baratti1997.pdf

ELSEVIER PII: SO954-1810(97)00002-2
Artificial Intelligence in Engineering 11 (1997) 40-412
0 1997 Elsevier Science Limited. All rights reserved
Printed in Great Britain
0954-1810/97/317.00
A feedforward control strategy for
distillation columns
Roberto Baratti,” Stefano Corti
Dipartimento di Ingegneria Chimica e Materiali, Universit6 di Cagliari, Piazza D’Armi, I-09123 Cagliari. Italy
Alberta Servida
Dipartimento di Chimica e Chimica Zndustriale, Universitci di Genova. via Dodecaneso 31. I-16146 Genova, Italy
(Received for publication 6 January 1997)
The prime objective of this work is to demonstrate the potential of neural network
modeling for advanced nonlinear control applications. In particular, for the case
of a single composition distillation column, a model-based neural controller
is developed to regulate the composition of the distillate stream. The neural
controller relies on process inversion for the evaluation of the actuator action on
the manipulated variable (reflux flowrate) to maintain the controlled variable
(distillate composition) at the prescribed value.
The performance of the neural controller is assessedand compared with that of
a conventional temperature control loop and of a neural inferential control
structure. The neural controller by far outperforms the other two in terms of the
response speed by which the upsetting loads are compensated. 0 1997 Elsevier
Science Limited.
Key words: model-based control, inferential control, process inversion, neural
networks, distillation columns.
INTRODUCTION
Nowadays, advanced control systems are playing a
major role in plant operations because they allow for
effective plant management. The prime advantage of a
full plant automation stands in the improvement of
plant profitability and productivity that lead to short
term payoffs of the investment required for the imple-
mentation of advanced automation systems.
Typically, advanced control systems rely heavily on
real-time process modeling, and this puts strong
demands on developing effective process models that,
as a prime requirement, have to exhibit real-time
responses. Because in many instances detailed process
modeling is not viable, efforts have been devoted
towards the development of approximate dynamic
models. The models suitable for real-time applications
result from a trade-off between model complexity and
degree of representability of the actual process
dynamics.
*To whom correspondence should be addressed.
405
Approximate process models can be classified in
structured and unstructured models. While the former
are based on first principles, and thus require good
understanding of the process physics, the latter are
based on some sort of black-box modeling and, in
principle, would not require any a priori knowledge of
the process. Neural network modeling represents an
effective framework to develop unstructured models
when relying on an incomplete knowledge of the process
under examination.‘-3 Because of the simplicity of
neural models, they exhibit great potentials in all those
model-based control applications that require real-time
solutions of dynamic process models. The better under-
standing acquired on neural network modeling has
driven its exploitation in many chemical engineering
applications: data reconciliation and rectification, pro-
cess identification and optimization, software sensor
development (inferential measurements), state estima-
tion, fault analysis, multivariable control (for a list of
references cf. the recent review by Stephanopoulos and
Han’). The common belief that effective and efficient
neural models can be developed without any a priori

406 R. Baratti et al.
I
11
3 ;
-I-
/
/
4
Fig. 1. Schematic of a distillation column showing the
controlled and manipulated variables used in a conventional
temperature control loop.
understanding of the investigated process is not fully
correct. This misconception may lead to neural applica-
tions exhibiting poor performance, and this may have
been the reason for preventing the full exploitation of
neural modeling potentials in real-world applications.
Indeed, the selection of the proper inputs and outputs to
be fed to the network represents a critical step when
developing a neural network model, and an adequate
understanding of the process under examination is
demanding.
Recently Baratti et af.415 have demonstrated the
potentials of neural modeling to develop effective infer-
ential control strategies of industrial multicomponent
distillation columns. Figure 1 schematically shows a tray
distillation column with the typical controlled (tempera-
tures on specified trays, the so called pilot temperatures)
and manipulated (reflux, Lo, and boilup, V, flowrates)
variables. Figure 1 refers to a case where the pilot
temperatures are the ones of the 16th and 3rd trays.
The prime aim of this work is to investigate and fully
exploit neural modeling in advanced model-based
control applications of distillation columns. The basic
idea is to construct a neural tool-box that, on the basis
of detected disturbances, is capable of acting properly
on the manipulated variable so as to compensate for the
upsetting loads as soon as they are detected. This action
relies on a process modeling that must provide the
description of the effect of disturbances and mani-
pulated variables on the process performance. The direct
identification of the action on the manipulated variable
so as to meet the specifications requires a process inverse
model. In this work, we discuss the use of neural
modeling to develop an inverse model of a multicom-
ponent distillation column. The neural model provides
the controller action law. The direct action on the
manipulated variable allows the control engineer to
overcome the problems associated with the optimal
tuning of the parameters of conventional feedback con-
trollers. Furthermore, the action of the neural controller
resembles that of a feedforward strategy, and thus it
may be expected to exhibit a good control performance
also in the case of a crude description of the process
inverse model.6
For the case of a single composition control problem
of a distillation column, neural network modeling is
used to develop a model-based neural control strategy
to compensate for upsets in the operating pressure,
feed flowrate, and feed composition so as to hold
constant the content of the key component in the
distillate stream. The neural controller is developed for a
five-component distillation column (propane, butane,
n-pentane, i-pentane and hexane) that is representative
of actual stabilizer units, so commonly encountered in
the refinery industries.
The performance of the neural controller is compared
with that of a conventional temperature control loop
and of a neural inferential controller.
CASE STUDY
The neural controller is developed for a multicomponent
distillation column for which the set of calibration
data are produced making use of a dynamic simulator.
A 18-trays (ideal stages), five component (propane,
hexane, butane, n- and i-pentane) distillation unit is
selected as prototype of a gasoline stabilizer tower. The
column is also equipped with a reboiler and a total
condenser. The feed (saturated liquid) is fed at the
11th stage from the bottom and is constituted of a
mixture of propane (lo%), butane (5%), n-pentane
(15%), i-pentane (30%) and hexane (40%) (the com-
position is given on a mole basis). The control objective
is to maintain as low as possible the content of i-pentane
(i-C,> is the distillate and that of propane (C,) is the
residue when the main upsets are due to variations or
fluctuations in the feed composition, operating pressure
and feed flowrate. The variation in the feed composition
is the result of mixing the primary stream with a
secondary one (slop) whose composition (on mole basis)
is: CX, 2%; c4, 2%; c6, 45%; t&5, 20%; i-C+ 31%.
Typically, the secondary stream accounts for about 5-
15% of the total feed flowrate entering the distillation
unit.
The dynamics of distillation columns can be simulated
by various model approaches, involving different levels
of complexity.7 Here, we are mainly concerned with
addressing the issue of process inversion via neural
modeling, thus we have adopted a simple description of
the distillation unit that relies on the following assump-
tions: each tray is an adiabatic ideal equilibrium stage,
all the involved species have the same latent heat of
vaporization, the vapor and liquid are ideal mixtures,

Feedforward control strategy 407
Table 1. Relevant operating parameters wed for the simulation
of the reference case
Feed flowrate: 0.229kmol/s Reflux flowrate: 0.035kmol/s
Boilup flowrate: 0.063kmol/s Pressure: 821kPa
Lb,: 5s
the vapor and liquid holdups are negligible. From these
assumptions it follows that the vapor and liquid
flowrates are constant along the column but different
in the stripping and enriching sections. Moreover, the
reboiler and the condenser dynamics have been
described as first order processes. Under the prescribed
assumptions, the mass and energy balance equations are
decoupled, and the process model is simply constituted
of the transient liquid mass balances. Within the
framework of this description the tray temperatures
are evaluated by computing the bubble points corre-
sponding to the composition of the liquids on the trays.
The equilibrium vapor pressures are calculated through
the Antoine equation.8 The adopted dynamic model of
the distillation column is rather simplified but, never-
theless, is capable of capturing the most significant
dynamical features of the unit that are mainly related
to the dynamics of the liquid compositions on the
tray. Indeed, these are known to exhibit the largest
characteristic times.
The design specifications are: i-Cs content in the
distillate less than 0.5% (on mole basis) and C3 content
in the residue less than O-1%. The parameters corre-
sponding to the reference case are summarized in
Table 1.
MODEL-BASED NEURAL CONTROLLER
The relative gain array analysis has been used to select
the most proper control strategy, and has indicated the
E (reflux ratio)/V (vapor boilup) strategy as the most
appropriate for the prescribed objectives. In particular,
the distillate prescription should be controlled with V
while E should be preferred to regulate the residue
specification.
The control of the residue specification exhibits rather
serious problems due to the high volatility of propane
that makes the residue C3 content almost insensitive to
E. This results in high gains that lead to serious pitfalls
when trying to implement the neural controller that is
based on process inversion. Overall, the dual compo-
sition control problem exhibits rather ill-conditioned
characteristics so as to prevent the neural controller
from achieving a reasonable performance. Indeed, for
the reference case, the condition number of the steady-
state gain matrix is >>1.
In this work, we use a multilayer feed-forward neural
network trained through the backpropagation algor-
ithm. The adopted network has three layers: the first
layer (input) contains nl nodes corresponding to the
actual net inputs, the second layer (hidden) contains n2
nodes, and the third layer (output) contains n3 nodes
that correspond to the number of the monitored state
variables. The configuration of such a network is simply
indicated as (nl - n2 - q). The feed and hidden layers
are also augmented with a bias unit that represents the
threshold value. The input value of the bias node is held
constant and equal to 1. Two pilot temperatures, located
at the 3rd and 16th tray from the bottom, are also used
as inputs to the neural network. It is worth pointing out
that we have selected as network inputs those that
closely resemble the field measurements available for
actual distillation units. Further details on the develop-
ment of the multilayer neural network, on the selection,
and on the pre-processing of the inputs are given
elsewhere.415
For the dual composition problem, three control
systems based on the El V strategy have been developed:
a conventional temperature control loop, a neural
inferential system, and a neural controller based on
process inversion. The inferential control strategy relies
on a neural observer that provides compositions as
setpoints to a conventional PI feedback controller. The
actions of the inferential and neural controllers are
schematically shown in Fig. 2.
In the neural controller the actuator actions (on V
and E) are modeled as first order filters, with a time
constant typical of the industrial control valves. The
actuator model is given by the following first order filter
equation:
dy
7-&+y=x
Manipulate
Variable Control Variable
* Process ---)
Inferred Control Variable
1
Disturbances
Manipulate
Variable
Neural Controller
Fig. 2. Schematic of the actions of the inferential (a) and neural
(b) controllers.

where r represents the time constant of the actuator
device (control valve) and is given by At,,/a; At_
represents the acquisition interval; y is the filtered signal
(actuation); x is the manipulated signal, that is, the
reflux ratio (E) or the boil-up rate (V) as predicted by
the process inverse model. The parameter Q governs the
response speed of the actuator, in other words, it
determines the rate at which the control action is able to
compensate for the upsetting loads to the unit.
The process model inversion is simply carried out
by inverting the trained neural network exchanging
the manipulated variables (E and V) with the con-
trolled ones (the C3 content in the residue and the
i-C5 content in the distillate). As anticipated earlier,
the dual composition control problem exhibited an
ill-character that has prevented us from efficiently
performing the process inversion, thus, making unsuc-
cessful any attempt to develop a model-based neural
controller.
The problem has been overcome by reconsidering the
control objectives and discovering that the ill-condition
character of the distillation unit is mainly related to the
weak sensitivity of the residue composition (C, content)
on the reflux ratio, E. This suggested releasing the
control specification on the residue by only maintaining
the one on the distillate, thus reducing the control
problem from dual composition to single composition,
Indeed, numerical experiments showed that, for the
investigated column upsets, the specification on the
residue composition was met anyway when controlling
the i-C5 content in the distillate. For the single com-
position control problem, the analysis of the steady-state
gain matrix has led to selection of the reflux flowrate
as a manipulated variable to control the distillate
specification. The neural network used for the single
composition problem has been furtherly simplified to a
(5-l-l) configuration that has proved to be very efficient
for process inversion, even though it was less accurate
than the (7-2-2) network in representing the distillation
unit.
DISCUSSION OF THE RESULTS
For the reference case, the relevant parameters used
for simulating the distillation unit are summarized in
Table 1. All the results discussed for the neural con-
troller refer to the single composition control problem
where the i-C5 content in the distillate (controlled
variable) is regulated by acting on the reflux flowrate
(manipulated variable).
The neural networks used to develop the inferential
control strategy and the neural controller were trained
using sets of calibration data spanning 50 h of operation.
The calibration data were generated by making use of
the dynamic simulator previously described. The simu-
lations have been designed so as to capture the principal
response features of the distillation units to variations in
feed flowrate, feed composition, distillate and residue
specifications, reflux and boil-up flowrates, and operat-
ing pressure.
At first, the performance of the neural controller has
been evaluated with respect to the value of the para-
meter Q, to test the response sensitivity of the neural
controller to changes in the time constant of the
actuator device. Increasing cr, the response speed of
the neural controller increases, at least up to a critical
value cr,, above which the controller response is not
significantly affected anymore. In reality, for values of
(Y very close to 1 the neural controller exhibits an
unstable behavior. For the case of a change of setpoint
in the i-C=,content in the distillate, the response of the
neural controller is shown in Fig. 3 for two values of a,
corresponding to valve time constants of 50s (a) and
125s (b). As expected, increasing the value of Q, the
response rate of the neural controller increases, and
thus it takes a shorter time to compensate for the set-
point change. Above the value of (Y= 0.1 no detectable
changes in the controller response occurred. By further
increasing CY,
the neural controller exhibited an unstable
behavior above the value of O-9. This is the reason
why in all the following simulations the value of a = 0.1
has been adopted. The results of Fig. 3 also show the
insensitivity of the residue composition (C, mole
fraction) that does not exhibit any significant change
when the operating conditions are changed to accom-
modate for the new setpoint in the i-C5 content in the
distillate stream. This explains why the neural con-
troller is capable of meeting both the specifications on
distillate and residue, even though it directly acts only
on the manipulated variable that regulates the distillate
composition.
A second series of tests have been carried out to
compare the performance of the neural controller with
that of the inferential strategy. The results, shown in
Fig. 4, refer to a setpoint change in the distillate com-
position. The neural observer is based on a (7-2-2)
neural network configuration, while the neural con-
troller relies on a (5-I- 1)network topology. The responses
of the two control systems (inferential and neural con-
troller) are compared in terms of the capability in
setpoint tracking. The neural controller exhibits a much
faster response than the inferential strategy. The first is
able to accommodate to the new operating conditions
within about 8min, while the second requires about
20min. To verify that the control action of the neural
controller is indeed achievable, the required load on the
manipulated variable is shown in Fig. 5 in terms of the
percentage load variation with respect to the reference
case. The results show that the neural controller is able
to achieve a good setpoint tracking performance by
requiring at most a 3% change in the manipulated
variable.
Finally, the performance of the three control structures
(temperature loop, inferential and neural controllers)
are compared in Fig. 6 in terms of the disturbance

Feedforward control strategy
5*10” I 1 1 I I
.............................................................. ...
8
‘E 3*10‘s_
E
p 2*10”-
a=O.l
8
72
3 3*10’-
rt:
b)
“a 2*1u3-
a=O.CM
409
z
1*10”- ____________________~~~~~~~~~~~~~~~-.
o*lo” I I I I I
0 5 10 15 20 25 30
Time, min
Fig. 3. Response of the neural controller to a change of the key component setpoint in the distillate stream; (-) i-C5 mole fraction
in the distillate stream; (- - - -) C3 mole fraction in the residue stream; (. . .) distillate composition setpoint.
rejection characteristics. The data of Fig. 6, spanning
over an operating window of about 1000h, refer to unit
upsets due to fluctuations in feed flowrate (?Ll-2%),
feed composition (& 5- 15%) and operating pressure
(k 3-4%). The results show that the control system
based on the temperature loop (Fig. 6(a)) fails to meet
the desired distillate specification (i-C5 composition),
even though the temperature setpoint, not shown here,
is met. The nonunique relationship between temperature
and composition, for multicomponent systems, is the
prime reason for the observed faulty performance of the
temperature control strategy. On the other hand, the C3
specification on the residue stream is well satisfied, and
this is due to its insensitivity to the operating conditions
rather than to an actual efficacy of the temperature loop
policy. The tall peaks exhibited by the i-CS response
correspond to perturbations in the feed composition
that represent the most critical type of upsets to be faced
by the temperature loop control.
The response of the inferential system is illustrated in
Fig. 6(b), which clearly shows the improvement achieved
over the temperature loop performance. The inferential
controller also performs well in rejecting the critical
disturbances in the feed composition, the response peaks
are attenuated by a factor of about 4, with respect to
those detected with the temperature loop strategy. The
neural controller response is illustrated in Fig. 6(c). The
results show that this control strategy by far out-
performs the previous two policies since it is capable
of completely compensating for the upsetting loadings.
In practice, with the neural controller a complete
disturbance rejection is achieved. This is mainly due
to the intrinsic characteristics of the neural controller
that acts on the manipulated variable as soon as the
upsetting conditions are detected, and this makes the
correction action much faster. It is worth pointing out
that the slight inaccuracy exhibited by the (5-l-l) net
topology has not affected at all the rejection disturbance
performance of the neural controller. This may be due to
the feedforward characteristics of the neural controller
structure that can also lead to good controller per-
formance in the case of an approximate description of the
process inverse model.5 The achievability of the control
action has been verified by checking the required load
on the manipulated varir.ble. The results are illustrated
in Fig. 7 and show that the neural controller is able to
achieve a good disturbance rejection performance by
requiring at most a 6% change on the manipulated
variable.
The case of Fig. 6(c) refers to the single composition
control problem, and again shows the weak sensitivity
of the residue composition on the operating conditions,
and this explains why the neural controller is capable of
meeting both specifications.

5*10’ I I I I I I I I I
(4
o*,o”l I I I I I I I I I
0 20 40 60 80 100 120 140 160 180
Time, min
Fig. 4. Response of the inferential control strategy (a) and of the neural controller (b) to changes in the distillate key component
setpoint; notation as in Fig. 3.
CONCLUSIONS (the inferential and neural controllers) perform better
than the conventional temperature control loop.
For a multicomponent distillation column, represen- The neural controller is based on process inversion
tative of an actual stabilizer unit, three control policies that is performed by simply inverting the neural network
are discussed: a conventional temperature loop, an model of the distillation unit. The results indicate that
inferential strategy and a neural controller. The results for ill-conditioned processes, such as the one investi-
show that control strategies based on process modeling gated here, the implementation of nonlinear control
2 I I I 1 I I I I I
-4 1 I I I I I I I I
0 20 40 60 80 100 120 140 160 180
Time, min
Fig. 5. Percentage load on the manipulated variable (reflux flowrate) for the neural controller; the data refer to the setpoint tracking
case of Fig. 4(b).

Feedforward control strategy 411
0.03 I I I I I I I I I
(a)
8
‘;: 0.02-
0
Lt
% 0.01 -
E
) (
o.op--“--+ ------- -B---s ------ __c-a-s
0.03 I I I f 1 I I I I
0)
r:
0.03 I l I I I I I I I
Cc)
400 600
Time, min
Fig. 6. Responses of the three control strategies to upsetting loads due to fluctuations in the feed flowrate, feed composition, and
operating pressure: (a) temperature loop; (b) inferential control strategy; (c) neural controller; notation as in Fig. 3.
strategies based on process inversion requires a careful
revision of the control objectives as well as of the input
set to be fed to the neural network.
The neural controller by far outperforms the
temperature loop and the inferential control strategies.
In reality, the neural controller is the only one cap-
able of achieving an almost complete disturbance
rejection, even for the most critical upsetting loadings
represented by fluctuations in the feed composition.
All this suggests that for multicomponent distillation
units the conventional temperature loop strategies
should be abandoned in favor of more advanced
control policies.
The results clearly demonstrate the potentials of
neural modeling for developing advanced nonlinear
control strategies based on process inversion. It is
remarkable that even for ill-conditioned processes,
such as the one investigated here, neural modeling
I I
400 600
Time, min
I I
800 1000
Fig. 7. Percentage load on the manipulated variable (reflux flowrate) for the neural controller; the data refer to the disturbance
rejection case of Fig. 6(c).

provides an efficient framework to perform process
inversion.
ACKNOWLEDGEMENT
Support from the Council for National Researches
(CNR) through the research contribution 96.02598.cTO3
is gratefully acknowledged.
REFERENCES
1. Stephanopoulos, G. and Han, C., Intelligent systems in
process engineering: a review. Comput. Chem. Engng, 1996,
20, 743-79 1.
2. Bulsari, A. B. (ed.), Neural Networks for Chemical
Engineers. Elsevier Science B.V., Amsterdam, 1995.
3. Page, G. F., Gomm, J. B. and Williams, D. (ed.),
Application of Neural Networks to Modelling and Control.
Chapman & Hall, London, 1993.
4. Baratti, R., Vacca, G. and Servida, A., Neural network
modeling of distillation columns. Hydrocarbon Processing,
1995, 74(6), 35-38.
5. Baratti, R., Vacca, G. and Servida, A., Control of
distillation columns via artificial neural networks. In
Engineering Applications of Arttficial Neural Networks, ed.
Bulsari, A. B. and Kallio, S. Finnish Artificial Intelligence
Society, Helsinki, 1995, pp. 13-16.
6. Luyben, W. L., Process Modeling, Simulation and Control
for Chemical Engineers. McGraw-Hill, New York, 1990.
7. Gani, R., Ruiz, C. A. and Cameron, I. T., A generalized
model for distillation columns - I. Comput. Chem. Engng,
1986, 10, 181-198.
8. Gmehling, J. and Onken, U., Vapour-Liquid Equilibrium
Data Collection - Chemistry Data Series, Vol. I/6.
DECHEMA, Frankfurt, 1977.

baratti1997.pdf

Recommended

Recommended

More Related Content

Similar to baratti1997.pdf

Similar to baratti1997.pdf (20)

More from karitoIsa2

More from karitoIsa2 (6)

Recently uploaded

Recently uploaded (20)

baratti1997.pdf