41

CHAPTER 3

MODELLING OF DISTILLATION COLUMN

In this chapter, the basis of distillation, need for distillation control

and different control techniques are described. Model of WB and Skogestad
column is presented. Neuro model and fuzzy model identifications are also
explained.

3.1 DISTILLATION

Distillation is one of the most important unit operations in chemical

engineering. The aim of a distillation column is to separate a mixture of
components into two or more products of different compositions. The
physical principle of separation in distillation is the difference in the volatility
of the components. The separation takes place in a vertical column where heat
is added to a reboiler at the bottom and removed from condenser at the top. A
stream of vapour produced in the reboiler rises through the column and is
forced into contact with a liquid stream from the condenser flowing
downwards in the column. The volatile (light) components are enriched in the
vapour phase and the less volatile (heavy) components are enriched in the
liquid phase. A product stream taken from the top of the column therefore
mainly contains light components, while a stream taken from the bottom
contains heavy components.
 42

3.1.1 Distillation Equipment

A simple continuous binary tray distillation column for separating a

feed stream into two fractions, an overhead distillate product and a bottoms
product is shown in Figure 3.1. The inside of the column is normally provided
with horizontal plates or trays. The liquid mixture to be separated is
introduced more or less centrally into a vertical cascade of trays. A reboiler is
provided at the bottom of the column to supply the heat required for the
vaporization involved in distillation and also to compensate for heat loss. A
water-cooled or air-cooled condenser is provided at the top of the column to
condense and cool the overhead stream. The purity of the top product can be
improved by recycling some of the externally condensed top product liquid as
reflux from the upper part of the column.

Figure 3.1 Control scheme of binary distillation column

 43

The more reflux that is provided, the better is the column separation
of the lower boiling from the higher boiling components of the feed. The feed
tray divides column into two parts namely rectifying section and stripping
section. In rectifying section, the vapour rising is rectified with liquid flowing
down from top to remove less volatile component and in stripping section the
liquid is stripped of volatile components by vapour produced at bottom by
partial vaporization of bottom liquid in reboiler. The condensed liquid that is
removed from reflux drum is known as distillate or top product and the liquid
removed from reboiler is known as bottom product.

3.1.2 Need for Distillation Control

Distillation is used in many process industries for separating feed

streams and for purification of final and intermediate product streams. There
are many reasons for the interest in distillation control. From an academic
point of view distillation control is an interesting multivariable problem, and
from an industrial point of view improved distillation control has a potential
to substantially increase the profit. Distillation accounts for approximately
95% of the separation systems used for refining and in chemical industries. It
has a major impact upon the product quality, energy usage, and plant
throughput of these industries. It consumes enormous amounts of energy, both
in terms of cooling and heating requirements. It can contribute to more than
50% of plant operating costs. The energy requirement may be reduced
significantly through improved operations. This is achieved not only through
optimal column design, but requires, in addition, a control system which is
able to maintain the optimal conditions.

Distillation control is a challenging endeavour due to (1) the

inherent nonlinearity of distillation, (2) multivariable interaction, (3) the non-
stationary behavior and, (4) the severity of disturbances (Shinskey 1984).
Tighter control of distillation columns is consequently important for energy
 44

savings, and will also yield increased profit through improved product
recovery. The major benefits of improved distillation control are reduced
energy consumption, increased yield and higher throughput.

3.1.3 Distillation Control Techniques

Distillation columns provide a very challenging example within the

field of process dynamics and process control. Traditionally PID controllers
were used in the process industries for control of the distillation column. The
main drawback of the conventional feedback PID control is that corrective
action for disturbances does not begin until the controlled variable deviates
from the set point (Skogestad 1997). In industry, most of the columns are
operated by SISO controllers and usually only one composition is
automatically controlled (one point control). This leads to waste of valuable
products and excessive energy. However, automatic control of both
compositions may be very difficult to obtain due to strong interaction between
top and bottom product compositions (Shinskey 1984).

Skogestad et al (1988) have reported that high purity columns, i.e.

columns where both top and bottom compositions are very pure, suffer from
strong interaction which makes the system very sensitive to small changes in
the manipulated variables (input uncertainty). Without a rigorous method for
dealing with uncertainty it may be practically impossible to tune a two point
controller for a system with strong interaction. This may in fact be one of the
reasons to why one point control is so commonly used. Another disadvantage
with such a decentralized (multiloop) control is that the control performance
may seriously deteriorate if the system hit some constraints. For example, if a
stabilizing loop saturates, the system goes unstable. To avoid this, the plant
has to be operated sufficiently far away from the constraints, or facilities for
reconfiguration have to be installed ‘on-top’ of the SISO controllers
(Lundstrom and Skogestad 1995).
 45

Configuration selection is an important aspect in the case of

multiloop controller design. Control configuration for a distillation column
can be selected from the knowledge of the thermodynamic parameters, reflux
ratio, vapor boil-up rate and distillate to bottoms ratio for binary and
multicomponent distillation (Stilchmar 1995). Improper choice of
manipulated/controlled variable pairings can result in poor control
performance. Decouplers are introduced into the multiloop configuration to
compensate for the process interactions and reduce the control loop
interactions. Hurowitz et al (2003) have used decouplers to control the top
product composition using reflux flow rate, bottom product composition using
vapour boil-up rate for the xylene/toluene column and the depropanizer. In
both cases, the decouplers resulted in improved control performance
compared to the feedback controllers without a decoupler.

The insufficient performance of SISO controllers lead to the

development of specialized single loop control strategies such as feed-forward
control (Broll et al 1995), inferential control (Zhang and Agustriyanto 2001),
cascade control (Kano et al 2000), adaptive control (Natarajan et al 2006) etc.
The abilities of the specialized single-loop control strategies and multiloop
controllers were not satisfactory for increasingly stringent performance
requirements of the chemical processes which led to the development of
multivariable control techniques.

3.1.4 Multivariable Controllers

Processes which are multivariable in nature, i.e. processes where

the variables to control and the variables available to manipulate cannot be
separated into independent loops, where one input only would affect one
output, constitute a major source of difficulty in process control. These
processes show a certain degree of interaction, i.e. the change in process
 46

variables of one control loop affects that of other loop. The complexity of the
control problem increases as this interaction increases (Luyben 1992).
Multivariable processes in industrial and other applications are often of higher
order, where there are many, possibly tens or hundreds, of control loops
interacting (Postlethwaite and Skogestad 1996).

The term multivariable control refers to the class of control

strategies in which each manipulated variable is adjusted on the basis of errors
in all of the controlled variables, rather than the error in the single controlled
variable, as in the case of multiloop control. Multivariable control is
particularly well-suited for controlling processes with several interacting
controls which need to be simultaneously decoupled.

An adequate model is generally considered as a prerequisite for

multivariable controller design. The model is used to predict the behaviour of
the controlled variables with respect to changes in the input variables (Sagfors
and Waller 1998). Established multivariable control techniques rely on the
availability of the linear system models. This is to ensure that the resulting
control scheme is closely matched to the dynamics of the process. The
multivariable system must therefore first be modelled either analytically using
set of differential equations to describe the system behaviour or empirically
by fitting experimentally obtained data to an assumed structure of the process
i.e. black-box modelling. Obviously, how well the resulting control strategy
performs depends on the accuracy of the model. In applications where the
physical and/or chemical characteristics of the system are well known, usually
the former approach is adopted. In the process industries, where the higher
degree of uncertainty prevails about the process behaviour, empirical
modelling approach is often employed. However for control system design
purposes, the input-output (transfer function) model obtained using later
 47

approach is generally adequate. Multivariable controls strategies can also be

developed that include integral, derivative and feed-forward control action.
Among the multivariable controllers, MPC is an important advanced control
technique which can be used for difficult multivariable control problems
(Perez et al 2001).

3.2 MODEL OF DISTILLATION COLUMN

In the present work, two distillation column models are taken for
case study. The first example is Wood and Berry (WB) column which is in
the form of transfer function model and second example is based on
Skogestad model.

3.2.1 Wood and Berry (WB) Model

The first 2 x 2 MIMO process is presented by Wood and Berry

(1973). The study was performed on a 9 inch diameter, 8 tray column
equipped with a total condenser and a basket type reboiler. The required
control action for the manipulative variables in the composition loops, reflux
and steam flow, were cascaded to the set points of the appropriate flow
controllers. The transfer function characterizing the column dynamics were
established by pulse testing. Parameters of the assumed first order plus time
delay transfer function were determined from the transient data. The time
delays were established and the gains and time constants are determined by
least squares fit employing Rosenbrok’s direct search technique. The process
transfer function matrix of the distillation process is given by

12.8e s 18.9e 3s 3.8e 8.1s

X D (s) 16.7s 1 21s 1 R(s) 14.9s 1
G(s) (3.1)
F(s)
X B (s) 6.6e 7s 19.4e 3s S(s) 4.9e 3.4s
10.9s 1 14.4s 1 13.2s 1
 48

where XD(s) and XB(s) are the overhead and bottom compositions of methanol,
respectively. R(s) is the reflux flow rate and S(s) is the steam flow rate to the
reboiler, F(s) is the feed flow rate, a load disturbance.

3.2.2 Skogestad Distillation Column Model

The distillation column used in this work is based on Skogestad

model (Skogestad 1997). The column consists of N stages (trays), numbered
from bottom to top. The feed enters the column at stage NF, with 1 < NF < N.
The feed flow, F is a saturated liquid with feed composition zF [mole fraction]
and feed liquid fraction qF. LT denotes the reflux flow rate of the condenser;
VB is the boilup flow rate of the reboiler. The top product consists of a
distillate flow rate D, with composition XD. Likewise the bottom product
consists of a bottom flow rate B, with composition XB.

The stages positioned above the feed stage define the enriching
section and those below are the stripping section of the column. The material
balance equations for the feed stage and the stages in the stripping section of
the column are affected by the continuous flow to the column and the
withdrawal of the bottom product from the reboiler.

The vapour flow rates are assumed to be constant molar flows with
no vapour dynamics. They are given by

VB (1 q F ) F
Vi (3.2)
VB

VB (1 q F ) F forN F i N
Vi (3.3)
VB for1 i N F
 49

The constant qF is the feed liquid fraction and is determined by the

feed thermal quality. The liquid flow rates in the stripping and enriching
section assume linearized tray hydraulics with a time constant . They are
defined as

LT top of column i N

Mi M i* (3.4)
Li L* enriching section NF i N

Mi M i*
L*B stripping section 1 i NF

where M*(i) is some nominal stage holdup at tray i, L* [kmol / min] is a

nominal reflux flow in the enriching section and L*B L * q * F * [kmol /
min] is a nominal liquid flow in the stripping section. The distillate and
bottom product flow rates are

D VT LT
(3.5)
B L B VB

Let Li and Vi [kmol / min] denote the liquid and vapour flow on
stage I of the column. Denote by Xi and Yi the liquid and vapour compositions
of the light component on stage i, respectively. Further, let Mi, denote the
liquid hold-up on the i-th stage. The vapour-liquid equilibrium describes the
relation between the vapour and liquid compositions Yi and Xi on each stage i
of the column and is given by the non-linear expression

Xi
Yi , i 1, . . . , N (3.6)
1 ( 1) X i

where is the so called relative volatility (dependent on the product) the total
material balance on the various stages is given by the differential equations
 50

dM N
VN 1 L N D condensor stage
dt
dM i
Li 1 Li Vi 1 Vi enriching section N F i N
dt
dM F (3.7)
L NF 1 LN F V N F 1 VN F F feeder stage
dt
dM i
Li 1 Li Vi 1 Vi stripping section 1 i NF
dt
dM 1
L2 V1 B reboiler stage
dt

here LN = LT, VN = VT and MN = MT at the top of the column and L1 = LB,

V1 = VB, M1 = MB at the bottom of the column.

The material balances for the component holdup of the column are
given as follows.

dMT X N
VN 1YN 1 (LN D) X N condensorstage
dt
dMi X i
Li 1 X i 1 Vi 1Yi 1 Li X i Vi Yi NF i N
dt (3.8)
dMF X NF
LN F 1 X N F 1 VN F 1YNF 1 LNF X N F LNF YNF FzF feed stage
dt
dMi X i
Li 1 X i 1 Li X i Vi 1Yi 1 Vi Yi 1 i NF
dt
dM1 X 1
L2 X 2 VB Y1 BX1 reboilerstage
dt

using the chain rule for differentiation, the liquid composition Xi on the i-th
stage then satisfies

d (M i X i ) d (M i X i )
Xi
dX i dt dt (3.9)
, i 1, . . . . , N
dt Mi

The distillation column specifications used for this study are listed
in Table 3.1.
 51

Table 3.1 Distillation column data

Number of trays 40
Relative volatility between light and heavy component 1.5
Feed tray (numbering from the bottom) 21
Feed composition (light component molar fraction) 0.5
Distillate composition (mole fraction) 0.99
Bottoms composition (mole fraction) 0.01
Feed flow rate (kmol/min) 1
Distillate flow rate (kmol/min) 0.5
Bottom flow rate (kmol/min) 0.5
Reflux flow rate (kmol/min) 2.706
Boilup flow rate (kmol/min) 3.206
Condenser holdup (kmol) 10.9
Reboiler holdup (kmol) 15.8
Fraction of liquid in feed 1
Average tray holdup (kmol) 0.5

3.2.3 Neural Modelling of the Distillation Column

Prior to design of the neuro-fuzzy controller an identification

process is accomplished in order to predict the plant dynamics (Fernandez et
al 2000). A simulated binary distillation column plant based on the nonlinear
equations has been used to obtain representative data for training process, and
an RBF neural network has been used as an identification model of the
distillation column shown in Figure 3.2.
 52

z
xB
R
V Distillation
Column xD

TDL

-
xB

Neural
Identifier xD

Figure 3.2 Neuro Identification scheme of distillation column

To obtain representative data, varying feed flows, initial liquid

composition values both in the column, rebolier and condenser along with
input values for the control actions were imposed on the model. The
identification model has been carried out using RBF given by
y[k 1] N 1 ( y[ k ], u[ k ]) , with N1 belonging to the class trained with the

Orthogonal Least Square (OLS) algorithm, using 750 patterns for

u[ k ] [ R[ k ], V [ k ]] and y[k ] [ x B [ k ], x D [ k ]] regularly spaced, covering the

operating range of the state variables. The neural identifier has been utilised
to approximate the Jacobian of the plant in order to adjust the neuro-fuzzy
controller’s parameter.
 53

3.2.4 Fuzzy Modelling of the Distillation Column

Fuzzy Logic does not deal with dynamic elements like other
conventional logic, and therefore careful consideration is required for
designers to construct such model rule forms that can express reasonably
dynamic elements of time-varying systems Along with this consideration, an
unique model is developed that imitates verbal understanding of operators
against dynamic behaviours of process (Yamazaki 1996). The rule form used
represents explicitly the relation between past control actions as causes and
process responses as results, and the control actions are defined by three fuzzy
variables, long past, medium past and short past actions. Another
consideration in the model rule is the partial pairing of j-th input and k-th
output among J sets of inputs and K sets of outputs in order to prevent an
explosive increase in the total number of model rules required for MIMO
systems.

The Fuzzy variables Uj[n+l] and Zj[n] denote a change in the j-th
control action and a change in the past j-th control action respectively. Yj,k[n+l
In] and Ek[n] denote a predicted change in k-th process by Uj and a control
error. The symbol [n] denotes a sampling instance, and Y[n+l | n] represents
the change in the 1 -sampling future predicted at n-sampling instance.

The Fuzzy labels ( Aij , B ij , C ij ) denote fuzzy labels of Z={z}, and

Q ij ,k denotes fuzzy labels of Y=(y}. Here, a fuzzy linear relation is assumed

between fuzzy labels Q ij ,k and { Aij , B ij , C ij }

The Past control actions zTi, are classified in accordance with fuzzy
labels Ti (Tl=S, T2=M, T3=L). Superscripts L, M, and S on Z denote the
change in long, medium and short past control actions respectively.
 54

Each past control action zT1 corresponding to fuzzy labels is

calculated by

M M
zTi [ n] {u[n m]. (m)} / Ti ( m) (3.10)
m 0 m 0

By means of inclusion of U[n+l] into the short past control action ZS[n+l], the
i-th rule for j-th process input and k-th process output among N sets of rules is
given by (3.11)

R(i j ,k ) : If ( Z Lj [n+l ] = Aij , Z M i S i

j [n+l ] = B j , Z j [n+l ] = C j

then Y j ,k [n+l | n] = Q ij, k (3.11)

Q ij ,k = { L
G ij ,k * Aij , M
G ij ,k * B ij + S
G ij ,k * C ij }

where L G ij ,k , M
G ij ,k , S
G ij ,k are scalar and the symbols “ + ” and “ * ”
represent fuzzy sum and multiplication respectively.

Now, for the calculation of predicted process change from the

model rules above defined, Sugeno's simplified method can be applied as
follows:

0
Q ij ,k [ n 1 | n] G ij ,k * 0 X j [n+1] (3.12)

and the degree of compatibility of the above rule is defined as

X ij [n] min{ Aij ( 0 z Lj [n]), B ij ( 0 z Mj [ n]), C ij ( 0 z Sj [ n])} (3.13)

The predicted k-th process output inferred from all the rules of j-th
input is calculated by

N N
0 0
Y j ,k [ n 1 | n ] { Q ij ,k [n 1 | n] X i
j [ n] / X ij [n]} (3.14)
j 1 j 1
 55

3.3 CONCLUSION

In this chapter, the need for distillation column control, different

control techniques used are presented. The mathematical model of distillation
column used in this work is also given. The WB column is presented in
transfer function form and Skogestad column is given in state space form. The
neural and fuzzy modelling approach is also briefly explained.