Improvement of Linear Distillation Column Control

Performance Using Fuzzy Self-Tuning PI Controller

Abdul Wahid1, a), Albert Harazaki Mendrofa2, b), Muhammad Fathi Fadlian2, c), and
Shafira Anandita1, d)
1
Department of Chemical Engineering, Faculty of Engineering University of Indonesia, Depok, West Jawa, 16424
2
Department of Electrical Engineering, Faculty of Engineering University of Indonesia, Depok, West Jawa, 16424
a)
Corresponding author: wahid@che.ui.ac.id
b)
albertmendrofa@gmail.com
c)
fadlianfathi00@gmail.com
d)
shafiranndt@gmail.com

Abstract. PID controller still plays an important role in industrial processes due to the structure’s simplicity and the
ability to be implemented in industrial scale. However, PID controller is unable to give a satisfactory control performance
in controlling nonlinear and multiple input and multiple output (MIMO) systems with strong loop interactions. To
improve the control performance, PID controller is made to be adaptive by adding fuzzy controller. Fuzzy controller will
calculate the values of ΔKp, ΔKi, dan ΔKd, so that the values of Kp, Ki, and Kd in PID controller will always be updated
based on the error input in both PID and fuzzy controller. Hence, the controller is named as fuzzy self-tuning PID, due to
the ability of calculating the optimum values of Kp, Ki, and Kd based on the system’s current condition. With the
potentials aforementioned, this controller will be implemented in one of the multivariable systems in chemical processes,
which is distillation column. The distillation column model used in this simulation is linear distillation column 2 × 2 by
McAvoy and Weischedel as mentioned in Marlin. The simulation will be done with the MATLAB/Simulink software and
the program routines are made with C (C MEX Sfunction). Verification of the control performance improvement will be
reviewed from three parameters, which are overshoot, settling time, and IAE (integral absolute error). Keywords: Fuzzy,
fuzzy self-tuning PID, C MEX, distillation column

INTRODUCTION
The fuzzy logic control has proved its broad potential in industrial applications because of its good performance
for controlling nonlinear and uncertain systems that could not be controlled satisfactorily by using conventional
controllers, such as a conventional PID controller [1]. With the addition of self-tuning to the fuzzy logic control, it
advances the fuzzy logic PID controller to overcome problems such as multiple-input and multiple-output (MIMO)
systems with strong loop interactions.
One of the most common multivariable system in industrial applications is the distillation column. There are a lot
of researches regarding the distillation column, including the areas of process synthesis, process dynamics, and
process control. Various process control methods, including optimization (tuning), have been studied and applied to
various distillation column systems. However, with the potentials aforementioned, the application of fuzzy self-tuning
PID controller for controlling distillation columns has not been studied further.
The only paper found to study this subject is by Almeida and Coelho [1]. With a Wood and Berry 2 × 2 distillation
column, they have proven that a fuzzy self-tuning PID controller results in a better controlling than using a
conventional PID controller (with Ziegler-Nichols tuning method). However, the simulation method proposed in the
paper is found to be unclear and unpractical due to the lack of a detailed discussion. To create a clear understanding
about the subject, this essay will discuss in detail the simulation method proposed for implementing a fuzzy self-
tuning PI controller for distillation column with MATLAB/SIMULINK. The distillation column used as a case study
in this paper is a 2 × 2 linearized distillation column model by McAvoy and Weischedel (1981) as shown in Marlin
[2].

LITERATURE REVIEW
Here we provide some basic advice for formatting your mathematics, but we do not attempt to define detailed
styles or specifications for mathematical typesetting. You should use the standard styles, symbols, and conventions
for the field/discipline you are writing about.

Structure of Fuzzy Self-Tuning PID Controller

According to Si and Wang [3], the fuzzy-PID controller takes conventional PID as the foundation, which uses the
theory of fuzzy reason and variable discourse of universe to on-line regulate the parameters of PID automatically. The
structure of fuzzy self-tuning PID controller is shown below.

Fig. 1. Structure of fuzzy self-tuning PID controller.

Here, e denotes the system error and ec denotes the system error changing rate. The error and error changing rate
are used as the input variables in control system, and the output variables are the parameters of PID control, those are
ΔKp, ΔKi, and ΔKd. The fuzzy sets of e, ec, ΔKp, ΔKi, and ΔKd are { NB, NM, NS, 0, PS, PM, PB }, where NB, NM,
NS, 0, PS, PM and PB are linguistic values, which represent “negative big”, “negative medium”, “negative small”,
“0”, “positive small”, “positive medium”, and “positive big” respectively.
The regions of e and ec are described by the following method.
 PB = error tolerance 1*set point 
 PM = error tolerance 2*set point 
PS = error tolerance 3*set point 
 ZO = 0 
 NS = -PS 
NM = -PM 
NB = -PB 
By describing the e and ec region this way, it is allowed to determine the system’s error tolerance. Meanwhile, the
regions of ΔKp, ΔKi, and ΔKd are described by the following method.
PB = PB / rescaler 
 PM = PM / rescaler 
PS = PS / rescaler 
 NS = NS / rescaler 
 NM = NM / rescaler 
NB = NB / rescaler 
 Which Microsoft Word Equation Editor Should I Use? (Second Level Heading)
The essential part of the fuzzy logic controller is a set of linguistic rules. In many cases, it is easy to translate an
expert’s experience into such rules. Any number of rules can be created to define the actions of the fuzzy controller.
In this paper, the fuzzy control rules design is based on the medical robot that can approach the target quickly and
stable. The application of conventional fuzzy conditions and fuzzy relations “If e is A and ec is B, then ΔKp is C, ΔKi
is D, and ΔKd is E” can establish fuzzy rules. The finally determined fuzzy rules are shown below.

TABLE I. FUZZY RULE FOR ΔKP

TABLE II. FUZZY RULE FOR ΔKI

TABLE III. FUZZY RULE FOR ΔKD

After constructing the table of fuzzy control rule (ΔKp, ΔKi, and ΔKd), the adaptive correction is made by the
following method.
 Kp = Kp’ + ΔKp 
Ki = Ki’ + ΔKi 
 Kd = Kd’ + ΔKd 
Based on the above analysis, according to the principles of self-tuning parameters of PID, the model of fuzzy self-
tuning PID controller is shown below.
Fig. 2. Model of fuzzy self-tuning PID controller.

C MEX
According to Yusivar and Wakao [4], C MEX is an add on feature of SIMULINK that enables S-function to be
written using C. A simple C MEX S-function can be composed of a S-function block with one input port and one
output port configuration. The number of I/O elements can be multiplexed or demultiplexed using the Mux or Demux
block. S-function is identified by a name that is actually same as the C program’s file name so that MATLAB can
recognize and link each other.
There are three simple basic structures of C MEX, where each structure consists of program routines that are
meant to do specific tasks. The initialization routine is dealt with setting of I/O port width number, sample time,
number of states (discrete, continuous, or none), and initialing the values of those states. A state is a calculated value
that will be used in the next iteration. The mdlOutputs is the area for formulating outputs, usually in a simple math
equation. The mdlUpdate routine is used to express the algorithm. Usually, the structure with a mdlUpdate routine is a
discrete function. The mdlDerivatives is required by a continuous system. In this routine, SIMULINK integrates the
differential equations using variety methods that can be chosen. In the Termination routine, may be performed any
actions that are necessary at the termination of a simulation, such as for free the memory that has been allocated.
Structure A is used for a function that has only feed through calculations for outputting data from inputs. The
sample time for a function with structure A can be discrete or continuous, but it has no state. Therefore, structure A
has no mdlUpdate nor mdlDerivatives routine. Structure B is usually for a discrete system. All algorithms are written
in the mdlUpdate routine that may also consist a discrete integration created by the user. This routine is only for
updating data or states every time the sample time is hit. SIMULINK doesn’t do any integration process. Structure C
is used for a continuous system. SIMULINK will automatically process the integration of any differential equations
written down in the mdlDerivatives area. The integration results will update the states automatically. Then, by
combining the simple C MEX S-functions with any structures, we can simulate dynamic systems such as motor drive
system and distillation process with a constant and scope blocks additions for parameter setting and data displaying
purposes.

Fig. 3. Simple C-MEX S-function configuration.

Fig. 4. Simple structure of C-MEX file.

PROCESS DESCRIPTION AND MODEL

The fuzzy self-tuning PID controller is implemented on a 2 × 2 linearized distillation column model by McAvoy
and Weischedel (1981) as shown in Marlin [2]. The model is based on a two-product distillation column separating a
binary feed as shown in Fig. 5 with the initial conditions for the model is shown in Table 6.
Fig. 5. Example of a distillation column.

TABLE IV. INITIAL CONDITIONS OF THE MODEL

Parameters Value
Relative volatility 2.4
Number of trays 17
Feed tray 9
Analyzer dead times 2 min
Feed light key (XF) 0.50
Distillate light key (XD) 0.98 mole fraction
Bottoms light key (XB) 0.02 mole fraction
Feed flow (FF) 10.0 kgmole/min
Reflux flow (FR) 8.53 kgmole/min
Distillate flow (FD) 5.0 kgmole/min
Reboiler flow (FV) 13.53 kgmole/min
Tray holdup (H) 1.0 kgmole
Holdup in drums (HD) 10.0 kg mole

The transfer functions are derived based on the schematic diagram of binary distillation column and the linearized
equation is shown below.

 

Based on Fig. 5 and Eq. 17, the manipulated variables are reflux and reboiler flow rates, the controlled variables
are distillate and bottoms compositions, and the disturbance is the feed composition. Meanwhile, other parameters
such as pressure and levels are controlled tightly.
The control configuration used in the distillation column is a two-point PI control with LV-configuration. A two-
point PI control means that both distillate and bottom compositions are controlled, especially when both distillate and
bottom products are valuable, so that the effect of disturbances will be small at low frequencies. While LV-
configuration means that the reflux flow (FR) and boil-up flow (FV) are the variables used to control both distillate and
bottom compositions, whereas the distillate flowrate (D) and the bottom flowrate (B) are used for level control [5].

SIMULATION
Figures, tables, and equations must be inserted in the text and may not be grouped at the end of the paper.
Important: A miscount of figures, tables, or equations may result from revisions. Please double check the numbering
of these elements before you submit your paper to your proceedings editor.
 A. Open Loop Simulation
The open loop simulation is done without any controller with the Simulink structure example and how the changes
are performed given in Giwa et al. [6]. The aim of the open loop simulation is to understand the effect of the
manipulated variables and disturbance towards the controlled variables. The open loop simulation is done by giving a
positive step of 10% for both manipulated variables and disturbance. For FR, the value is changed from 8.53 to 9.383.
For FV, the value is changed from 13.53 to 14.883. For XF, the value is changed from 0.4 to 0.44.

Fig. 6. Open loop / dynamic simulation of the system.

If you need to arrange a number of figures, a good tip is to place them in a table, which gives you additional
control of the layout. Leave a line space between your figure and any text above it, like this one:

(a) (b)
FIGURE 1. To format a figure caption use the Microsoft Word template style: Figure Caption. The text “FIGURE 1,” which
labels the caption, should be bold and in upper case. If figures have more than one part, each part should be labeled (a), (b), etc.
Using a table, as in the above example, helps you control the layout

Cite all figures in the text consecutively. The word “Figure” should be spelled out if it is the first word of the
sentence and abbreviated as “Fig.” elsewhere in the text. Place the figures as close as possible to their first mention
in the text at the top or bottom of the page with the figure caption positioned below, all centered. Figures must be
inserted in the text and may not follow the Reference section. Set figure captions in 9 point size, Times Roman font.
Type the word “FIGURE 1.” in bold uppercase, followed by a period.

B. Closed Loop Simulation

The closed loop simulation is done with the controller scheme purposed in this paper as shown in Fig. 7. To have a
clear understanding whether the controller scheme purposed in this paper has a better performance, a closed loop
simulation with the PI controller is also done as shown in Fig. 8. The tuning constants used for the PI controller are
calculated by the MATLAB auto tuner. The performance of each controller is analyzed from various parameters such
as overshoot, settling time, and IAE (Integral Absolute Error). Smaller overshoot, smaller settling time, and smaller
IAE value are desired from a controller.

Fig. 7. Closed loop simulation of the system with fuzzy self-tuning PI controller.

Fig. 8. Closed loop simulation of the system with PI controller.

C. Simulation Testing
Simulation testing is done by setpoint changes and disturbance rejection testing to see the aggressiveness (how
aggressive the controller behaves towards new setpoint value) and the robustness (how robust the controller behaves
to handle disturbance) of the controllers.
Setpoint change testing is done by two methods: (1) changing ±5% of the XD setpoint to see the changes in both
XD and XB and (2) changing ±5% of the XB setpoint to see the changes in both XD and XB. Changing the XD setpoint
will result in a disturbance for XB, while changing the XB setpoint will result in a disturbance for XD. By changing the
XD and XB setpoint in different simulations and not done simultaneously, the simulation results will show pure effect
of the XD setpoint change towards XB and pure effect of the XB setpoint change towards XD respectively.
Disturbance rejection testing is done by changing ±10% of the XF value to see the changes in both XD and XB.
 Due to the wide range and complexity of tables, we simply offer an example for guidance. Please follow the
style for table (and figure) captions.
TABLE 1. To format a table caption, use the Microsoft Word template style: Table Caption. The text
“TABLE 1,” which labels the caption, should be bold and all letters capitalized. Center this text above the
Table. Tables should have top and bottom rules, and a rule separating the column heads from the rest of the table
only.
Column Header Goes Here Column Header Goes Here Column Header Goes Here
Row Name Here x x
Row Name Here x x
Row Name Here x x

RESULTS AND DISCUSSION

Here are the main points you need to follow (the AIP author template packages contain comprehensive
guidance):

A. Open Loop Simulation

Fig. 9 and Fig. 11 show the dynamic response of the system to positive steps in reflux flow rate (FR) and feed
composition (XF) for distillate and bottom composition respectively. Fig. 10 and Fig. 12 show the dynamic response
of the system to positive steps in reboiler flow rate (FV) and feed composition (XF) for distillate and bottom
composition respectively. The results show that when the reflux flow rate and feed composition are increased, both
distillate and bottom composition are increased. Meanwhile, when the reboiler flow rate and feed composition are
increased, both distillate and bottom composition are decreased. The results are in accordance with the transfer
functions model given by the case study used in this paper.

Fig. 9. Dynamic response of the system to positive steps in reflux flow rate (FR) and feed composition (XF) for distillate composition (XD).
Fig. 10. Dynamic response of the system to positive steps in reboiler flow rate (FV) and feed composition (XF) for distillate composition (XD).

Fig. 11. Dynamic response of the system to positive steps in reflux flow rate (FR) and feed composition (XF) for bottom composition (XB).

Fig. 12. Dynamic response of the system to positive steps in reboiler flow rate (FV) and feed composition (XF) for bottom composition (XB).
 B. Closed Loop Simulation
Fig. 13 and Fig. 14 show the comparison of the system’s response between PI and fuzzy PI controller for both
distillate and bottom compositions respectively. From Fig. 13, the fuzzy PI controller shows better control
performance for XD because it can reach the setpoint on a much faster rate (at around t = 600 s) than the PI controller
(at around t = 1500 s). From Fig. 14, the fuzzy PI controller also shows better control performance for XB because its
overshoot is three time less than the PI controller. From Fig. 13 and Fig. 14, the IAE values are evaluated based on the
area under the setpoint graph. The calculation result can be seen on Table V, where the fuzzy PI controller has smaller
IAE values for both XD and XB, or in other words, has a better control performance.

Fig. 13. Comparison of the system’s dynamic response between PI and fuzzy PI controller for XD.

Fig. 14. Comparison of the system’s dynamic response between PI and fuzzy PI controller for X B.

TABLE V. IAE VALUES FOR FUZZY PI AND PI CONTROLLER

AE of Fuzzy PI Controller IAE of PI Controller
XD XB XD XB
68.5967 15.2723 105.6351 107.4742

C. Setpoint Change Testing

Fig. 15 and Fig. 16 show the dynamic response of the system towards +5% XD setpoint change at t = 2000 s and -
5% XD setpoint change at t = 3200 s for XD and XB respectively. Fig. 17 and Fig. 18 show the dynamic response of
the system towards +5% XB setpoint change at t = 2000 s and -5% XB setpoint change at t = 3200 s for XD and XB
respectively.
Fig. 15. Dynamic response of the system towards +5% XD setpoint change at t = 2000 s and -5% XD setpoint change at t = 3200 s for XD

Fig. 16. Dynamic response of the system towards +5% XD setpoint change at t = 2000 s and -5% XD setpoint change at t = 3200 s for XB

Fig. 17. Dynamic response of the system towards +5% XB setpoint change at t = 2000 s and -5% XB setpoint change at t = 3200 s for XD
Fig. 18. Dynamic response of the system towards +5% XB setpoint change at t = 2000 s and -5% XB setpoint change at t = 3200 s for XB

D. Disturbance Rejection Testing

Fig. 15 and Fig. 16 show the dynamic response of the system towards +10% disturbance at t = 2000 s and -
10% disturbance at t = 3200 s for XD and XB respectively.

Fig. 19. Dynamic response of the system towards +10% disturbance at t = 2000 s and -10% disturbance at t = 3200 s for XD

Fig. 20. Dynamic response of the system towards +10% disturbance at t = 2000 s and -10% disturbance at t = 3200 s for XB

CONCLUSION
Ut wisi enim ad minim veniam, quis nostrud exerci eliton ullamcorper suscipit lobortis nisl ut aliquip ex en
commodo consequat. Duis te feugifacilisi per suscipit lobortis nisl ut aliquip ex en commodo consequat.Lorem
ipsum dolor sit amet, consectetuer adipiscing elit, sed diem nonummy nibh euismod tincidunt ut lacreet dolore
magna aliguam erat volutpat.
 REFERENCES
1. Almeida, O. D. M. and Coelho, A. A., 2002. A fuzzy logic method for autotuning a PID controller: SISO and
MIMO systems. In 15th IFAC World Congress on Automatic Control.
2. Marlin, T. E. 1999. Process Control: Designing Processes and Control Systems for Dynamic Performance, 2 nd
Edition, New York: McGraw-Hill
3. Si, X. and Wang, J., 2010, November. A Vector-control System Based on Fuzzy Self-tuning PID controller for
PMSM. In E-Product E-Service and E-Entertainment (ICEEE), 2010 International Conference on (pp. 1-4).
IEEE.
4. Yusivar, F. and Wakao, S., 2001. Minimum requirements of motor vector control modeling and simulation
utilizing C MEX S-function in matlab/simulink. IEEE journal.
5. M. Thone, M. Potters and S. Baldi, "Control configurations in distillation columns: A comparative study,"
2016 European Control Conference (ECC), Aalborg, 2016, pp. 37-42.
6. Giwa, A., Owolabi, J.O. and Giwa, S.O., 2017. System identification and IMC-based PID control of a reactive
distillation process: a case study of n-butyl acetate production. In International Journal of Engineering
Research in Africa (Vol. 31, pp. 104-119). Trans Tech Publications.