Professional Documents
Culture Documents
net/publication/291659272
CITATIONS READS
2 1,232
3 authors, including:
Some of the authors of this publication are also working on these related projects:
Modeling of High-pressure multiphase behaviour of ternary systems carbon dioxide–water–polar solvent View project
All content following this page was uploaded by Binous Housam on 24 January 2016.
The exclusive license for this PDF is limited to personal website use only. No part of this digital document
may be reproduced, stored in a retrieval system or transmitted commercially in any form or by any means.
The publisher has taken reasonable care in the preparation of this digital document, but makes no expressed
or implied warranty of any kind and assumes no responsibility for any errors or omissions. No liability is
assumed for incidental or consequential damages in connection with or arising out of information contained
herein. This digital document is sold with the clear understanding that the publisher is not engaged in
rendering legal, medical or any other professional services.
Chapter 9
ABSTRACT
In the present study, a binary mixture, composed of ethanol/water, is considered.
This mixture presents deviation from ideal behavior, which is taken into account using
the Wilson model. The dynamic behavior of the distillation columns that separate this
mixture is studied and transfer functions are obtained for a step in the reflux ratio, the
reboil ratio and the feed composition. Then, the control of the columns using Model
Predictive Control (MPC) is performed. The authors show that both distillate and bottom
compositions are controlled while feed composition disturbances are rejected readily
using MPC. Such calculations can be readily performed, using mathematical software
available in most universities (Mathematica® and Matlab©), by students under the
direction of a chemical engineering faculty in the context of a three-month senior term
project.
INTRODUCTION
In order to control a distillation column, one has to obtain transfer functions by
performing dynamic simulations. Such simulations have been performed for both non-
reactive (Holland, 1966) and reactive (Alejski and Duprat, 1996) distillation columns. A
profusion of techniques have been applied to control distillation columns. Indeed, for binary
mixtures for example, Wood and Berry (1973) has used a decoupling technique in order to
control both distillate and bottom purities. More recently, an educational paper treated the
control of binary distillation column with both Wood and Berry (1973) technique using
SIMULINK© and Rijnsdorp (1965) method using Mathematica® (Binous et al., 2011). Our
objective, in the present study, is to use a modern tool namely model predictive control
256 Abdallah Al-Shammari, Naim Faqir and Housam Binous
(MPC) to control binary distillation columns. Authors feel that this will help chemical
engineering students and educators learn about this challenging field of process dynamic and
control through this simple worked out example; especially that only few recent papers have
addressed MPC from a simple educational view point. Indeed, one paper by Ricardez-
Sandoval et al. (2010) presented educational material about an experiment where MPC is
used in order to control a double pipe heat exchanger. Another paper, presented more
advanced educational material where a bioreactor is controlled using neural networks and
MPC (Mete et al., 2010). In an earlier paper dating back to 1998, Doyle et al. described an
educational tool based on several process control modules including an application of MPC
but this tool focused on various unit operations and did not treat in great details binary
distillation. Model predictive control is a class of advanced process control strategies that has
been widely accepted in industries since it was proposed by Culter and Ramaker (1980) and
Richalet et al. (1978). In fact, it is the most commonly applied advanced control strategy in
oil refining and petrochemical industries. One main advantage of MPC is its ability to handle
process constraints for multi-variable systems. In other words, MPC can simultaneously
control multiple key variables in the process by adjusting others variables within their
operational constraints. A detailed review of MPC, which discussed work done on the area of
MPC and current challenges that face on-line MPC such as feasibility, control stability, and
MPC performance is given in Morari and Lee (1999). Qin and Badgwell (2003) conducted a
survey of industrial model predictive control technology and discussed the history,
formulation, application, limitations and future perspectives of MPC. Maciejowski (2002)
discussed different aspects of constrained MPC.
Since the last two decades, MPC has found a wide range of industrial applications such as
distillation columns, FCCs, batch reactors and other chemical processes. Distillation columns
are important separation units in all refineries and petrochemical plants. Distillation columns
need to be maintained at optimal operation in order to maximize the net profit of the plant.
The main challenge in controlling distillation columns is the existence of multiple variables
and products that need to be maintained at certain set points without violating process
constraints. Such challenge makes the MPC a good candidate to be used in distillation
control. One of the early applications of the MPC to control a binary distillation column was
done by Martin-Sanchez and Shah (1984). They demonstrate the potential and scope of
applications of MPC to industrial processes. Many studies were proposed to use different
schemes of MPCs to various types of distillation columns including C3/C4 splitter (Porfirio et
al., 2003), crude distillation column (Hovd et al., 1997), superfractionators (Pannocchia and
Brambilla, 2005), and reactive distillation columns (Kawathekar and Riggs, 2007). Moreover,
Volk et al. (2005) presented practical study about multivariable predictive control of an
industrial distillation column considering different types of constraints.
The chapter contains the following sections:
1. The first part of the present chapter studies the dynamic behavior of distillation
columns that separates a mixture of ethanol/water. In this section, the dynamic
responses to a step in the reflux ratio, reboil ratio or feed composition are given and
corresponding transfer functions are determined.
2. The second part is an application of MPC in order to drive product purities to
specification and to reject any feed composition disturbance.
Model Predictive Control of a Binary Distillation Column 257
3. The authors conclude the chapter with some remarks concerning possible extension
of their work to multi-component mixtures.
The outputs are the distillate and bottom compositions, and xb , respectively. The
manipulated inputs are the reflux and reboil ratios, R and S, respectively. The disturbance
input is the light component mole fraction in the feed stream, x f .
Figure 1. Behavior of distillate and bottom compositions for a +1% step (dashed curve) and -1% step
(continuous curve) change in the reflux ratio.
Model Predictive Control of a Binary Distillation Column 259
Figure 2. Behavior of distillate and bottom compositions for a +1% step (dashed curve) and -1% step
(continuous curve) change in the reboil ratio.
The form of the obtained transfer functions (see Table above) is similar to that given by
Wood and Berry (1973) (i.e., first-order plus delay). One major distinction is that we consider
that the unmeasured disturbance variable is the feed composition while Wood and Berry
(1973) take the feed flow rate as a disturbance variable. Indeed, because of the lack of
composition analyzers in a chemical plant, one must consider feed composition as an
unmeasured disturbance. On the other hand feed flow rate is easily determined using a
flowmeter. Also, the manipulated variable in Wood and Berry treatment are the reflux flow
rate and the reboiler steam flow rate while we take more natural manipulated variables
namely the reflux and reboil ratios. We feel that our approach is more appropriate for the
subsequent process control case studies. Indeed, feed composition disturbances are far more
frequent in a chemical plant than flow rate disturbances and it is simpler to set bounds
(constraints) on the reflux and reboil ratios rather on the reflux and steam flow rates.
260 Abdallah Al-Shammari, Naim Faqir and Housam Binous
Figure 3. Behavior of distillate and bottom compositions for a +1% step (dashed curve) and -1% step
(continuous curve) change in the feed composition.
P
Min F yk 1 y s Q yk 1 y s
T
yk 1 , uk 1 k 1
M 1
u k 1 u s R u k 1 u s u k 1 S u k 1
T T
k 1
where ys and us are the targeted steady-state output and input trajectory. The change in input
variable from one time step to the next is given by: u k 1 u k 1 u k . This optimization
problem is solved at each time step k to obtain output predictions, {yk+1}, over a prediction
horizon (P) and the best moves of manipulated variables, {uk+1}, over a control horizon (M).
Then, the first control move is implemented and this calculation is repeated at each step time.
Thus, it is obvious that the success of MPC heavily depends on the accuracy of the process
model which represents the relation between the inputs and the outputs and which is used to
predict future moves. Moreover, it also depends on proper tuning of some parameters in Eq.
(1) such as M and P and also weighting matrices Q and R. The matrices Q and R allow the
output and input variables to be weighted according to relative importance. With improper
values of control or prediction horizons, the MPC controller action can be aggressive or
sluggish. The controller tends to be more aggressive with increasing control horizon (M) or
decreasing prediction horizon (P). In other words, increasing control horizon would raise the
computation effort while decreasing prediction horizon reduces the number of moves which
lead to aggressive control action. Generally tuning these parameters need some experience
but there are some quick rules to get good estimates such that 5 ≤ M ≤ β0 and P =M+N where
N is the model horizon, which usually taken as the settling time of the open-loop response
over the sampling time (Seborg et al. 2005). In MIMO systems, each MV or CV may have
specific values of M or P. Regarding weighting matrices such as Q for outputs and R for
inputs, they are used to allow weighting process variables (CVs and MVs) based on their
relative importance or their effect on the process. Most important variables usually have the
largest weight values in order to give them high priority in control calculation and to drive
them first to their targets. In many plants, selection or tuning of weight matrices is done on-
line and requires consulting the process engineer because it totally depends on the nature of
the process. A broad review of many tuning methods for several classes of MPC formulations
from theoretical and practical perspectives is presented in Garriga and Soroush (2010).
Indeed, we are interested in both product purities. In addition, since the degree of freedom is
two for this two-product column, one can achieve product purity specifications for both the
distillate and the residue.
For all scenarios the sampling time is taken equal to 2 min. The first scenario consists of
a step in the feed composition (+0.2) with P and M set equal to 35 and 5, respectively. No
constraints were applied to the manipulated variables. Figure 5 shows that the controller
rejects the disturbance and brings back the product purities to their set points. Here, the MVs
behave in a non-intuitive manner. Indeed, the reflux ratio increases while the reboil ratio
decreases. This behavior is not shown but a similar result can be seen in Figure 7. One would
expect the reflux ratio to decrease since the feed contains more ethanol. In reality, the bottom
flow rate increases leading to (1) a decrease in the reboil ratio and (2) a decrease in the
distillate flow rate (due to mass balance considerations: F=D+B). and thus to an increase in
the reflux ratio. A steady-state simulation done separately using Mathematica® confirms this
finding since the bottom flow rate increases, from 4.11 kmol/min to 4.20 kmol/min.
A similar scenario (called scenario 2) applies, in addition to the feed disturbance (+0.2),
the following constraints on the manipulated variables (-2.5 R, S 2.5). Here, the controller
is unable to reject the disturbance in feed composition (see Figure 6) and to take products
purity back to the set points because of the saturation of MVs (see Figure 7). However, when
constrains on the MVs were relaxed (-3.0 R, S 3.0), the controller was able to reject the
disturbance.
In another scenario, a pulse change in the feed composition (+0.2) was applied at t=10
min and lasted 20 minutes. Prediction and a control horizons were equal to P=10 and M=5,
respectively. Here, no constraints were applied on the MVs. Again, the controller is capable
of rejecting this disturbance in feed composition and taking products purity back to their set
points (see Figures 8-9).
Finally, a step change in the set points of distillate purity (+0.05) and bottom purity (-
0.05) at t=10 min without any change in feed composition was performed. Here, we select
prediction and control horizons such that: P=35 and M=5, respectively. Constrains on MVs
have been applied as follows: -0.25 R, S 0.25. Although there are restrictions on the
movements of MVs, the controller achieved the new set points (see Figure 10) by increasing
both of reflux and reboil ratios (both reached to saturation limit as can be seen in Figure 11).
It is intuitive to expect that if one wants to achieve both a higher mole fraction of ethanol and
water in the distillate and bottom streams, respectively, then both the reflux and reboil ratios
should increase.
The dynamic simulations were performed using Mathematica® (the commented code is
available upon request from the corresponding author). Such computations can also be
performed using Matlab© making it possible to perform both dynamic and control simulations
using the same platform. Indeed, MPC calculations were conducted using MPC toolbox with
either mpctool, which is based on the Graphical User Interface of Matlab© called GUI
(Bemporad, 2009) or writing appropriate commands in m-files. The code can be readily
changed to study any other binary mixture and can also be extended to tackle more
challenging problems such as multi-component distillation.
264 Abdallah Al-Shammari, Naim Faqir and Housam Binous
Figure 6. Disturbance is not rejected for this constrained scenario (scenario 2).
Figure 8. MVs and disturbance versus time. (A pulse in the feed composition is applied at t=10 min).
Figure 9. Behavior of the output variables when column is subject to a pulse disturbance in feed
composition.
266 Abdallah Al-Shammari, Naim Faqir and Housam Binous
Figure 10. Both product purities reach new set point (i.e., +0.05 for distillate and -0.05 for bottom).
CONCLUSION
Dynamic simulations have been performed and allowed the determination of the different
transfer functions using Mathematica®. These functions can be used to perform model
predictive control simulations with the help of the Matlab© MPC toolbox. Various scenarios
have been studied. Open-loop simulations are possible with mpctool and show the behavior
of the distillation column when subject to a unit-step in the reflux ratio for example. Closed-
loop model predictive control show simulations show that the set points can be readily
attained and that the feed disturbances can be efficiently rejected. Constraints are easily set
for the manipulated variables (i.e., reflux and reboil ratios). In few simulations, the MVs
reached saturation and disturbances could not be rejected. Similarly, calculations can be
performed for ternary systems (e.g., benzene/toluene/p-xylene). Such multi-component
problems will be the focus of a separate manuscript.
ACKNOWLEDGMENT
The support of King Fahd University of Petroleum & Minerals is duly acknowledged.
This work has been carried out during the sabbatical leave granted to Dr. Naim M. Faqir from
the University of Jordan in the academic year 2012/2013.
APPENDIX 1
The governing equations that allow the dynamic simulations to be performed are given
below. They are composed of the total and partial mass balances and the energy balance
around the feed stage, the partial reboiler, the total condenser and any tray other than the feed
tray. These balance equations, written separately for all the column sections mentioned above,
are the following:
dM f
F L f 1 V f 1 L f V f (A-1)
dt
d ( M f x f ,i )
F z f ,i L f 1 x f 1,i V f 1 y f 1,i L f x f ,i V f y f ,i for i=1,2,…,Nc (A-2)
dt
d (M f h f )
F hF L f 1 h f 1 V f 1 H f 1 L f h f V f H f (A-3)
dt
268 Abdallah Al-Shammari, Naim Faqir and Housam Binous
dM k
Lk 1 Vk 1 Lk Vk (A-4)
dt
d ( M k x k ,i )
Lk 1 xk 1,i Vk 1 y k 1,i Lk xk ,i Vk y k ,i for i=1,2,…,Nc (A-5)
dt
d ( M k hk )
Lk 1 hk 1 Vk 1 H k 1 Lk hk Vk H k (A-6)
dt
dM 1
V2 ( L1 D) (A-7)
dt
d ( M 1 x1,i )
V2 y 2,i ( L1 D ) x D,i for i=1,2,…,Nc (A-8)
dt
d ( M 1 h1 )
V2 H 2 ( L1 D) h1 Q C (A-9)
dt
dM N
LN 1 LN V N (A-10)
dt
d ( M N x N ,i )
LN 1 x N 1,i LN x N ,i VN y N ,i for i=1,2,…,Nc (A-11)
dt
d ( M N hN )
LN 1 hN 1 LN hN VN H N Q B (A-12)
dt
APPENDIX 2
The Wilson model, which allows the prediction of the activity coefficients, is given by
the following equations:
ln k ln j c1 x j Akj 1 i c1
N N xi Aik
Nc
j 1
x j Aij
,
Model Predictive Control of a Binary Distillation Column 269
vj ij ii
where the binary interaction parameter is: Aij v exp RT
.
i
The molar volumes (in cc/mole) and binary interaction coefficients (in kcal/kmole) for
the ethanol/water mixture are given in the two tables below:
REFERENCES
Alejski, K; Duprat, F. ―Dynamic simulation of the multicomponent reactive distillation,‖
Chemical Engineering Science, 51, 4237-4252, (1996).
Bemporad, A; Morari, M; Ricker, NL. Model Predictive Control Toolbox 3, User‘s Guide,
Mathworks, Inc., Natick, Massachusetts, U.S.A. (2009).
Binous, H; Al-Mutairi, E; Faqir, N. ―Study of the separation of simple binary and ternary
mixtures of aromatic compounds,‖ Computer Applications in Engineering Education,
DOI: 10.1002/cae.20533, (2011).
Doyle, FJ III; Gatzke, EP; Parker, RS. ―Practical Case Studies for Undergraduate Process
Dynamics and Control Using Process Control εodules,‖ Computer Applications in
Engineering Education, 6, 181–191, (1998).
Garriga, Jδ; Soroush, ε. ―εodel Predictive Control Tuning εethodsμ A Review,‖ Industrial
Engineering Chemistry Research, 49, 3505–3515, (2010).
Holland, CD. Unsteady State Processes with Applications in Multicomponent Distillation.
Prentice-Hall, Engleewood Cliffs, New Jersey, U.S.A. (1966).
Hovd, ε; εichaelsen, R; εontin, T. ―Model predictive control of a crude oil distillation
column,‖ Computers & Chemical Engineering, 21, S893-S897, (1997).
Kawathekar, R; Riggs, JB. ―Nonlinear model predictive control of a reactive distillation
column,‖ Control Engineering Practice, 15, 231-239, (2007).
Maciejowski, JM. Predictive Control with Constraints, Prentice Hall, Upper Saddle River,
New Jersey, U.S.A. (2002).
Martin-Sanchez, Jε; Shah, Sδ. ―εultivariable Adaptive Predictive Control of a Binary
Distillation Column‖, Automatica, 20, 607-620, (1984).
Mete, T; Ozkan, G; Hapoglu, H; Alpbaz, ε. ―Control of dissolved oxygen concentration
using neural network in a batch bioreactor,‖ Computer Applications in Engineering
Education, DOI: 10.1002/cae.20430 (2010).
εorari, ε; δee, JH. ―εodel Predictive Controlμ Past, Present, and Future,‖ Computers &
Chemical Engineering, 23, 667-682, (1999).
270 Abdallah Al-Shammari, Naim Faqir and Housam Binous