European Symposium

European Symposium onon Computer

Computer Arded
Aided Process
Process Engineering
Engineering –– 15
15
L. Puigjaner
L. Puigjaner and
and A.
A. Espuña
Espuña (Editors)
(Editors)
© 2005
© 2005 Elsevier
Elsevier Science
Science B.V.
B.V. All
All rights
rights reserved.
reserved.

First Principles Model Based Control

Manuel Rodrígueza*, David Péreza
a
Universidad Politécnica de Madrid
José Gutiérrez Abascal, 2. Madrid 28006, Spain

Abstract
Model Based Control is an important and widely used (mainly MPC) technique. This
paper provides a new control architecture based on the use of physical models. In this
preliminary work the architecture is applied for unconstrained multivariable control. It
has been applied on several standard process units obtaining encouraging results. It has
some advantages over MPC as it can use non linear rigorous models and it doesn’t need
any identification step.
Keywords: Model Based Control, Model Predictive Control, Internal Model Control.

1. Introduction
The interest in Model Predictive Control (MPC) started to increase after the presentation
of IDCOM (Identification and Command) (Richalet, 1978) and DMC (Dynamic Matrix
Control) (Cutler, 1979). After 25 years MPC has become a widely used technology in
process control. Nowadays, a new crude distillation unit in a refinery is not conceived
with other control scheme but MPC (and the same happens in many other processes).
The technology applied is usually based on a previous identification step (which is of
the most importance) to get a linear model of the unit and, then, on an implementation
step (usually more simple and less time consuming). Although a lot of research has been
done regarding Nonlinear MPC using different approaches, differential equations,
neural nets (Temeng, 1995), Hammerstein models (Fruzzetti, 1997), Volterra equations
(Maner, 1996), fuzzy models (Sousa, 1997) … it is still an open area where many
problems arise.

The purpose of this work is the use of first principles models for model predictive
control. To achieve this goal, a new architecture has been developed and tested on some
simulated process units.

The remaining paper is organised as follows: Section two describes in detail the new
architecture, how it works and its components, and explains the software
implementation. Section three shows the results obtained when it is applied to control
some operation units. Finally, section four presents the conclusions and future steps of
this work.

*
Author to whom correspondence should be addressed: mrod@diquima.upm.es
2. Physical Model Based Control Architecture.
The PMBC architecture is composed of two modules: the Physical Model Based (PMB)
Controller Module and the PMB Model Module. Figure 1 shows the proposed
architecture:

Figure 1. PMBC Architecture

The models used in the components of the architecture have been developed with some
degree of error with respect to the process, as it is impossible to get a perfect model of a
real process. To this purpose, some parameters have been slightly changed.
2.1 PMB Controller Module
This module is the core of the architecture, it predicts the values of the manipulated
variables that provide the desired performance (set point change or disturbance
rejection).
This module has a model of the unit to control. It has a first principles model comprised
of a set of DAEs. This module has the following variables as inputs:
x The error of the controlled variables, it can be referred to the controlled variable
directly or to its derivative. It is the desired response or the target curve.
x Measured disturbances.
The outputs of this module are the set of manipulated variables (which are commonly
inputs in a model of the process).
If the controlled variable is at its set point, the input of the module will be the derivative
set to zero or the value set to that constant (so it remains in that value). If a disturbance
happens, the controller will produce an output to compensate that disturbance and keep
the controlled variable at the desired set point value. If a set point change is desired, the
input will be an exponential (softened step) in the controlled variable. If the controlled
variable acts like an integrator, such as controlling the level with an output stream, a
ramp has to be used instead of a step. The time used for the step or ramp is an adjustable
parameter.
The time scale of this module is the simulation execution time of the model in the
selected platform. It is an adjustable parameter, it can be slowed down to avoid
excessive control actions that sometimes will provide a worse system performance.
2.2 PMB Corrector Module
This module has a model of the unit to be controlled. The model is exactly the same as
above except for which are the input variables to it. This model is used in the standard
way, so the inputs to it are the disturbances and the inputs to the real process. The
outputs of the model are the controlled and output variables of the process.
The purpose of this module is to correct, in some way, the prediction made by the
Controller Module. As no model is perfect, the controller output will lead to a control
action that doesn’t set the controlled variable to the desired value. This means that some
feedback corrections are necessary. In order to accelerate this procedure, this module
compares its output with the actual process output and sets a correction factor to be
applied to the control action. So, this module has two components: the model
component and the comparison component. The final output of this module, with the
correction factor, is the control action to be applied to the real process.
The time scale of this module has to be that of the real process, so it has to be
synchronized to it as accurate as possible, in order to be able to compare the same
variables.
2.3. Software implementation
The implementation has three components, the Controller Module, the Corrector
Module and the Process Module. All the components have been implemented in the
same machine, a PC running under Linux OS. The model used in the Controller and
Corrector Modules is not perfect, some changes have been made in several parameters
of them in order to check the proposed architecture.

The models have been developed using the ABACUSS II simulator (Barton,2003). This
software allows to embed the simulation code in other application. Using the C++
programming language, different executables have been generated for every component
of the architecture. The information flow has been implemented through shared memory
procedures.

3. Applications
In this section the results of the application of this architecture are presented. Set point
and load changes are applied to all the tested units and the performance of the controller
is evaluated.

3.1. Stirred tank heater

This unit is a perfectly stirred tank heater with a jacket. It has two input streams and one
output stream. There are level and temperature control loops. The manipulated variables
are the tank input flow and the jacket input flow. These two variables are calculated in
the Controller Module. This module has as inputs the heater temperature and the tank
level. To achieve a good control on the tank, the following control equations are used:
t
-
dH et
= (SPH - PH )× (1)
dt t
SPH: Level set - point, H: level of the model, PH: level of the actual process.
 t
-
dTO et
= K ×(SPT - PT )× (2)
dt t
SPT: Temperature set - point, To: jacket temperature of the model, Pt: temperature of
the actual process.

The jacket temperature is used as the forcing equation of control instead of the tank
temperature because the use of the latter poses a 2 - index problem. Due to this problem,
the gain between the two variables needs to be added to the control equation. In this
model some parameters are changed about 5% with respect to the actual process.
Figure 2 shows the control of the unit in the presence of a flow disturbance and a level
set point change.

Tout Tout

Level
Jacket Flow

Flow in Flow in

Level

Jacket Flow

Flow disturbance

Figure 2. Heater response after a flow disturbance on the left. Heater response after a level
set point change on the right.

In these tests both loops are affected as the temperature changes with a change in any of
the stream flows.
3.2. Continuous Stirred Tank Reactor (CSTR)
This unit has the same equations as the heater but a first order reaction, following the
Arrhenius expression, is added. In this case, the same two loops (level and temperature)
are controlled. But, in this case, the level is controlled with the output stream, which
means that the control action has to be changed. An exponential (step) cannot be used
and a ramp is used instead. The following equation shows the level control action
implemented:
dH SPH - PH
= (3)
dt t

The temperature control equation is the same as in the heater application. Figure 3
shows the performance of the system in the presence of concentration and temperature
disturbances. In this model, the parameter that is changed with respect to the actual
process is the Arrhenius pre-exponential constant in 5%.
T out
T out

Conc. out Conc. out

Conc. disturbance T disturbance

Jacket Flow
Jacket Flow

Figure 3. CSTR response after a concentration disturbance on the left. CSTR response after a
temperature disturbance on the right.

3.3. Distillation Column

This unit is a binary distillation unit. It assumes equimolal overflow, 20 theoretical
trays, simple liquid tray hydraulics and constant relative volatility (ideal mixtures) for
the vapour - liquid equilibrium in every stage. Feed enters above 10th tray (feed stage).
Total overhead condenser and partial reboiler.

The implementation of this unit for a continuous test is not finished at the time of this
paper. The only implementation available is just the first control action to be applied to
the process. The model in this initial implementation is assumed perfect (this means that
just a single control action gets the controlled variable to the desired value). The control
equations used are similar to the ones presented in the previous applications. Next figure
shows that this control architecture can be applied to the distillation column as well.

4. Conclusions
This paper has presented a new architecture for model based control. Regarding the
obtained results, this physical model based control seems to be a usable technology for
 Distillate Flow
Bottoms Flow

Vapor Flow

Reflux Flow

Bottoms composition
Feed composition

Figure 4. Distillation column performance with a composition disturbance and a bottoms

composition set point change.

any process unit. This new approach has some clear advantages over the ones used
currently in the industry. It uses a first-principles model that is suited to any operating
region, so, it is a non-linear MPC. It doesn’t need any identification step which is very
resource consuming for any process unit.
Although the results presented are promising, a lot of work still has to be done to
establish the availability of this technology. First of all, the initial approach taken is for
unconstrained predictive control, additional steps taking into account constraints (like
valve range,…) have to be made. New issues have to be studied as dead-time,
stability,…The next step in this work will be to test this architecture with a real system.
References
Barton, P.I, 2003 http://yoric.mit.edu/abacuss
Cutler, C. R. and Ramaker B. L. ,1979. Dynamic matrix control—a computer control algorithm.
AIChE 86th National Meeting. Houston,TX.
Fruzzetti et al., 1997. Nonlinear control using Hammerstein models. J. Proc. Control, 7, n1, 31-4
Maner et al, 1996. Nonlinear model predictive control of a simulated multivariable
polymerisation reactor using second order Volterra models.
Richalet, J., Rault, A., Testud, J. L., and Papon, J. ,1978. Model predictive heuristic control:
applications to industrial processes. Automatica, 14(5), 413–428.
Sousa et al, 1997. Fuzzy predictive control applied to an air conditioning system. Control
Engineering Practice, 5. n10, 1395-1406.
Temeg et al, 1995. Model predictive control of an industrial packed bed reactor using neural
networks. J. Proc. Control, 5, n1, 19-27

Acknowledgements
The project is sponsored by Repsol-YPF foundation.