Model Predictive Control: Theory and Applications

Professor. Niket S. Kaisare

Department of Chemical Engineering
Indian Institute of Technology Madras
Lecture 1
Course Background: Model Predictive Control

So, welcome to the course on Model Predictive Control Theory and Applications. We
will just go over the course overview for this course.

(Refer Slide Time 00:25)

The first objective, though oftentimes it is not talked in terms of a control system, is
safety. There should not be any explosion or accidents. We would want to avoid as much
unplanned shutdowns as possible. So, there are going to be various systems in place in
the DCS itself, which will do this action of ensuring that the system and the plant, is
operated in a safe manner.

The next set (this is the traditional control we have always talked about) is, one whether
or not your product meets the desired specs and second is whether or not there are going
to be any problems with respect to waste or environmental issues. So, therefore, we need
to minimize the formation of the waste.
So, what do we mean by product specifications? That basically means, for example, if
there is a reactor; there would be certain set points that are present in that reactor. These
would be set points for product compositions but more often than not these are set point
for flow rates temperatures and pressures that are there. So, you will have a control
system that would be able to meet those set points. Have you heard the term set points in
your courses?

Student: Yes.

Professor: So, set points are basically where you want to take your system to. Then there
are disturbances. You would have also read about or done disturbances in your earlier
control course. Disturbances are nothing but changes in various inputs of the system;
those are other types of inputs to the system, where you do not have any kind of hold on.

So, there are certain inputs, for example, flow rates, and so on and so forth, which you
can change; you can increase the flow rate or decrease it. But let us say the flowing fluid
is stored in a tank. In the morning that tank might be at 25-26oC. By the time it comes to
mid-afternoon in Chennai, the temperature would be as high as 35-36oC. So the inlet
temperature of the system is also an input, but it is an input that you do not have any
handle on; that becomes disturbance to the system.

So, you need to get rid of those disturbances and you need to operate the system at their
desired conditions to meet the product specs and to minimize wastage. In addition to that,
the other aims are to minimize energy consumption, to maximize the production rate.
These last two come under the category of economic optimization. So, this comes under
the category of safety, this comes under the category of economic optimization, this
comes under the category of what is traditionally covered in a control course.

As I have said before, where the current state of the art is going towards is to ask whether
or not we can meet the economic requirements along with product specification
requirements. So, that would be, some kind of a merger of the role between different
traditional roles of a control system.
(Refer Slide Time 04:00)

As I said, just to summarize, the control system does multiple tasks; does the task of
safety to prevent catastrophic failure by identifying and isolating fault. Monitoring is
where you are probing the parameters so that you identify if the faults are there, to
identify the problem areas, to analyze how your system is working, whether it is working
in a reasonable manner or not.

Then there is control and optimization; control is to maintain the process within a
desirable operation range and optimization is to meet certain goals, which is
minimization of cost or maximization of profit, minimization of energy consumption or
maximization of productivity. So, these are the main goals that we will be interested in
this course.

Higher level goals are also there, which is planning and scheduling, which is basically
decision making for an entire plant. Let us say we have to plan based on market
requirements. For the next month, we predict that there is a whole lot of ethylene that is
required, because industry is you know really getting into productive mode. Or maybe six
months later because of, let us say, economic downturns and so on, we expect ethylene
production to go down. Or we expect ethylene production to go down because
polyethylene is going down. Why? Because there are regulations on plastic bags.
So, you need to plan how the overall factory or industries are working or how overall,
you know, multiple locations are working and that would come under the umbrella of
planning and scheduling. And these are the various things and all of this comes under the
umbrella called process systems engineering for chemical processes. We are more
concerned with control systems and control systems will fall under this box that is shown
over here.

(Refer Slide Time 06:07)

Model Predictive Control is one of the most popular advanced control strategies in the
industry. It was first started in the process chemical industry. And the main thing about
MPC is the term Model Predictive Control; all these three terms are quite important. It
uses a model to predict the behavior in future. Control strategies such as PID control do
not do that, MPC does this. It uses a model to predict the behavior in the future and based
on that, it determines what would be the best control actions.

So, it is a model based control strategy, it is a predictive control strategy and it is an

optimization based control strategy because it determines the best steps that you are
going to take in the future. It is a feedback control strategy which means that when we go
on to the next step, this overall process is repeated all over again.

(Refer Slide Time 07:10)

The traditional view of control, and this is what you would have seen in your process
control lectures, think of this as a block diagram. You have your Gp, your plant that you
want to control. Plant has inputs, outputs and there are disturbances.

Lot of times, you would have had this disturbances just being added to the output. I am
making it more general by just having this as one more sets of inputs into the plant and
then there are measurements. You compare the measured variable with the set point and
based on the error, you determine the control actions. That is the traditional view of a
control system.
(Refer Slide Time 07:49)

The modern control theory view, or the MPC view of the control system, is decision
making under constraints with feedback. Which means there are certain constraints such
that temperature cannot go below so and so value, temperature should not go above so
and so value. Valve cannot be less that 0 percent open, valve cannot be more than
hundred percent open. Pressure should not go above such and such values. All these are
process constraints or input constraints that are present in the system. So, you want to
make those decisions in presence of these constraints.

So, you may not always be able to measure what you want to control. For example, you
want to control composition from a distillation column, but you can only measure
temperature and pressure. So, you infer the composition from the temperature and
pressure measurements.

The technical term for doing this is an observer; an observer takes the measurements,
uses the model and based on the model, it computes the information state of the system.
Based on the information state and the model that is present in the system, based on an
objective function that you have defined MPC will optimize the control moves. Based on
this optimization of the control moves you will implement a particular control move and
go on into the future.
So, these are the various components of the control system, you have the plant and you
have a model for the plant. So, we will be talking about model. You have an observer,
which is going to convert the measurements and the model into useful information. And
you have the controller which consists of the objective function, an optimizer and a
model. All these components we are going to talk about in the next twelve weeks of the
course.

(Refer Slide Time 09:42)

The popularity of MPC is because it is a computer based control, it parallels the

popularity of controls. As I said it is model based it uses a dynamic model and it uses
predictions of this model for future behavior. It is optimization based as I just said. The
real reason for the popularity of MPC is, unlike a lot of control algorithms, MPC is
naturally a MIMO multi input multi output algorithm.

SISO, is just a special case for MPC; SISO is just a MIMO with one input one output. As
against that in your undergraduate control, you dealt with SISO for a whole lot of time
but now let us say if we had four inputs, four outputs, how do we handle them? But here
we talk about n inputs p outputs naturally, where n equal to 1, p equal to 1 becomes a
SISO system.
And the other thing is ability to handle constraints. It is not forced upon into an MPC it is
inbuilt into the way MPC was built up. So, these are the two most important reasons why
it has become popular.

(Refer Slide Time 11:01)

And we will talk about the idea of receding horizon control. So, I will finish off with this
particular example. Let us say you are driving a car and you are going behind another
vehicle. This vehicle is going at a certain speed, you are going at a certain speed. Based
on these conditions, you determine that you are going to take this red route in order to
overtake this vehicle.

You are considering the fact that this particular vehicle is moving. So, you are
considering all that into the account. So, you already have predictive behavior in your
decision making. You have accelerated your car. Now, let us say that you are actually
going downhill; because you are going downhill this vehicle has started accelerating. If
you do not do feedback, you will take this red trajectory and you might actually collide
with this driver. But because there is a feedback correction you will take this green
trajectory.

So, what is happening is when you are over here, you have decided a red trajectory, when
you move on over here, you will re compute the entire thing all over again. When you
move over here you will re-compute once again everything, when you move over here
you will re-compute once again everything. But each time you are re-computing you are
taking the feedback measurements into action, into your view.

That is what receding horizon means, that you are predicting into the future every time
and that prediction keeps moving as your vehicle is moving and you make all these
corrections. That is one of the key features of MPC algorithm. So, with that, I will come
to end of this overview.