Proceedings of the 6th World Congress on Intelligent Control

and Automation, June 21 - 23, 2006, Dalian, China

State Space Modeling and Predictive Control of a Binary

Batch Distillation Column *
Zhiyun Zou and Dehong Yu Zhen Hu , Ning Guo, Luping Yu and Wenqiang Feng
School of Mechanical Engineering The Sixth Department
Xian Jiaotong University Beijing Research Institute of Pharmaceutical Chemistry
Xian 710049, China P. O. Box 1043, Beijing 102205, China
zouzhiyun@[Link], dhyu@[Link] guoninggn@[Link], pana0813@[Link]

Abstract - A linear discrete state-space model of a Frattini-Fileti et al [3] developed a self-adjusting gain-
methanol/water binary batch distillation column is developed scheduling PI control method and used it to the control of a
based on theoretical analysis of dynamic mass balance and binary batch distillation column. This method reached the
vapor-liquid phase balance, and this state-space model is used to control goal by continuously adjusting the gain to higher
design a Model Predictive Control (MPC) strategy. The
value. Li and Wozny [4] introduced an adaptive control
composition of methanol inside the distillation column is
estimated using an empirical temperature-composition strategy to follow the pre-defined optimal reflux-ratio profiles.
relationship model. The state space model based MPC algorithm Ref. [5] uses a robust nonlinear control method to compensate
is presented in detail, and the MPC strategy is implemented on modeling errors, and adjusts the reflux ratio to control the
an industrial control computer to directly control the estimated composition of a batch distillation process. Up to now,
composition of the batch distillation column. Control although researchers around the world have put forward many
experiments of the batch distillation column show that MPC advanced control strategies for the control of batch distillation
gives smooth and accurate control results, and its control results process, most of these control strategy research only carried
are much better than the commonly used PI control. out theoretical investigation or simulation study, and few of
them are applied to practical industry batch distillation process
Index Terms - Batch Distillation, State Space Model, Model
or tested by experiment.
Predictive Control, Inferential Control.
Model predictive control (MPC) is a widely used
I. INTRODUCTION advanced control strategy in the process industry. MPC
possesses good stability and robustness [6-7], and it is
Batch chemical processes are suitable for the production desirable to use MPC to batch distillation process control.
of small-scale, multi-product and high-value added fine Dechechi et al [8] applied a nonlinear model predictive
chemicals, so the research and development of batch chemical control algorithm to the composition control of a batch
processes are getting much more attention from the process distillation column, and achieved satisfactory simulation
industry in recent years. Batch distillation processes are very results.
important separation processes in the process industry. In this paper, based on theoretical analysis of dynamic
Because single batch distillation equipment is able to separate mass balance and vapor-liquid phase balance, a linear discrete
multi-component chemicals, high-viscosity low-volatility state-space model of a methanol/water binary batch distillation
components, and mixtures with large amounts of solid-state column is developed, and this state-space model is used to
substances, the research on modeling and advanced control of design a MPC strategy. The composition of methanol inside
batch distillation process possess very important application the distillation column is estimated using an empirical
value. temperature-composition relationship model, and the MPC
Traditionally, the development of batch distillation strategy is implemented on a EuroBEEB industrial computer
process control strategy was centered on the calculation of with real-time BASIC program to directly control the
open-loop optimal reflux-ratio profile [1]. However, with the estimated composition of the batch distillation column.
influence of modeling error and process disturbance, the result Control experiments illustrate that MPC gives smooth and
of open-loop optimal reflux-ratio control was not satisfactory. accurate control results, and its control results are much better
Recently, the trend of batch distillation process control than the commonly used PI control.
research is to develop effective feedback control strategies.
Control experience shows that, although PID control is useful II. BRIEF DESCRIPTION OF THE METHANOL/WATER BATCH
for the control of continuous distillation process, it’s not DISTILLATION COLUMN
suitable for the control of the nonlinear non-stationary batch The flow-sheet diagram of this methanol-water binary
distillation process. Quintero-Marmol et al [2] designed an batch distillation column is illustrated as Fig.1. All the feed of
extended Luenberger observer to estimate the composition of methanol-water is pumped into the still pot of the reboiler in
a batch distillation process, and carried out inferential control. one time before distillation. The reboiler is heated up by

*
This work is supported by China Scholarship Council Grant #21302095 to Z. Zou.

1-4244-0332-4/06/$20.00 ©2006 IEEE

6252
Authorized licensed use limited to: Politechnika Gdanska. Downloaded on June 10,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
 Fig. 1 The flow-sheet diagram of the methanol-water binary batch distillation column

flowing steam through its jacket, and the flow-rate of steam is dX 1

M1 VYs  LX 2  VY1  LX 1
controlled by a flow controller. In total, 6 sieve-plate trays are dt (2)
mounted inside the column. The methanol vapor at the top of dX 2
the column is cooled down by flowing through a cooling M2 V (Y1  Y2 )  L( X 3  X 2 )
dt (3)
water condenser. The flow direction of the condensed
dX 3
methanol liquid is time-sharing controlled by an electric M3 V (Y2  Y3 )  L( X 4  X 3 )
butterfly valve. dt (4)
dX 4
M4 V (Y3  Y4 )  L( X 5  X 4 )
dt (5)
III. DEVELOPMENT OF STATE SPACE MODEL
dX 5
The model of the methanol/water binary batch distillation M5 V (Y4  Y5 )  L( X 6  X 5 )
dt (6)
column has been developed using a number of model dX 6
assumptions [1]. Some of these model assumptions are listed M6 VY5  LY6  VY6  LX 6
dt (7)
as below:
1) Constant vapor flow rate (V, mole/h) and reflux flow Here Ms is the molar holdup of the reboiler still. Xi
rate (L, mole/h) through the column. No entrainment and (i=1,2,…,6,s) is the liquid composition of methanol (mole
weeping effects, and total condensation with no sub-cooling in fraction) on tray i (i=1,2,…,6) and the reboiler still (s). Yi
the condenser. (i=1,2,…,6,s) is the vapor composition of methanol (mole
2) Constant molar holdup on each tray (Mi, i=1, 2,…, 6). fraction) on tray i (i=1,2,…, 6) and the reboiler still (s).
Negligible vapor holdup. The equilibrium relationship between the vapor
3) Theoretical plates, constant tray efficiency, phase composition Y and liquid composition X of methanol in the
equilibrium and ideal vapor phase. binary mixture can be approximated using the constant
4) Adiabatic operation, perfect mixing on the trays and in relative volatility (Į), and is given by:
the reboiler still and etc.
Based on these assumptions, the dynamic composition Yj DX j /(1  (D  1) X j ) j=1,2,…,6,s. (8)
mass balance equations of the reboiler still and tray 1 to tray 6
are as follows: As the above model are nonlinear differential equations,
directly implementing predictive control strategies with this
dX s (1) model will result in a complex nonlinear programming
Ms L( X1  X s )  V (Ys  X s )
dt problem with large on-line real-time computation

6253
Authorized licensed use limited to: Politechnika Gdanska. Downloaded on June 10,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
requirements. In order to implement a linear state space model prediction model is redefined at each sampling instant from
based predictive control algorithm [6] on the EuroBEEB the actual state vector
LC486/33 industrial computer with computational efficiency,
a linear discrete-time state space model is needed [9]. x p (t / t ) x (t )
(16)
Substituting (8) into (1) to (7), and linearizing these
nonlinear differential equations around the steady state of total And the modification of the control action is only limited in p2
reflux operation, a linear continuous-time state space model of steps (p2<p1), so
the distillation column is given below: u (t  p1  1) u (t  p1  2)  u (t  p2 ) (17)
dX / dt Ac X  BcU (9)
Applying (15) recursively to the initial conditions and
Y Hc X
(10) combining (16) and (17), the following output prediction
equations can be obtained:
Where:
y p (t  1 / t ) HAx(t )  HBu (t ),
X >x1 , x2 , x3 , x4 , x5 , x6 , xs @T (18)
T
y p (t  2 / t ) HA 2 x(t )  HABu (t )  HBu (t  1)
dX ª dx1 dx 2 dx3 dx 4 dx5 dx6 dx s º 
dt « dt , dt , dt , dt , dt , dt , dt »
¬ ¼ y p (t  p1 / t ) HA p1 x (t )  HA p1 1 Bu (t )    HA p1  p2 1 Bu (t  p 2 )
U >l , v@ Y > y1 , y 2 , y3 @ >x2 , x4 , xs @
T T T

Here yp(t+j/t) denotes the output vector prediction at instant t

X is the state variable vector, U is the control variable vector for instant t+j . A composite form of (18) can be written as
and Y is the output variable vector; xi (i=1,2,…,6,s) and l,v is follows:
the perturbation values of compositions, reflux flow rate and
vapor flow rate from their steady state counterparts. x2, x4 and Yp Gx (t )  FU
(19)
xs are chosen as output variables because they can be inferred
from their measured temperatures T2, T4 and Ts using the Where
following temperature-composition relationship equation:
Yp [ y p (t  1 / t ), y p (t  2 / t ), , y p (t  p1 / t )]T
,
Tj 80.1  15.44x j  20 exp( x j / 0.1277) , j=1,2,…,6,s. (11)
ª HA º ª HB 0  0º
« HA 2 » « HAB »
The linearized model equation (9) and (10) can be converted G « », F « HB  0»,
to a discrete-time form by using the Matlab numerical tool «  » «    »
function C2D. The following discrete-time state space model « p1 » « p1 1 p1  2 »
¬ HA ¼ ¬ HA B HA B  HA p1  p2 1 B ¼
is obtained by setting the sample time to 10 seconds.
U [u (t ), u (t  1), , u (t  p2 )]T .
X (t  1) A X (t )  B U (t ) (12)
The predictive control law is formulated to minimize a
Y (t ) HX (t ) (13) cost function given by
p1
IV. THE MODEL PREDICTIVE CONTROL ALGORITHM (20)
J ¦[ y (t  j )  y
j 1
r p (t  j / t )]T Q j [ yr (t  j )  y p (t  j / t )] / 2

Consider a system described by the linear state equations:

or J (Yr  Y p )T Q (Yr  Y p ) / 2,
x(t  1) A x(t )  B u (t ) (14-1)
where yr(t+j), j=1,2,…, p1, is a reference trajectory (or set
y (t ) Hx(t ) (14-2)
point) for the output vector which may be redefined at each
sampling instant t, Yr=[yr(t+1), yr(t+2),…, yr(t+p1)]T. Qj is a
Where x(t)ęRn, y(t)ęRr, and u(t)ęRm are the state, output
non-negative definite weighting matrix , and Q is a non-
and control vectors respectively; A,B and H are matrices with
negative definite matrix composed by Qj as
appropriate dimensions.
The structure of the model (14) is used for formulating ªQ1 0 0 0 º
predictive controllers. First, define a state prediction model of «0 Q 0 0 »»
the form: Q « 2

«    »
x p (t  j / t ) A x p (t  j  1 / t )  B u (t  j  1) , « »
«¬ 0 0 0 Q p1 »¼ .
j=1,2,…,p1 ; (15)
The solution minimizing the cost function (20) can be
Where xp(t+j/t) denotes the state vector prediction at instant t obtained by solving ΩJ/ΩU=0, from which direct computations
for instant t+j , and p1 is the prediction horizon. This may be obtained explicitly as:

6254
Authorized licensed use limited to: Politechnika Gdanska. Downloaded on June 10,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
 U ( F T QF ) 1 F T Q[Yr  Gx(t )] T [Yr  Gx(t )] . (21) Here V0 and L0 is the steady-state value of V and L at total
reflux operation.
Usually the matrix T=(FTQF)-1FTQ is calculated offline, so the From simulation results by implementing control law (27)
online real time computation requirement of (21) is very low. using some history batch experiment data, it shows that the
As the predictive control law is a receding horizon strategy, in changing trend of the reflux ratio profile is in accordance with
order to reject disturbance and compensate for modeling error, past operation experiences, but its numerical value is a little
usually only the first vector u(t) of U is implemented, rejecting higher. So a scaling factor or proportional gain K is added to
the rest and repeating the calculations at the next sampling (27), resulted in the following final control law applied to the
instant [7]. binary batch distillation column using the Real Time BASIC
program:
V. EXPERIMENTAL IMPLEMENTATION AND EXPERIMENTAL RESULTS
Setting the tuning parameters of the above model RR(t ) K (l (t )  L0 ) /(v(t )  V0  (l (t )  L0 )) . (28)
predictive control algorithm as: p1=3 , p2=2, and Usually K is set at 0.2-0.5 according to simulation results [10].
Qj=diag(3,6,1), j=1,2,3, which means the output prediction One batch experimental run of the methanol/water
horizon is 3, the control horizon is 2, and the weighting factor distillation column using the above model predictive control
for the composition of tray 2 (lower tray), tray 4 (top tray) and strategy was carried out. In this batch run, K is taken as 0.5,
reboiler still is 3, 6 and 1 respectively. So here put more the minimum and maximum limit for reflux ratio is set at 0.5
weighting for the top tray composition, in order to reach the and 10.
specification set for the top distillate product. Then taking the The temperature profile of tray 2 (T-2) and tray 4 (T-4),
set point for the composition of tray 2, tray 4 and reboiler still and reflux ratio profile of this batch run is illustrated in Fig.2.
as 0.6, 0.9 and 0.1 respectively, and substituting the linear From Fig. 2, we can see that the changing of top tray
discrete-time state space model (12) and (13) of the distillation temperature T-4 is kept within 0.4ć, which infers that, the
column into the above model predictive control algorithm, the top tray composition, the main control target, is also kept
following control law for the reflux flow rate and vapor flow within a desired narrow range.
rate is obtained:
l (t ) 0.0144  0.2949 x1 (t )  0.013 x 2 (t )  0.0076 x3 (t )  0.0004 x 4 (t ) 7HPSHUDWXUH3URILOH
 0.0094 x5 (t )  0.0215 x6 (t )  0.0329 x s (t )
(22) 
7HPSHUDWXUH ć

v(t ) 0.0231  0.4078 x1 (t )  0.0197 x2 (t )  0.0143 x3 (t )  0.007 x4 (t ) 

 0.0029 x5 (t )  0.0153 x6 (t )  0.027 x s (t ) 
(23) 

As only the temperature of tray 2, tray 4 and reboiler still
      
is directly measured, so only the composition of tray 2, tray 4
7LPH VHF
and reboiler still can be estimated using (11). In order to
implement the control law (22) and (23) on this batch 7 7
distillation column, the following approximation is made for
estimating the composition of trays with no temperature a. Temperature profile
measurements using the nearby trays with temperature
measurements, let: 5HIOX[5DWLR3URILOH

x1 (t ) x2 (t ), x3 (t ) x2 (t ) / 2  x4 (t ) / 2,

x5 (t ) 2 x4 (t ) / 3  xs (t ) / 3, x6 (t ) 2 xs (t ) / 3  x4 (t ) / 3. (24)
5HIOX[5DWLR


Substituting (24) into (22) and (23), the following control law

for the reflux flow rate and vapor flow rate is obtained:

l (t ) 0.0144  0.3117 x2 (t )  0.0092 x4 (t )  0.0504 x s (t )
(25)       

v(t ) 0.0231  0.4346 x2 (t )  0.0071x4 (t )  0.0382 xs (t )
(26) 7LPH VHF

From (25) and (26), the reflux ratio RR (t) can be calculated 5HIOX[5DWLR
as:
b. Reflux ratio profile
RR(t ) (l (t )  L0 ) /(v(t )  V0  (l (t )  L0 )) . (27)
Fig. 2 Temperature profile and reflux ratio profile of model predictive control
experiment

6255
Authorized licensed use limited to: Politechnika Gdanska. Downloaded on June 10,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
 In order to compare the above MPC strategy with the
commonly used PI controller, a batch distillation run of PI 

control at the same operation conditions as MPC batch run 
was carried out. Here the parameters of PI control were 

7 ć˅

chosen by simulation experiments using Matlab [10], the

proportional gain K=100, the integration time TI=20 s, and 
the set point of tray 4 composition SP=0.9. The tray 4 (T-4) 

temperature profiles of the PI control batch run and the MPC 
batch run are presented together as shown in Fig.3. Fig.3       

shows that the temperature deviation of PI control is 0.7-0.9 7LPH VHF

ć, and the temperature deviation of MPC is only 0.2-0.3 ć.

3, 03&
The final measurement report proved that the methanol
composition of top distillate product of MPC batch run is
98.9%, fairly above the 98% product specification, while the Fig.3 Temperature profile of MPC and PI control experiment
methanol composition of top distillate product of PI batch run
is only 93.43%.
REFERENCES
VI. CONCLUSIONS [1] U.M. Diwekar, Batch distillation, Simulation, Optimal design and
The above experimental results show that this linear Control, London:Taylor & Francis, 1995.
[2] E. Quintero-Marmol, W. L. Luyben, and C. Georgakis, “Application of an
discrete-time state space model based MPC strategy can extended Luenberger observer to the control of multicomponent batch
perform smooth and accurate control on the nonlinear and distillation”, Industrial and Engineering Chemistry Research, vol.30,
non-stationary distillation process of this methanol/water no.8, pp.1870–1880, 1991.
binary batch distillation column. Using MPC, the deviation of [3] A. M. Frattini-Fileti and J. A. F. Rocha-Pereira, “The development and
experimental testing of two adaptive control strategies for batch
tray 4 temperature is much smaller than the deviation of PI
distillation”, in Distillation and absorption ’97, R. Darton Ed., Rugby,
control at the same operation condition. Furthermore, by UK: IChemE, 1997, pp. 249-259.
conducting MPC, more qualified methanol product was [4] P. Li and G. Wozny, “Tracking the predefined optimal policies for
collected due to its narrow control deviations. This MPC multiple-fraction batch distillation by using adaptive control”, Computers
and Chemical Engineering, vol.25, no.1, pp. 97–107, 2001.
strategy is very feasible and effective to control the batch
[5] R. Monroy-Loperenaa and J. Alvarez-Ramirezb, “A note on the
distillation column. This also proves that the linear discrete- identification and control of batch distillation columns”, Chemical
time state space model of the distillation column, which is Engineering Science, vol.58, pp.4729 – 4737, 2003.
built by analysis, linearization and discretization of dynamic [6] S. Huang, K. K. Tan and T. H. Lee, Applied Predictive Control, London:
mass balance and vapor-liquid phase balance equations of the Springer, 2001.
[7] E.F. Camacho and C. Bordons, Model Predictive Control. London:
distillation column, can represent the dynamics of the batch Springer, 1998.
distillation column quite well around the steady-state total [8] E. C. Dechechi, L. F. L. Luz Jr, A. J. Assis, M. R. W. Maciel, and R.
reflux distillation operation conditions. Maciel Filho, “Interactive supervision of batch distillation with advanced
However, in order to control the nonlinear and non- control capabilities”. Computers and Chemical Engineering, vol.22,
Suppl., pp.S867–S870, 1998.
stationary batch distillation process better, future work will [9] N.L. Ricker and J.H. Lee, “Nonlinear model predictive control of the
concrete on the building of nonlinear process model, on-line Tennessee Eastman challenge process”, Computers and Chemical
real-time composition estimation using extended nonlinear Engineering, vol.19, no.9, pp.961-981, 1995.
Kalman filter, and the development and implementation of [10] Z. Y. Zou, D. H. Yu, “Modeling and advanced control of a
methanol/water binary batch distillation column”, Control Engineering of
more effective nonlinear MPC strategies. China, vol.12, No.7, pp.166-170, 2005.

6256
Authorized licensed use limited to: Politechnika Gdanska. Downloaded on June 10,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.