17th European Symposium on Computer Aided Process Engineering – ESCAPE17

V. Plesu and P.S. Agachi (Editors)

© 2007 Elsevier B.V. All rights reserved. 1

Advanced Control of a Reactive Distillation

Column

Zoltan K Nagya, Reinhardt Kleinb, Anton A Kissc, Rolf Findeisenb

a
Chemical Engineering Department, Loughbborough University,Loughborough
LE11 3TU, United Kingdom, [email protected]
b
Univeristy of Stuttgart, Stuttgart, Germany
c
Akzo Nobel Chemicas, Arnhem, Nederland

Abstract

The paper presents a detailed analysis of the dynamic behavior of a reactive

distillation column. A control relevant dynamic model is derived using first-
principles modeling and it is used to study the dynamic behavior of the process
at high and low purity operating regimes. The results are used to analyze the
performance of linear and nonlinear model predictive control in comparison to
coupled PID control.

Keywords: reactive distillation, nonlinear model predictive control.

1. Introduction

Reactive distillation processes have received a tremendous industrial interest

over the last decade. Reactive distillation combines both separation and reaction
in one unit, offering significant economic advantages in some systems,
particularly for reversible reactions, which are limited by equilibrium. Despite
the complex nonlinear dynamics of these systems caused by the coupling
between the reaction and separation processes, research has been focused
mainly on the steady-state design of the reactive distillation system [1,2].
Several control approaches have been proposed based on linear [3], or nonlinear
control approaches [4]. The objective of the paper is to evaluate the feasibility
2 Z.K. Nagy et al.

of different control techniques from both linear and nonlinear controllers by

considering the application of nonlinear model predictive control (NMPC),
linear model predicative control (LMPC) and coupled PI control to a sample
reactive distillation process. The main contribution of the paper besides the
simulation of the different control strategies consists in the systematic analysis
of the control performances based on the detailed open-loop analysis of the
dynamic behavior of the system at low and high purity operating conditions.

2. Mathematical model of the reactive distillation column

Fig. 1. Reactive distillation column Fig. 2. Tray-by-tray model of the column

This case study is based on a general mathematical representation of a reactive

distillation column (Fig. 1), where a reaction of the form:
r
A + B ←⎯⎯ → C +D (1)

takes place in the liquid phase. The column is considered to have 41 theoretical
trays. The feed of component A and B enter the column in tray number 10 and
30, respectively. The products C and D are removed as top and bottom
products, respectively. The process control objective is to maximize conversion
of the reactants and separation of the products. As manipulated variables the
reboiler vapor flow and the liquid reflux in the condenser are used. The distillate
flow and the bottom liquid flow are used to control the total liquid holdup in the
reboiler and the condenser, respectively. The reaction is assumed to take place
only in the liquid phase. The tray-to-tray model equations (Fig. 2) for the
column trays are given by:
dM il NC
= Fi + Li +1 +Vi −1 − Li −Vi +Voli ∑ μj ri (2)
dt j =0

d (M il x i, j )
= Fi x f + Li +1x i +1, j +Vi −1yi −1, j − Li x i, j −Vi yi, j +Voli μj ri (3)
dt
Advanced Control of a Reactive Distillation Column 3

for i = {2, ", 40} , j = {1, ", 4} . Here, x i, j and yi, j are the mol fractions of
component j in the liquid and in the vapour of tray i . M il is the total molar
liquid hold-up of a tray i . The model equations for the condenser and the
reboiler are as follows:
dMbl NC
= L2 − Bf −Vb +Volb ∑ μj rb (4)
dt j =1

d (M cl ) NC
= V40 − Lc − Df +Volc ∑ μj rc (5)
dt j =1

d (Mbl x i, j )
= L2x 2, j − Bfxb, j −Vb yb, j +Volb μj rb (6)
dt
d (M cl xc, j )
= V40y 40, j − Lc x c, j − Dfx c, j +Volc μj rc (7)
dt
where Bf and Df are the product flow from the reboiler and the condenser,
respectively. The reaction rate is given by:
⎛ 1 ⎞⎟
ri = kor ⎜⎜⎜x i,Ax i,B − x i,C x i,D ⎟⎟ (8)
⎜⎝ keq ⎟⎠
The vapour mol fraction is calculated based on constant volatilities by:
αj ⋅ x i, j
yi , j = NC
(9)
1 + ∑ (αj − 1) ⋅ x i, j
j =1

Finally, the change in component composition is given by:

dx i, j
1 ⎛d (M l x ) dM il ⎟⎟⎞
⎜⎜ i i, j
= ⎜⎜ − x i, j ⎟ (10)
dt Mi ⎜⎝ dt dt ⎟⎟⎠

The reactive distillation column model was implemented in the

Matlab/Simulink environment. The column is represented by a Simulink S-
Function, which is written in C-language and compiled with the Watcom
compiler to a dll file for fast simulation.

3. Open-loop dynamic analysis

Reactive distillation is an extremely nonlinear process with highly interacting

dynamics. The characteristics of the process change under different operating
conditions, thus, linear models are only good approximations of the real process
for very small deviations of the steady state. Prior to design a controller we have
done several simulations to characterize the sensitivity of the control variables
with respect to small input changes and disturbances. Steady state deviations
have been analyzed starting from two different steady states.
4 Z.K. Nagy et al.

Fig. 3. Steady state analysis within the low Fig. 4. Steady state analysis within the high
purity operating regime (0.79, 0.81) (open purity operating regime (0.99, 0.99) (open
loop simulations). Small changes in the input loop simulations). Small changes in the input
result in significant steady state changes. result in drastic steady state changes.

Fig. 5. Step response of the reactive distillation column for very small steps in the inputs in the
low purity operating region (left) and high purity region (right).

The reboiler vapour flow V and the condenser reflux L have been varied with
± 10% of the respective steady stated values. Figures 3 and 4 show new steady
states resulting from the input variation. The trend is different for small regions
of the steady state for operation in a low purity region and a high purity region,
however the overall trend for larger input changes are similar. A high
composition for both C41 in the condenser and D1 in the reboiler is achievable
only in a narrow regime, thus, disturbances are expected to have high influences
in the composition of the outflows. Figure 5 presents the step responses of the
column w.r.t. very small steps (at t=1000 min) in the vapour and liquid flow.
The responses are completely different in the two operating regimes. Note the
sign reversal in the gain, also the longer response time for the high purity
regime. This already suggests, that a linear controller would have to be tuned
quite sluggish if it is to be used in the whole operating region.
Advanced Control of a Reactive Distillation Column 5

Fig. 6. Performance of the coupled PI control Fig. 7. NMPC performance for setpoint
for setpoint changes and various disturbances change from low purity to high purity
using coupled PID control. Disturbances in performed in two steps. Arrows indicate
feeds A and B occurred at times 1000, 20000 unmeasured disturbances in the flow rate for
and 40000 minutes (+20%, -20%, -20% feeds A and B: +20%, nominal, -20%.
disturbance in the nominal flow).

4. Comparative simulation of coupled PI, LMPC and NMPC of the system

Different controller and control structures have been tested. Among them,
decentralized PID, coupled PID, linear and nonlinear model predictive control.
The decentralized PID couldn’t cope with large setpoint changes and any
disturbances in the high purity region. The PID-like control structure was based
on open-loop analysis, RGA calculations and trial and error simulations
(intuition) with the nonlinear model. The controller parameter found by
considerations of a linear model (Taylor approximation around a steady state)
failed to stabilize the system over the whole operating regime. Since we want to
operate the column under different conditions, we have analysed two different
steady states: one at low product purity (0.8, 0.8) and one at high product purity
(0.99, 0.99). Linear models were derived by first order power series expansion
and controllability analysis performed using RGA. The values of the entries in
the RGA were quite large and negative at frequency zero (steady state),
indicating difficult control problem. In addition, we observed sign reversal and
inverse response. For systems with negative entries in the RGA it is not
recommended to choose diagonal PID, since the system might be unstable if a
loop becomes open (e.g. due to saturation). Interaction is taken into account
using a coupled PID structure. Problems due to saturation were handled by
considering anti-windup structure and less aggressive tuning. If the controller is
tuned more aggressive, problems with saturations are more pronounced, and for
different cases (e.g. large disturbances or setpoint changes), the controller fails.
The control performance of the best PI structure and tuning (coupled with anti-
6 Z.K. Nagy et al.

windup) is shown on Figure 6. The closed-loop response is very slow, however

a faster controller would lead to a highly oscillating response and even
instability of the closed system. LMPC is regarded as more advanced control
technique than simple PID’s, however for such a highly nonlinear process with
a wide operating regime and large disturbances, the LMPC is not able to
stabilize the process at the desired setpoint using a single linear model. Multiple
linear models, identified in different operating conditions, or a model of the
measured disturbances can be employed to enhance the performance, but since
the effects are expected to vary over the operating regime, this control technique
has been abandoned in favour to the NMPC, where the aforementioned
problems are inherently considered in the controller. The NMPC controller uses
a multiple shooting optimisation solver called HQP (huge quadratic
programming) and an optimization package OptCon. Due to the multiple
shooting approach a much larger optimisation problem arises which is,
however, highly structured. Therefore sparse techniques can be applied, such
that the optimisation problem is solved very fast. The excellent control
performance achieved by the NMPC controller is shown in Figure 7. The results
evidence that in comparison with a well-tuned linear controller the nonlinear
model predictive controller shows a superior performance, by keeping a tight
product composition, with respect to setpoint-changes and disturbances.

5. Conclusions

The application of several advanced control approaches have been illustrated in

the case of a reactive distillation column. The dynamic and steady state analysis
of the open loop system illustrates the difficulty of the control depends on the
operating regime. Linear MPC was not able to control the system, whereas a
coupled PID gave acceptable results, with very slow response time in the high
purity operating zone. A nonlinear MPC based on efficient multiple shooting
algorithm and the software package OptCon has been implemented, which
showed excellent control performance.

References

1. Taylor, R., & Krishna, R. (2000). Modelling reactive distillation. Chemical Engineering
Science, 55, 5183–5229.
2. Malone, M. F., & Doherty, M. F. (2000). Reactive distillation. Industrial and
Engineering Chemistry Research, 39, 3953–3957.
3. Sneesby, M., Tade, M., & Smith, T. (1999). Two-point control of a reactive distillation
column for composition and conversion. Journal of Process Control, 9, 19–31.
4. Kumar, A., & Daoutidis, P. (1999). Modeling, analysis and control of ethylene glycol
reactive distillation column. AIChE J, 45(1), 51–68.