Chemcal Enyinveriny S~ienr-r, Vol. 44, No. 11, pp. 2571-2581, 1989. C005-2509/89 53.00+0.

00
Printed in Great Britain. ,C 1989 Pergamon Press plc

DYNAMICS AND IDENTIFICATION OF A BINARY

DISTILLATION COLUMN

HENRIK WEISBERG ANDERSEN, MOGENS KUMMEL and STEN BAY JBRGENSEN

Department of Chemical Engineering, The Technical University of Denmark, DTH-229, DK-2800 Lyngby,
Denmark

(First receitxd i August 1988; accepted in recised@m 1 February 1989)

Abstract-Recent work in the area of control

design has shown that the degree of gain directionality has a
significant impact on achievable robust
performance. Thus, it is important to obtain process models which
give a satisfactory description of gain directionality. In this paper a parametric identification algorithm is
applied to a case study of binary distillation. When the actuators are perturbed simultaneously this
approach yields models which give a satisfactory description of the high- and low-gain directions. However,
a common practical approach for modeling of distillation columns is to perturb the inputs separately and
then combine the models into an overall description of the plant dynamics. It is shown that this method
might have the drawback that the obtained model yields a poor description of the low-gain direction in the
distillation plant. This indicates that modelling of multivariable processes can be improved by application of
multivariable identification algorithms.

1. INTRODUCTION changes in the internal flows and changes in the

In this paper a multivariable frequency domain tech- external flows for the column.
nique is used for analysis of: The mode1 used in the first part of the paper is based
upon direct linearization around the steady-state op-
-open-loop dynamics erating point. This model is well suited for theoretical
-model parameter determination analysis of the dynamics, but is not applicable for
practical modeling. Therefore the second part of the
for a binary distillation column. Recent work in the paper is devoted to the practical problem of how to
area of distillation column control has shown that the obtain a model suitable for control design. A discrete-
degree of gain directionality has a significant impact time parametric identification algorithm proposed by
on the control design problem (Skogestad and Morari, Goldschmidt et al. (1985) is applied to the case study of
1988b). This holds both for the selection of control binary distillation. The basic idea in the approach is to
strategy and for the tuning of the corresponding obtain the structure in a MLMO ARMAX model. This
control parameters. Thus, it is important to insure that structure is obtained by analysis of the fundamental
the model used for control design gives a satisfactory physical laws which determines the dynamics of the
description of the gain directionality in the plant. process. Hereafter the parameters are identified by
Previous work has shown that frequency domain pseudo random-binary perturbations of the inputs
singular value decomposition (SVD) is a suitable tool and application of an extended, recursive, least-
for analysis of gain directionality (Lau et al., 1985; squares algorithm.
Moore, 1986; Skogestad and Morari, 1988a; Bequette For experimental determination of model par-
et al., 1987). The method applied in the present paper ameters for chemical processes it is common to per-
is based on singular value decomposition of transfer- turb the inputs separately and then combine the
function matrices as originally proposed by obtained STMO models into one multi-model which
MacFarlane and Scott-Jones (1979). The basic idea is describes the multivariable couplings in the process. In
to obtain upper and lower bounds on the gain of a this paper this approach is applied in the parametric
transfer-function matrix. Moreover, the analysis out- identification by estimation of model parameters for
lines which combinations of the inputs will cause separate perturbations of the actuators. As discussed
respectively the upper and lower gain in the plant and in Skogestad and Morari (1988a) the question is
how these affect the outputs. whether independent determination of the individual
In order to describe the impact of gain directionality transfer-functions is sufficient in order to obtain a
on the quality of identified models the first part of the suitable model of the gain directionality. Thus, it is
paper reviews the fundamental dynamic behavior of interesting to examine whether simultaneously pertur-
binary distillation as described in Skogestad and bations are necessary in order to obtain reliable
Morari (1988a). The model used is the continuous- parameter estimates for description of the dynamics of
time linear model based on ideal component balances. the binary column.
It is obtained by direct linearization around the The discrete-time version of the open-loop fre-
operating point. The analysis reveals the fundamental quency analysis can be used to address this problem.
difference between the open-loop response towards This is due to the fact that the analysis shows to which

2571
 2572 HENRIK WEISEIERG ANDERSEN et al.

extent the obtained models describe the multivariable WE crnx m: left singular matrix: WH W = I,
dynamics of the process. In this case where the best
VE Ckxk: right singular matrix: VHV = It
discrete-time model is known, this is done by evalu-
ating to which extent the obtained models describe the E E CmXk: singular value matrix
high-respectively low-gain directions in the distil-
where E is a diagonal type of matrix containing the
lation column example. A major conclusion is that the
singular values in decreasing order. The gain of the
multi-SIMO identification approach only yields a
mapping in eq. (2) is then limited by:
poor description of the low-gain direction whereas the
MIMO approach yields a satisfactory description of
the low-gain direction. The significance of this result is
that identification by simultaneous perturbation of the
(4)
inputs, may be needed in order to obtain models which
gives a correct description of the multivariable dy-
namics, which again is a prerequisite for the success of
the subsequent control-design. Moreover, the fre- where C?is the maximum and a the minimum singular
quency analysis is used to check whether the selected value of G.
sample time is suitable. This is a non-trivial problem Note that in the case where kr m, one cannot use
since separate perturbations in the inputs, which are the minimum singular value as a lower bound since
often used to select the sample time, not necessarily there exists a vector space of dimension k - m which is
affects the fast time constants in the process. mapped into zero and thus the minimum gain is not
This case study reveals that singular-value de- given by the minimum singular value.
composition provides a useful tool for the under- From eqs (2) and (3) we have:
standing of multivariable process dynamics and there-
fore also yields guidelines about which identification Gvi = rsiwi (5)
and control schemes that should be used.
where triis the ith right singular vector corresponding
to bi. Thus, vi gives information about the changes in
2. MULTIVARIABLE FREQUENCY-DOMAIN ANALYSIS the inputs which achieve the gain bi. In a similar way
This section contains a brief introduction to the wi gives information about the effect that the gain vi
multivariable frequency domain analysis method used has on the outputs. Thus, the right respectively left
in the case study of distillation column dynamics and singular vectors form a basis for a directional gain
identification. A more detailed description is given in analysis. These vectors are interpreted by means of
Andersen (1988). gain and phase.

Input gain:
2.1. System representation
The system must be described in a standard state- 5: rj = lV(j)lz, j = 1, k (6)
space form. 0: xj = Ip(j)12,j= 1, k. (7)
6 {x(t) > = Ax(t) + Bu(t) + Dq(t)
output gain:
y(t) = z-x(t) + Fu(t) (1)
5: Tj = I*(j)l*, j = 1, m 6)
where S is the differential operator in continuous time
and the difference operator in discrete time. a: ‘yj = Iw(j)(*, j = 1, m (9)
where V(j) denotes thejth element of the right singular
x : n : states.
tl: k : actuators.
vector V and W(j) the jth element of the left singular
vector W corresponding to singular value 5. In a
4 : ad : external disturbances.
y : m : outputs. similar way v(j) denotes the jth element of the right
singular vector v and w(j) the jth element of the left
A : n x n : state matrix. singular vector i corresponding to singular value 0.
B : n x k : control-input matrix.
D:nxn d : disturbance-input matrix. Input phase lag:
T: m x II : output matrix.
5: fj= arctan{im[5(j)+H]/re[fi(j)(pH]),
F : m x k: direct couplings from actuators.
j=l, k (10)
F: I-i = arctan {im[~(j)$x]/reCo(j)~H]},
2.2. Multivariable gain and phase lag
Consider the following linear equation: j=l,k (11)
y = Gu,yEcm, UE@k, GECrnXlr. (2) Output phase lag:
The singular value decomposition of G is given as: 5: Vi= arctan (im[*(j)$r]/re[*(j)(PH]},
G= WEl’ (3) j = 1, m (12)
 Dynamics and identification of a binary distillation column 2573

I: yj = arctan {im[w(j)$r]/re[r(j)yH]},

j = 1, m (13)

where (p and q are reference sine functions, i.e. they

have phase lag zero. They are selected to be the first
element of v respectively g which has a nonzero gain,
i.e.:

G = U(fi) (14)

40
- = r(n)
- - ‘6
fi: first element in U such that Ta # 0

p: first element in g such that y_”# 0.

2.3. Open-loop frequency analysis

The open-loop frequency analysis is performed by
means of the transfer-function matrices from actuators
respectively disturbances to the outputs, i.e.:

y(cIwI) = G,q,(c{wI) u(c{wI) (15)

Fig. 1. Distillation column configuration.
y(c{oI) = G,O,(c{4)q(c{oI). (‘6)

These transfer-function matrices are easily deduced Table 1. Operating point

from the state-space representation given in eq. (1).
C(W) is a frequency dependent function. The function H, reboiler holdup, 150 mole
depends on whether the analysis is performed in H, tray holdup. 15 mole
H, condenser holdup, 75 mole
continuous or discrete time.
L, feed flow, 10 mole/min
X, feed concentration, 0.5 mole-ratio methanol
Continuous time:
L reflux, 20.15 molejmin
cfw} = jw, i.e. the imaginary axis. (17) V boilup, 25.15 mole/min
B bottom product, 5 mole/min
Discrete time: D top product, 5 mole/min
x=: 0.03 14 x6: 0.5952
c{w> = exp(jwT), w = [0, n/T], i.e. the unit circle x, : 0.0700 x,: 0.7123
x2: 0.1354 x8: 0.8079
where T is the sample time. .x3: 0.2334 x9: 0.8790
By means of the method outlined in Section 2.2 the x4: 0.3542 xc-: 0.9300
upper and lower singular values of G:” and G$ are x5: 0.4722
calculated as functions of the frequency. Moreover,
the directional gain and phase lag analysis, eqs (6x13)
are performed as functions of the frequency.
-the thermodynamics involved in the vapour-liquid
equilibrium is described by UNIFAC.
3. CASE STUDY: BINARY DISTILLATION
3.1. The distillation column model Reflux and boilup are selected as manipulated vari-
This case study considers a binary distillation ables. Furthermore disturbances in feed-flow rate and
column separating methanol and isopropanol. The feed concentration are modeled. The top product
column consists of nine trays, reboiler and total concentration equals the accumulator concentration,
condenser. The coiumn configuration is shown in Fig. and the bottom product concentration equals the
1 and the steady-state column operating point is listed concentration on tray 1 (see Fig. 1).
in Table 1. The non-linear simulations are based on
the following assumptions: 3.2. Open-loop frequency analysis
A linear continuous time state-space model was
Nonstant molar flow rates, obtained by linearization around the operating point
-liquid feed at its bubble point, listed in Table 1. The state-space matrices of the linear
-a tray efficiency of 100%, model are given in Table 2. The transfer-function
-the liquid holdsup on each tray, in reboiler, and matrix between actuators (reflux and boilup) and the
condenser are constant and perfectly mixed, product compositions were evaluated by means of the
-the holdup of vapour is neglected, open-loop frequency analysis outlined in Section 2.3.
-the reflux stream is not subcooled, The results of this analysis is shown in Fig. 2(Aa).
-no deadtimes are included, The gain plots in Fig. 2(A) show that there are a
-the distillation is performed at 1 atm. No pressure significant difference between the high- (maximum
gradient from the reboiler to the condenser is singular value) and low-gain direction (minimum
included, singular value) in the plant. The difference in gain is of
2574 HENRIK WEISBERG ANDERSEN et al.

Table 2. Linear continuous time state-space model

Y,: liquid-mole ratio methanol in condenser

x2: liquid-mole ratio methanol on tray 9 (top tray)
x3: liquid-mole ratio methanol on tray 8
x4: liquid-mole ratio methanol on tray 7
x5: liquid-mole ratio methanol on tray 6
x6: liquid-mole ratio methanol on tray 5, feed tray
x7: liquid-mole ratio methanol on tray 4
r,: liquid-mole ratio methanol on tray 3
xg: liquid-mole ratio methanol on tray 2
xX0: liquid-mole ratio methanol on tray 1
xI1: liquid-mole ratio methanol on in reboiler
ul: reflux in mole/min
uZ: boilup in mole/min
4 I : feed-flow rate, mole/min
q2: mole ratio methanol in feed
-
-0.335 0.192
1.343 - 2.301 0.973
1.343 -2.317 1.035
1.343 - .2.378 1.174
1.343 -2.518 1.407
A: 1.343 -3.417 1.736
2.010 - 3.746 2.223
2.010 -4.233 2.776
2.010 -4.786 3.252
2.010 5.262 3.584
0.168 - 0.358
J
Unit: min-’

o.oooo
0.0034
0.0000
- 0.0027
o.oooo
o.oooo
o.oooo
o.oooo
0.0047 -0.0038 o.oooo o.oooo
0.0064 -0.0051 o.oooo o.oooo
0.0078 -0.0063 o.oooo o.oooo
B: 0.0082 -0.0073 D: 0.0019 0.6667
0.0079 -0.0094 0.0079 o.OOOo
0.0081 - 0.0097 0.0081 o.Oooo
0.0065 -0.0078 0.0065 o.ooOO

l-0.0044
0.0000
-0.0052
0.0000

Unit: mole ratio methanol/mole Unit: [mole

-
0.0044
o.OoOO
o.Oooo
00000
-
ratio methanol/mole, min - ‘1

the size 100 and the dominant time constant of the of the product flows are constant but the ratio of the
maximum gain is in the region of 48 min whereas it is 8 top-product flow relative to the bottom-product flow
min for the minimum gain curve. In order to examine is changed. From Fig. 2(D) it is seen that the minimum
the directional dependence of the maximum and mini- gain is achieved when the changes of equal size in
mum gain curves [Fig. 2(A)] on changes in the actu- reflux and boilup are in phase, i.e. when the reflux is
ators, a directional analysis as given in eqs (6) and (7) increased (decreased) the boilup must increase (de-
and eqs (10) and (11) were performed. The results of crease) with the same amount. This corresponds to
this analysis are shown in Fig. 2(8-D). Figure Z(B) changes in the internal flows keeping the individual
shows the size ofchanges in the two actuators required product flows constant. These results approximately
in order to reach the maximum gain curve in Fig. 2(A). holds for the frequency range of interest, i.e. up to the
The size of changes required in order to reach the maximum bandwidth.
minimum gain curve is not shown explicitly since for a In order to examine how the maximum respectively
2 x 2 system these are obtained by reversing the minimum gain affects the product concentrations a
indices in Fig. 2(B). Thus, both the maximum and directional analysis according to eqs (8) and (9) and
minimum gain occur for equal changes in the actu- eqs (12) and (13) were performed. The results are
ators. Figure 2(C) shows the phase between actuator shown in Fig. 2(E-G). As seen from Fig. 2(E) the effect
changes which are required in order to obtain the of the maximum gain is relative high on the bottom
maximum gain. As seen the changes must have reverse product concentration (I W(Z)l’ = 0.72) compared to
phase, i.e. whenever the reflux is increased (decreased) the bottom product concentration 1ti( 1) I* = 0.28). As
the boilup must be decreased (increased) by the same described for the input analysis the size of the effect of
amount. This corresponds to the case where the sum the minimum gain is obtained by exchange of the
 Dynamics and identification of a binary distillation column 2575

0
Al

100 180 la0

SO 6 140 D

I=4
140

E3
a0 100 100
70
60 60
60 2
20 20 1
SO
--2o -20
00
1 2
30
-60 -60

20 -100 100
10 140
0 i-0
3 -2 -1 0 3 -2 -1 s

100 lB0 180

90 E 140 F la0

El --I-
.SO 2 100 100
70
60 60
60
20 20
SO
-20 -20
00
30 -60 2 -60

20 1 -¶OO LOO

10 -100 100
t -
0 -1eo 190
3 -2 --i 0 - 7 --z -1 0-

Fig. 2. Open-loop analysis of outputs towards actuators. x-Axis =log frequency. y-Axis A =log gain,
BE= %, C, D, F and G =degree phase lag. Index 1: B, C and D = reflux, E, F and G = top-product quality.
Index 2: B, C and D = boilup, E, F and G = bottom-product quality. (A) Upper and lower gain bound.
(B) Input gain in the direction of the upper gain bound. (C) Input phase in the direction of the upper gain
hound. (D) Input phase in the direction of the lower gain bound. (E) Ouiput gain in the direction of the
upper gain bound. (F) Output phase in the direction of the upper gain bound. (G) Output phase in thr
direction of the lower gain bound.

indices in Fig. 2(E). Thus, the minimum gain affects the boilup by equal size and of opposite phase so that the
top product more than the bottom product concentra- sum of the product flows are constant. The minimum
tion (l~(l)(~ = 0.72 and Iw(2)1’ = 0.28). Moreover, direction is relatively fast and occurs when the shape
Fig. 2(F) shows that the changes in the outputs are in of the profile is changed (as for the maximum gain this
phase for the maximum gain (i.e. the position of the only holds approximately) which is accomplished by
concentration profile is changed) and Fig. 2(G) that changing the reflux and boilup by equal size in phase
the changes for the minimum gain are in reverse phase so that the individual product flows are constant.
(i.e. the concentration profile is either stretched or
compressed). These results are summarized in Table 3. 3.3. Modeling and identification
Thus, the maximum gain direction is relative slow and A prerequisite for multivariable control design is a
occurs when the position of the profile is changed (this plant model of a satisfactory quality. Moreover, the
only holds approximately since some change in the approach for modeling and identification must be
shape of the profile is achieved due to the fact that the suited for practical application. A standard approach
effect is highest in the bottom concentration, see for modeling of distillation columns is to perform
Table 3) which is accomplished by changing reflux and open-loop step changes in the manipulated variables

CES 44:11-K
2.576 HENRIK WEISBERG ANDERSEN et al.

Table 3. Results of open-loop frequency analysis

G ‘c CL, VI [Top, bottom]

Max. gain 0.1757 48 co.707, -0.7071 cO.529, -0.8491

Min. gain 0.0018 8 co.707, 0.7077 cO.849, 0.5291

G =gain, mole-ratio methanol/(mole/min).

z = time constant, min.
L = reflux, molejmin.
V = boilup, mole/min.
[L, V] =relative changes required for the corresponding gain.
Top = top product concentration, mole ratio methanol.
Bottom=bottom product concentration, mole ratio methanol.
[Top, bottom] = relative changes accomplished by the corresponding gain.

and thereby obtain approximate parameters in selec- Figure 3 shows a simplified continuous-time model
ted transfer-functions. When the selected manipulated for the column. The main reason for this input-output
variables are reflux and boilup this approach has a selection is that the endpoints should be included in
major drawback. Changes in either reflux and boilup order to obtain satisfactory control of the product
will affect the sum of the product flows. As described in concentrations. Two internal concentrations are in-
Section 3.2 this will affect the maximum gain direction cluded since the sensitivity of the concentration profile
in the plant. Since the difference between the maxi- is relatively high inside the column compared to the
mum and minimum gain directions is a factor 100 the sensitivity of the product concentrations. The feed tray
output responses will only give information about the is included, due to the fact that the disturbances enter
maximum gain direction. Other selections of mani- the column at the feed tray. The reboiler concentration
pulated variables might solve this problem. Section 3.2 is included since the reboiler is a major capacity in the
suggests that one should use the sum and difference column (due to the relatively large holdup). The
between reflux and boilup as manipulated variables. structure of the A, 5 and D matrices in Fig. 3 follow
This is due to the fact that changing L + V keeping L directly from the equations which describe the funda-
- V constant will affect the minimum gain direction mental dynamic behaviour of binary distillation
whereas changing L- V keeping L-t V constant will columns (Nielsen et al., 1988). The corresponding
affect the maximum gain direction. Thus, in principle discrete-time structure is shown in Fig. 4. The intro-
step changes in these manipulated variables will yield duced interactions in Fig. 4 compared to Fig. 3 are due
a model of both the maximum and the minimum to the sample delay. The numerals indicate the inverse
direction. However, in practice it will be difficult to order of interaction, i.e. the interaction is decreased as
change L + V without affecting L - V, i.e. keeping the the number of capacities through which it couples
product flows constant, and since the difference be- increases. The number of parameters to be estimated
tween the maximum and minimum gain is a factor may be reduced by neglecting parameters with the
100, even small changes in the product flows will cause inverse order of 2 or higher. The size of the sample
a large uncertainty in the modeling of the minimum time relative to the characteristic response time deter-
gain direction. mines how suitable this approximation is.
In this case study a parametric identification The following ARMAX structure was used in the
method proposed in Goldschmidt et al. (1985) is identifications:
applied. The basic idea in the approach is to obtain a
qualitative understanding of the interactions between
x(t+T)= Ax(t)+B + Ee(t) + e(t + T) (18)
the capacities in the process. This is based on the
physical laws which determine the dynamics of the
process. This can be used to obtain a structural state- where E is a diagonal matrix and e(t) is an error vector
space representation and therefore also the par- which contains the difference between predicted and
ameters which should be estimated. This represen- measured values of x. The structure of A and B are
tation is transferred into discrete time (by inclusion of shown in Fig. 5. This structure is obtained from Fig. 4
couplings introduced by the sample delay) and a by neglecting couplings with inverse order 2 and
model-reduction technique is applied in order to higher.
reduce the state-vector to a vector of measurable
process outputs. This yields a MIMO ARMAX model 3.3.1. MIMO identification. The parameters in the
(Auto Regressive Moving Average with eXogene in- model in Fig. 5 were estimated with a sample time of
puts). The identification is performed by pseudo- respectively 1 min and 10 min. The corresponding
random binary perturbation of the input variables and state-space models are given in Table 4. The open-
the parameters are identified by a recursive extended loop discrete-time frequency analysis of the models for
least-squares algorithm. T = 1 min and T = 10 min are shown in Fig. 6(A, B).
 Dynamics and identification of a binary distillation column 2577

x : Liquid-methanol composition in condenser, top product.

1
.x : Liquid-methanol composition on tray 7.
2
.x : Liquid-methanol composition on tray 5, feed tray.
3
x4: Liquid-methanol composition on tray 3.
x5: Liquid-methanol composition on tray 1, bottom product.
x : Liquid-methanol composition in reboiler.
6

Reflux.
Boilup.
Feed-flow rate.
Feed composition.

Fig. 3. Structure of 6-order approximate continuous-time

1 1:u(t) + ;
X
x

model. Y indicates a nnn-7ero element

q(t)

xl2345
1x1234 1 IIT1

Fig. 4. Structure of 6-order approximate discrete-time model. x indicates a direct interaction of a quantity
upon itself. The numerals indicate the inverse order of the interaction.

x 1
1 x 1 1 1
1 x 1 11 1 1
x(t+l) = x(t) +
1x1
1x1 :: :
lx A

Fig. 5. Structure of 6-order approximate discrete-time model.

The maximum frequency for which the singular values with a sample time of 1 min yields a satisfactory
are plotted equals the Nyquist frequency, oN = K/T. description of the dynamics and thus it is used for
As seen from the minimum gain curves the sample control design in Andersen and Kiimmel (1989).
time of 10 min is too slow since there is no significant
bandwidth. However, the sample time of 1 min is 3.3.2. Multi-SIMO identification. In Section 3.3.1
sufficient. In Fig. 7 the frequency analysis of the the inputs to the process where perturbed simulta-
identified reduced order model with T = 1 min is neously. A more common approach is to perturb the
compared with the “ideal” discrete time model. This inputs separately and then combine the models into
model was obtained by discretization of the original one which describes the overall interconnections. As
11 order linear continuous-time model. Figure 7(A) an example of this approach the reflux and bollup are
shows that the maximum and minimum gain proper- perturbed separately. The individual model structures
ties of the column dynamics towards actuator changes are given as:
are satisfactory described by the identified reduced
order model. However, from Fig. 7(B) it is seen that the x,(t+ T) = A,x,(t)+B,u,(t) ul: reflux. (19)
approximated model-structure does not allow an ex- x2(t + T) = A*x,(t) + B,u,(t) u2: boilup. (20)
act determination of the maximum and minimum gain
properties of the column dynamics towards disturb- In eqs (19) and (20) A, and A, both have the structure
ance inputs. However, the MTMO identified model shown in Fig. 5. The structure of B, [eq. (19)] is
2578 HENRIK WEISBERG ANDERSEN et al.

Table 4. Identified parameters for T= 1 min and T= 10 min

r
T= 1 min

I
0.8589 0.0367
0.2736 0.6894 0.1829
A: 0.3 157 0.3273 0.3540

1 -0.0052
0.2398 0.4327
0.2468 -
0.7670
0.7202
0.1292
0.8996
0.7239

-0.0073 0.0031 0.2724

B:
- 0.0094 0.0067
-0.0033 0.0027
1

r 0.23 16
-0.1474

1
0.1204
C:
0.0067
0.5825

T= 10 min
0.1477
-0.1557 0.5781
0.5827 -0.1222 0.3580
A:
0.3298 -0.5491 2.3380
0.0286 0.5 199
0.3664

- 0.0498
-0.0582 0.0286 0.9704
B:
- 0.0643 0.038 5
- 0.0228 0.0149

-0.0518
-0.2047
0.0416
C:
0.0178
0.1532
- 0.0436
i

0 2
A

0
--i

-%

--z
--o

-S
-3

-4
-.

-5’

Fig. 6. Open-loop analysis of identified discrete-time models. x-Axis=log frequency, y-Axis =log gain.
(A) Product compositions towards actuators. (B) Product compositions towards disturbances. (---- )
Discrete-time model for T= 1 min. (- - - -) Discrete-time for T= 10 min.
 Dynamics and identification of a binary distillation column 2579

--L

-a -
--1

-_) -
-_)

-. -
-1

I
-a 1 I L

-3 --I --1 0 I -3 --P --i 0 I

Fig. 7. Comparison of “true” and identified model. x-Axis = log frequency. y-Axis =log gain. (A) Product
compositions towards actuators. (B) Product compositions towards disturbances. (------- ) Discrete-time
“true” model for T= 1 min. (- - - -) Discrete-time identified for T= 1 min.

similar to the structure of the first column of B in Fig. 5 0

and the structure of S, is similar to the structure of the

second column. By pseudo-random binary perturba- --I
tion of respectively u1 and u2 with a sample time of
1 min the parameters in A,, B,, A, and B, where
estimated. A model which describes the overall -a
input-output relations is obtained as:

[::::z
;;I=[iIT1
OAJ[:::::I
-3

--4 c

+[OS’I [::i:;l
OS* (21)
-S
x,(t) --z

[
-3 --1 0 1
x(t) = [I I]
x2(t) 1 Fig. 8. Comparison of “true”, MIMO and multi-SIMO
The open-loop frequency analysis between the out- identified models. x-Axis = log frequency. y-Axis = log gain.
Product compositions towards actuators. (-) Discrete-
puts (top- and bottom-product concentrations) for the time “true” model for T=l min. (- ~ ~ -) Discrete-time
“true” discrete-time model, the MIMO identified and MIMO identified model for T= 1 min. (- -) Discrete-time
the multi-SIMO identified reduced-order model are multi-SIMO identified model for T= 1 min.
shown in Fig. 8. As seen the high-gain direction is
equally well described by the reduced order models.
However, the multi-SIMO identification yields a poor 4. CONCLUSIONS
description of the low-gain direction. The reason for This paper has addressed the problem of estimating
this is that no perturbations are made entirely in the models for distillation columns which adequately de-
low-gain direction by the multi-SIMO approach, and scribe the strong gain directionality. Such models are
since the gain ratio between the high- and low-gain important from a control design point of view since
directions is approximately 100 the high-gain direc- gain directionality has a significant impact on the
tion is relatively easy to estimate whereas the low-gain closed-loop control performance.
direction only can be estimated as the difference It is well known that binary distillation shows
between large deviations in the outputs. Thus, estima- strong gain directionality. This is explicitly described
tion of the low-gain direction is ill-conditioned in the by the frequency domain singular-value decompo-
multi-SIMO case. This means that if the process is sition which reveals that there is a significant differ-
linear and the selected model order and model struc- ence between the response towards changes in the
ture is correct, then multi-SIMO identification will internal flows (the individual product flows are kept
give a correct estimate of the low-gain direction. constant) and changes in the external flows (the sum of
However, non-linearities in the process and approxi- the product flows is kept constant). Thus, frequency
mations in model order and model structure causes domain singular-value decomposition is a suitable
that the estimate of the low-gain direction will be poor tool for examination of how well the identified models
in the multi-SIMO case since estimation of this direc- describes the gain directionality. The gain direc-
tion is ill-conditioned. tionality properties of the column dynamics leads to
2580 HENRIK WEISBERG ANDERSEN et al.

an ill-conditioned estimation of the low-gain direction direction of the minimum singu-

in the multi-SIMO case. In the case where only one lar value
product concentration is to be controlled this is of no Fj input phase for input j in the
practical importance for the control design since the direction of the maximum singu-
major part of the control action will be in the high- lar value
gain direction which is properly described by the rj input phase for input j in the
process model. However, in the case where both direction of the minimum singu-
product concentrations are to be controlled, the ill- lar value
conditioning will have a significant impact on the Tj output phase for output j in the
performance since the control design is based on a direction of the maximum singu-
poor estimate of the low-gain direction. Thus, when- lar value
ever control action is required in this direction the. 'flj output phase for outputj in the
control performance may be arbitrarily poor. These direction of the minimum singu-
problems are solved by application of MIMO identifi- lar value
cation where inputs are perturbed simultaneously and reference sinosoid, i.e. zero phase-
the identification yields a proper description of both lag, for maximum singular value
the high- and low-gain direction. This result is signifi- reference sinosoid, i.e. zero phase-
cant from a practical point of view since it shows that lag, for minimum singular value
MIMO identification may be necessary in order to open-loop transfer function
obtain a suitable model quality, depending of the matrix between outputs and
degree of gain directionality of the plant. actuators
G,“(c{w>) open-loop transfer function ma-
NOTATION trix between outputs and external
x dim(x)=n states in linear state-space model disturbances
u dim(u) = k manipulated variables in linear cIw> continuous time: the imaginary
state-space model axis
CJdim (4) = n,, external disturbances in linear discrete time: the unit circle
state-space model T sample time
y dim(y)=m outputs in linear state-space L reflux: liquid flow above feed tray
model L liquid flow below feed tray
A dim(A)=nxn state matrix V boilup
B dim(B) = n x k control input matrix D distillate flow
D dim (D) = n x nd disturbance-input matrix B bottom product flow
T dim (t) = m x n output matrix & feed flow rate
F dim(F)=mxk direct couplings from actuators XF feed composition
d(x) differential operator in continu- xc condenser composition
ous time, difference operator in XR reboiler composition
discrete time xi composition on tray i
G dim(G)=mxk transfer-function matrix HC condenser holdup
Wdim(W)=mxm left singular matrix HR reboiler holdup
Vdim(V)=kxk right singular matrix ff, tray holdup
Z dim (Z) = nr x k singular value matrix
64) maximum singular value of A
REFERENCES
c(A) minimum singular value of A
Andersen, H. W., 1988, Advanced control of distillation
singular value number i where
ci
columns. Ph.D. thesis, Department of Chemical Engineer-
the singular values are ordered ing, Technical University of Denmark, DTH-229, DK-
after decreasing size 2800 Lyngby, Denmark.
Wi left singular vector correspond- Andersen, H. W. and Kiimmel, M., 1989, Discrete-time
ing to ci control of a binary distillation column. Chem. Engrag Sci.
44, 2583-2595.
right singular vector correspond-
Vi Bequette, B. W., Horton, R. R. and Thomas, T. F., 1987,
ing to ci Resilient and robust control of an energy integrated distil-
7j input gain for inputj in the direc- lation column. Proc. 1987 Am. Control Co& pp.
tion of the maximum singular 1027-1033.
Goldschmidt, L. B., Jorgensen, S. B. and Hallager, L., 198.5,
value
Sparse process modelling for robust adaptive surveillance
Yj input gain for inputj in the direc- and control of a binarv distillation column. IFAC Work-
tion of the minimum singular shop on Adaptive Control of Chemical Processes,
value Frankfurt, F.R.G., pp. 167-172, 2 l-22 October 1985.
output gain for output j in the Lau, H., Alvarez, J. and Jensen, K. F., 1985, Design of control
3j
structures by singular value analysis. A.1.Ch.E. .I. 31,
direction of the maximum singu- 427439.
lar value MacFarlane, A. G. J. and Scott-Jones, D. F. A., 1979, Vector
output gain for output j in the gain. Int. J. Control 29, 65-91.
 Dynamics and identification of a binary distillation column 2581

Moore, C. F. 1986, Application of singular value decompo- Skogestad, S., 1987, Studies on robust control of distillation
sition to the design, analysis and control of industrial columns. Ph.D. Thesis, California Institute of Technology,
processes. Proc. 1986 Am. Control Cc& pp. 634-650. Pasadena, California, U.S.A.
Nielsen, C. S., Andersen, H. W., Brabrand, H., Tofteglrd, B., Skogestad, S. and Morari, M., 1988a, Understanding the
Clement, K. and Jsrgensen, S. B., 1988, Identification and dynamic behaviour of distillation columns. Ind. Engng
adaptive control of a binary distillation column with a heat Chew Res. 27, 184881862.
pump. CHEMDATA 88, 13-15 June 1988, GBteborg, Skogestad, S. and Morari, M., 1988b, LV-control of a high-
Sweden. purity distillation column. Chem. Engng Sci. 43, 3348.