A Lecture on Model Predictive Control

Jay H. Lee
School of Chemical and Biomolecular Engineering Center for Process Systems Engineering Georgia Inst. of Technology

Prepared for Pan American Advanced Studies Institute Program on Process Systems Engineering

Schedule

Lecture 1: Introduction to MPC Lecture 2: Details of MPC Algorithm and Theory Lecture 3: Linear Model Identification

Lecture 1
Introduction to MPC
- Motivation - History and status of industrial use of MPC - Overview of commercial packages

Key Elements of MPC

Formulation of the control problem as an (deterministic) optimization problem

min i ( xi , ui )
ui

g i ( xi , ui ) 0 xi +1 = F ( xi , ui )

i =0

u0 = ( x0 ) HJB Eqn.

On-line optimization Receding horizon implementation (with feedback update)

At t = k , Set x0 = xk (Estimated Current State) Solve the optimization problem numerically Implement solution u0 as the current move. Repeat!

Popularity of Quadratic Objective in Control

Quadratic objective Linear State Space System Model

x Qx + u Ru
i =0 T i i i =0 T i

m1

xk +1 = Axk + Buk yk = Cxk

Fairly general
State regulation Output regulation Setpoint tracking

Unconstrained linear least squares problem has an analytical solution. (Kalmans LQR) Solution is smooth with respect to the parameters Presence of inequality constraints no analytical solution

Classical Process Control

t 1 de 1 + e(t ' )dt '+ D p = kc dt I 0

Ad Hoc Strategies, Heuristics

PID Controllers Lead / Lag Filters Switches Min, Max Selectors If / Then Logics Sequence Logics

Regulation Constraint handling Local optimization

Model is not explicitly used inside the control algorithm No clearly stated objective and constraints

Other Elements

Inconsistent performance Complex control structure Not robust to changes and failures Focus on the performance of a local unit

Example 1: Blending System

Control rA and rB Control q if possible Flowrates of additives are limited Classical Solution

Model-Based Optimal Control

& x = f ( x , u ), y = g ( x , u )
Controller Plant

Set-point

Input

Output

r(t)

u(t)

y(t)

Open-Loop Optimal Control Problem

p 1 ( x i , u i ) + p (x p ) min u 0 ,K , u p 1 i = 0
stage-wise cost

terminal cost

g i ( xi , ui ) 0 g p (xp ) 0 & x = f ( x, u )

Path constraints Terminal constraints Model constraints

Model-Based Optimal Control

& x = f ( x , u ), y = g ( x , u )
Plant

Controller

Set-point

Input

Output

r(t)

u(t)
Measurements

y(t)

Open-loop optimal solution is not robust

Must be coupled with on-line state / model parameter update p 1 ( x i , u i ) + p (x p ) Requiresuminu p 1 solution for each updated problem on-line 0 i= 0 ,K ,
stage-wise terminal Analytical solution possible only in a few cases (LQ control) cost cost

Open-Loop Optimal Control Problem

g i(xi,u ) 0 Computational i limitation for numerical solution, esp. back in Path constraints p ( x p ) the 50sgand 60s 0 Terminal constraints & x = f (x,u )
Model constraints

Model Predictive Control (Receding Horizon Control)

At time k, solve the open-loop optimal control problem on-line with x0 = x(k) Apply the optimal input moves u(k) = u0 Obtain new measurements, update the state and solve the OLOCP at time k+1 with x0 = x(k+1) Continue this at each sample time Implicitly defines the feedback law u(k) = h(x(k))

Analogy to Chess Playing

my move his move my move

Opponent (The Plant) My Move I (The Controller)

The Opponents Move New State

Operational Hierarchy Before and After MPC

Unit 1 - Conventional Structure
Plant-Wide Optimization

Unit 2 - MPC Structure Global Steady-State Optimization (every day) Local Steady-State Optimization (every hour) Dynamic Constraint Control (every minute) Supervisory Dynamic Control (every minute) Basic Dynamic Control (every second)
LC

Unit 1 Local Optimization

Unit 2 Local Optimization

High/Low Select Logic

PID

Lead/Lag

PID

Model Predictive Control (MPC)

SUM

SUM

Unit 1 Distributed Control System (PID)

Unit 2 Distributed Control System (PID)

FC PC

TC LC

FC PC

TC

Example: Blending System

Control rA and rB Control q if possible Flowrates of additives are limited Classical Solution
p = Size of prediction window

MPC: Solve at each time k

p

(rA (k + i | k ) rA *)2 u ( j ),u ( j ),u ( j )

min
2 1 3

j = k ,L, k + p 1

i =1

(ui ) min

+ (q(k + i | k ) q *) ui ( j ) (ui ) max , i = 1, L,3, << 1

+ (rB (k + i | k ) rB *)

2 2

Optimization and Control

An Exemplary Application (1)

An Exemplary Application (2)

Industrial Use of MPC

Initiated at Shell Oil and other refineries during late 70s. The most applied advanced control technique in the process industries. >4600 worldwide installations + unknown # of in-house installations (Result of a survey in yr 1999). Majority of applications (67%) are in refining and petrochemicals. Chemical and pulp and paper are the next areas. Many vendors specializing in the technology
Early Players: DMCC, Setpoint, Profimatics Todays Players: Aspen Technology, Honeywell, Invensys, ABB

Models used are predominantly empirical models developed through plant testing. Technology is used not only for multivariable control but for most economic operation within constraint boundaries.

MPC Industry Consolidation

CPC-V (late 1995) Adersa Setpoint CCI DMCC DMCC DOT Products Honeywell Honeywell Litwin (Profimatics) Neuralware Pavilion Predictive Controls Profimatics Litwin Setpoint Treiber Control MDC Technology Aspentech Neuralware Treiber Update (Late 2000) Adersa ABB DOT CCI GE

Pavilion SimSci Foxboro Invensys PCL FRSI MDC Emerson Elec.

Linear MPC Vendors and Packages

Aspentech
DMCplus DMCplus-Model

Honeywell
Robust MPC Technology (RMPCT)

Adersa
Predictive Functional Control (PFC) Hierarchical Constraint Control (HIECON) GLIDE (Identification package)

MDC Technology

(Emerson)

SMOC (licensed from Shell) Delta V Predict

Predictive Control Limited (Invensys)

Connoisseur

ABB
3d MPC

Result of a Survey in 1999 (Qin and Badgwell)

Nonlinear MPC Vendors and Packages

Adersa
Predictive Functional Control (PFC)

Aspen Technology
Aspen Target

Continental Controls DOT Products

xk+1 = Axk + Buuk + Bv v k yk = g(xk ) = Cxk + NN(xk )

Multivariable Control (MVC): Linear Dynamics + Static Nonlinearity NOVA Nonlinear Controller (NLC): First Principles Model

Pavilion Technologies
Process Perfecter: Linear Dynamics + Static Nonlinearity

Results of a Survey in 1999 for Nonlinear MPC

Controller Design and Tuning Procedure

1. Determine the relevant CVs, MVs, and DVs 2. Conduct plant test: Vary MVs and DVs & record the response of CVs 3. Derive a dynamic model from the plant test data 4. Configure the MPC controller and enter initial tuning parameters 5. Test the controller off-line using closed loop simulation 6. Download the configured controller to the destination machine and test the model predictions in open-loop mode 7. Commission the controller and refine the tuning as needed

Role of MPC in the Operational Hierarchy

Determine plant-wide the optimal operating condition for the day Make fine adjustments for local units Take each local unit to the optimal condition fast but smoothly without violating constraints
Multivariable Control
Plant-Wide Optimization

Local Optimization

MPC
Distributed Control System (PID)
FC PC TC LC

Local Optimization
A separate steady-state optimization to determine steady-state targets for the inputs and outputs; RMPCT introduced a dynamic optimizer recently Linear Program (LP) for SS optimization; the LP is used to enforce input and output constraints and determine optimal input and output targets for the thin and fat plant cases The RMPCT and PFC controllers allow for both linear and quadratic terms in the SS optimization The DMCplus controller solves a sequence of separate QPs to determine optimal input and output targets; CVs are ranked in priority so that SS control performance of a given CV will never be sacrificed to improve performance of lower priority CVs; MVs are also ranked in priority order to determine how extra degrees of freedom is used

Dynamic Optimization
At the dynamic optimization stage, all of the controllers can be described (approximately) as minimizing a performance index with up to three terms; an output penalty, an input penalty, and an input rate penalty:

J = j =1 e
P

y k+ j

2 Qj

+ j =0 uk + j
M 1

2 Sj

+ j =0 e
M 1

u k+ j

2 Rj

A vector of inputs uM is found which minimizes J subject to constraints on the inputs and outputs:

u = u , u ,...u
M T 0 T 1

T T M1

u uk u
u u k u
y yk y

x k +1 = f (x k , u k )
y k +1 = g ( x k +1 ) + b k +1

Dynamic Optimization
Most control algorithms use a single quadratic objective The HIECON algorithm uses a sequence of separate dynamic optimizations to resolve conflicting control objectives; CV errors are minimized first, followed by MV errors Connoisseur allows for a multi-model approach and an adaptive approach The RMPCT algorithm defines a funnel and finds the optimal trajectory yr and input uM which minimize the following objective:
y k + j ,u

minM J = r

P j =1

r k+ j

yk+ j

2 Q

+ u M 1 u ss

2 S

subject to a funnel constraint

Output Trajectories
Aspen Techs DMC ID-COM, Adersas

quadratic penalty Setpoint past future past future

Soft Constraint, Zone Control

quadratic penalty Reference trajectory

Honeywells RMPCT

quadratic penalty Zone past future past future Funnel quadratic penalty

Move suppression is necessary when reference trajectory is not used

Output Horizon

Finite horizon past future prediction horizon P

Coincidence points

Input Parameterization
u Multiple moves (with blocking)

control horizon u Single move (extreme blocking)

u Basis function (parametrized)

Process Model Types

Model Type Differential Equations State-Space Origin Linear/Nonlinear Stable/Unstable S,U

physics L,NL

physics L,NL data

S,U

Laplace Transfer physics L Function data ARMAX/NARMAX data Convolution data (Finite Impulse or Step Response) Other (Polynomial, Neural Net) data L,NL L

S,U

S,U S

L,NL

S,U

Identification Technology
Most products use PRBS-like or multiple steps test signals. es non-PRBS signals Most products use FIR, ARX or step response models
Glide uses transfer function G(s) RMPCT uses Box-Jenkins SMOC uses state space models

Glide us

Most products use least squares type parameter estimation:

prediction error or output error methods RMPCT uses prediction error method Glide uses a global method to estimate uncertainty

Connoisseur has adaptive capability using RLS A few products (DMCplus, SMOC) have subspace identification metho ds available for MIMO identification Most products have uncertainty estimate, but most products do not m ake use of the uncertainty bound in control design

Summary
MPC is a mature technology!
Many commercial vendors with packages differing in model form, objective function form, etc. Sound theory and experience

Challenges are
Simplifying the model development process plant testing & system identification
nonlinear model development

State Estimation
Lack of sensors for key variables

Reducing computational complexity

approximate solutions, preferably with some guaranteed properties

Better management of uncertainty

creating models with uncertainty information (e.g., stochastic model) on-line estimation of parameters / states robust solution of optimization

FCCU Debutanizer

~20% under capacity

Debutanizer Diagram
PC
190 lb

Pressure
PCT Pre-Heater

160 F

Fan

Reflux

To Deethanizer

From Stripper
Feed

TC

TC

Flooding

Tray 20 Temp.

TC
400 F

RVP

Slurry Pump Around

Gasoline to blending

Process Limitation Operation Problems: Overloading -- over design capacity. Flooding -- usually jet flooding, causing very poor separation. Lack of Overhead Fan Cooling -- especially in summer. Consequences: High RVP, giving away Octane Number High OVHD C5, causing problems at Alky.

Control Objectives

Constrained Control: Preventing safety valve from relieving Keep the tower from flooding Keep RVP lower than its target. Regulatory Control: Regulate OVHD PCT or C5 at spec. Rejecting disturbance not through slurry, if possible.

Real-Time Optimization

Optimization Objectives: While maintaining PCT, RVP on their specifications Minimizing energy consumed Minimizing overhead reflux Minimizing overhead cooling required Minimizing overhead pressure Maximizing separation efficiency.

MPC Configuration

Controlled
PCT Fan RVP Diff. Pressure Feed Temp. Tray 20 Temp. OVHD C5 % OVHD Pressure R-PID Internal Reflux

Manipulated
Fan Output Reflux Pre-heater By-Pass Reboiler By-Pass

MPC

R-PID Pre-heater By-Pass Reboiler By-Pass

Disturbances
Feed Stripper Bottom Temp.

MPCs MV Moves
Optimal Point Reached
200 180 160 140 120 100 80 60 40 20 0 0

Pressure (lb)

Feed-heater by-pass

Reboiler by-pass Reflux

500 1000 1500 2000 2500 3000

minute

Reflux vs. Feed

32 30 28 26 24 22 20 18 16 14 12 0 500 1000 1500 2000 2500 3000

Feed An indication that separation efficiency increased. Reflux

minute

Reflux-to-Feed Ratio

0.54

0.52

0.5

0.48

Handling Flood
0.46 0.44

0.42

500

1000

1500

2000

2500

3000

minute

Handling Flood

3.2 3 2.8 2.6

Potential Flooding

Differential Pressure Across Tower

2.4 2.2 2 1.8 1.6 1.4 1.2 0 500 1000 1500 2000 2500 3000

minute

Product Spec's

180 160 140 120 100 80 60 40 20 0 0 500

Top pressure compensated temperature (PCT)

Bottom Reid Vapor Pressure (RVP)

1000 1500 2000 2500 3000

minute

MPC Operation Overview

190 lb Pressure PCT Pre-Heater

160 F

Fan Top Tray

Reflux Feed Tray

Feed Flooding Tray 20 Temp. Tray 20 After MPC is on 400 F RVP Temperature

Gasoline to blending

Tray #

MPC Control Results (1)

C4- Total 5 % 4.5

4 3.5 3 2.5 2 1.5 1 0.5 0 1 2 3 4 5 6 7 8 9 10

MPC When RMPCT was turned on

11

12

13

14

15

16

17

April, 1991

Date

Debutanizer Bottom C4- Total (I-C4, N-C4 & C4=-)

MPC Control Results

(2)

RVP

7.5

When RMPCT was turned on MPC

6.5

5.5

5 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

April, 1991

Date

Debutanizer Bottom RVP Daily Average

MPC Control Results

(3)

C5+ Total 20 % 18
16 14 12 10 8 6 4 2 1 2 3 4 5 6 7 8 9 10

MPC When RMPCT was turned on

11

12

13

14

15

16

17

April, 1991

Date

Debutanizer Top C5+ Total (I-C5, N-C5 & C5=+)

Other Benefits Reflux is 15% lower than before Separation efficiency is increased Now have room for lower RVP or PCT, if needed Variance on PCT and RVP is reduced; Note: Variance on the tray-20 temperature is increased, and it should be!! Energy saving (~10%) OVHD fan maximum bound may be avoided Flooding is eliminated.

Acknowledgment

Professor S. Joe Qin, UT, Austin Dr. Joseph Lu, Honeywell, Phoenix, AZ Center for Process Systems Engineering Industrial Members

Lecture 2

Details of MPC Algorithms and Theory

- Impulse and step response models and the prediction equation - Use of state estimation - Optimization - Infinite-horizon MPC and stability - Use of nonlinear models

Important Ingredients of an MPC Algorithm

Dynamic Model Prediction Model

Predicted future outputs = Function of current state (stored in memory) + feedforward measurement + feedback measurement correction + future input adjustments

Objective and Constraints Optimization Algorithm Receding Horizon Implementation

General Setup
Dynamic Model State: Compact representation Of the past input record Previous State (in Memory) State Update New Input Move (Just Implemented) Current State Prediction Future Input Moves (To Be Determined) Prediction Model for Future Outputs
?

Measurement Correction Feedback / Feedforward Measurements

To Optimization

Options

Model Types
Finite Impulse Response Model or Step Response Model State-Space Model Linear or Nonlinear

Measurement Correction
To the prediction (based on open-loop state calculation) To the state (through state estimation)

Objective Function
Linear or Quadratic Constrained or Unconstrained

Prediction Model for Different Model Types

Finite Impulse Response Model Step Response Model State-Space Model

Sample-Data (Computer) Control

Model relates input samples to output samples

{y0 , y1 , y2 ,L}
v can be a MV (u) or a Measured DV (d)

{v0 , v1 , v2 ,L}

Finite Impulse Response Model (1)

Assumptions: H0 = 0: no immediate effect The response settles back in n steps s.t. Hn+1 = Hn+2 = = 0: Finite Impulse Response (reasonable for stable processes)

Finite Impulse Response Model (2)

Linear Model Superposition Principle

y (k ) = H1v(k 1) + L + H n v(k n)

Finite Impulse Response Model (3)

State x( k ) = v ( k 1),L , v ( k n)
T T

n past input samples (includes both MVs and measured DVs)

State Update: Easy!

x(k + 1) = M 0 x(k ) + v(k ) { { insertion

shift

v(k ) v(k 1) M v(k n + 2) v(k n + 1)

insert

v(k 1) v(k 2) M v(k n + 1) v ( k n)

drop

Finite Impulse Response Model (4)

Prediction
Model prediction of y Model prediction error

y (k ) = H1v(k 1) + L + H n v(k n) = Cx (k )

Model error+Unmeasured disturbance

e( k ) = y m ( k ) y ( k )

Measured output Future input samples (to be decided)

Predicted future output samples

Past input samples stored in memory

yk +1|k u (k 1) u (k ) M = u M + u M 1 2 Dynamic Matrices yk + p|k u (k n) u (k + p 1) (made of d (k 1) d ( k ) e( k ) impulse + M response 1d M + 2d M coefficients) d ( k n) d (k + p 1) e(k )

Feedforward term: Past disturbance samples stored in memory

Feedback Error Correction

Feedforward term: new measurement (Assume d(k)=d(k=1)=.=d(k+p-1))

Step Response Model (1)

Assumptions: S0 = 0: no immediate effect The response settles in n steps s.t. Sn= Sn+1 = = S: the same as the finite impulse response assumption Relationship with the impulse response coefficients:

Step Response Model (2) State

x(k ) = y0T (k ),L , yn 1T (k )

n future outputs assuming the input remains constant at the most recent value

yi (k ) = y (k + i ) w/ u (k ) = u (k + 1) = L = 0

x(k)

Note yn 1 (k ) = yn (k ) = L = y (k ) Also y (k ) = y0 (k )

Pictorial Representation of the State Update

x(k+1)

M1 x(k)

Step Response Model (3)

State Update

x(k + 1) = M 1 x(k ) + {
shift

step response

S {

v(k )
drop

y0 (k + 1) y (k + 1) 1 M yn 2 (k + 1) y (k + 1) n 1

S1

S2

Sn-1 Sn

y0 ( k ) y (k ) 1 M yn 2 (k ) y (k ) n 1

Step Response Model (4)

Prediction
x(k)

Effect of future input moves (to be determined)

Step Response Model (4)

Prediction
Model prediction of y(k) Model prediction error Predicted future output samples

e( k ) = y m ( k ) y ( k )
The state stored in memory

y (k ) = y0 (k ) = [1,0, L,0]x(k )
Future input moves (to be decided)

Dynamic Matrices (made of step response coefficients)

yk +1|k y1 (k ) u (k ) M = M + u M y k + p| k y p ( k ) u (k + p 1) d (k ) e(k ) + M d M d (k + p 1) e(k )

Feedforward term: new measurement (Assume d(k+1)== d(k+p-1)=0)

Feedback Error Correction

State-Space Model (1)

z (k + 1) = Az (k ) + B u u (k ) + B d d (k ) y (k ) = Cz (k )
Discretization State-Space Realization

~ ~ ~ & z = Az + B u u + B d d ~ y = Cz
Linearization

y ( z ) = G u ( z )u ( z ) + G d ( z )d ( z ) or y ( s ) = G ( s )u ( s ) + G d ( s )d ( s )
I/O model Identification

State-Space Identification

& z = f ( z, u, d ) y = g ( z)
Fundamental Model

{y(i), u (i), i = 1,L N }

Test Data

State-Space Model (2)

z (k + 1) = Az (k ) + B u u (k ) + B d d (k ) y (k + 1) = Cz (k + 1) z (k ) = Az (k 1) + B u u (k 1) + B d d (k 1) y (k ) = Cz (k ) z (k + 1) = Az (k ) + B u u (k ) + B d d (k )

y (k + 1) = Cz (k + 1) y (k + 1) = y (k ) + C Az (k ) + B u u (k ) + B d d (k )

Bd z (k + 1) A 0 z (k ) B u y (k + 1) = CA I y (k ) + u u (k ) + d d (k ) CB CB State Update z (k ) y (k ) = [0 I ] x(k + 1) = x(k ) + u u (k ) + d d (k ) y (k )

y ( k ) = x ( k )

State-Space Model (3)

Prediction
Model prediction of y(k) Model prediction error Predicted future output samples

y (k ) = x(k )

e( k ) = y m ( k ) y ( k )
The state stored in memory Future input moves (to be decided)

Dynamic Matrix (made of step response coefficients)

yk +1|k u (k ) x(k ) + u M = M M yk + p|k p u (k + p 1) d (k ) e(k ) + M d M d (k + p 1) e(k )

Feedforward term: new measurement (Assume d(k+1)== d(k+p-1)=0)

Feedback Error Correction

Summary

Regardless of model form, one gets the prediction equation in the form of

yk +1|k u (k ) M = Lx x(k ) + Ld d (k ) + Le e(k ) + Lu M 14444 4444 2 3 known b ( k ) y k + p| k u (k + p 1) 4 4 4 14 244 3 123

Y (k ) U ( k )

Assumptions
Measured DV (d) remains constant at the current value of d(k) Model prediction error (e) remains constant at the current value of e(k)

Ramp Type Extrapolation

For Integrating Processes, Slow Dynamics

e(k + i ) = e(k ) + i (e(k ) e(k 1)) d (k ) = d (k + 1) = L = d (k + p 1)

e(k)-e(k-1)

Use of State Estimation

Measurement Correction of State

Previous State Estimate (in Memory) State Update New Input Move (Just Implemented) Current State Estimate Prediction Future Input Moves (To Be Determined) Prediction Model for Future Outputs Measurement Correction Feedback / Feedforward Measurements State Estimate: Compact representation Of the past input and measurement record

To Optimization

State Update Equation

x(k + 1) = x(k ) + u (k ) + d (k ) + K ( ym (k ) x(k ))

u d

K is the update gain matrix that can be found in various ways

Pole placement: Not so effective with systems with many states (most chemical processes) Kalman filtering: Requires a stochastic model of form

x(k + 1) = x(k ) + u u (k ) + d d (k ) + w(k ) y ( k ) = x ( k ) + ( k )

Can be obtained using, e.g., subspace ID White noises of known covariances Effect of unmeasured disturbances and noise

Prediction Equation
Contains past feedback measurement corrections

yk +1|k u (k ) u M M = M x(k ) + yk + p|k p u (k + p 1) d (k ) e(k ) + M d M d (k + p 1) e(k )

Additional measurement correction NOT needed here!

What Are the Potential Advantages?

Can handle unstable processes
Integrating processes, run-away processes

Cross-channel measurement update

More effective update of output channels with delays or measurement problems based on other channels.
Unmeasured inputs Measured inputs Early, robust update through modeled correlation

y1

Process

Delays, Measurement Difficulties, Slow Sampling

y2

Systematic handling of multi-rate measurements Optimal extrapolation of output error and filtering of noise (based on the given stochastic system model)

Optimization

Objective Function
Minimization Function: Quadratic cost (as in DMC)
p m 1 i =0

V (k ) = ( yk +i|k y*) ( yk +i|k y*) + u T (k + i )u u (k + i )

T y i =1

Consider only m input moves by assuming u(k+j)=0 for jm Penalize the tracking error as well as the magnitudes of adjustments

Use the prediction equation.

First m columns of Lu
T V (k ) = (Y (k ) Y *) diag(y )(Y (k ) Y *) + U m (k )diag(u )U m (k ) T

Substitute
T m

Y (k ) = b(k ) + Lu U m (k ) m
T

V (k ) = U (k ) HU m (k ) + g (k )U m (k ) + c(k )

constant

Constraints

Substitute the prediction equation and rearrange to

CU m (k ) h(k )

Optimization Problem

Quadratic Program
U m ( k )

min U (k ) HU m (k ) + g (k )U m (k )
T m T

such that CU m (k ) h(k )

Unconstrained Solution

1 1 U m (k ) = H g (k ) 2
Constrained Solution
Must be solved numerically

Quadratic Program

Minimization of a quadratic function subject to linear constraints Convex and therefore fundamentally tractable Solution methods
Active set method: Determination of the active set of constraints on the basis of the KKT condition Interior point method: Use of barrier function to trap the solution inside the feasible region, Newton iteration

Solvers
Off-the-shelf software, e.g., QPSOL Customization is desirable for large-scale problems

Two-Level Optimization
Steady-State Optimization (Linear Program) min L( y|k , u s (k ))
us ( k )

y | k Cs cs ( k ) u s (k ) u s (k ) = u (k 1) + u (k ) + L + u (k + m 1) y|k = bs (k ) + Ls u s (k )
Optimal Setting Values (setpoints)
Steady-State Prediction Eqn. State Feedforward Measurement Feedback Error Dynamic Prediction Eqn.

y , u (k )
* | k * s

(sometimes input deviations are included in the quadratic objective function)

To Dynamic Optimization (Quadratic Program)

Use of Infinite Prediction Horizon and Stability

Use of Prediction Horizon Why?

Stability guarantee
The optimal cost function can be shown to be the control Lyapunov function

Less parameters to tune More consistent, intuitive effect of weight parameters Close connection with the classical optimal control methods, e.g., LQG control

Step Response Model Case

V (k ) = ( yk +i|k y*) ( yk +i|k y*) + u T (k + i )u u (k + i )
T y i =1 i =0 m 1

V (k ) =

m + n 1 i =1

( yk +i|k y*) ( yk +i|k y*) + u T (k + i )u u (k + i )

T y i =0

m 1

with extra constraint yk + m + n 1|k = y *

Must be at y* for the cost to be bounded

Additional Comments

Previously, we assumed finite settling time. Can be generalized to state-space models

More complicated procedure to turn the -horizon problem into a finite horizon problem Requires solving Lyapunov equation to get the terminal cost matrix Also, must make sure that output constraints will be satisfied beyond the finite horizon construction of output admissible set

Use of a sufficiently large horizon (p m+ the settling time) should have a similar effect Can we always satisfy the settling constraint?
y=y* may not be feasible due to input constraints or insufficient m use two-level approach

Two-Level Optimization

Steady-State Optimization (Linear Program or Quadratic Program)

Optimal Setting Values (setpoints)

y , u (k )
* | k * s

Dynamic Optimization (-horizon MPC)

* Constraint yk + m + n 1|k = y|k is guaranteed to be feasible.
* * Constraint u(k) + L + u(k + m - 1 ) = us yk + m + n-1|k = y|k

Use of Nonlinear Model

Difficulty (1)

& x = f ( x, u , d ) y = g ( x)

Discretization?

x(k + 1) = F ( x(k ), u (k ), d (k )) y (k ) = g ( x(k ))

Orthogonal Collocation

yk + 2|k = g o F (F ( x(k ), u (k ), d (k )), u (k + 1), d (k ) ) + e(k ) M yk + p|k = g o F p ( x(k ), u (k ),Lu (k + p 1), d (k ) ) + e(k )

yk +1|k = g o F ( x(k ), u (k ), d (k )) + e(k )

The prediction equation is nonlinear w.r.t. u(k), , u(k+p-1)

Nonlinear Program (Not so nice!)

Difficulty (2) State Estimation

& x = f ( x, u , d ) + w y = g ( x) +
Extended Kalman Filtering
k +1

x(k + 1) =

f ( x , u , d ) + K ( k )( y
k

(k ) g ( x(k ) )

Computationally more demanding steps, e.g., calculation of K at each time step Based on linearization at each time step not optimal, may not be stable Best practical solution at the current time Promising alternative: Moving Horizon Estimation (requires solving NLP) Difficult to come up with an appropriate stochastic system model (no ID technique)

Practical Algorithm
EKF

x(k)

Dynamic Matrix based on the linearized model at the current state and input values

k +1 f ( x, u , d ) yk +1|k k u (k ) y k +2 u (k + 1) k + 2|k = f ( x, u , d ) + u (k ) M k M k + p M + p 1 y k + p| k u (k444 ) 144 2 3 f ( x, u , d ) Linearized Effect 4 Input Adjustments of Future k

Model integration with constant input u=u(k-1) and d=d(k) Linear prediction equation

Linear or Quadratic Program

Additional Comments / Summary

Some refinements to the Practical Algorithm are possible

Use the previously calculated input trajectory (instead of the constant input) in the integration and linearization step Iterate between integration/linearization and control input calculation

Full-blown nonlinear MPC is still computationally prohibitive in most applications

Lecture 3 Linear Model Identification

- Model structure - Parameter/model estimation - Error analysis - Plant testing - Data pretreatment - Model validation

System Identification

Building a dynamic system model using data obtained from the plant

Why Important?

Almost all industrial MPC applications use an empirical mo del obtained through system identification Poor model Poor Prediction Poor Performance Up to 80% of time is spent on this step Direct interaction with the plant
Cost factor, safety issues, credibility issue

Issues and decisions are sufficiently complicated that syst ematic procedures must be used

Steps and Decisions Involved

Plant Test
Raw Data Test signal (shape, size of perturbation) Closed-loop or open-loop? One-input-at-a-time or simultaneous? How long? Etc. Outlier removal Pre-filtering

Pretreatment
Conditioned Data

ID Algorithm
Model

Model structure (Parameter) estimation algorithm

No

Validation
Yes

Source of validation data Criterion

End Objective: Control

Model Structure

Model Structure (1)

Model
Process Noise Disturbance Model Output Noise Measured Outputs

Inputs

Plant Dynamics

I/O Model
White noise sequence

y ( k ) = G ( q )u ( k ) + 1 24 4 3
effect of inputs

effect of disturbances, noise

H ( q ) e( k ) 14 3 24

Models auto- and cross-correlations of the residual (not physical cause-effect)

Assume wolog that H(0)=1

SISO I/O Model Structure (1)

FIR (Past inputs only)

y (k ) = b1u (k 1) + L + bmu (k m) + e(k )

ARX (Past inputs and outputs: Equation Error)

G (q ) = b1q 1 + L + bm q m , H (q ) = 1

y (k ) = a1 y (k 1) + L + an y (k n) + b1u (k 1) + L + bmu (k m) + e(k )

b1q 1 + L + bm q m 1 G (q) = , H (q) = ARMAX (Moving average of the noise term) q 1 L a q n 1 a1q 1 L an q n 1 a1 n

y (k ) = a1 y (k 1) + L + an y (k n) + b1u (k 1) + L + bmu (k m) + e(k ) + c1e(k 1) + L + cn e(k n)

b1q 1 + L + bm q m 1 + c1q 1 + L + cn q n G (q) = , H (q) = 1 n 1 a1q L an q 1 a1q 1 L an q n

SISO I/O Model Structure (2)

Output Error (No noise model or white noise error)

~ (k ) = a ~ (k 1) + L + a ~ (k n) + b u (k 1) + L + b u (k m); y 1y ny 1 m y ( k ) = ~ ( k ) + e( k ) y
b1q 1 + L + bm q m G (q) = , H (q) = 1 1 n 1 a1q L an q

Box Jenkins (More general than ARMAX)

b1q 1 + L + bm q m 1 + c1q 1 + L + cn q n G (q) = , H (q) = 1 n 1 a1q L an q 1 d1q 1 L d n q n

MIMO I/O Model Structure

Inputs and outputs are vectors. Coefficients are matrices. For example, ARX model becomes

y (k ) = A1 y (k 1) + L + An y (k n)
G (q) = I A1q L An q
1 1

( H (q) = (I A q

+ B1u (k 1) + L + Bmu (k m) + e(k )

1 n 1

L An q

n 1

) (B q )
1

+ L + Bm q m

Identifiability becomes an issue

A is an sets n matrix and Bi is an n y same G(q) iDifferentn y of ycoefficient matrices giving exactlynu matrixand H( q) through pole/zero cancellations Problems in parameter estimati on Requires special parameterization to avoid problem

State Space Model

Deterministic

x(k + 1) = Ax(k ) + Bu (k ) y (k ) = Cx(k ) + e(k )

Combined Deterministic / Stochastic

Output Error Structure

x(k + 1) = Ax(k ) + Bu (k ) + Ke(k ) y (k ) = Cx(k ) + e(k )

xd (k + 1) = Axd (k ) + Bu (k ) yd (k ) = Cx(k )
Effect of deterministic input

ARMAX Structure

xs (k + 1) = Axs (k ) + Ke(k ) y s (k ) = Cxs (k ) + e(k )

Auto- and cross-correlation of the residual

Identifiability can be an issue here too

State coordinate transformation does not change the I/O relationship

Parameter (Model) Estimation

Overview

Data

Model Structure Selection

Model Structure

Parameter (Model) Estimation

Model For Validation

Prediction Error Method

IV Method

Statistical Method
MLE Bayesian

Subspace ID Method ETFE

Frequency Domain

Prediction Error Method

Predominant method at current time Developed by Ljung and coworkers Flexible
Can be applied to any model structure Can be used in recursive form

Well developed theories and software tools

Book by Ljung, System ID Toolbox for MATLAB

Computational complexity depends on the model st ructure

ARX, FIR Linear least squares ARMAX, OE, BJ Nonlinear optimization

Not easy to use for identifying multivariable models

Prediction Error Method

Put the model in the predictor form

y (k ) = G (q, )u (k ) + H (q, )e(k ) yk |k 1 = G (q, )u (k ) + I H 1 (q, ) ( y (k ) G (q, )u (k ) ) 14 244 4 3

Contains at least 1 delay

e(k ) = y (k ) yk |k 1 = H 1 (q, )( y (k ) G (q, )u (k ) )

Choose the parameter values to minimize the sum of the pr ediction error for the given data

1 min N

k =1

e( k ) 2
2

e(k ) = H 1 (q, )( y (k ) G (q, )u (k ) )

ARX, FIR Linear least squares, ARMAX, OE, BJ Nonlinear least squares

Subspace Method
More recent development Dates back to the classical realization theories but rediscov ered and extended by several people Identifies a state-space model Some theories and software tools Computationally simple
Non-iterative, linear algebra

Good for identifying multivariable models

No special parameterization is needed

Not optimal in any sense May need a lot of data for good results May be combined with PEM
Use SS method to obtain an initial guess for PEM

Main Idea of the SS-ID Method (1)

Assumed Form of the Underlying Plant

x(k + 1) = Ax(k ) + Bu (k ) + w(k ) y (k ) = Cx(k ) + (k )

Innovation Form (Steady-State Kalman Filter) Equivalent to the above in I/O sense

x(k + 1) = Ax(k ) + Bu (k ) + Ke(k ) y (k ) = Cx(k ) + e(k )

Identify {A, B, C, K, Cov(e)} within some similarity transformation
We are free to choose the state coordinates

Main Idea of the SS-ID Method (2)

y (k 1) M y (k n )
Past outputs An upper bound of the state dimension n

L1

x1 (k ) M xn ( k )

C CA Extended Observability M Matrix n 1 CA

u (k 1) M u (k n )
Past inputs

L2

State (Minimum Storage)

yk |k 1 ek |k 1 M + M yk + n 1|k 1 ek + n 1|k 1
Future Output Prediction Prediction Error

u (k ) M u (k + n 2)
Future inputs

Main Idea of the SS-ID Method (3)

Y 0+ Y = o [L1 L2 ] + L3U 0+ + E 0+ 1 24 U 4 3 M

Find M through linear least squares

Consistent estimation since E0+ is independent of the regres sors Oblique projection of data matrices

Perform SVD on M and find n as well as o

M = [Q1

n Q2 ] 0

0 P1T T 0 P2

o = Q11/ 2 n
Some variations exist among different algorithms in terms of picking the state basis

The column space of Q1 is the state space

Main Idea of the SS-ID Method(4)

Obtain the data for x(k)

Y (k ) x(k ) = o { U (k ) pseudo inverse

Obtain the data for x(k+1) in a similar manner Obtain A,B and C through linear regression
Consistent estimation since the residual is independent of the regre ssor

x(k + 1) A B x(k ) Ke(k ) y (k ) = C 0 y (k ) + e(k ) 4 4 123

Residual

Obtain K and Cov(e) by using the residual data

Properties
N4SID (Van Overschee and DeMoor)
Kalman filter interpretation Proof of asymptotically unbiasedness of A, B, and C Efficient algorithm using QR factorization -

CVD (Larimore)
Founded on statistical argument Same idea but the criterion for choosing the state basis (Q1) diff ers a bit from N4SID based on correlation between past I/O data and future output data, rather than minimization of the pre diction error for the given data

Alternative

o C , A B
MOESP (Verhaegen)

Error Analysis

Error Types
Bias: Error due to structural mismatch
Bias = the error as # of data points Independent of # of data points collected Bias distribution (e.g., in the frequency domain) depends on the input spectrum, pre-filtering of the data, etc. Frequency-domain bias distribution under PEM - by Ljung

Variance: Error due to limited availability of data

Vanishes as # of data points Depends on the number of parameters, the number of data poin ts, S/N ratio, etc. but not on pre-filtering Asymptotic distribution (as n, N ):

Main tradeoff
Richer structure (more parameters) Bias, Variance
n cov(vec(GN )) ( u ) T d N 144 44 2 3
Noise to signal ratio

Plant Test for Data Generation

Test Signals
Very Important
Signal-to-noise ratio Distribution and size of the variance Bias distribution

Popular Types
Multiple steps: Power mostly in the low-frequency region Good e stimation of steady-state gains (even with step disturbances) but g enerally poor estimation of high frequency dynamics PRBS: Flat spectrum Good estimation of the entire frequency re sponse, given the error also has a flat spectrum (often not true) Combine steps w/ PRBS?

Multi-Input Testing (MIT) vs. Single Input Testing (SIT)

MIT gives better signal-to-noise ratio for a given t esting time Control-relevant data generation requires MIT MIT can be necessary for identification of highly i nteractive systems (e.g., systems with large RGA) SIT is often preferred in practice because of the m ore predictable effect on the on-going operation

Open-Loop vs. Closed-Loop

Open-Loop Testing

Perturbation Signal r G0 u
Closed-Loop Testing

Dither Signal r
Location 1 Location 2

d y G0

Pros and Cons of Closed-Loop Testing Pros

Safer, less damaging effect on the on-going operation Generates data that are more relevant to closed-loop c ontrol

Cons
Correlation between input perturbations and disturbanc es / noise through the feedback. Many algorithms can fail or give problems They give bias unless the assumed noise structure is perfect

Important Points from Analysis

External perturbations (dither) are necessary.
Perturbations due to error feedback hardly contributes to variance reduction (since they are correlated to the errors)

)) n ( r ) T cov(vec(GN u d N

Portion of the input spectrum due to the dithering

Output error spectrum

The level of external perturbation signals also contribute to the siz e of bias due to the feedback-induced correlation
e eu u 1 = ( Pe u 1 ) ( u u 1 ) 123 1 24 4 3 Noise to signal ratio

E{G0 GN } = ( H 0 H ) eu u 1

Relative contribution of noise feedback to input spectrum

Specialized algorithm may be necessary to avoid bias

Different Approaches to Model Identification with Closed-Loop Data Direct Approach

D N = { y (i ), u (i ), i = 1,..., N } G N ( and H N )

Indirect Approach
T I C) yr G =(T G D = { y (i ), r (i ), i = 1,..., N } TN N

yr N

yr N

Joint I/O Approach

yr ur T (T ) D = { y (i ), u (i ), r (i ), i = 1,..., N } (TN , TN ) G = GN

yr N

ur 1 N

Two-Stage or Projection Approach

~N D = {u(i), r(i),i = 1,...,N} D1 = {u(i)} GN ~N D2 = { y(i)}
N 1

Data Pretreatment

Main Issues (1) Time-consuming but very important Remove outliers Remove portions of data corresponding to u nusual disturbances or operating conditions Filter the data
Affects bias distribution (emphasize or de-emphasize different frequency regions) Does NOT improve the S/N ratio often a misconcept ion

Main Issues (2) Difference the data? (y = y(k) y(k-1), u(k) = u(k) u(k-1))
Removes trends (e.g., effect of step disturbances, set point changes) that can destroy the effectiveness of m any ID methods (e.g., subspace ID) Often used in practice Also removes the input power in the low-frequency re gion. (PRBS zero input power at = 0) Amplifies high-frequency parts of the data (e.g., noise ), so low-pass filtering may be necessary

Model Validation

Overview
Use fresh data different from the data used for m odel building Various methods
Size of the prediction error Whiteness of the prediction error Cross correlation test (e.g., prediction error and inputs)

Good prediction with test data but poor predictio n with validation data
Sign of overfit Reduce the order or use more compact structures like ARM AX (instead of ARX)

Concluding Remarks on Linear ID

System ID is often the most expensive and difficult par t of model-based controller design Involves many decisions that affect
Plant operation during testing Eventual performance of the controller

Good theories and systematic tools are available System ID can also be used for constructing monitorin g models
Subspace identification Trend model, not a causal model Active testing is not needed

Deterministic Multi-Stage Optimization

p 1 min1 ( x j , u j ) + p (x p ) u0 ,K,u p j = 0
stage-wise terminal cost cost

g j (x j ,u j ) 0 g p (xp ) 0 x j +1 = f ( x j , u j )

Path constraints Terminal constraints Model constraints

General formulation for deterministic control and scheduling problems. Continuous and integer state / decision variables possible In control, p= case is solved typically. Uncertainty is not explicitly addressed.

Solution Approaches
Analytical approach: 50s-70s
Derivation of closed form optimal policy ( requires solution to HJB equation (hard!)

u j = * (x j )

Numerical approach: 80s-now

Math programming (LP, QP, NLP, MILP, etc.):
Fixed parameter case solution Computational limitation for large-size problem (e.g., when p= ).

Parametric programming:
General parameter dependent solution (e.g., a lookup table) Significantly higher computational burden

Practical solution:
Resolve the problem on-line whenever parameters are updated or constraints are violated (e.g., in Model Predictive Control or Reactive Scheduling).

Stochastic Multi-Stage Decision Problem with Recourse

j min) E ( x j , u j ) u j = ( x j j =0
0 < <1 x j +1 = f h ( x j , u j , j ) Pr g ( x j , u j ) 0
Discount factor Markov sys. model Chance constraint

Next holy-grail of control: A general form for control, scheduling, and other real-time decision problems in an uncertain dynamic environment. No satisfactory solution approach currently available.

Limitation of Stochastic Programming Approach

Simple case of 2 scenarios ( or) per stage

x k+2

xk+1
xk
uk
u k +1

u 2 k+ x k+2 u k+2

x k +2

Total number of decision variables = (1+2+4++2p-1) nu Number of branches to evaluate for each candidate decisions = 2p

x+1 k
u k +1

u k +2 x k +2 u k +2

Stochastic Programming Approach

General case (S number of scenarios per stage)

Total number of decision variables = (1 + S + S2 ++ Sp-1) nu Number of branches to evaluate for each decision candidate = Sp Not feasible for large S (large number of scenarios) and/or large p (large number of stages)
practically limited to two stage problems with a small number of scenarios.

Current practical approach: Evaluate most likely branch(es) only. BUT highly limited!

Dynamic Programming (DP)

The concept of cost-to- * J ( xk ) go
Represents future costs under (optimal) control Parameterizes the solution as a function of state x
u

j = k +1

j k 1 (x j , * ( x j ) )

( x ) = arg min { ( x, u ) + J * ( fh ( x, u) )}

J*(x) is a solution to Bellman Equation

J * ( x ) = min E ( x, u ) + J * ( f h ( x, u ) )
uU

Value Iteration Approach to Solving DP

J * ( xk ) = min E ( xk , u k ) + J * ( f h ( xk , u k )
u(k )

Discretization of entire state space

ali ) ion ces s +a en J ip1( ) = min E ( , u ) + J i () S m u Di tion f c eo A s ur te & sampling & C ta discretization (S

Valuety iteration

i = i+1

Approximate Dynamic Programming (ADP)

Bellman equation needs to be solved x X
Curse of dimensionality! Not suitable for high dimens. sys.

Key idea of ADP

To find approximate cost-to-go function Use simulations under known suboptimal policy to sample a very small relevant fraction of the states and initialize cost-to-go value table. Iteratively improve the policy and cost-to-go function
Iterate over only the sampled points in the state space Use interpolation to evaluate the cost-to-go values for nonsampled points.

Approximate Dynamic Programming (ADP)

Approximate Value Iteration

On-line Decision ~ u = ( x) = arg min E ( x,u ) + J ( f ( x,u ) )

u

Monte-Carlo Simulations Closed-loop w/ suboptimal policies MPC, PI, etc.

Iteration on the Bellman Equation

~ J i +1 ( x) = min E ( x,u ) + J i ( f ( x,u ) )

uU

Sample Avg. sampled state points Sampled State Points

i i +1

converged solution

~ J*

Approximation of Cost-to-Go State and input trajectories: xk , u k Initial cost-to-go: J ( x) = i i

~ J ( x) := xk J ( xk )