4. Determine region where profiles are valid:


Feasibility condition
 *  *
g ( x , t )  0  G1 ( x )  max {g ( x , t )}  0
*
t[ t o ,t f ]

Optimality condition
~ ~ *
 ( x , t )  0  G2 ( x )  min { ( x , t )}  0
* *
t[ t1 ,t 2 ]
Control Law
vˆ  Ac (t , x* )  x*  bc (t , x* )
if Applied for
0  CR1 ( x* )  x*  CR 2 ( x* ), t* t t*+Δt
c  1,, N c
OR

Implement continuously


ˆv( x* )  lim Ac (t , t k ( x* ))  xo  bc (t , t k ( x* ))
t 0

Continuous Control Law via
mp-DO
• Property 1:

v continuous ,  ,  are nonlinear function of x*


• Property 2:

f ( x* ) is continuous , convex
• Property 3:
t kt (x * ), t kx (x * ) continuous NL
• Property 4:
Feasible region: X* convex but each critical region non-convex
2 - state Example
open-loop unstable system

0.0116 0.0099
tf

f ( x )  min x Px f   [ x 
* T
f
T
 x  10 4 2
v ]dt
v *
t 0.0099 0.0084
1.63  0.63 1
x    x   v
 1 0  0 
g  [1.5 1]x  0
 100  v  100
xo  x *

t  [0,1.5]
mp-DO Result
Region
1.4e 10.77t  (1.09 x1  0.94 x2 )  0.22e 0.86t  (0.13x1  1.43x2 )  2.4  0
1.09 x1  0.94 x2
where : t  0.1 ln( )
0.13x1  1.43x2

v(t )  12.4e 10.76t  (1.09 x1  0.93x2 )  21.0e 0.86t  (0.13x1  1.43x2 );


mp-DO Result
Results for constrained region:
mp-DO Result
Results for constrained region:
mp-DO Result
Complexity
 q  N  10 10  (1  2) 
nv N
mp-QP: Max number of regions
 
i 0  i 
   
i 0 

i 

 q  1 1  2 
nv
mp-DO: Max number of regions      
i  i   4
i 0   i 0  

Reduced space of optimization variables and


constraints
mp-DO Result - Simulations

Constrained

Unconstrained
mp-DO Result - Suboptimal
Feature: 25 regions correspond to the same active
constraint over different time elements

Merge and get convex Hull

Compute
feasible
Control law
v = -6.58x1-3.02x2
In Hull

v = -6.92x1-2.9x2-1.59
mp-DO Result - Suboptimal
Brachistrone Problem
Find trajectory of a gravity driven ball such that it will reach
the opposite wall in minimum time
x

γ wall- target
g

x=l
plane-obstacle

y=xtanθ+h
y
Brachistrone Problem
f ( y0 , x0 )  min t f

x  2 gy  cos( )
y  2 gy  sin(  )
y  x tan( )  h, tan ( )  0.5, h  1
x(t f )  l , l  1
x(to )  xo
y (to )  yo
t  [0, t f )
Brachistrone Problem - Results
Brachistrone Problem - Results
Absence of disturbance: open=closed-loop profile
Brachistrone Problem - Results
Presence of disturbance
Concluding Remarks
Advantages
• Improved accuracy and feasibility over discrete time case

• Suitable for the case of model – based control

• Reduction in number of polyhedral regions

• Relate switching points to current state


Issues

• Unexplored area of research


• Non-linearity in path constraints even if dynamics are linear
• Complexity of solution

You might also like