2018 Indian Control Conference (ICC)

January 4-6, 2018. IIT Kanpur, India

Design of Feedforward-Feedback Controller for Reactive Distillation

Column having Inverse Response
Gaurav Kataria1 and Kailash Singh2

Abstract— Inverse response occurs due to the two opposing effect of right plane zeroes into dead time and then tuned
processes working at the same time. This inverse response can PID accordingly as for dead time compensation. [9] removed
occur because of a step change either in a load variable or inverse response and dead time in a boiler drum by using
a manipulated variable. In this paper inverse response due
to a load variable is considered for a reactive distillation a auto tuned PID controller. The design of the controller is
column carrying an esterification reaction. A feedforward- based on h∞ theory and internal model control (IMC). While
feedback controller is designed to eliminate the effect of inverse everybody was working on 1 degree of freedom (DoF) PID,
response in the process. The results showed a decrease in [10] worked on PID with 2 DoF. They tuned the controller
sluggishness and integral errors while using a feedforward- using model reference technique to eliminate the inverse
feedback controller instead of a feedback controller in the
closed loop. response from a second order process transfer function.
Recently [11] used model reference technique based PI
I. INTRODUCTION controller for the removal of inverse response and dead time
in a second order process with two first order opposing
During inverse response initially, the output of an open
processes. [12] used a Smith predictor accommodated with
loop goes in the opposite direction to that direction where
PID with advanced tuning method to remove the inverse
the response ends. The reason for the inverse response is
response. They used Maclaurin series approach to tune PID
when two opposing processes work simultaneously, and one
for solving the purpose.
process is dominating in the start of the process but the
In literature, major work is done on the removal of inverse re-
second process prevails in the end. The main characteristic of
sponse because of manipulated variable, but inverse response
the inverse response is the presence of a positive zero in the
due to load variable has not been discussed yet. The problem
process transfer function [1]. There is a need of removing the
of inverse response due to load variable was observed by [13]
inverse response as it makes closed loop difficult to control.
when they gave a step change in one of the feed variable
When the inverse response is occurring due to the step
which caused sluggish response of the controlled output.
change in manipulated variable the problem can be solved
In this paper, the inverse response due to a step change in
using a Smith predictor [1], [2] or by keeping inverse
load variable is considered in a reactive distillation column.
response as a function of tuning parameters for PID[3],
The effect of inverse response has been minimized using
[4]. [1] discussed simple Smith predictor for two 1st order
a feedforward-feedback (FF-FB) controller and the results
processes and [2] extended the technique by designing the
of the closed loop were compared with process having FB
Smith predictor using h∞ theory for getting more robustness
controller only.
in the process. [3] considered continuous stirred tank reactor
(CSTR) with the inverse response and dead time. He tuned II. INVERSE RESPONSE
the controller parameters of PI by taking them as a function Lets take an open loop with a second order process
of dead time and positive zero. [4] derived a PI controller carrying two single order opposing processes as shown in
by comparing the coefficients of the closed loop transfer Fig. 1 .  
function and then tuning it accordingly. Some authors used K1 K2
y(s) = − m(s) (1)
other tuning techniques for the purpose ([5], [6], [7], [8], τ1 s + 1 τ2 s + 1
[9], [10] and [11]) . [5] used an internal model control for (K1 τ2 − K2 τ1 )s + (K1 − K2 )
an inverse response in two first order opposing processes. [6] y(s) = (2)
(τ1 s + 1)(τ2 s + 1)
removed inverse response and dead time problem in a CSTR
using sliding model controller constructed from a first order Where K1 and K2 are processes gains and τ1 and τ2 are the
plus dead time model. [7] had taken process with the inverse time constants of their respective processes.
response and a dead time and then eliminated both the The process shows inverse response when process 2 is
problems by tuning a PID controller using direct synthesis dominating in the start of the process but process 1 dominates
controller technique. [8] opted technique by converting the in the end. This happen when
τ1 K1
> >1 (3)
1 Gaurav Kataria is a Ph.D Research Scholar at Department of Chem-
τ2 K2
ical Engineering, Malaviya National Institute of Technology, Jaipur,India
gkataria64@gmail.com and a zero, z, of the process transfer function, y(s), is greater
2 Kailash Singh is a Faculty at Department of Chemical En-
than 0
gineering, Malaviya National Institute of Technology, Jaipur,India K1 − K2
ksingh.mnit@gmail.com z= <0 (4)
K1 τ2 − K2 τ1

978-1-5386-2904-8/18/$31.00 ©2018 IEEE 246

This inverse response can be either because of step change To design the feedforward controllers for both the distur-
given in a manipulated variable or step change given in a bances, the coefficients of d1 and d2 has to be zero.
load variable. Gd1
Gc 1 = (7)
Gp
Gd2
Gc2 = (8)
Gp
And any PID tuning technique like Ziegler-Nicholas, Cohen-
Coon etc. can be used for the feedback controller.
The performance of the both the FF and FF-FB controllers
needs to be evaluated on some error criteria. The integral
errors such as integral of square of error (ISE), integral of the
absolute value of error (IAE) and integral of time-weighted
absolute error (ITAE) are used for the purpose of evaluating
Fig. 1. Open loop with inverse response
controllers performance[14]. In the case of ISE, the square
of error is integrated over period from 0 to ∞. While in the
case of IAE and ITAE, an absolute error and time weighted
III. FEEDFORWARD-FEEDBACK CONTROL absolute error is integrated with respect to time. The ISE
criteria is used when the value of error is large, IAE is used
Let us take a process, G p , with two load variables and one in the case of small errors and ITAE is used when error is
manipulated variable. The open loop of the process can be persistent over the time.
seen in Fig. 2. Where the output ȳ(s) of the process can be
IV. CASE STUDY
controlled by a feedback controller by manipulating variable
m . But when process, G p is showing inverse response due A reactive distillation column (RDC) is taken for the n-
to a load variable, a feedforward controller coupled with Butyl Acetate esterification reaction.
feedback controller is needed for the removal of disturbance
C2H4O2 + C4H9OH C6H12O2 + H2O (9)
effect before it enters the plant. The closed loop with FF-FB
controllers can be seen in Fig. 3. Here Acetic acid and n-butanol are reacting to give n-butyl
The output of the closed loop is acetate and water as a reaction products. The RDC is used to
restrict the backward reaction by removing products of the
ȳ = mG p + d1 Gd1 + d2 Gd2 (5) forward reaction simultaneously as reaction is going on. A
nonlinear MATLAB/SIMULINK model has been developed
by taking acetic acid and butanol feed flow rate as load
ȳ(s) = ysp1 Gc1 G p + ysp2 Gc2 G p + d1 (Gd1 − Gc1 G p ) + d2 (Gd2 − Gc2 G p ) (6)
variables and reboiler duty of the column as a manipulated
variable.
The transfer functions of process and both the disturbances
are estimated using data obtained by giving a step change in
manipulated variable and both the load variables.
0.077s2 + 17.46s + 27.25
Gp = (10)
s3 + 78.11s2 + 2116s + 2851
−2.73 × 10−03 s2 − 0.132s + 0.13
Gd1 = (11)
s3 + 19.38s2 + 115.35s + 49.05
Fig. 2. Open loop of the process
6.77 × 10−03 s2 − 0.549s − 3.045
Gd2 = (12)
s3 + 53.25s2 + 536s + 944.6
The percent fit and mean square error between RDC
SIMULINK model and transfer function model are given in
Table. I. From these transfer functions all the controllers are
tuned. PID controller for the feedback loop has been tuned
using MATLAB automatic tuning with the PID parameters
as P = 166.27, I = 112.6 min−1 and D = 18.4 min . According
to (7) and (8) feedforward controllers are estimated as
−2.72 × 10−03 s5 − 0.35s4 − 6.67s3 + 1.93s2 + 237.24s + 370.34
Gc1 = (13)
7.71 × 10−02 s5 + 18.95s4 + 374.52s3 + 2545.88s2 + 3999.7s + 1336.61
6.77 × 10−03 s5 − 0.02s4 − 31.6s3 − 1380.22s2 − 8008.4s − 8681.29
Gc2 = (14)
7.71 × 10−02 s5 + 21.57s4 + 998.35s3 + 10882.5s2 + 31098.7s + 25740.3
Fig. 3. Closed loop of the process

247
 0.9905 0.9925

Butyl Acetate Bottoms Composition

Butyl Acetate Bottoms Composition

0.99 0.992

0.9895 0.9915

0.989 0.991

0.9885 0.9905

0.988 0.99

0.9875 0.9895

0.987 0.989

0.9865 0.9885
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time (hr) Time (hr)

Fig. 4. Open Loop Response for Step Change in Acetic Acid Feed Flow Rate by (a) −10% (b) 10%

0.993 0.9895
Butyl Acetate Bottoms Composition

Butyl Acetate Bottoms Composition

0.9925 0.989

0.992
0.9885

0.9915
0.988
0.991
0.9875
0.9905

0.987
0.99

0.9895 0.9865

0.989 0.986
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time (hr) Time (hr)

Fig. 5. Open Loop Response for Step Change in Butanol Feed Flow Rate by (a) −10% (b) 10%

TABLE I TABLE II
P ERCENT F IT AND MSE BETWEEN RDC MODEL AND T RANSFER I NTEGRAL E RRORS DURING L OAD C HANGE IN ACETIC ACID F EED
F UNCTION M ODEL F LOW R ATE
-10% 10%
Transfer function Percent fit MSE FB FF-FB FB FF-FB
Gp 99.03 1.58 × 10−12 ISE 1.62 × 10−09 1.18 × 10−09 1.62 × 10−09 1.18 × 10−09
Gd1 99.14 4.18 × 10−11 IAE 6.90 × 10−05 3.43 × 10−05 6.90 × 10−05 3.43 × 10−05
Gd2 99.56 2.52 × 10−12 ITAE 1.34 × 10−04 4.17 × 10−05 1.34 × 10−04 4.17 × 10−05

V. RESULTS 1) Load Change in Acetic Acid Feed Flow Rate: The

closed loop response of the process when step change of
The open and closed loop dynamics were studied by giving ±10% is given in the feed flow rate of acetic acid is shown
step changes in load and manipulated variables. Target of the in Fig. 6 and its change on the manipulative variable is shown
closed loop is to control the composition of n-butyl acetate in Fig. 7. It can be observed that FF-FB controller is show-
in the bottoms of the column. ing better control results than FB controller and removing
the sluggishness problem of the process by controlling the
A. Open Loop Response composition 2-2.5 hr earlier than its counterpart. The integral
errors for both the closed loop responses are shown in Table.
The step change was given in both the load variables and
II and it can be seen that integral errors are decreasing while
its behavior was studied for the control purpose. The open
using FF-FB controller.
loop response when step change of ±10% is given in acetic
2) Load Change in Butanol Feed Flow Rate: The closed
acid feed flow rate and butnaol are shown in Fig. 4 and 5.
loop response for the second load variable is studied by
It can be noticed that open loop response during the step
giving a step change of ±10% in feed flow rate of butanol.
change in acetic acid is showing an inverse response.
The closed loop response is shown in Fig. 8 and 9 and change
in reboiler duty during the change is shown in Fig. 10. While
B. Closed Loop Response
FB controller was controlling the output in 1.5-2 hr, FF-FB is
The bottoms composition of butyl acetate was controlled controlling the same output in half hour. The integral errors
using closed loop with FB and FF-FB controllers. As adding as shown in Table. III are also getting decreased to high
a feedforward controller does not affect the servo problem extent when FF-FB controller is used.
of the closed loop, so only regulatory problem in the loop 3) Process Transfer Function Mismatch: The process
was studied. model robustness was checked by changing the coefficient of

248
 0.9895 0.9894

Butyl Acetate Bottoms Composition

Butyl Acetate Bottoms Composition
FB FB
0.9895 0.9894
FF-FB FF-FB

0.9894 0.9894

0.9894 0.9894

0.9894 0.9893

0.9894 0.9893

0.9894 0.9893

0.9893 0.9893
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time (hr) Time (hr)

Fig. 6. Closed Loop Response for Step Change in Acetic Acid Feed Flow Rate by (a) −10% (b) 10%

1.95 1.8

1.75
FB
1.9
FB FF-FB
1.85 FF-FB 1.7
Reboiler Duty (kW)

Reboiler Duty (kW)

1.65
1.8
1.6
1.75
1.55
1.7
1.5

1.65
1.45

1.6 1.4

1.55 1.35
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time (hr) Time (hr)

Fig. 7. Reboiler Duty during Step Change in Acetic Acid Feed Flow Rate by (a) −10% (b) 10%

0.9896 0.9894
Butyl Acetate Bottoms Composition

Butyl Acetate Bottoms Composition

0.9896
0.9894
0.9895

0.9895
0.9894
0.9895

0.9895 0.9894

0.9895
0.9894
0.9894

0.9894
0.9894
0.9894

0.9894 0.9894
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time (hr) Time (hr)

Fig. 8. Closed Loop Response for −10% Step Change in Butanol Feed Flow Rate havinng (a) FB Controller (b) FF-FB Controller

0.9894 0.9894
Butyl Acetate Bottoms Composition

Butyl Acetate Bottoms Composition

0.9894
0.9893

0.9894
0.9893

0.9894

0.9892
0.9894

0.9892
0.9894

0.9891 0.9894
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time (hr) Time (hr)

Fig. 9. Closed Loop Response for 10% Step Change in Butanol Feed Flow Rate havinng (a) FB Controller (b) FF-FB Controller

249
 1.8 2.05

1.75 FB 2
FF-FB
1.7 1.95 FB
FF-FB

Reboiler Duty (kW)

Reboiler Duty (kW)

1.65 1.9

1.6 1.85

1.55 1.8

1.5 1.75

1.45 1.7

1.4 1.65

1.35 1.6

1.3 1.55
0 2 4 6 8 10 12 0 2 4 6 8 10 12
Time (hr) Time (hr)

Fig. 10. Reboiler Duty during Step Change in Butanol Feed Flow Rate by (a) −10% (b) 10%

0.9895 0.9894
Butyl Acetate Bottoms Composition

Butyl Acetate Bottoms Composition

Coeff = 3.51e-04 0.9894 Coeff = 3.51e-04
0.9895 Coeff = 5.26e-04 Coeff = 5.26e-04
0.9894
Coeff = 1.75e-04 Coeff = 1.75e-04
0.9894
0.9894
0.9894

0.9894 0.9894

0.9894
0.9894
0.9894

0.9894
0.9894
0.9894

0.9894 0.9894
0 2 4 6 8 10 12 0 0.5 1 1.5 2 2.5 3 3.5
Time (hr) Time (hr)

Fig. 11. Process model mismatch by ±50% (a) −10% step change in Acetic Acid Feed Rate (b) −10% step change in Butanol Feed Rate

0.9895
Butyl Acetate Bottoms Composition

Coeff = 0.02 0.9894

Butyl Acetate Bottoms Composition

0.9895 Coeff = 0.03 Coeff = 1.10e-03

0.9894
Coeff = 0.01 Coeff = 1.59e-03
0.9894
0.9894 Coeff = 5.29e-04
0.9894
0.9894
0.9894

0.9894 0.9894

0.9894
0.9894
0.9894

0.9894 0.9894
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3
Time (hr) Time (hr)

Fig. 12. Disturbances model mismatch (a) ±50% mismatch in GD1 (b) ±50% mismatch in GD2

TABLE III
and it can be observed that errors are not changing much
I NTEGRAL E RRORS DURING L OAD C HANGE IN B UTANOL F EED F LOW
during acetic acid feed rate change but a high change can be
R ATE
observed during a step change in butanol feed rate.
-10% 10%
FB FF-FB FB FF-FB
ISE 4.47 × 10−09 1.27 × 10−17 4.47 × 10−09 1.28 × 10−17
IAE 5.07 × 10−05 5.24 × 10−09 5.07 × 10−05 5.24 × 10−09
ITAE 1.71 × 10−05 2.42 × 10−08 1.71 × 10−05 2.42 × 10−08
4) Disturbances Transfer Function Mismatch: The model
robustness has also been tested by changing the coefficient
of s3 in the characteristic equation of both the disturbances.
s3 of the characteristic equation by ±50% and then studying The response of the closed loop was noticed by giving a
the regulatory problem. A step change of −10% was given step change of −10% in acetic acid feed rate when GD1
in both the feed flow rates individually, and the response was coefficient is given a change of ±50% and −10% step change
observed as shown in Fig. 11. During step change in acetic in butanol feed rate when coefficient of GD2 was changed
acid flow rate, the change in coefficient does not affect the by ±50%. The closed loop response is shown in Fig. 12. In
process much. The step change in butanol feed flow rate is this also the acetic acid mismatch is not affecting the process
affecting the response and it ends up with some oscillations, much but butanol mismatch showed a quite difference in
and finally, output settles at its set point while taking some closed loop response but staying within the manageable
more time. The integral errors are shown in Table. V-B.3 limits. The integral errors are shown in Table. V.

250
 TABLE IV
I NTEGRAL E RRORS D URING P ROCESS M ODEL M ISMATCH
Step Change in Acetic Acid Feed Rate Step Change in Butanol Feed Rate
Coeff. Value 3.51 × 10−04 5.26 × 10−04 1.75 × 10−04 3.51 × 10−04 1.75 × 10−04 5.26 × 10−04
ISE 1.18 × 10−09 1.22 × 10−09 1.13 × 10−09 1.95 × 10 −17 3.16 × 10 −11 4.11 × 10−11
IAE 3.43 × 10−05 3.46 × 10−05 3.42 × 10−05 6.03 × 10−09 1.98 × 10−06 2.43 × 10−06
ITAE 4.17 × 10−05 4.17 × 10−05 4.16 × 10−05 2.48 × 10 −08 2.22 × 10 −07 2.89 × 10−07

TABLE V
I NTEGRAL E RRORS D URING M ODEL M ISMATCH OF GD1 AND GD2
Change in GD1 coefficient Change in GD2 coefficient
Coeff. 0.02 0.03 0.01 1.10 × 10−03 1.59 × 10−03 5.29 × 10−04
ISE 1.18 × 10−09 1.01 × 10−09 1.31 × 10−09 1.95 × 10−17 9.48 × 10−11 5.14 × 10−11
IAE 3.43 × 10−05 3.23 × 10−05 3.65 × 10−05 6.03 × 10−09 3.75 × 10−06 3.68 × 10−06
ITAE 4.17 × 10−05 4.09 × 10−05 4.27 × 10−05 2.48 × 10−08 5.61 × 10−07 6.58 × 10−07

VI. CONCLUSION [11] J. A. Martinez, O. Arrieta, R. Vilanova, J. D. Rojas, L. Marin, and

M. Barbu, Model reference PI controller tuning for Second Order
In this paper, inverse response due to load variable was Inverse Response and Dead Time Processes, in IEEE Int. Conf. Emerg.
removed using a feeforward-feedback controller. The in- Technol. Fact. Autom. ETFA, vol. 2016 Nov, pp. 2–6, 2016.
[12] J.C. Jeng and S.W. Lin, PID controller tuning based on Smith-type
verse response in a reactive distillation column carrying compensator for second-order processes with inverse response and
an esterification reaction was taken which was showing time delay, in 8th Asian Control Conf., pp. 1147–1152, 2011.
inverse response because of the step change in acetic acid [13] S. R. V. Raghavan, T. K. Radhakrishnan, and K. Srinivasan, Soft
sensor based composition estimation and controller design for an ideal
feed flow rate causing a slow response in the closed loop. reactive distillation column, ISA Trans., vol. 50, no. 1, pp. 61–70,
When results for FF-FB and FB controllers were compared, 2011.
FF-FB removed the sluggish behavior of the closed loop [14] D. R. Coughanowr, Process Systems Analysis and Control (2nd ed.).
McGraw-Hill International Education,1991.
and a decrease in the integral errors was observed. The
robustness of the process transfer function was also checked
by changing the coefficient of the characteristic equation
of all the transfer functions involved. The model mismatch
response also showed adequate control results making the
process robust to changes.

R EFERENCES
[1] G. Stephanopoulos, Chemical process control : an introduction to
theory and practice. Prentice-Hall, 1984.
[2] W. Zhang, X. Xu, and Y. Sun, Quantitative Performance Design for
Inverse-Response Processes,Ind. Eng. Chem. Res., vol. 39, no. 6, pp.
2056–2061, 2000.
[3] W. L. Luyben, Tuning Proportional Integral Controllers for Processes
with Both Inverse Response and Deadtime, Ind. Eng. Chem. Res., vol.
39, no. 4, pp. 973–976, 2000.
[4] R. P. Sree and M. Chidambaram, Simple method of tuning PI con-
trollers for stable inverse response systems, J. Indian Inst. Sci., vol.
83, no. 3–4, pp. 73–85, 2003.
[5] C. Scali and A. Rachid, Analytical Design of Proportional Integral
Derivative Controllers for Inverse Response Processes, Ind. Eng.
Chem. Res., vol. 37, no. 4, pp. 1372–1379,1998.
[6] O. Camacho, R. Rojas, and W. Garca, Variable structure control
applied to chemical processes with inverse response, ISA Trans., vol.
38, no. 1, pp. 55-72, 1999.
[7] I.L. Chien, Y.C. Chung, B.S. Chen, and C.Y. Chuang, Simple PID
Controller Tuning Method for Processes with Inverse Response Plus
Dead Time or Large Overshoot Response Plus Dead Time,Ind. Eng.
Chem. Res., vol. 42, no. 20, pp. 4461–4477, Oct. 2003.
[8] P. Y. Chen, Y. C. Tang, Q. Z. Zhang, and W. D. Zhang, A new design
method of PID controller for inverse response processes with dead
time, in Proc. IEEE Int. Conf. Ind. Technol., vol. 2005, no. 2, pp.
1036–1039, 2005.
[9] D. Gu, L. Ou, P. Wang, and W. Zhang, Relay feedback autotuning
method for integrating processes with inverse response and time delay,
Ind. Eng. Chem. Res., vol. 45, no. 9, pp. 3119-3132, 2006.
[10] V. M. Alfaro and R. Vilanova, Robust tuning of 2DoF five-parameter
PID controllers for inverse response controlled processes, J. Process
Control, vol. 23, no. 4, pp. 453–462, 2013.

251