See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/278627260

Applied Model Predictive Control - a brief guide do MATLAB/Simulink MPC

toolbox

Technical Report · August 2014

DOI: 10.13140/RG.2.1.2334.7369

CITATIONS READS
3 5,475

2 authors, including:

Stefano Ciannella
Universidade Federal da Paraíba
7 PUBLICATIONS   14 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

A practical guidebook for MPC applications in MATLAB for Chemical Engineers View project

All content following this page was uploaded by Stefano Ciannella on 17 June 2015.

The user has requested enhancement of the downloaded file.

APPLIED MODEL PREDICTIVE CONTROL
A brief guide to MATLAB/Simulink® MPC toolbox

by

STEFANO CIANNELLA
under the guidance and supervision of Professor W.R. Cluett, Ph.D.

August 2014
Contents

1. Introduction………………………………………………………………………5

2. Model predictive control overview…………………………………................6

3. MATLAB/Simulink® MPC Toolbox……………………………………………9

3.1. MPC Toolbox interface overview………………………………….……...8

3.1.1. Controller tab.......................................................................…....9

3.1.2. Constraints tab...................................................................……10

3.1.3. Weight tuning tab………………………………………….….……12

3.1.4. Scenarios tab……………………………………………….……...13

3.1.5. Exporting a controller…………………………………………......13

3.2. Setting up and starting a simulation…………………………………….14

3.2.1. Example 1: First Order Process…………………………………14

3.2.2. First Order Process: Effect of Prediction horizon P…………...20

3.2.3. First Order Process: Effect of Control horizon M………………23

3.3. General instructions to implement MPC……………………………….24

4. Case study problems....……..……………………………………………….25

4.1. Case study #1: Continuous Heat Transfer System………………….25

4.1.1. CHTS – servo problem………………………………………….26

4.1.2. Effect of P and M on servo problem…………………………...31

4.1.3. CHTS – regulation problem…………………………………….33

4.1.4. Effect of P and M on regulation problem……………………...37

1
 4.1.5. Effect of input rate weight on regulation problem…..………..38

4.1.6. CHTS – adding a feedforward controller routine…………….40

4.2. Case study #2 – Evaporator System…………………………………43

4.2.1. Effect of P and M on servo problem...………………………..47

4.2.2. Effect of input rate and output weights on servo problem….48

4.2.3. Regulation problem on evaporator system……...…………...52

4.2.4. Effect of P and M on regulation problem………...…………...53

4.2.5. Effect of input rate and output weights on reg. problem.……56

4.2.6. Controller validation ultimate test……………..………….……….61

5. References………………………...…………………………………………63

2
List of symbols

𝑦 Measured output

𝑦̂ Predicted output

̂𝑐
𝑦 Predicted output after model correction

𝑢 Manipulated variable

∆𝑢 Change in the manipulated variable from one sampling instant to the next

𝑘 Current sampling instant

𝑀 Control horizon, or number of control moves

𝑃 Prediction horizon

𝑟 Setpoint

𝑑 Difference between predicted and measured outputs

𝑤 Weight for changes in the manipulated variable

𝑆𝑁 N step-response coefficients

𝑀𝑉 Manipulated variable

𝐶𝑉 Controlled variable

𝐷𝑉 Disturbance variable

3
Acknowledgments

In first place, I would like to acknowledge Professor W.R. Cluett at the University of Toronto

whom taught Process Dynamics and Control as an undergraduate course and supervised me

during my Summer research, stimulating and encouraging my interest in this field of study. Also,

the Chemical Engineering and Applied Chemistry department staff for orientation and support

over all the courses and academic work I have engaged during my stay in Canada.

Secondly, my scholarship agency CNPq for financial support and the Science without

Borders academic exchange program, which offered the opportunity to study abroad and learn

about North America high education level and lifestyle.

Finally, my friends and colleagues whom provided me a comfortable and enjoyable working

experience, and naturally the support from my family whom, even standing far in Brazil,

supported me in all possible ways.

4
 1. Introduction

Model Predictive Control (MPC) is a class of control techniques first derived from Internal

Model Control, or IMC, and is widely applied in the process industries due to its capability to

deal with constraints in an optimal fashion; as the name suggests, MPC is based on predictions

of setpoint tracking behavior or disturbance rejection over both past controlled and manipulated

variables measurements, in which each prediction is followed by an optimization routine to find

the optimal input for the closed loop response imposed by a certain criteria, such as maximizing

a profit function or production rate.

This text provides a succinct background on the MPC philosophy and modeling equations,

followed by a step-by-step guide to how to implement predictive techniques using

MATLAB/Simulink for SISO and MIMO systems. Furthermore, the role of the various

parameters within MPC Simulink toolbox and how to find proper values to obtain a desired

closed response shall be addressed.

After reading this text material, one should become familiar with the basic ideas behind

SISO and MIMO MPC, including the following:

 MPC block insertion and set-up in a Simulink diagram;

 Evaluate parameters such as prediction and control horizons and input/output

weights;

 The effect of such parameters on the control performance and system stability.

5
 2. Model predictive control overview

As briefly discussed in the introduction, MPC controllers are widely applied over process

and oil/petrochemical industries due to its capability to deal in an optimal form with input/output

process constraints (upper or lower values for a specific variable, for instance); in this sense,

one should be able to safely operate their system by restricting it to be conducted in a limited

region of operation, such as maximum and minimum liquid level within a distillation column or

a maximum opening degree in a valve.

Basically, MPC calculations are performed at each sampling time which can be also set by

the control designer; these calculations are based on current measurements and predictions of

future output values. Two types of computation are primordial in a MPC controller: setpoint

calculations and control calculations, this last one includes process constraints and other

parameters that are able to be manually specified. The main task of a MPC controller is to

determine a sequence of control moves in the manipulated variable, so the system can be

tracked to its setpoint in an optimal fashion.

Figure 1 - Basic concept of MPC

6
 The control calculations are based on minimizing the predicted deviations from the

reference trajectory. The key idea behind a MPC controller is illustrated in Figure 1. When using

MPC, an optimization problem is solved at each time step 𝑘 through an objective function based

on output predictions over a prediction horizon of 𝑃 time steps; this objective function (usually

a quadratic one) is minimized by a selection of manipulated variables moves over a control

horizon of 𝑀 control moves. It’s important to emphasize that, even though at each time step a

group of 𝑀 moves is calculated, only the first one 𝑢𝑘 is implemented. After this step, the

measurement at the next time instant 𝑦𝑘+1 is obtained, followed by a correction due to model

error, and then a new optimization problem is solved again. These procedures are carried every

for every time step 𝑘.

Before starting to enlighten the reader with MPC equations, it’s important to assert the

definition of a finite step response (FSR) model; these models are used in almost all MPC

modeling equations and are obtained by simply introducing a unit step change to a process

operating at a steady state. The model coefficients (𝑆1 , 𝑆2 , 𝑆3 , … , 𝑆𝑁 ) are the output values at

each time step.

The predicted output for an instant time 𝑘 is calculated through the following equation:

̂𝑘 = ∑𝑁−1
𝑦 𝑖=1 𝑆𝑖 ∆𝑢𝑘−𝑖 + 𝑆𝑁 𝑢𝑘−𝑁 (i)

The model predicted output is hardly equal to the actual measured output at a certain time

step 𝑘; this discrepancy is denoted by:

𝑑𝑘 = 𝑦𝑘 − 𝑦
̂𝑘 (ii)

The corrected prediction is then:

̂𝑐 = 𝑦
𝑦 ̂𝑘 + 𝑑𝑘 (iii)
𝑘

By combining equations (i), (ii) and (iii), the corrected prediction for the jth step into the

future can be demonstrated as:

7
 ̂ 𝑐 𝑗 𝑁−1
𝑦𝑘+𝑗 = ∑𝑖=1 𝑆𝑖 ∆𝑢𝑘−𝑖+𝑗 + ∑𝑖=1 𝑆𝑖 ∆𝑢𝑘−𝑖+𝑗 + 𝑆𝑁 𝑢𝑘−𝑁+𝑗 + 𝑑𝑘+𝑗 (iv)

Equation (iv) can be easily seen as a “compilation” of the effect of future control moves (first

term to the right), and past control moves (second and third terms), as well as a correction term

𝑑𝑘+𝑗 .

Furthermore, the difference between setpoint trajectory and future predictions in jth step,

with j ≤ P, is given by:

̂ 𝑗
𝑐
𝑟𝑘+𝑗 − 𝑦𝑘+𝑗 = 𝑟𝑘+𝑗 − [∑𝑁−1
𝑖=1 𝑆𝑖 ∆𝑢𝑘−𝑖+𝑗 + 𝑆𝑁 𝑢𝑘−𝑁+𝑗 + 𝑑𝑘+𝑗 ] − ∑𝑖=1 𝑆𝑖 ∆𝑢𝑘−𝑖+𝑗 (v)

Equation (v) is essential for the optimization problem and can be used in a quadratic

objective function for a prediction horizon 𝑃 and a control horizon of 𝑀 moves:

2
𝑓𝑜𝑏𝑗 = ∑𝑃𝑖=1(𝑟𝑘+𝑖 − 𝑦̂
𝑐 𝑀−1 2
𝑘+𝑖 ) + 𝑤 ∑𝑖=0 ∆𝑢𝑘+1 (vi)

For MIMO systems, the derivation of prediction equations is primarily based over the

previous analysis for SISO systems. As an example, let’s consider a 2x2 system: the predictive

control model will consist of two equations and four individual step-response models,

𝑦1 + 𝑗) = ∑𝑁−1
̂(𝑘 𝑁−1
𝑖=1 𝑆11,𝑖 ∆𝑢1 (𝑘 − 𝑖 + 𝑗) + 𝑆11,𝑁 𝑢1 (𝑘 − 𝑁 + 𝑗) + ∑𝑖=1 𝑆12,𝑖 ∆𝑢2 (𝑘 − 𝑖 + 𝑗) + 𝑆12,𝑁 𝑢2 (𝑘 − 𝑁 + 𝑗) (vii)

And,

𝑦
̂(𝑘
2 + 𝑗) = ∑𝑁−1 𝑁−1
𝑖=1 𝑆21,𝑖 ∆𝑢1 (𝑘 − 𝑖 + 𝑗) + 𝑆21,𝑁 𝑢1 (𝑘 − 𝑁 + 𝑗) + ∑𝑖=1 𝑆22,𝑖 ∆𝑢2 (𝑘 − 𝑖 + 𝑗) + 𝑆22,𝑁 𝑢2 (𝑘 − 𝑁 + 𝑗) (viii)

where 𝑆12,𝑖 denotes the ith step-response coefficient for the model that relates 𝑦1 and 𝑢2 , and

vice-versa. It can be easily noticed that a 2x2 system will feature two objective functions, one

for 𝑦1 and another for 𝑦2 , each one with their own parameters and input weights 𝑤1 and 𝑤2 .The

reader should feel encouraged to derive such objective functions by following the same

procedure performed to SISO systems, obtaining two equations that are analogous to equation

(vi).

8
 3. MATLAB/Simulink® MPC Toolbox

3.1. MPC Toolbox interface overview

To open MPC toolbox in MATLAB, type the following command at the command window:

>>mpctool

After the toolbox is loaded, the following window becomes visible:

This is the “MPC Design Task” tab, in which one may import a plant from MATLAB

workspace or even a controller which was previously designed; the plant can be a linearized

model obtained from a Simulink block diagram (right click on a block or subsystem then choose

“Linear analysis” to linearize it), or even a transfer function or state space model. Upon

importing a plant or a controller, one may set input and output signal properties, such as name,

type of variable (manipulated, measured disturbance or unmeasured disturbance) and their

description along with physical units and a nominal value.

In order to enlighten the reader about the various tabs and options in MPC toolbox, a

generic single 1x1 linearized model collected from a first-order transfer function will imported

as a plant, so we can start to design a controller. After clicking on “Import Plant” and choosing
9
a model from MATLAB workspace, the following interface becomes visible:

3.1.1. Controllers tab

By clicking on “Controllers” tab, you can notice the existence of a rough MPC controller

archive which was created when the plant was imported into the toolbox (MPC1). In this tab,

one may import or export and create or delete controllers, as well as check its configuration

and control/prediction horizon values.

To configure the controller, click on “MPC1”. This interface is the most important when it

comes to design controllers and set all of its parameters; the toolbox provides default values

for prediction and control horizons as well as for sampling time, however in most of the cases
10
they are not appropriate for a first guess when designing a controller. The recommendation is

to set all the parameters to one except for the sampling time, which must be equal to integration

step size, and that may differ between each system.

3.1.2. Constraints tab

In the “Constraints” tab it is possible to define maximum and minimum values for each

one of the MV’s and CV’s, including their maximum down and up rates. This is a very important

component when designing a controller for the MPC approach is able to deal with process

constraints in an optimal fashion, so one may feel free to test and evaluate closed-loop

responses with different constraints.

11
 3.1.3. Weight tuning tab

The “Weight Tuning” tab provides an interface for one to choose values for input and

output weights according to their desired closed-loop response. If the system is multivariable,

it is possible to set a tight control routine over a certain CV by tuning output weights, or if the

controller is being too aggressive or too sluggish when dealing with MV’s, tuning the input

weights may turn to be a good option. Moreover, this tab provides a quick way to choose

between a more robust or faster response by managing the “Overall weight” (upper slide bar);

one can designate values between 0 (more robust response) and 1.0 (faster response), being

aware that overall weight’s values close to one may destabilize the system depending on which

values were set to prediction and control horizons.

12
 3.1.4. Scenarios tab

Finally, the tab “Scenarios” provides a quick simulation of your system before exporting

the controller to Simulink, thus being a good way to evaluate the effect of each parameter in

the closed-loop response in a quick and practical form. In this tab, one can set the simulation

length and a variety of types of setpoint changes (step, ramp, impulse, etc) over the CV’s, their

size and time instant in which they are applied. By pressing the green arrow button on the upper

left corner, the toolbox simulates the system and shows two graphs: one for the MVs and

another one for the CVs; the next diagrams illustrate this feature for a step change over CV’s

setpoint with P = 5, M = 1, sample of 1.0 second and a length of 25 seconds,

3.1.5. Exporting a controller

Once the designer is satisfied with their controller, it must be imported to MATLAB

workspace. To do so, click on “Controllers” tab and then on “Export”; it is possible to assign a

name or choose between various controllers that were designed within the toolbox, if that’s the

case.

13
 3.2. Setting up and starting a simulation

In order to introduce the reader to MPC toolbox and its configurations, a first example

which covers a simple first-order transfer function will be explored and the effect of model and

prediction horizons shall be addressed.

3.2.1. Example 1: First Order Process

The first-order process, where the time unit is minutes, is:

1
𝐺𝑝 (𝑠) = 5𝑠+1 (ix)

For its simulation, we will keep a time sample of 1 minute so the effect of other tuning

parameters should become more visible and easy to understand. First, equation (ix) must be

inserted in a Simulink block diagram along with a MPC controller block; in order to do so, one

must type the following command into MATLAB’s command window:

>>simulink

This command will open the Simulink Library Browser; next, click on “File”  “New” 

“Model”. The following window should become visible:

Several simulation parameters are able to be set in order to turn the calculations faster

14
and more accurate; to do so, click on “Simulation”  “Model Configuration Parameters”, then

below “Solver options” choose “Fixed step” as the type of integration step, and then choose

“ode4 (Runge-Kutta)” as a solver for the simulation. Also, for this case a fixed-step size of 1

can provide a fair accuracy upon the solution of equation (ix).

Once the basic simulation settings are defined, a block diagram can be built. For this

example just a few blocks are necessary, which are listed below:

 Step block

 Transfer function block

 Scope block

 MPC controller block

15
The non-controlled open-loop block diagram is showed below:

To set the coefficients within transfer function block, double click on it and specify values

for numerator and denominator (zeros and poles):

Finally, after all these steps the simulation is ready to be run. To do so, click on the green

circle button on the upper bar. Once the simulation is done, the results for the non-conrolled

16
 open-loop response due to a unitary step change over the MV are as follows:

The next step is to introduce a MPC controller block in the diagram. The original diagram

must be altered in a couple of ways to well insert the controller; the closed-loop block

diagram is demonstrated below (the setpoint for the system will be denoted by zero):

To configure the controller, double click on the MPC controller box (uncheck the

“Measured disturbance” box first) and the click on “Design”. A small window will show up asking

the number of manipulated and measured variables, as well the sample time for the simulation;

17
because this is just a 1x1 system, leave the default values for manipulated and measured

variables, however for the sample time choose a value of 1.0 (same value as the integration

step size).

After clicking OK, the MPC block will linearize the system and compute operating points,

this phase can take a few seconds or a couple of minutes depending on how complex is the

system. Once the linearization is done, the following window should appear:

This window features a diagram providing the number of MVs, measured outputs,

measured disturbances (for a feedforward approach), among others. One can also define

nominal values for input and output signals and its properties, such as name, description and

units; this first window is basically to check whether the provided settings in the block diagram

are correct or not, and primarily characterize the system. After feeling done with this tab, one

can advance to the controller parameters.

18
 Click on “Controllers” tab and then on “MPC1”; the following window should become

visible:

Here one can set values for basic parameters within a MPC controller, such as prediction

and control horizons. The control designer is also able to define constraints for manipulated

and CVs besides input and output weights; it can be affirmed that all the turning work is done

in this tab, hence one should feel encouraged to further explore the effect of such parameters

when working with any control design and simulation.

19
 3.2.2. First Order Process: Effect of Prediction horizon P

In order to explore the effect of prediction horizon in this example, let’s set the sample

time as 1.0 (same as the integration step size) and a control horizon of 1 move. For a first

simulation the value of 1.0 time unit for the prediction horizon is suggested, so the controller

will predict the behaviour of the system along one sample time and perform the control move

calculation in the same step.

After setting values above, the controller must be exported to the workspace; click on

the “Controller” tab and then on “Export”. One can provide a name for the controller if necessary,

then click again on “Export”.

To perform a unitary step change over the CV, double click on the “Step” block and

choose its final value as 1.0, leaving the initial value as zero; also, you can set the time instant

in which this step change will be applied on the “Step time” bar. After performing these

procedures, click “Ok”. The results from this first controller design for such unitary setpoint

change are as follows:

20
 Essentially, the closed-loop response can be improved by increasing the prediction

horizon; by changing the value of P to 3, the following response is obtained (after changing any

parameter within the “Controller” tab, remember to export the controller again to the workspace

for there is not a link between MPC toolbox and the workspace to provide a simultaneous

update of such values):

21
For a prediction horizon of 5:

And finally for a prediction horizon of 10:

22
 It can be noticed that, as the prediction horizon is increased, the magnitude of the control

moves becomes smaller due to the optimization routine which is embedded in the controller;

moreover, the system is smoothly tracked towards its setpoint with less variation. The reader

should feel encouraged to try more substantial values for P and check the effect over both MV

and CV, as well for system stability.

3.2.3. First Order Process: Effect of Control horizon M

The outcome of varying the prediction horizon in a MPC simulation was briefly addressed

above and the reader should feel introduced to how to start designing this type of controller.

The control objective is to calculate a set of control moves (MV changes) which make the

corrected predictions as close to a reference trajectory as possible. The next step is to evaluate

the effect of the control horizon M, which is related to the number of control moves calculated

per each sample time.

For this specific example, after trying few different values for M, the reader should notice

that the control horizon has a tiny effect over this type of system, and distinct values of M lead

to practically the same closed-loop response. The following figures show the results for a

controller with prediction horizon of 5 and control horizons of 2 and 4, respectively.

23
 In some cases, the prediction horizon has a more significant effect in the closed-loop

response performance than the control horizon; for more complex systems, such as

multivariable ones, both P and M are generally equally relevant when designing a controller,

thus one should feel inspired to spend more time on trying to find a successful combination of

these parameters to obtain a satisfactory control performance.

3.3. General instructions to implement MPC

Two basic properties of any system are not independent: model-length and sample-time.

The model length must be chosen as approximately the process settling time, which is time

required for the system to reach a new steady state after a step input change. Generally, one

should set the sample time as one tenth the dominant time constant, so the model length is

roughly the settling time of the process.

Prediction and control horizons also differ in length; for most of the cases, prediction

horizon is selected to be greater in value than the control horizon. This is notably true if the

there is no control weighting over inputs or outputs. Usually, if the prediction horizon is longer

than the control horizon, the control system is less sensitive to model error. As the control

horizon becomes more substantial, the control moves tend to become more aggressive and a

smaller weight on the input is needed to soften the control moves.

The following procedures summarize the steps involved in implementing MPC on any

process:

 Develop a discrete step response model with length 𝑁 and sample time ∆𝑡;

 Specify the prediction and control horizons in such a way that: 𝑁 ≥ 𝑃 ≥ 𝑀;

 If a dominant time delay exists in the process, the prediction horizon must be longer

or equal to the time delay, otherwise the close-loop response may become unstable;

24
  Specify weighting 𝑤 on the control action;

 Convert all variables to deviation form.

4. Case Study Problems

4.1. Case study #1: Single variable control – Continuous Heat Transfer System

(CHTS)

Consider the Continuous Heat Transfer System composed of one shell-tube heat

exchanger. Hot glycol circulates through the shell (which is assumed to be perfectly mixed),

and then heat flow between hot media and tube increases the energy amount inside the tube,

in which flows water. The following scheme depicts such heat transfer unit:

TIC4 (glycol temperature setpoint)

Hot glycol stream
TE21
TIC-36 (outlet water temperature)
(valve)

CHTS
Hot/cold
Water stream water stream

Figure 2 - Continuous Heat Transfer System

The objective is to control outlet water temperature (TE21) by manipulating water flow

rate through TIC36 valve opening, thus consisting on a 1x1 system; furthermore, both servo

(outlet water temperature setpoint change) and regulation problems (disturbance rejection) are

able to be performed using MPC. In sum, the variables in a CHTS unit are the following:

 Manipulated variable: TIC36 (valve);

 Controlled variable: TE21 (outlet water temperature);

25
  Disturbance variable: TIC4 (glycol temperature setpoint).

The transfer function model which relates MV and CV is a first-order plus delay type:

𝑇𝐸21(𝑠) 0.7105
= 𝑒 −5.64𝑠 (x)
𝑇𝐼𝐶36(𝑠) 5.966𝑠+1

And the one relating DV and CV is third-order plus delay and numerator dynamics type:

𝑇𝐸21(𝑠) 2.516𝑠+0.2605
= 𝑒 −24𝑠 (xi)
𝑇𝐼𝐶4(𝑠) 3145𝑠3 +426.6𝑠2 +62.4𝑠+1

Steady state values for each variable are listed below in Table 1:

Table 1 – Steady state values for CHTS

TIC36 (percentage of closure) TIC4 (°C) TE21 (°C)

75.02 60.0 16.45

4.1.1. CHTS – servo problem (changes on TE21 setpoint)

This topic will discuss how to implement equation (x) in a Simulink block diagram in order

to simulate step changes over the outlet water temperature (TE21); later on, equation (xi) will

be also included to simulate disturbances in the system as a setpoint change over the hot glycol

stream. To introduce the plant in Simulink, only two blocks will be necessary:

 Transfer function block

 Transport delay block

To create a new Model in Simulink, review the steps on topic 3.2.1. and do not forget to

set the simulation parameters by using the “Simulation” menu. For this case, we will use a fixed

step integration step-size of 1.0 second and ode4 (Runge-Kutta) solver; one is welcomed to

test other step-size values or solvers, however let us stick with these configuration for the time

being. After creating a blank model, insert the blocks previously indicated and connect then

26
with a straight arrow. Double click on “Transfer Fcn” block and set poles and zeros as follows

and click “Apply”:

This will properly define equation (x); the next step is to set the time delay. To do so,

double click on “Transport delay” block and in the Time delay bar, set its value as 5.64, leaving

the other setting as their default values. Finally, click “Apply” and then “Ok”.

27
 The diagram should look like the following:

Next step is to connect the MPC block and start to design a controller. The final diagram

which will be used for this example is demonstrated below:

28
 To designate signal names, right-click on the arrow which corresponds to a signal that

you want to name and then click on “Properties”; on the “Signal name” bar, type the desired

name and then click “Apply” and “Ok”.

Before starting to design the controller, it is important to reassure that the system is

primarily in its steady state; double click on “TE21 setpoint” step block and set both initial and

final values as zero, and the Step time as 100. Double click on the “MPC Controller” block and

then click on “Design”. A small window will show up asking the number of MVs and CVs, as

well as the sample time; leave the default values of 1 and then click “Ok”.

The following window should become visible:

We are ready to start our controller design. First, we have to set prediction and control

horizons as one and check if a stable close-loop response is obtained. Click on “Controllers”

tab on the left-side menu and then on “MPC1”. On “Model and Horizons” tab, leave the Control

interval as 1.0 and set prediction and control horizons as 1; also, click on “Constraints” and set

the minimum value as -75 and the maximum as 25 (remember that all the variables within MPC

29
are in deviation form, thus -75 corresponds to zero opening and 25 corresponds to total

opening). After all these configurations are done, export the controller to MATLAB’s workspace

and run the simulation with a TE21 setpoint change of 2°C.

The following closed-loop response is obtained:

It is clear that the system became unstable with such parameters by being too fast, still

this situation can be easily fixed by looking at the Overall Weight; on the “Weight tuning” tab

within “MPC1”, the default value for Overall Weight is 0.8 which coincides with a faster

response. To turn the system slower and more robust, set its value to 0.4 (or lower values) and

export the controller. The response for this case should be similar to the following:

30
 By a simple change over the speed of the system, it was able to obtain a stable yet

sluggish closed-loop response.

4.1.2. CHTS – Effect of prediction and control horizons on servo problem

It is possible to turn the system response faster by adjusting accordingly the value for

prediction horizon. Let us try several different values and choose one which fits a middle ground

between robustness and speed. First, try a prediction horizon of 5 seconds by going to “MPC1”

 “Model and horizons” tab; after exporting the controller again and running the system, the

response should look similar to the next:

31
 According to the previous graph, a better response for a TE21 setpoint change was

obtained featuring a settling time of around 120 seconds and no oscillations over both MV and

CV; it is possible to obtain a faster response yet oscillatory when the prediction horizon is set

as 10 seconds or huger values:

and for a prediction horizon of 15 seconds:

32
 Again, one is encouraged to try different values for the prediction horizon and further

explore system’s stability by increasing P.

In the same way as example 1, the control horizon M has a very slight effect in the

system, thus being the prediction horizon the main parameter in this case. Of course, the “Rate

weight” over TIC36 in the “Weight tuning” tab has a significant relevance when tuning the

controller for a certain desired response, and one is welcomed to explore the effect of such

parameter by increasing (slower response) and decreasing (faster response) its value. The

effect of weight tuning will be better explored when simulating a multivariable system.

4.1.3. CHTS – regulation problem (TIC4 disturbance rejection)

In order to include a disturbance model into CHTS, it is necessary to consider equation

(xi) during the simulation, which is the transfer function model relating TE21 (outlet temperature)

to TIC4 setpoint (hot glycol temperature). Usually, glycol temperature is kept constant during

the process by another heat transfer apparatus which is connected to the heat exchanger shell

33
outlet, however in case of failure in this apparatus the glycol temperature may be altered,

leading the system to a new steady state. This topic will simulate the disturbance rejection in

the system as a step change over TIC4, again exploring the effects of prediction and control

horizons.

To insert equation (xi) add another Transfer Fcn, Transport Delay and Steps blocks into

the diagram as demonstrated below:

To properly define equation (xi) in Transfer Fcn block along with the Transport Delay,

double click on such blocks and define its coefficients as showed below:

34
 The final diagram including the disturbance model has the following appearance:

Again, prior to designing the controller, it is important to reassure that the system is

primarily in its steady state; double click on “TE21’ setpoint” and “TIC4’ setpoint” step blocks

and set both initial and final values as zero, and the Step time as 0 and 100, respectively.

Double click on the “MPC Controller” block and then click on “Design”. A small window will show

up asking the number of MV’s and CV’s, as well as the sample time; leave the default values

of 1 and click “Ok”.

35
 As for the servo problem, the configuration listed below must be set before exporting the

controller to MATLAB workspace:

 On “Model and Horizons” tab, leave the Control interval as 1.0 and set prediction

and control horizons as 1 move;

 Click on “Constraints” and set the minimum value as -75 and the maximum as 25

for the MV;

 Go for “Weight tuning” tab and slide the Overall Weight bar until a value of 0.4 is

reached.

Once these steps are done, export the controller and run the Simulink diagram; the

close-loop response for a setpoint change over TIC4 of 5°C should look similar to the following:

36
 4.1.4. CHTS – Effect of prediction and control horizons on regulation problem

The next figure depicts the influence of prediction horizon when rejecting a disturbance

in the system (hot glycol temperature setpoint); values within a range of 50 seconds were tested

and results are showed below:

To apply changes in any parameter when designing a controller, make sure to re-export

the controller to MATLAB’s workspace so that Simulink will be able to update the controller.

It is clear that a prediction horizon greater than 20 seconds undergoes minimal

improvements to TE21 setpoint tracking; this sort of behavior is common in MPC controllers

both to prediction and control horizons, however such “steady” values varies from system to

system and one should look for these values in order to assess sensibility of a plant regarding

predictive parameters.

Let’s choose a value of 20 seconds for the prediction horizon and start to explore the

role of the control horizon when a disturbance input are introduced in CHTS. The next diagram

37
demonstrates control horizon’s influence when a +5°C step change in TIC4 setpoint is detected;

values within a range of 15 control moves were tested (remember to set 𝑀 ≤ 𝑃):

4.1.5. CHTS – Effect of input rate weight on regulation problem

As previously discussed in topic 3.3, increasing the control horizon turns the controller

to be more aggressive, thus the control moves assume more considerable magnitudes when

dealing with MV’s, undergoing considerable overshoots and undershoots; a simple way to

compensate such variations is to introduce a smaller weight on the input by going to “Weight

tuning” tab and try rate weight values less than 0.1. For instance, you may choose a control

horizon of 2 moves and test rate weights of from 0.2 to 0.01 and check for oscillations in the

closed-loop response as well as its stability. The following figure shows such sensitivity

analysis:

38
 As can be noticed that, as the rate weight is decreased, the control moves (and so the

controlled variable behaviour) have a role of diminishing overshoots, yet the system becomes

more oscillatory until the point of turning completely unstable, which is case for input rate weight

values less than 0.02.

4.1.6. CHTS – adding a feedforward controller routine

The MPC block features a feedforward controller add-on which measures important

disturbance variables and take corrective action before they lead the process to a new steady

state. To make an effective use of feedforward control, an approximate process model should

be available, in this case such model is equation (ix). To activate feedforward control, go back

to the original Simulink diagram and double on the MPC block; the following window should

become visible:

39
 Check the “Measured disturbance” box and connect TIC4 signal to a new port which

becomes available in MPC block:

Then start to design the controller by clicking on “Design”; again start with prediction and

horizon values as 1 to check system stability; remember to include constraints (-75 and +25) in

the manipulated variable and set the overall weight 0.4. Also check system specification

through this window:

40
 The control designer is encouraged to explore the effect of prediction and control

horizons over a MPC/feedforward controller by introducing several distinct values for these

parameters, as well as the influence of different input rate weights in the system. The final

designing task is to properly tune all these parameters inside a MPC controller to obtain a

specific closed-loop response within a range determined by input/output constraints.

41
 4.2. Case study #2: Multivariable control – Evaporator system

The objective of this case study is to simulate an evaporator plant and develop a MPC

controller for a multivariable system, which consists of an evaporator unit connected to a gas-

liquid separator. The system consists of 3 inputs and 3 outputs, defined as follows:

 Manipulated variables:

o Product flow rate (F2);

o Steam pressure (P100) – MV#1;

o Cooling water flow rate (F200) – MV#2;

 Main disturbance variables:

o Feed flow rate (F1) – DV#1;

o Feed composition (X1) – DV#2;

o Feed temperature (T1) – DV#3;

 Controlled variables:

o Separator level (L2).

o Product composition (X2) – CV#1;

o Operating pressure (P2) – CV#2;

The control performance can be evaluated by considering positive and negative

disturbances on these three main disturbance variables and checking if the manipulated

variables were able to keep the desired outputs by managing their values within a reasonable

range.

42
 Figure 3 - Simplified process flow diagram of evaporator system

The steady state information for evaporator system is given in Table 2:

Table 2 – Evaporator system steady state values

X2 (%) P2 (kPa) P100 (kPa) F200 (kg/min) F1 (kg/min) X1 (%) T1 (°C)

25.0 50.5053 194.7 208.0 10.0 5.0 40.0

It will be considered that F200 steady state value corresponds approximately to a linear

half-opened valve and P100 may vary within a range of more or less its steady state value; thus

the constraints for the MV’s are +208/-208 kg/min and +194.7/-194.7 kPa, respectively.

In order to simplify the control problem, it will be assumed that the liquid level inside the

evaporator unit is controlled and its value is kept constant through time; as a matter of fact, the

transfer functions for this system were obtained with data collected when a level controller (PI-

type) was on automatic. Futhermore, the relative gain array for the process indicated that it is

more feasible to control Product Composition (X2) by manipulating Steam Pressure (P2), and

43
to control Operating Pressure (P2) by manipulating Cooling Water Flow Rate (F200) in the

condenser.

System’s dynamics is described by transfer function models below:

𝑋2′(𝑠) −0.0001041∗𝑠4 −0.005142∗𝑠3 −0.1022∗𝑠2 +0.0293∗𝑠+0.0004008

= 𝑒 −4∗𝑠 (xii)
𝑃100′(𝑠) 𝑠5 +0.6715∗𝑠4 +987.1∗𝑠3 +149.2∗𝑠2 +16.89∗𝑠+0.3213

𝑃2′(𝑠) −0.04646
= 11.51∗𝑠2 +32.38∗𝑠+1 𝑒 −0.5∗𝑠 (xiii)
𝐹200′(𝑠)

𝑋2′(𝑠) 0.2111∗𝑠−0.04504
= 47.73∗𝑠2 +13.31∗𝑠+1 𝑒 −3∗𝑠 (xiv)
𝐹1′(𝑠)

𝑋2′(𝑠) 213.9∗𝑠+3.681
= 236∗𝑠2 +80.62∗𝑠+1 𝑒 −3.74∗𝑠 (xv)
𝑋1′(𝑠)

𝑋2′(𝑠) 0.04059∗𝑠+0.0005693
= 2014𝑠3 +378.3∗𝑠2 +47.09∗𝑠+1 𝑒 −4.78∗𝑠 (xvi)
𝑇1′(𝑠)

𝑃2′(𝑠) 1.996
= 209.7∗𝑠2 +28.97∗𝑠+1 𝑒 −4∗𝑠 (xvii)
𝐹1′(𝑠)

𝑃2′(𝑠) −133.8
= 33.81∗𝑠+1 (xviii)
𝑋1′(𝑠)

𝑃2′(𝑠) 0.609∗𝑠+0.06062
= 552.6∗𝑠3 +263.3∗𝑠2 +54.92∗𝑠+1 (xix)
𝑇1′(𝑠)

It is important to emphasize the nonlinear behaviour which is embedded in such

equations, or in other words, the gain of each variable is different according to the magnitude

of a MV step change. Also, the system may become unstable when distinct step changes are

applied in DV’s and CV’s setpoints; for instance, a controller which has a decent performance

when dealing with positive changes over F1 (feed flow rate) may not be suitable for a negative

disturbances over the same variable, or even a controller which does a fine job for regulation

problems may not be satisfactory when dealing with a servo problem.

The main design task is to design and deploy a controller which well corresponds to all

disturbance types and regulation/servo problems; such demand may show up as a challenging

and intriguing work when using linear MPC, which is the type of controller explored in this text.

The focus for this control problem is to keep product composition (X2) as close as

44
possible to its setpoint due to product quality, as well as trying to reduce its settling time (turning

the system faster), with operating pressure showing a larger variance yet still being tracked

towards its setpoint. Simulation length now is 500 minutes and the fixed step-size for ode4

solver is 0.1 minutes.

A nonlinear mathematical model has been derived in the literature to describe this

process and a Simulink model has been developed from this mathematical model . Because

such modeling is quite complex to implement in Simulink, a simplified block diagram is

demonstrated below in order to show how to connect MPC block to a multivariable plant. The

following block diagram will be used for this case study:

45
 Nevertheless, one is suggested to implement equations (xii) to (xix) as schemed below

and run the simulations:

For this configuration, remember to configure all step blocks for both initial and final

values as zero before starting to design the controller. Double click on “MPC Controller” block

and click “Design”; a small window will show up asking the number of MV and CV’s, as well as

the sample time. Set 2 manipulated variables, 2 controlled variables and 0.1 minutes,

respectively. Then click “Ok” and wait for the toolbox to load.

46
 4.2.1. Evaporator system: effect of prediction and control horizons on servo

problem

Essentially the system traits a slow dynamics for product composition X2 but a faster

one for operating pressure P2 when a MPC controller is on automatic, thus initial values of 1.0

for both prediction and control horizons are way from being satisfactory when setpoint changes

are applied over these CV’s, especially due to a predominant delay or dead time of 4.0 minutes

(check equation xii); the next graph provides a sensitivity analysis for prediction horizon when

+10% step changes are applied in X2 and P2 setpoints:

P-values less than the system time delay (4.0 minutes) destabilize both product

composition and operating pressure; moreover, it can be noticed that a setpoint change on X2

minimally affected operating pressure. Also, the prediction horizon was unable to provide a

decent X2 setpoint tracking (offset is present), however it did a fair job when tracking P2 setpoint

for P-values above 5.0 minutes. Let us choose a value of 15 minutes for P and start analyzing

control horizon’s role in this model.

47
 Operating pressure is still stable and well controlled yet product composition still features

an offset, huge deviation from its setpoint and a lengthy settling time. An acceptable option

would be a value of 5 to the control horizon, so let us choose this value and try to improve

control performance of product composition by analyzing how input rate and output weights can

boost X2 setpoint tracking.

4.2.2. Evaporator system: effect of input rate and output weights on servo

problem

When prediction and control horizons are unable to provide a decent closed-loop

response by themselves, one should start to deal with input and output weights. By decreasing

or increasing their values, it is possible to accelerate or slow down a CV without substantially

affect others; essentially, by decreasing an input rate weight a certain CV can be accelerated,

with the opposite being valid, and by increasing output weights the control performance of a

single CV can be put in first place, with the opposite again valid. A sensitivity analysis for these

parameters can be performed on the following tab:

48
 Let us first deal with input rate weights and try to improve product composition control

performance by increasing its closed-loop response speed; this task can be performed by

decreasing its input rate weight (P100’) from the default value of 0.1.

Product composition control was improved by managing P100 rate weight however still

presents an offset; such situation can be solved by dealing with the remaining MV rate weight:

49
cooling water flow rate F200. For its sensitivity analysis, let us choose a P100 rate weight value

of 0.08 and try F200 rate weights over and below the default value of 0.1.

By managing new values for F200 rate weights, it became possible to achieve a stable

and fair closed-loop response for both CVs. Nevertheless, operating pressure seems to present

an excessively fast response which is not necessary; to solve this situation and improve even

more product composition control performance, let us keep a value of 0.01 for F200 rate weight

and we can start to deal with output weights over X2 and P2.

50
 Values less than 1.0 for X2 weight led to offsets and poor performance for its closed-

loop response; in contrast, values above 1.0 provided a faster setpoint tracking actions even

though turning the system oscillatory in face of a P2 setpoint change. Now let us choose an X2

weight value of 2.0 take a quick look on how we can improve composition control by trying

different values for P2 weight.

51
 Certainly the optimum value for P2 weight for this case is 0.005 due slight variations in

product composition for either a step change on its setpoint or rejecting/minimizing deviances

originated from P2 variations, no offsets in both variables and a smooth transition to P2 setpoint.

Finally, the MPC parameters for this controller are listed in Table 3:

Table 3 – MPC parameters for servo problem

Prediction Control Sample P100 F200 X2 P2

horizon horizon timing (min) rate rate weight weight

(min) (# of moves) weight weight

15.0 5 0.1 0.08 0.01 2.0 0.005

4.2.3. Regulation problem on evaporator system (disturbance rejection)

The previously designed controller well succeeded when dealing with setpoints changes

over product composition and operating pressure, however this is just half part of the whole

control problem. The second half consists basically of testing control performance face a

disturbance, which for this case are based on fluctuations in feed flow rate, feed composition

and feed temperature, and redesign its parameters if a poor disturbance rejection performance

is obtained.

First, we will apply setpoint changes in all DVs and check controller performance, then

a sensitivity analysis will be performed among prediction and control horizons to verify whether

disturbance rejection can be improved or not. Finally, we will proceed with a sensitivity analysis

involving input rate weights and output weights, and testing the new parameters for the servo

problem.

52
 4.2.4. Evaporator system: effect of prediction and control horizon on

regulation problem

Simulations performed with the previously designed controller (Table 3 values) for +10%

setpoint changes over F1, X1 and T1 provided the following results:

53
 These are fair closed-loop responses for disturbance rejection of all DVs, featuring

acceptable over/undershoots and relatively small settling times, still only a sensitivity analysis

will enlighten the designer to a sense of whether such responses can be enhanced or not. For

the next analysis, we will stick with just product composition control which is the focus for this

case study, and explore the role of prediction and control horizon in a regulation problem (all

the other parameters will be kept as their previous values except for the control horizon, which

must be inferior to the prediction horizon). The following graphs show such sensitivity analysis

for P:

54
 A value of 10 minutes for the prediction horizon seems to be enough to achieve a

satisfactory closed-loop response for product composition, thus let us keep this value and check

for the control horizon influence when rejecting disturbances:

55
 The control horizon did not seem to provide any sort of improvement in product

composition control since all the curves are superposed; these results are due to input and

output weights being already tuned, therefore the closed-loop response shapes are already

56
defined by such values regardless the control horizon magnitude. One is welcomed to test this

hypothesis in the servo problem.

4.2.5. Evaporator system: effect of input rate and output weights on regulation

problem

Once the sensitivity analysis for prediction and control horizons are accomplished, we

can start to evaluate input and output weights role in a disturbance rejection routine. Let us

choose a value of 5 control moves and vary P100 rate weight:

57
 Let us choose a value of 0.005 for P100 rate weight (yellow line) and look for F200 rate

weight effect in X2 close-loop response:

58
 Let us keep a value of 0.001 (yellow line) for F200 rate weight due to its improvement

over settling time, turning the system faster yet oscillatory. The last sensitivity analysis will

investigate output weights part in the whole controller design; again, keep in mind that P2 weight

is 0.005 (servo problem design) and let us start with X2 weight:

59
 By increasing X2 weight it became possible to reduce over/undershoots when dealing

with all disturbance inputs without destabilizing the control loop; let us choose a new value for

product composition weight of 4.0 and diagnose operating pressure weight influence:

60
 Both values of 0.1 or 0.05 for P2 weight seemed to reduce oscillations in X2 closed-loop

response when rejecting variations in feed flow rate and feed composition, the other values

provided a sluggish and oscillatory responses, demonstrating a poor control performance.

61
 Ultimately, the MPC parameters for this controller are listed in Table 4:

Table 4 – MPC parameters for regulation problem

Prediction Control Sample P100 F200 X2 P2

horizon horizon timing (min) rate rate weight weight

(min) (# of moves) weight weight

10.0 5 0.1 0.005 0.001 4.0 0.10

4.2.6. Evaporator system: controller validation final test

A final test to properly validate and deploy the controller designed for regulation problem

would be to apply positive and negative step changes over all DVs setpoints and check for

system stability, as well for X2 and P2 setpoint changes. The next closed-loop responses were

obtained using parameter values listed in Table 4:

A quick comparison between these results and the ones depicted in topic’s 4.2.4. first

graph induces the control designer to conclude that there were no improvements among any

62
disturbance rejection, thus the parameters utilized in servo problem are satisfactory for both

control problems. Nevertheless, it is important to check for any enhancements in servo problem

performance with regulation problem parameters:

Certainly, the last graphs corroborated the conclusion drown from closed-loop responses

in regulation problem using Table 4 parameters: the first tuning action (Table 3 parameters)

stated satisfactory results for both servo and regulation problem. This is quite remarkable for

such a nonlinear system as an evaporation assembly, demonstrating MPC controller’s power

and capability to deal with linear and nonlinear constrained multivariable systems.

63
 5. References

Bequette, B. (2003). Process control: Modeling, design, and simulation. Upper Saddle River,

N.J.: Prentice Hall PTR.

Seborg, D. (2011). Process dynamics and control: International student version (3.rd ed.).

Hoboken, NJ: John Wiley and Sons.

Morari, M., & Zafiriou, E. (1989). Robust process control. Englewood Cliffs, N.J.: Prentice Hall.

Prett, D., & A, C. (1988). Fundamental process control. Boston: Butterworths.

64

View publication stats