Expert Systems With Applications 42 (2015) 8533–8549

Contents lists available at ScienceDirect

Expert Systems With Applications

journal homepage: www.elsevier.com/locate/eswa

A fractional order fuzzy PID controller for binary distillation

column control
Puneet Mishra∗, Vineet Kumar, K.P.S. Rana
Division of Instrumentation and Control Engineering, Netaji Subhas Institute of Technology, Azad Hind Fauj Marg, Sector-3, Dwarka, New Delhi, PIN-110078, India

a r t i c l e i n f o a b s t r a c t

Keywords: Expert and intelligent control schemes have recently emerged out as a promising solution with robust-
Binary distillation column ness which can efficiently deal with the nonlinearities, along with various types of modelling uncertainties,
Fractional order control
present in different real world systems e.g. binary distillation column. This paper is an attempt to propose an
Fuzzy logic control
intelligent control system which takes the form of a fractional order fuzzy proportional–integral–derivative
Genetic algorithm
Robust control (FOFPID) controller which is investigated as a solution to deal with the complex dynamic nature of the distil-
lation column. The FOFPID controller is an extension of an existing formula based self tuning fuzzy propor-
tional integral controller structure, which varies its gains at run time in accordance with the instantaneous
error and rate of change of error. The FOFPID controller is a Takagi–Sugeno (TS) model based fuzzy adaptive
controller comprising of non-integer order of integration and differentiation operators used in the controller.
It has been observed that inclusion of non-integer order of the integration and differentiation operators made
the controller scheme more robust. For the performance evaluation of the proposed scheme, the performance
of FOFPID controller is compared with that of its integer order counterpart, a fuzzy proportional–integral–
derivative (FPID) controller. The parameters of both the controllers were optimized for minimum integral
of absolute error (IAE) using a bio-inspired global optimization algorithm, genetic algorithm (GA). Intensive
LabVIEWۛ simulation studies were performed which included setpoint tracking with and without uncertain-
ties, disturbance rejection, and noise suppression investigations. For testing the parameter uncertainty han-
dling capability of the proposed controller, uncertain and time varying relative volatility and uncertain tray
hydraulic constant were applied. Also, for the disturbance rejection studies, intensive simulations were con-
ducted, which included two most common causes of disturbance i.e. variation in feed composition and vari-
ation in feed flow rate. All the simulation investigations clearly suggested that FOFPID controller provided
superior performance over FPID controller for each case study i.e. setpoint tracking, disturbance rejection,
noise suppression and parameter uncertainties.
© 2015 Elsevier Ltd. All rights reserved.

1. Introduction textile industries to chemical industries etc. These industries usually

include processes which exhibit highly nonlinear, uncertain and cou-
For past two decades, research in expert and intelligent systems pled behaviour and hence need an expert and intelligent solution for
and their applications in the industrial environment have soared to their efficient operation. Out of these, chemical industries are the one
a great extent. Artificial intelligence is one of the tools of expert and which is associated to a large proportion of the world population
intelligent system which is being widely studied to exploit its poten- and have a high impact on their lives, due to their nature of opera-
tial in the fields of robotics, pattern recognition, decision making etc. tion and products which include food, pharmaceutics, beverages and
It is not a question anymore that the use of expert and intelligent sys- petroleum products etc. One of the main processes, on which chem-
tems can make the industrial environment more profitable and effi- ical industries potentially rely, is distillation process. Distillation is
cient to a considerable amount with the inclusion of human knowl- used for the separation and purification of mixtures containing two
edge in the test and automation systems. Various industries, which or more substances. However, distillation process is characterized by
can be (or are being) benefitted from the use of expert and intelli- highly nonlinear and uncertain dynamics and thus their control be-
gent systems, range from automotive industries, food industries, and comes a challenging task. Since distillation processes usually require
large amount of heat for their operations and the thermodynamic
∗
efficiency is lesser than 10% (Smith, 2012; Tham, 2009), distillation
Corresponding author. Tel.: +91 11 25099050; fax: +91 11 25099022.
E-mail addresses: puneet.mishra@ymail.com (P. Mishra),
control becomes a necessity. However, use of an appropriate control
vineetkumar27@gmail.com (V. Kumar), kpsrana1@gmail.com (K.P.S. Rana). algorithm, employing expert and intelligent techniques, along with

http://dx.doi.org/10.1016/j.eswa.2015.07.008
0957-4174/© 2015 Elsevier Ltd. All rights reserved.
8534 P. Mishra et al. / Expert Systems With Applications 42 (2015) 8533–8549

optimization methods can lead to an increment in the productivity for optimization of operation policy of reactive batch distillation. The
and profitability up to 25% (Liptak, 2009). Conventional controllers, highest yield and mole fraction of ethyl acetate was claimed to be
such as proportional–integral–derivative (PID), fail to provide satis- achieved through the use of the obtained optimization policy. It was
factory performance for such processes with nonlinear and uncertain also claimed that reduced model (ANFIS) was able to reduce CPU use
dynamics (Miccio & Cosenza, 2014), thereby application of an expert up to 1/18,000 times that of a real mathematical model (Khazraee,
and intelligent system is desired for effective control of these pro- Jahanmiri, & Ghorayshi, 2010). Moghadam et al. proposed the ap-
cesses. To cater to these issues and needs, several control strategies plication of LQR for controlling concentration profiles along a cat-
have been presented over the time for the control of distillation pro- alytic distillation column (Moghadam, Aksikas, Dubljevic, & Forbes,
cess. Following is a chronological survey conducted on this subject 2011).
for the past decade. Urselmann et al. proposed memetic algorithm specifically for
Recent developments in this field are more inclined to the use of the design of chemical processes. The application of the algorithm
intelligent controllers, adaptive controllers, nonlinear controllers or was studied for the design of a RD column and it was claimed
a combination of these such as fuzzy logic control, nonlinear model that memetic algorithm is the only one that achieves the optimal
predictive control (NLMPC) etc. Intelligent control structures using design solution in reasonable computation times even with large
fuzzy logic (Zadeh, 1965) have been successfully applied to the con- number of design parameters and nonlinear constraints (Urselmann,
trol of several nonlinear systems and they have been found to be Barkmann, Sand, & Engell, 2011). Karimi and Salahshoor proposed
quite effective. Kawathekar and Jiggs applied NLMPC to a highly non- a new fault detection and diagnosis approach for a distillation col-
linear reactive distillation (RD) column. NLMPC provided a factor umn based on combined principal component analysis and ANFIS
of 2–3 times better performance than the corresponding diagonal scheme and it was found that the proposed approach could detect
proportional-integral (PI) controller. The effect of process/model mis- the faults in a wider range, when compared with the existing ap-
match on the performance of the NLMPC controller was also eval- proaches (Karimi & Salahshoor, 2012). Szabó et al. proposed a three
uated (Kawathekar & Jiggs, 2007). Sanandaji et al. proposed a non- level hierarchical controller for binary distillation column. The hier-
linear fuzzy identification approach based on genetic algorithm (GA) archical controller was based on the relationships between tempera-
and Takagi–Sugeno (TS) fuzzy system for multi input multi output ture of trays and product purities (Szabó, Nemeth, & Szeifert, 2012).
(MIMO) dynamical system and established that the proposed ap- Mishra et al. designed and studied the application of model predic-
proach provided better results than the conventional TS fuzzy iden- tive control (MPC) for a binary distillation column. In this work a
tification approach (Sanandaji, Salahshoor, & Fatehi, 2007). Singh et method of removing ringing effect generated in MPC is also addressed
al. proposed an approach for inferential control of distillation col- (Mishra, Khalkho, Kumar, & Dan, 2013). Rani et al. developed a novel
umn based upon Levenberg–Marquardt (LM) algorithm. The pro- approach for the design of soft sensor based upon LM approach and
posed scheme was compared with that of an ANN based estimator adaptive linear network. Inferential controllers using different soft
using steepest descent back propagation algorithm and it was found sensors including the proposed approach i.e. dynamic adaptive lin-
that the problems associated with the older estimator such as satu- ear network based inferential controller (DADIC) are tested for the
rated outputs and erroneous results, were suppressed by using devel- control of multi-component distillation column. It was claimed that
oped approach (Singh, Gupta, & Gupta, 2007). DADIC was more robust and performed better for the setpoint track-
Kumar and Kaistha studied the impact of steady-state multiplici- ing and disturbance rejection (Rani, Singh, & Gupta, 2013). Safe et
ties on the control of a simulated industrial scale methyl acetate RD al. implemented an optimization strategy based upon response sur-
column. The study brings out the importance of input–output rela- face methodology involving differential evolution to carry out reac-
tions for control system design and helps to understand the complex tion and separation simultaneously in a reactive dividing wall batch
dynamic behaviour of RD systems (Kumar & Kaistha, 2008). Jana and distillation column (Safe, Khazraee, Setoodeh, & Jahanmiri, 2013).
Adari proposed an advanced adaptive control structure for a batch RD More recently type 2 fuzzy logic controllers (FLC) are being stud-
column. The nonlinear adaptive control law consisted of the generic ied for counteracting the effects of the uncertainties in the plant
model controller and an adaptive state estimator. This nonlinear con- model, a typical reason for the control performance deterioration.
trol strategy shows comparatively better closed-loop performance Use of type 2 FLC can be seen successfully applied for various
than the gain-scheduled PI controller (Jana & Adari, 2009). Jones and chemical processes such as nonlinear bioreactor (Galluzo & Cosenza,
Hengue proposed the use of GA to determine the optimum weight- 2009; Galluzzo, Cosenza, & Matharu, 2008), nonisothermal continu-
ing matrices for linear quadratic regulator (LQR) to control a binary ous stirred tank reactor (Galluzo & Cosenza, 2011), and binary distil-
distillation column nonlinear model (Jones & Hengue, 2009). Tan et lation column (Miccio & Cosenza, 2014). The advent of type 2 FLC was
al. presented a novel approach for designing sets of ternary periodic more attracted by researchers due to the lack of uncertainty handling
signals with different periods for MIMO system identification. Three capability, and difficulties related to the optimal choice of the type 1
of these signals were applied to identify the transfer function matrix fuzzy sets for type 1 FLCs. Despite the fact that type 2 fuzzy logic can
as well as the singular values of a simulated distillation column (Tan, handle uncertainties in the plant model to some extent, there is al-
Godfrey, & Barker, 2009). Zakeri and Sadeghi proposed a robust H∞ /L1 ways a need of more robust and adaptive controller for the nonlinear
control system for a high purity distillation column. The authors used complex processes.
multi-objective performance criterion i.e. H∞ and L1 with the inclu- The control engineers always strive for inclusion of additional pa-
sion of linear matrix inequalities. Simulation studies suggested that rameters in the controller as it offers more design freedom. Imple-
the proposed controller adequately performed for reference tracking mentation of FLC essentially requires the derivative and the inte-
and disturbance rejection (Zakeri & Sadeghi, 2010). Kariwala and Cao gral mathematical operators. In simulation, these operators can be
proposed novel bidirectional branch and bound approach for con- easily implemented using conventional Laplace operator ‘s’. In many
trolled variable selection using the minimum singular value rule and recent research woks, a more flexible variant of this operator, frac-
the local average loss criterion in the framework of self-optimizing tional order operator sμ , where ‘μ’ is a non integer, has been used.
control. Random matrices and binary distillation column case study This implementation gives an additional degree of freedom, i.e. ‘μ’,
are used to demonstrate the computational efficiency of the proposed to the control engineer. The control algorithm implemented using
method (Kariwala & Cao, 2010). Khazraee et al. proposed the appli- fractional order operator is called fractional order control. Fractional
cation of adaptive neuro fuzzy inference system (ANFIS) instead of order modelling and control recently has emerged as one of the inno-
the highly nonlinear model of a reactive batch distillation column vative tool for the control designers to make control systems more ro-
for optimization. Differential evolution algorithm has been employed bust as these are claimed to represent the natural systems with more
 P. Mishra et al. / Expert Systems With Applications 42 (2015) 8533–8549 8535

reality. Use of fractional order control has been recommended to artificial neural network, can be seen in a recent work by Ramli, Hus-
systems with time varying parameters (Sabatier, Oustaloup, Iturricha, sain, Jan, & Abdullah (2014). Ramli et al. implemented an equation
& Lanusse, 2002) and in some other fields, such as, automatic voltage based neural network (NN) model for the prediction of composition
regulator (AVR) (Das, Pan, & Das, 2013; Das, Pan, Das, & Gupta, 2012; of top and bottom products of a debutanizer column. The NN based
Pan & Das, 2012), coupled tank system (Delavari, Ghaderi, Ranjbar, & model was compared with two other methods based on partial least
Momani, 2010), robotic manipulator control (Delavari, Ghaderi, square method and regression analysis method. This work is a nice
Ranjbar, HosseinNia, & Momani, 2010; Sharma, Rana, & Kumar, 2014). example of use of intelligent techniques for effective operation of an
Another instance of use of fractional order differ integrators in chem- industrial system (Ramli et al. 2014). These recent works clearly de-
ical process control can be seen in 13 C separation process (Dulf, pict that how expert and intelligent systems are being explored for
Clement, Dulf, Pop, & Both, 2011). Literature presented here clearly the process industries to enhance the efficiency of the system under
suggests that the application of fractional order control for uncertain consideration. The current work is also an attempt in the same trend
plants can provide greater robustness. which thrives to improve the plant profitability by the use of expert
More recently, some other controllers using expert or intelligent system which is based on the intelligent technique i.e. fuzzy control.
techniques have also been proposed to deal with the nonlinear and The expertness in the system is due to the inclusion of human intel-
uncertain nature of various industrial processes. Savran and Kahra- ligence in the system using fuzzy logic system through its rule base,
man developed an adaptive PID control technique which used fuzzy which changes its gains in run time without human intervention or
process model for the adaptation of the gains of PID controller (Savran supervision. The current work is an extension of the control scheme
& Kahraman, 2014). They used a fuzzy predictor to obtain multi-step proposed by Ding, Ying, and Shao (1999) and is quite different and
ahead output of the process, which was further used to adapt the simple from the recent reported approaches. Ding et al. presented the
gains of the PID controller by minimizing the sum of the squared er- analytical structure of a TS based fuzzy PI controller with nonlinear
rors between the predicted output and the reference input. The fuzzy gains having a unique feature of self tuning capability i.e. it changes
predictor was trained on-line to keep a track of variations in param- its gains in run time in accordance with the instantaneous error (e)
eters of a plant, which in this case was a bioreactor. The scheme was and rate of change of error (r). The authors tested the controller to
able to compensate for large parametric variations and also was ro- control temperature in a tissue hyperthermia therapy in simulation
bust to noisy measurements. Fereidouni et al. presented two struc- environment (Ding et al., 1999). Having self tuning capability and non
tures of fuzzy PID controllers, out of which one was said to be non- linear gains this controller can effectively deal with complex plants
adaptive PID fuzzy logic controller (NA-PID-FLC) while the other one with strong nonlinear and uncertain dynamics.
was adaptive PID fuzzy logic controller (A-PID-FLC). Both of these The present work proposes to utilise the base innovative design by
controllers derived their gains from a fuzzy system on the basis of Ding et al. (1999) with an enhancement in terms of PID action and in-
error and rate of change of error. The only difference between these clusion of the fractional order operators for integral and derivative
two was that the A-PID-FLC had a provision for the adjustment of the operations to control binary distillation column. This combination
output scaling factor using another fuzzy system, while the NA-PID- can be expected to inherit the best features of self tuning and the
FLC did not have such mechanism (Fereidouni, Masoum, & Moghbel, fractional order operators, thereby yielding a superior performance.
2015). The parameters were tuned by two different optimization al- It may be noted that for the distillation control undertaken here a
gorithms, i.e. particle swarm optimization (PSO) and hybrid bacterial fuzzy PID (FPID) and a fractional order fuzzy PID (FOFPID) structure
foraging optimization. The developed controllers were tested in sim- based on the above design were implemented wherein both FPID
ulation mode on different systems including linear systems, systems and FOFPID actions were realised as ‘fuzzy PI + fuzzy PD’ and ‘frac-
with time delay, and a simple nonlinear system also. It was claimed tional order fuzzy PI + fractional order fuzzy PD’ respectively. The
that the A-PID-FLC was more capable than NA-PID-FLC, to control performance of the investigated FPID and FOFPID controller for main-
the systems for servo and regulation problems. Gizi et al. presented taining the desired product specification for some common practical
an integrated approach which used Sugeno fuzzy model in coordi- cases, such as variation in feed flow rate (FL ), feed composition (ZL )
nation with radial basis function neural network to determine the and relative volatility (α ) has been evaluated in detail in this work.
optimal parameters of a PID controller. The authors have claimed to Additionally, the effect of uncertainty of the tray hydraulic constant
achieve a high sensitive response of an AVR system and have com- (β ) has also been investigated, as usually in practice the β is as-
pared the proposed algorithm with the real-coded genetic algorithm sumed to be constant for each tray of distillation column, but in the
tuned PID controller, PSO tuned PID controller and a linear-quadratic real world scenario it may not be true and the value of β may also
regulator method to control the AVR system (Gizi, Mustafa, & Jebur, vary in a range. Measurement noise being one of the important pa-
2014). Xu et al. also developed a data driven adaptive control ap- rameter to have deteriorating effect on the product quality, simula-
proach for the MIMO systems. The controller is based on the mul- tion study has also been conducted to investigate the effectiveness
tiobserver models which were online derived from the input–output of the designed controller for noise suppression. Overall, the moti-
data of the controlled plant. Closed loop stability was also assured by vation behind this study is to show the requirements and the effec-
using Lyapunov stability criterion. The control strategy was validated tiveness of a self-tuning intelligent robust controller to achieve the
on a linear model of a MIMO system i.e. Wood–Berry distillation col- desired top and bottom product specification for binary distillation
umn and it was found that the control strategy was able to handle column.
process dynamic variations and provided satisfactory setpoint track- This paper is organized as follows. Section 2 presents a brief prob-
ing and disturbance rejection (Xu, Jiang, & Shi, 2014). Dounis et al. lem formulation and put an overview of the dynamic model of bi-
proposed an adaptive fuzzy gain scheduled PID scheme for designing nary distillation column used in this paper. Section 3 presents the
maximum power point tracking controller for photovoltaic systems. structure of the FPID and FOFPID controllers along with the imple-
The adaptation was incorporated in the control scheme by adjust- mentation aspects of the fractional order calculus used in this work.
ing the input scaling factors of the fuzzy gain scheduler using a fuzzy This section also presents the tuning criteria for the parameters of
system itself. It was claimed that the proposed approach was able both the controllers. The performance comparison of FPID and FOF-
to provide good maximum power operation under variety of con- PID controllers in terms of setpoint tracking, disturbance rejection
ditions such as different levels of solar radiation, photovoltaic (PV) and robustness testing are shown in Section 4. Finally, concluding re-
cell temperature for different PV sources (Dounis, Kofinas, Alafodi- marks along with the discussion on the findings and future research
mos, & Tesels, 2013). Another instance of use of intelligent technique, scope are presented in Section 5.
8536 P. Mishra et al. / Expert Systems With Applications 42 (2015) 8533–8549

Table 1
Parameters’ values for the simulation of binary distillation column (Jana,
2010).

Binary system (1-propanol/ethanol) Value Unit

Flow rate of liquid distillate, DL 27.8 lbmol/h

Composition of liquid distillate, XD 0.899 mol fraction
Flow rate of vapour distillate, DV 200 lbmol/h
Composition of vapour distillate, YD 0.947 mol fraction
Liquid holdup in reflux drum, MD 8.056 lbmol
Reflux flow rate, R 400.864 lbmol/h
Flow rate of liquid feed, FL 800 lbmol/h
Composition of liquid feed, ZL 0.60 mol fraction
Flow rate of vapour feed, FV 200 lbmol/h
Composition of vapour feed, ZV 0.53 mol fraction
Composition of bottom product, XB 0.481 mol fraction
Liquid holdup in column base, MB 4.897 lbmol
Vapour boil-up rate, VB 428.66 lbmol/h
Hydraulic time constant, β 3.6 s
Relative volatility, α 2.0 –
ODE solver Runge–Kutta-4 –
Step size 0.001 h
Total number of trays, NT 15 –
Feed tray location, NF 5 –

Fig. 1. Schematic diagram of binary distillation column.

Based upon the above factors it is quite obvious that there is a
need of an intelligent and robust controller which can maintain the
system performance in case of model uncertainties occurrences along
with the capability to handle various disturbances in this type of
2. Problem formulation highly interacting system. Table 1 shows the simulation parameters
and operating conditions used in this study. The differential equa-
The distillation column used in this paper is a binary distilla- tions related to the plant modelling can be stated as follows (Jana,
tion column used to separate 1-propanol and ethanol. Fig. 1 shows 2010).
a schematic diagram of a conventional binary distillation column Varying liquid holdups are assumed on each tray and are calcu-
which depicts some of the required inputs and components present lated through a linearized form of the Francis weir formula,
in a large scale distillation column. Feed contains a mixture of the
two components to be separated. Separation of the components in Mn − Mn0
Ln = Ln0 + (1)
the feed is achieved on the basis of the fact that the vapour of a boil- β
ing mixture will be richer in the components that have lower boil-
where Ln , Ln 0 , and Mn 0 are internal liquid flow rate, reference value of
ing points. Therefore, when this vapour is cooled and condensed, the
internal liquid flow rate, and reference molar holdup for the nth tray
condensate will contain more volatile components. At the same time,
respectively.
the original mixture will contain more of the less volatile material.
Condenser and reflux drum:
Since this separation process requires large amount of heat in terms
Total continuity:
of boiling the mixture and condensing the vapour, there is a need of
an efficient control so as to increase the productivity and profitability dMD
= VNT − (R + DL + DV ) (2)
of plant. dt
There are various factors due to which the required product purity Component continuity (more volatile component):
is affected in a distillation column. Feed disturbances, such as change
in FL and ZL can cause a deviation in the distillate composition (XD ) d(MD XD )
= VNT YNT − (R + DL )XD − DV YD (3)
and bottom composition (XB ) from their respective setpoints. Also the dt
distillation column is a MIMO process, and there is a strong interac- Top tray (n = NT):
tion in different loops present in the system. Total continuity:
Any deviation from the steady state value in one loop will require
a suitable change in the controller action of other interacting loops to dMNT
= R + VNT −1 − LNT − VNT (4)
maintain their respective setpoints. Moreover, α is a major source of dt
the uncertainty in the distillation process. It is a measure of the differ- Component continuity (more volatile component):
ence between the vapour pressure of the more volatile components
in liquid mixture (Miccio & Cosenza, 2014). Since this parameter sig- d(MNT XNT )
= RXD − LNT XnT + VNT −1YNT −1 − VNT YNT (5)
nificantly varies with pressure, temperature and feed conditions, it dt
is not unrealistic to see controller’s capability to handle its variation nth tray:
and keeping XB and XD close to their setpoints. Besides that the value Total continuity:
of tray hydraulic constant can also cause the plant to behave differ-
ently from the expected response. The value of time constant typ- dMn
= Ln+1 − Ln + Vn+1 − Vn (6)
ically ranges from 3 to 6 s per tray (Jana, 2010; Luyben, 1990). All dt
the simulation studies in this paper were carried out by simulating Component continuity (more volatile component):
the dynamic model of binary distillation column described in detail
d(Mn Xn )
in Jana (2010). The change in the dynamic behaviour due to estima- = Ln+1 Xn+1 − Ln Xn + Vn−1Yn−1 − VnYn (7)
tion uncertainties in β can also account for significant changes in the dt
closed loop control performance. Feed tray (at n = NF):
 P. Mishra et al. / Expert Systems With Applications 42 (2015) 8533–8549 8537

Fig. 2. (a) Parallel structure of FPID. (b) An alternative structure of FPID (Mann, Hu, & Gosine, 1999).

Total continuity: accomplish this task, different control structures are available in the
literature as shown in Fig. 2(a) and (b). Fig. 2(a) shows parallel struc-
dMNF
= LNF +1 − LNF + VNF −1 − VNF + FL (8) ture of a fuzzy PID controller and Fig. 2(b) is another solution to
dt
achieve fuzzy PID action. The structures shown in Fig. 2(a) and (b)
Component continuity (more volatile component): are seldom used to achieve a direct action fuzzy PID controller. The
d(MNF XNF ) foremost reason being the complex implementation of the parallel
= LNF +1 XNF +1 − LNF XNF + VNF −1YNF −1 structure shown in Fig. 2(a), very high computational complexity and
dt
− VNF YNF + FL ZL (9) requirement of rigorous intuitive knowledge about the design of each
fuzzy system. Similarly, the controller structure shown in Fig. 2(b)
Bottom tray: with three linguistic variables as ‘Error’, ‘Rate of change of the error’,
Total continuity: and ‘Integral of error’, would be again quite complex to implement
dM1 and can have as many as 73 = 343 rules if linguistic variables are de-
= L2 − L1 + VB − V1 (10) fined using 7 membership function.
dt
A simpler control structure shown in Fig. 3 is quite often used in
Component continuity (more volatile component): literature to directly generate a fuzzy PID action and is also employed
d(M1 X1 ) in the current work. The advantage of this structure is that it requires
= L2 X2 − L1 X1 + VBYB − V1Y1 (11) only two linguistic variables ‘Error’ and ‘Rate of change of error’ and
dt
enjoys the liberty to share same rule base for the generation of fuzzy
Reboiler and column base:
PI and fuzzy PD component to form a cumulative fuzzy PID action. Re-
Total continuity:
duction in the number of linguistic variables catastrophically reduce
dMB the number of rules in the fuzzy system and hence the computational
= L1 − VB − B (12)
dt complexity.
Component continuity (more volatile component):
3.1. FPID controller description
d(MB XB )
= L1 X1 − VBYB − BXB (13)
dt The key element in FPID controller is formula based fuzzy con-
Other than the equations written above vapour liquid equilibrium troller (FBFC) (Ding et al., 1999), which is inherently a variable gain TS
was also used for relating the vapour phase composition to liquid fuzzy PI controller using two linguistic variables, i.e. ‘e’ and ‘r’. These
phase composition of more volatile component taking part in the dis- linguistic variables are defined over two membership functions, viz.
tillation process. ‘positive (P)’ and ‘negative (N)’. The two membership functions are
α Xn characterized by a real number ‘L’ and are symmetrical in the uni-
Yn = (14) verse of discourse. Typical variation in these membership functions
1 + (α − 1)Xn
can be shown as in Fig. 4.
where Vn is vapour phase flow rate, Yn is vapour phase composition, The rule base of the FBFC has rules of the form,
Mn is the molar holdup, Ln is liquid phase flow rate, and Xn is liquid Ri : If ’e’ is Pj and ’r’ is Pk then qi = ai e + bi r
phase composition for nth tray in distillation column, and n = 1, 2, … , where, i = 1, 2, 3, 4, j = 1, 2 and k = 1, 2.
NF - 1, NF, NF + 1, … , NT. B and D are bottom and distillate flow rate Ri is the ith rule; Pj and Pk are the membership functions defined
respectively. for ‘e’ and ‘r’; and qi is the consequent part of the ith rule. ai and bi
are the real constants. The output of FBFC is computed in the same
3. Design of FPID and FOFPID controllers way as suggested in Ding et al. (1999). As shown in Fig. 3, the out-
put of FBFC is calculated though the instantaneous value of the error,
It is often desirable that a fuzzy controller produces the ‘three ‘e(t)’ and rate of change of error ‘r(t)’. To generate the controller out-
term’ action i.e. ‘proportional’, ‘integral’ and ‘derivative’. In order to put the entire two dimensional input space formed by e(t) and r(t) is
8538 P. Mishra et al. / Expert Systems With Applications 42 (2015) 8533–8549

Fig. 3. Structure of FPID controller.

Fig. 4. Membership function variation of ‘e’ and ‘r’.

divided in 20 input combination (IC) regions based upon the mem-

bership functions shown in Fig. 4. Depending upon the point location
i.e. (e(t), r(t)), the formulas assigned for each IC calculates the con-
troller gains and hence the output (Ding et al., 1999). This interme-
diate output of FBFC is further used to get the final output of FPID or
FOFPID controller. Conventionally, the linear form of a PI controller in
velocity form or a PD controller in position form can be written as,
   
uPI (nT ) (or uPD (nT )) = Ki or Kp e(nT ) + Kp or Kd r(nT ), (15)
where, K p , Ki and Kd are the proportional, integral and derivative gains
respectively, of a linear PI or PD controller. However, the output of
the FBFC, as shown in Eq. (16), can also be related to the PI or PD
controller form but with nonlinear gains, i.e., K1 and K2 . The output
of FBFC can be taken, either in velocity form (PI controller) or position
form (PD controller), i.e. uFPI or uFPD . The evaluation of uFPI or uFPD
Fig. 5. K1 variation for bottom composition loop using FPID controller.
is done as follows (Ding et al., 1999),
4
u μ  μ It is instructive to show the expert system behaviour of the pro-
uF PI (or uF PD ) = 
i=1 i Ri
= 4i=1 4 i (ai e(nT ) + bi r(nT ))
R

i=1 μRi μR i
4
  
i=1 posed controller by showing the variation of the gains in the error
= 4i=1 K1i (e, r)e(nT ) + K2i (e, r)r(nT ) = K1 e(nT ) + K2 r(nT ) and rate of change of error space. It can be easily seen from the sur-
(16) face plot of the gains varying in the error and rate of change of error
space. Figs. 5 and 6 show the variation of K1 and K2 for the FBFC used
in bottom composition loop using FPID controller. Figs. 7 and 8 show
μRi ai μRi bi
K1i = 4 and K2i = 4 , i = 1, 2, 3, 4, and, (17) the variation of K1 and K2 for the FBFC used in distillate composition
i=1 μRi i=1 μRi loop for FPID controller respectively.
The adaptation in the gains of FBFC comes from the fact that these The output of the FPID controller is the aggregation of two com-
gains change from point to point as e(nT) and r(nT) change, according ponents i.e. fuzzy PI and PD, and can be written as,

to the rule base of the controller. The values of the nonlinear gains K1
uF PID = KPI (uF PI )dt + K u
 PD F PD

(18)
and K2 can be written as,
 
Position f orm
4
K1 (e(nT ), r(nT )) = K1i (e(nT ), r(nT )) and
Velocity f orm
i=1
4 where, uFPID is the output of FPID controller, uFPI is the fuzzy PI
K2 (e(nT ), r(nT )) = K2i (e(nT ), r(nT )) component and uFPD is the fuzzy PD component generated by the
i=1
 P. Mishra et al. / Expert Systems With Applications 42 (2015) 8533–8549 8539

Fig. 6. K2 variation for bottom composition loop using FPID controller. Fig. 9. K1 variation for bottom composition loop using FOFPID controller.

Fig. 10. K2 variation for bottom composition loop using FOFPID controller.
Fig. 7. K1 variation for distillate composition loop using FPID controller.

3.2. FOFPID controller description

The FOFPID controller is similar to that of a FPID controller shown

in Fig. 3. The only difference between FPID and FOFPID is that the
latter employs fractional order differentiator to generate the rate of
change of error and a fractional order integrator to integrate the out-
put of FBFC to generate the fuzzy PI action. The structure of the FOF-
PID is shown in Fig. 13. Further, the FOFPID controller also uses the
same FBFC structure for its implementation but uses different values
of the parameters than the FPID controller. Due to this reason, the
variation of the K1 and K2 gains of the FBFC used in FOFPID controller,
have different nature of variation in the error and rate of change of er-
ror space. Figs. 9 and 10 show the variation of the K1 and K2 gains for
the FBFC used in bottom composition loop using FOFPID controller.
Further, Figs. 11 and 12 show the variations for K1 and K2 for FBFC
used in distillate composition loop using FOFPID controller.
The output of FOFPID controller is calculated in the same manner
Fig. 8. K2 variation for distillate composition loop using FPID controller. as that of FPID. The only difference is that now the controller output
equation is now stated as,
 
d −λ
FBFC. KPI and KPD are the scaling factors and provide an additional de- uF OF PID = KPI (uF OF PI ) + K u
 PD  F OF PD

(19)
dt −λ
gree of freedom in the FPID controller design. The structure of FBFC  
Position f orm
taken in this work is inherently a variable gain fuzzy PI controller Velocity f orm
(Ding et al., 1999). The reason for taking fuzzy PI controller rule base
in FBFC is to emphasize on improvement of the steady state perfor- It is worth mentioning here that since FPID and FOFPID con-
mance with greater efficiency since this FPID structure can practi- troller have nearly same implementation, FPID controller structure
cally reiterate the nature of controllers shown in Fig. 2(a) and (b) (Li & was achieved by putting the values of μ and λ as ‘1’ in FOFPID con-
Gatland, 1996). troller.
8540 P. Mishra et al. / Expert Systems With Applications 42 (2015) 8533–8549

former renders the freedom to the user to employ the non-integer or-
der differential calculus (Sharma et al., 2014; Valério & Costa, 2013).
Different researchers have described and used the fractional order
calculus in various forms in control and other fields. Though numer-
ical methods are available to solve fractional order differential calcu-
lus but it is easier and cost effective to approximate the behaviour of
the fractional order operators using integer order operators (Valério
& Costa, 2013) and hence approximations are widely used to repli-
cate the behaviour of fractional order operators. Many approxima-
tion methods are present in the literature for continuous time do-
main as well as for discrete time domain. Some of the widely used
continuous time domain approximations are, ‘Crone or Oustaloup ap-
proximation’ (Oustaloup, Levron, Matthieu, & Nanot, 2000; Sabatier
et al., 2002), ‘Carlson approximation’ (Carlson & Hajlijak, 1964), ‘Mat-
suda approximation’ (Matsuda & Fujii, 1993) etc. For discrete time
Fig. 11. K1 variation for distillate composition loop using FOFPID controller.
approximation generally different techniques or operators for s →
z transformation are used and then the fractional power is raised
on the approximation function, by which s → z transformation was
achieved. Further, for the realization, the approximation can be ex-
panded using various available methods, such as, MacLaurin series
expansion, Taylor’s series expansion, continued fraction expansion
etc. Another widely used discrete time approximation of fractional
order operator is Grünwald–Letnikoff approximation and has a struc-
ture of finite impulse response filter (Valério & Costa, 2013).
In this paper, ‘Oustaloup approximation’ is used for the fractional
order calculus implementation. This approximation uses a higher or-
der filter having an order of 2N + 1 and fits the approximation within a
given frequency range (ωL , ωH ). The approximation of sμ can be given
as (Oustaloup et al., 2000; Pan & Das, 2012),

k=N
s + ω Zk
sμ ≈ K  (20)
s + ωPk
k=−N

where K is gain, ωPk and ωZk are the poles and zeros of analog filter
Fig. 12. K2 variation for distillate composition loop using FOFPID controller. respectively, and are defined as,
 ω  k+N+ (2N+1
1/2)(1+ μ)

ωPk = ωL H
(21)
ωL
3.3. Fractional calculus implementation
and,
Involvement of fractional order calculus has recently got attention  ω  k+N+ 2N+1
(1/2)(1−μ)

in the field of control theory. Fractional order control is in fact, con- ω Zk = ω L H

(22)
trol of plants described by fractional order dynamics, control of plants ωL
with fractional order controllers or control of plants such that the The choice of the value of N is an important deciding factor in
overall controlled plant behave as the fractional order system (Valério the performance of the fractional order operator approximation. High
& Costa, 2013). The reason of inclusion of fractional order calculus value of N produces lesser ripples in the magnitude and phase re-
in the control systems is the fact that it can provide such a level of sponse of the filter, but contributes to complex hardware imple-
performance, in terms of robustness, that integer order differential mentation. With lesser value of N, ease in hardware implementa-
calculus cannot, which is obviously by the virtue of the fact that the tion increases but the approximation performance deteriorates due

Fig. 13. Structure of FOFPID controller.

 P. Mishra et al. / Expert Systems With Applications 42 (2015) 8533–8549 8541

Fig. 16. Variation in XB for a setpoint change of -0.081 mol fraction.

Fig. 14. Cost function vs. iteration curve for both controllers.

to ripple formation in magnitude and phase response. For the current

work, N is taken as 5 and fitting frequency range [ωL , ωH ] is taken as
[10−2 , 102 ] which is quite common for process control applications
(Das et al., 2012).

3.4. Optimization of FPID and FOFPID controller parameters

The parameter values of the controller plays a vital role in per-

formance determination of the closed loop control. Different meth-
ods are available for the tuning of linear controllers depending upon Fig. 17. Variation in XD for a step setpoint change of -0.099 mol fraction.
either the time response or the frequency response. But for nonlin-
ear controllers and especially intelligent controller, such as based on
fuzzy logic, there is no straight forward tuning method for setting For tuning purpose, the calculation of IAE_XB and IAE_XD was done
their parameters. In such cases use of time integral performance cri- for a period of 3 h after giving a setpoint change of -0.081 mol frac-
terions, such as integral of squared error (ISE), integral of the absolute tion and -0.099 mol fraction in the bottom and top composition loops
error (IAE) or integral of the time-weighted absolute error (ITAE), are respectively. The expression for the calculation of IAE_XB and IAE_XD
much helpful. For the same cause the controllers used in this study can be written as,
 3
were tuned using a weighted sum of the IAE produced in the two con-
trol loops i.e. IAE produced in XB loop (IAE_XB ) and XD loop (IAE_XD ). IAE_XB = |eXB (t )|dt (24)
0
This weighted sum, to be minimized, was assumed as the cost func-
 3
tion (J) for the tuning problem and was optimized using GA to find a
global optimum solution for the cost function minimization problem.
IAE_XD = |eXD (t )|dt (25)
0
The reason for employing GA to tune the controllers’ parameters in
this work is its ability to deal with large number of decision variables The cost function vs. iteration curve for both the controllers is
simultaneously. GA uses mutation and crossover operators to achieve shown in Fig. 14.
global optimum solution. These operators essentially help GA to avoid Block diagram depicting the implementation of both controllers
premature convergence. The performance of GA largely relies on fac- is shown in Fig. 15. It is obvious that two separate controllers ded-
tors such as population size, mutation rate, crossover probability etc. icated for each loop has to be employed since both loops require
(Kumar & Chakarverty, 2011). For the current work the population different manipulated variables. Also, in real world the value of ma-
size was taken as 20, crossover rate was 0.85, and the mutation rate nipulated variables cannot go beyond a certain limit hence satura-
was adaptive feasible, with maximum iteration as 100. tors with a limit of 0–2000 lbmol/h were employed in every simu-
The cost function used in this study to tune the controller gains is lation study carried out in this paper. The behaviour of both control
defined as, loops are shown in Figs. 16 and 17. It is seen from the time trajecto-
ries of XB and XD that after tuning, both controllers show adequate
J = 0.5(IAE_XB + IAE_XD ) (23) performance for setpoint change and reaches to new setpoint with

Fig. 15. Block diagram for the implementation of controllers applied to binary distillation column.
8542 P. Mishra et al. / Expert Systems With Applications 42 (2015) 8533–8549

Table 2 Table 4
Performance index comparison setpoint tracking. Performance index comparison for variation in ZL of -0.1 mol frac-
tion at t = 5 h.
Controller Cost function (J) IAE_XB IAE_XD
Controller Cost function (J) IAE_XB IAE_XD
FPID 11.157E-4 6.400E-4 15.914E-4
FOFPID 10.260E-4 2.607E-4 17.912E-4 FPID 101.61E-4 170.84E-4 32.387E-4
FOFPID 2.9229E-4 2.7998E-4 3.0460E-4

Table 3
Tuned controller parameters’ values. Table 5
Performance index comparison for variation in ZL of 0.1 mol frac-
Parameters XB loop XD loop tion at t = 5 h.
FPID FOFPID FPID FOFPID Controller Cost function (J) IAE_XB IAE_XD

a1 42.880 91.120 5.099 12.160 FPID 31.068E-4 26.173E-4 35.963E-4

a2 99.570 96.250 2.400 0.928 FOFPID 2.3803E-4 1.7377E-4 3.0228E-4
a3 94.090 78.900 1.400 6.691
a4 95.790 183.950 32.900 192.651
b1 16.165 0.669 7.301 8.362
b2 71.520 86.595 4.741 14.659 as their gains change at run time according to the instantaneous error
b3 4.459 9.912 4.469 0.545
and rate of change of error.
b4 40.00 47.467 1.440 2.424
L 0.101 0.878 15.501 3.134
KPI 371.995 7929.72 1882 9989.93 4. Comparative study of FPID and FOFPID controllers
KPD 2.800 2.651 0.399 195.703
μ 1.000 0.526 1.000 0.354 For the performance comparison of the effectiveness of the two
λ 1.000 0.976 1.000 0.084
controllers various intensive simulation studies have been con-
ducted. The studied controllers are rigorously tested for disturbance
rejection, parameter uncertainties, and measurement noise suppres-
sion. Besides that since the plant under study is a multivariable
process, so it is highly desirable that controller must be able to
suppress the influence of other interacting loops if a load change or
setpoint change affects one or more loops present in the plant. The
(a) disturbance has been given in the system through many variables
and for the uncertainty analysis, relative volatility and variation in
tray hydraulic constant are considered. Further the controllers’ abil-
ity to suppress the noise in process variable is also studied. The basis
of comparison throughout this work is taken as the value of cost func-
tion i.e. a weighted sum of IAEs generated in XB and XD loop. Further
in each investigation study the system was at same initial conditions
(b) given in Table 1 and then at t = 0, setpoint of 0.4 mol fraction and
0.8 mol fraction was imposed on XB and XD loop respectively.
Fig. 18. Variations of gains for FPID controller (a) XD loop (b) XB loop.

4.1. Disturbance rejection

A distillation column is subjected to variations in large number of

parameters such as FL , ZL , FV , ZV , α etc. These variations can disrupt
the steady state operations of the plant and severely degrade the per-
(a) formance of the column. In order to neutralise the effects of distur-
bances on controlled variables and attaining the performance spec-
ifications, the controller must make corrective action appropriately.
This subsection particularly focuses on the performance study of FPID
and FOFPID controller to suppress the disturbances in a distillation
column caused by two major sources i.e. FL and ZL . The disturbance
in each case was introduced at t = 5 h and the IAE was calculated for
(b)
the next 5 h i.e., up to t = 10 h.
Fig. 19. Variations of gains for FOFPID controller (a) XD loop (b) XB loop.
4.1.1. Variation in feed composition
Since the feed composition in the distillation column can change
comparable cost function values. The values of cost function and in- any time depending upon the input feed and other conditions, the
dividual values of IAE_XB and IAE_XD for both controllers are listed controller must reject the disturbance caused by this change effec-
in Table 2. An improvement of 8.74% is achieved in cost function ‘J’ tively so that end compositions do not change from their setpoint
for FOFPID controller over FPID controller. Though there is a slight significantly. In order to perform this study, a step change of ±0.1 mol
increment in IAE_XD for FOFPID controller than FPID controller, but fraction at t = 5 h was applied in the liquid phase feed composition
since the cost function for the tuning of the controller was a weighted respectively. It can be easily inferred through simulation studies that
sum of individual IAEs, so this marginal increase in IAE_XD constitutes both controllers have a tendency to suppress the disturbance caused
lesser importance. The values of different tuned parameters used in due to change in ZL , but FOFPID produced far better performance than
the controllers are listed in Table 3. Figs. 18 and 19 show the variation FPID. This is further validated by the quantitative comparison on the
of the gains K1 and K2 for FPID and FOFPID controller respectively to basis of cost functions shown in Tables 4 and 5. The corresponding
show the expert and intelligent system behaviour of the controllers variation in XB and XD are shown in Figs. 20 and 21 for decrement in
 P. Mishra et al. / Expert Systems With Applications 42 (2015) 8533–8549 8543

Table 6
Performance index comparison for step change in FL of 300 lb-
mol/h at t = 5 h.

Controller Cost function (J) IAE_XB IAE_XD

FPID 44.302E-4 53.826E-4 34.777E-4

FOFPID 4.7482E-4 2.6081E-4 6.8883E-4

Table 7
Performance index comparison for impulse change in FL of 300 lb-
mol/h at t = 5 h for a period of 0.1 h.

Controller Cost function (J) IAE_XB IAE_XD

Fig. 20. Variation in XB for a step change of -0.1 mol fraction in ZL .

FPID 63.902E-4 116.09E-4 11.716E-4

FOFPID 2.2129E-4 2.4999E-4 1.9260E-4

stream conditions. To mimic the real world disturbances, two sep-

arate cases were considered in which the value of FL was varied in
step as well as an impulse. The variation magnitude in FL for both
cases was 300 lbmol/h and the time of application was t = 5 h. Du-
ration of the impulse disturbance was taken as 0.1 h. Correspond-
ing results are shown through Figs. 25–28. Simulation results for this
study revealed that better regulation can be achieved through FOF-
PID controller than FPID controller when dealing with variations in
Fig. 21. Variation in XD for a step change of -0.1 mol fraction in ZL .
FL . Tables 6 and 7 list the values of cost function for step and impulse
disturbance respectively. Fig. 29(a) and (b) presents the cost function
comparison graphically. It can be shown that an improvement factor
of around 9.33 and 28.87 times is achieved using FOFPID controller
than FPID controller for step and impulse disturbance in FL .

4.1.3. Unknown sources causing disturbance in XB and XD

It is quite hard in practice to account for all the dynamics in the be-
havioural modelling of a plant and these unmodelled dynamics also
might affect the final system performance. To investigate this issue
it was assumed that some unknown dynamics have caused XB and
Fig. 22. Variation in XB for a step change of 0.1 mol fraction in ZL .
XD to deviate from their respective setpoints. Since, distillation col-
umn studied in this paper is a MIMO plant and a change in one loop
will cause a disturbance in the other loop. So, this study helps in in-
vestigating about the controllers’ capability in handling multivariable
processes.
For the first case it was assumed that XB has deviated from its set-
point i.e. at XB has encountered a sudden increase by 0.05 mol frac-
tion moving from 0.40 mol fraction to 0.45 mol fraction at t = 5 h.
For making XB to track setpoint again the controller for XB loop must
take corrective action, but since the plant in interacting so the other
loop concerning with XD will also take corrective action so as to keep
XD unaffected from the variation in XB loop. Simulation results for
this case are shown in Figs. 30 and 31. Considering another case in
which XD suffers from a step disturbance of 0.05 mol fraction and
Fig. 23. Variation in XD for a step change of 0.1 mol fraction in ZL .
deviates from its setpoint value of 0.80 mol fraction to 0.75 mol frac-
tion at t = 5 h. The variations in XB and XD for this case are shown in
ZL and Figs. 22 and 23 for increment in ZL . For the sake of easy in- Figs. 32 and 33. These simulation studies clearly reveal that though
terpretation, the cost function comparison is also shown graphically both, FPID and FOFPID controllers rejected the disturbance in both
in Fig. 24(a) and (b). It can be deduced from the performance index loops but FOFPID controller outperformed the FPID controller with a
comparison analysis that the cost function was improved for the FOF- significant improvement in the cost function. The calculation of cost
PID controller by an approximate factor of 34.77 and 13.05 times for function in this study was done from t = 5 h to the end of simulation
the decrement and increment in ZL respectively with respect to FPID i.e. t = 10 h. Fig. 34(a) and (b) presents the cost function compar-
controller. ison graphically for the step disturbance in XB and XD respectively.
Tables 8 and 9 show the cost functions and the value of IAE for each
4.1.2. Disturbance through feed flow rate loop in this study. It can be shown that, for the case when XB loop was
Change in feed flow rate is another major source of disturbance disturbed, the FOFPID controller produced approximately 8.06 times
through which product quality of a distillation column deteriorates. better performance than FPID controller, and similarly for the other
The value of FL can vary significantly depending upon the instanta- case when XD loop was disturbed, the cost function was improved by
neous conditions of the plant and some other factors such as up- approximately 3.61 times for FOFPID controller.
8544 P. Mishra et al. / Expert Systems With Applications 42 (2015) 8533–8549

Fig. 24. Performance index comparison for step change of (a) -0.1 mol fraction in ZL (b) 0.1 mol fraction in ZL .

Table 8
Performance index comparison for step disturbance in XB of
0.05 mol fraction at t = 5 h.

Controller Cost function (J) IAE_XB IAE_XD

FPID 24.970E-4 29.197E-4 20.743E-4

FOFPID 3.0948E-4 1.3288E-4 4.8609E-4

Table 9
Performance index comparison for step disturbance in XD of -
0.05 mol fraction at t = 5 h.

Controller Cost function (J) IAE_XB IAE_XD

Fig. 27. Variation in XB for an impulse variation of 300 lbmol/h in FL for 0.1 h.
FPID 26.502E-4 20.789E-4 32.214E-4
FOFPID 7.3316E-4 2.2282E-4 12.435E-4

Fig. 28. Variation in XD for an impulse variation of 300 lbmol/h in FL for 0.1 h.

Fig. 25. Variation in XB for a step change of 300 lbmol/h in FL . ways taken while modelling a process. Since controllers are designed,
tuned and tested primarily on the simulated model rather than the
original plant, it becomes an integral part of the study on the con-
troller’s performance that what happens if the plant parameters’ val-
ues deviate from the ones, which were considered while designing
and tuning the controller. The capability of the controller to deal with
this type of problem is of prime importance because it is quite com-
mon in practice that the values of different parameters related to the
plant can deviate from their nominal values. Present study focuses on
uncertainty in two important parameters i.e. relative volatility and
tray hydraulic constant. Apart from this, measurement noise, another
major source of uncertainty is also considered in this work.

Fig. 26. Variation in XD for a step change of 300 lbmol/h in FL . 4.2.1. Uncertainty in relative volatility (α )
The value of α may change throughout the process operation de-
pending upon many factors such as described in Section 2. Hence for
4.2. Robustness testing the robustness testing of the both controllers, the value of α is varied
during run time. Two different profiles were chosen for the variation
This section mainly concentrates upon the effect of model un- in α . First one was a step variation applied at t = 5 h and the second
certainties on the plant operation and capability of the controller to one consisted of a linear variation of α from t = 4 to 7 h. In both pro-
handle such parametric variations in the plant. In practice, as it is files the value of α was varied from 2 to 3 mol fraction. Fig. 35 shows
not possible to find an exact plant model, some assumptions are al- the profiles for the variation in α for the step and linear variation.
 P. Mishra et al. / Expert Systems With Applications 42 (2015) 8533–8549 8545

Fig. 29. Performance index comparison for (a) step increment in FL of 300 lbmol/h (b) impulse change in FL of 300 lbmol/h.

Fig. 30. Variation in XB for a step disturbance of 0.05 mol fraction in XB . Fig. 33. Variation in XD for a step disturbance of -0.05 mol fraction in XD .

Table 10
Performance index comparison for step variation in α from 2 to 3
at t = 4 h.

Controller Cost function (J) IAE_XB IAE_XD

FPID 131.66E-4 161.68E-4 101.63E-4

FOFPID 9.2014E-4 2.1716E-4 16.231E-4

Table 11
Performance index comparison for linear variation in α from 2 to
3 at t = 4 h to t = 7 h.

Controller Cost function (J) IAE_XB IAE_XD

Fig. 31. Variation in XD for a step disturbance of 0.05 mol fraction in XB .
FPID 82.096E-4 65.294E-4 98.898E-4
FOFPID 7.1280E-4 2.0886E-4 12.167E-4

the step and linear change in α . It is evident from the data analysis
presented in Tables 10 and 11 that the cost function improvement for
step change in α is nearly 14.30 times and for the linear variation in
α , the improvement factor is nearly 11.51 times for FOFPID controller
when compared to FPID controller.

4.2.2. Uncertainty in tray hydraulic constant (β )

For distillation columns it is generally assumed that all the trays
Fig. 32. Variation in XB for a step disturbance of -0.05 mol fraction in XD . have same hydraulic constant but practically it can be different. In
present study the nominal value of the β was taken as 3.6 s, while
practically it can vary in a range of 3–6 s (Jana, 2010; Luyben, 1990).
The results are shown in Figs. 36–39. It can be clearly inferred from To test the robustness of the controllers the value of β was varied
the simulation results that FOFPID controller maintained the product randomly in the range of 3–6 s for each tray. It was observed from
composition at the desired value with a little variation at the chang- the simulation results that for other value of β than the nominal one,
ing instant or period of α , but FPID controller was not as effective as value of the cost function was degraded by a higher factor for FPID
FOFPID controller in maintaining the compositions close to the set- controller than the FOFPID controller. The change in the cost function
points. The results are further supported by the cost function analy- for uncertain β over nominal β was 31.05% for FPID controller and
sis given in Tables 10 and 11 and Fig. 40(a) and (b). The calculation the for FOFPID controller it reduced to only 12.29%. The variation in
for cost function was done from the application of change in α till XB and XD for uncertain β case is shown in Figs. 41 and 42 respec-
the end of simulation study, i.e. from t = 4 h to t = 10 h, for both tively. Table 12 shows the variation in cost function values for the
8546 P. Mishra et al. / Expert Systems With Applications 42 (2015) 8533–8549

Fig. 34. Performance index comparison for step disturbance (a) in XB of 0.05 mol fraction and (b) in XD of -0.05 mol fraction.

Table 12
Performance index comparison for uncertain β .

Controller Cost function (J) for nominal case Cost function (J) for uncertain case % Variation from nominal value

FPID 11.157E-4 14.622E-4 31.05

FOFPID 10.260E-4 11.521E-4 12.29

Fig. 38. Variation in XB for a linear change in α from 2 to 3 at t = 4 h to 7 h.

Fig. 35. Profiles for variation in α for robustness test of controllers.

Fig. 39. Variation in XD for a linear change in α from 2 to 3 at t = 4 h to 7 h.

Fig. 36. Variation in XB for a sudden step change in α at t = 4 h.

nominal case and the uncertain case. The cost function calculation
for this study was done for 3 h i.e. from t = 0 to 3 h.

4.2.3. Uncertainty due to measurement noise

Measurement noise can also severely degrade the performance of
the controller. A good controller is also expected to suppress the level
of noise introduced in the loop. To test the noise suppression ability
of the designed controller, 5% random noise was introduced in the
bottom and distillate composition measurement simultaneously. The
noise application in the measurement of XB and XD was simulated as,

XB (XD ) = XB (XD ) × (1 + 0.05 × (randn( − 1,1))) (26)

Fig. 37. Variation in XD for a sudden change in α at t = 4 h. Figs. 43 and 44 show the simulation results. It can be clearly
seen that FPID was not able to sustain the compositions to the
 P. Mishra et al. / Expert Systems With Applications 42 (2015) 8533–8549 8547

Fig. 40. Performance index comparison for (a) step change in α from 2 to 3 at t = 4 h (b) linear variation in α from 2 to 3 at t = 4 h to t = 7 h.

Fig. 43. Variation in XB for 5% random noise in measurement of XB and XD simultane-

Fig. 41. Variation in XB for different β for all the trays.
ously.

Fig. 44. Variation in XD for 5% random noise in measurement of XB and XD .

Fig. 42. Variation in XD for different β for all the trays.

Table 13 noisy measurement, the FOFPID controller provided approximately

Performance index comparison under noisy measurement of XB 6.19 times better performance than FPID controller, in terms of im-
and XD . provement in cost function.
Controller Cost function (J) IAE_XB IAE_XD
5. Conclusions and discussions
FPID 398.46E-4 743.02E-4 53.895E-4
FOFPID 64.492E-4 67.446E-4 61.538E-4
In this paper a successful attempt has been made to design an
intelligent controller which assumes the form of a fractional order
fuzzy proportional–integral–derivative (FOFPID) controller. The pro-
desired value when noisy composition measurement propagated to posed control scheme was motivated as a solution to deal with com-
the controller and the value of XB deviated with a significant amount plex dynamics of the binary distillation column. Essentially, it is a
from the setpoint but FOFPID retained XB close to the setpoint. The Takagi–Sugeno (TS) model based nonlinear fuzzy adaptive controller
cost function comparison, presented in Table 13 further supports the comprising of non-integer orders of integration and differentiation
finding that FOFPID provided much robust performance than FPID operators. These operators were used in the controller to compute
in terms of noise suppression. Again, the value of cost function was the fractional order integration change in controller output and frac-
evaluated for the duration from t = 0 to t = 3 h. Fig. 45 shows the tional derivative of error. The proposed expert and intelligent con-
cost functions and individual loop’s IAE comparison pictorially. It can trol scheme has offered five advantages like: First, it varied its gains
be shown through the data analysis presented in Table 13, that for at run time in accordance with the instantaneous error and rate of
8548 P. Mishra et al. / Expert Systems With Applications 42 (2015) 8533–8549

scheme i.e. one FOFPID controller for each loop, and thereby requiring
total 26 parameters. Tuning of large number of parameters is quite
cumbersome and may not be manageable manually. However, with
the help of bio-inspired optimization algorithms, this task can be eas-
ily accomplished, provided that there is a near exact estimation of the
model. Further, to make the control action more precise and smooth,
number of input combination regions can be increased by increas-
ing number of membership functions. But again the controller will
further have large number of the parameters to be tuned, which is a
tedious task. In other words, there has to be a trade-off between the
implementation complexity and the efficiency of the controller.
As a future scope of this study, various extensions of the current
work can be investigated to make the work more enlightening to the
scientific community. The most important task which can be carried
Fig. 45. Cost function comparison for 5% random noise in measurement of XB and XD .
out is the real time evaluation of the proposed control scheme on a
real distillation column. This will enable the user to know any de-
viation from the theoretical results presented in this work. A real
change of error. Second, inclusion of non-integer order of the inte- time framework for building proposed expert system based control
gration and differentiation operators made the proposed controller scheme can also be tried out, and the use of FPGA for the purpose may
scheme more robust while simultaneously providing increased de- serve as a nice solution for the same as there are multiple loops in-
sign freedom to the designer. Third, unlike some other classical in- volved. Since, the proposed controller is formula based it can be easily
direct adaptive techniques, which usually required continuous run implemented on FPGA, which is equipped with several features like
time identification of the plant parameters, the FOFPID controller en- parallel loop executions. The controller gains used in this work can be
joys direct adaptive control structure and hence has a simpler im- tuned using some other recent and efficient optimization algorithms
plementation. Fourth, the proposed controller has fewer membership for critical evaluation of the efficiency of the optimization algorithms
functions than other approaches and has a formula based structure, as the controller used in this paper has large number of parameters
thus reduces the complexity of the controller which allows its rel- to be tuned and solving a large dimensional optimization problem
atively easy implementation. Fifth, the proposed controller directly is a critical task. The proposed controller can also be explored for
produces the control output, in contrast to some other adaptive con- other types of chemical processes such as, azeotropic distillation, re-
troller schemes which have a hierarchical structure and generate or active distillation, multi-component distillation etc., in order to criti-
adjust gains of a second level controller. cally evaluate the performance of the expert control system proposed
For the proposed scheme, the implementation of gain adaptation in this paper for such systems. Though the proposed scheme is an
was demonstrated with the help of surface plots of different gains in expert system based approach which efficiently handled the prob-
the error and rate of change error space, as shown in Figs. 5–12. This lem of controlling nonlinear, uncertain and coupled system, but the
property of the FOFPID controller was due to the formula based fuzzy performance of the controller may be enhanced by using more num-
controller (FBFC) which served as the core of FOFPID controller. For ber of membership functions in the error and rate of change of er-
the performance evaluation and comparison of the proposed scheme, ror universe. An increment in the number of membership functions
the performance of FOFPID controller was compared with that of its might lead to smoother control action than obtained from the cur-
integer order counterpart, a fuzzy proportional–integral–derivative rent one. Lastly, extension of FOFPID controller in the Type-2 fuzzy
(FPID) controller. The parameters of both the controllers were op- domain may also be tried out, as it is assumed that the Type-2 fuzzy
timized for minimum integral of absolute error (IAE) using Genetic sets are more robust to plant uncertainties.
Algorithm (GA). Intensive LabVIEWTM simulation studies were per-
formed which included setpoint tracking with and without uncer-
tainties, disturbance rejection, and noise suppression investigations. References
For testing the parameter uncertainty handling capability of the pro-
Carlson, G. E., & Hajlijak, C. A. (1964). Approximation of fractional capacitors (1/s)1/ n by
posed controller, uncertain and time varying relative volatility and a regular Newton process. IEEE Transactions on Circuit Theory, 11(2), 210–213.
uncertain tray hydraulic constant were applied. Also, for the distur- Das, S., Pan, I., & Das, S. (2013). Performance comparison of optimal fractional order hy-
bance rejection studies, rigorous simulations were conducted, which brid fuzzy PID controllers for handling oscillatory fractional order processes with
dead time. ISA Transactions, 52(4), 550–566.
included two most common causes of disturbance i.e. variation in Das, S., Pan, I., Das, S., & Gupta, A. (2012). A novel fractional order fuzzy PID controller
feed composition and variation in feed flow rate. All the investi- and its optimal time domain tuning based on integral performance indices. Engi-
gations clearly suggested that FOFPID controller provided superior neering Applications of Artificial Intelligence, 25(2), 430–442.
Delavari, H., Ghaderi, R., Ranjbar, N. A., & Momani, S. (2010a). Fuzzy fractional order
performance over FPID controller for each case study i.e. setpoint sliding mode controller for nonlinear systems. Communication in Nonlinear Science
tracking, disturbance rejection, noise suppression and parameter un- Numerical Simulation, 15(4), 963–978.
certainties. Delavari, H., Ghaderi, R., Ranjbar, N. A., HosseinNia, S. H., & Momani, S. (2010b). Adap-
tive fractional PID controller for robotic manipulator. In Proceeding 4th IFAC work-
Further, it may be noted that every technique has its own pros shop fractional differentiation and its applications (pp. 1–7). Badajoz, Spain.
and cons and they must be kept in mind to critically evaluate it in Ding, Y., Ying, H., & Shao, S. (1999). Structure and stability analysis of a Takagi–Sugeno
comparison to others. The controller discussed in this work is not an PI controller with application to tissue hyperthermia therapy. Soft Computing, 2(4),
183–190 1999.
exception and also has some limitations. The FBFC is a mere build-
Dounis, A. I., Kofinas, P., Alafodimos, C., & Tesels, D. (2013). Adaptive fuzzy gain schedul-
ing block of the FOFPID and FPID controller, which use 9 parameters ing PID controller for maximum power point tracking for photovoltaic system. Re-
(a1 – 4 , b1 – 4 , L) for its implementation. Other than that, the controller newable Energy, 60, 202–214.
Dulf, E. H., Clement, F., Dulf, F.-V., Pop, C.-I., & Both, R. (2011). Fractional order controller
also uses two scaling factors at output of FBFC and two more param-
design for 13 C separation column. In Proceeding of 19th Mediterranean conference on
eters in the form of order of integration and differentiation in case of control and automation (pp. 582–587). Corfu, Greece.
FOFPID. Thus in total, a single FOFPID controller uses 13 parameters Fereidouni, A., Masoum, M. A. S., & Moghbel, M. (2015). A new adaptive configuration
for its implementation. Since, the problem undertaken in this work for PID type fuzzy controller. ISA Transactions, 56, 222–240.
Galluzzo, M., & Cosenza, B. (2009). Control of the biodegradation of mixed wastes in a
is of MIMO nature containing two control loops, two different con- continuous bioreactor by a type-2 fuzzy logic controller. Computers and Chemical
trollers were needed for the complete implementation of the control Engineering, 33(9), 1475–1483.
 P. Mishra et al. / Expert Systems With Applications 42 (2015) 8533–8549 8549

Galluzzo, M., & Cosenza, B. (2011). Control of a non-isothermal continuous stirred tank Oustaloup, A., Levron, F., Matthieu, B., & Nanot, F. M. (2000). Frequency–band complex
reactor by a feedback-feedforward structure using type-2 fuzzy logic controllers. noninteger differentiator: Characterization and synthesis. In IEEE Transactions on
Information Sciences, 181(17), 3535–3550. Circuits and Systems – I: Fundamental Theory and Applications: 47 (pp. 25–39).
Galluzzo, M., Cosenza, B., & Matharu, A. (2008). Control of a nonlinear continuous Pan, I., & Das, S. (2012). Chaotic multi-objective optimization based design of fractional
bioreactor with bifurcation by a type-2 fuzzy logic controller. Computers and Chem- order PIλ Dμ controller in AVR system. Electrical Power and Energy Systems, 43(1),
ical Engineering, 32(12), 2986–2993. 393–407.
Gizi, A. J. H. A., Mustafa, M. W., & Jebur, H. H. (2014). A novel design of high-sensitive Ramli, N. M., Hussain, N. A., Jan, B. M., & Abdullah, D. (2014). Composition prediction of
fuzzy PID controller. Applied Soft Computing, 24, 794–805. a debutanizer column using equation based artificial neural network model. Neu-
Jana, A. K. (2010). Chemical process modelling and computer simulations (1st ed.). New rocomputing, 131, 59–76.
Delhi: PHI. Rani, A., Singh, V., & Gupta, J. R. P. (2013). Development of soft sensor for neural network
Jana, A. K., & Adari, P. V. R. K. (2009). Nonlinear state estimation and control of a batch based control of distillation column. ISA Transactions, 52(3), 438–449.
reactive distillation. Chemical Engineering Journal, 150(2), 516–526 2009. Sabatier, J., Oustaloup, A., Iturricha, A. G., & Lanusse, P. (2002). CRONE control: Prin-
Jones, K. O., & Hengue, W. (2009). Distillation column control using genetic algorithms ciples and extension to time-variant plants with asymptotically constant coeffi-
for LQR design. In International conference on computer systems and technologies and cients. Nonlinear Dynamics, 29, 363–385.
workshop for PhD students in computing ACM. Safe, M., Khazraee, S. M., Setoodeh, P., & Jahanmiri, A. H. (2013). Model reduction and
Karimi, I., & Salahshoor, K. (2012). A new fault detection and diagnosis approach for a optimization of a reactive dividing wall batch distillation column inspired by re-
distillation column based on a combined PCA and ANFIS scheme. In Chinese control sponse surface methodology. Mathematical and Computer Modelling of Dynamical
and decision conference (pp. 3408–3413). Taiyuan. Systems,, 19(1), 29–50.
Kariwala, V., & Cao, Y. (2010). Bidirectional branch and bound for controlled variable Sanandaji, B. M., Salahshoor, K., & Fatehi, A. (2007). Multivariable GA-based identifi-
selection. Part III: Local average loss minimization. IEEE Transactions on Industrial cation of TS fuzzy models: MIMO distillation column model case study. In IEEE
Informatics, 6(1), 54–61. international fuzzy systems conference (pp. 1–6). London.
Kawathekar, R., & Jiggs, J. B. (2007). Nonlinear model predictive control of a reactive Savran, A., & Kahraman, G. (2014). A fuzzy model based adaptive PID controller design
distillation column. Control Engineering Practice, 5(2), 231–239. for nonlinear and uncertain processes. ISA Transactions, 53, 280–288.
Khazraee, S. M., Jahanmiri, A. H., & Ghorayshi, S. A. (2010). Model reduction and opti- Sharma, R., Rana, K. P. S., & Kumar, V. (2014). Performance analysis of fractional order
mization of reactive batch distillation based on the adaptive neuro-fuzzy inference fuzzy controller applied to robotic manipulator. Expert Systems with Applications,
system and differential evolution. Neural Computing and Applications, 20(2), 239– 41(9), 4274–4289.
248. Singh, V., Gupta, I., & Gupta, H. O. (2007). ANN-Based estimator for distillation using
Kumar, A., & Chakarverty, S. (2011). Design optimization using genetic algorithm and Levenberg–Marquardt approach. Engineering Applications of Artificial Intelligence,
cuckoo Search. In Proceedings of IEEE International Conference on Electro/Information 20(2), 249–259.
Technology (EIT) (pp. 1–5). Mankato, MN. Smith, C. (2012). Distillation control: An engineering perspective. New Jersey: John Wiley
Kumar, M. V. P., & Kaistha, N. (2008). Role of multiplicity in reactive distillation control and Sons.
system design. Journal of Process Control, 18(7), 692–706. Szabó, L., Nemeth, S., & Szeifert, F. (2012). Three-level control of a distillation column.
Li. , H. X., & Gatland, H. B. (1996). Conventional fuzzy control and its enhancement. IEEE Engineering, 4(10), 675–681.
Transactions on Systems, Man and Cybernetics – Part B, 26(5), 791–797. Tan, A. H., Godfrey, K. R., & Barker, H. A. (2009). Design of ternary signals for MIMO
Liptak, B. G. (2009), Distillation Control and Optimization. (Ebook), Itasca, IL: Putman identification in the presence of noise and nonlinear distortion. IEEE Transactions
Media, www.ControlGlobal.com. on Control Systems Technology, 17(4), 926–933.
Luyben, W. L. (1990). Process modeling, simulation, and control for chemical engineers Tham, M. T. (2009). Distillation: An introduction. (Online). Available: http://lorien.
(2nd ed.). Singapore: McGraw-Hill Company. ncl.ac.uk/ming/distil/distilpri.htm
Mann, G. K. I., Hu, B. G., & Gosine, R. G. (1999). Analysis of direct action fuzzy PID con- Urselmann, M., Barkmann, S., Sand, G., & Engell, S. (2011). A memetic algorithm for
troller structures. IEEE Transactions on Systems, Man, Cybernetics – Part B, 29(3), global optimization in chemical process synthesis problems. IEEE Transactions on
371–388. Evolutionary Computation, 15(5), 659–683.
Matsuda, K., & Fujii, H. (1993). H∞ optimized wave-absorbing control: Analytical and Valério, D., & Costa, J. S. D. (2013). An introduction to fractional control. London, United
experimental results. Journal of Guidance, Control and Dynamics, 16(6), 1146–1153. Kingdom: IET.
Miccio, M., & Cosenza, B. (2014). Control of a distillation column by type-2 and type-1 Xu, D., Jiang, B., & Shi, P. (2014). A novel model-free adaptive control design for mul-
fuzzy logic PID controllers. Journal of Process Control, 24(5), 475–484. tivariable industrial processes. IEEE Transactions on Industrial Electronics, 61(11),
Mishra, R. K., Khalkho, R., Kumar, B., & Dan, T. K. (2013). Effect of tuning parameters of 6391–6398.
a model predictive binary distillation column. In IEEE International Conference on Zadeh, L. A. (1965). Fuzzy sets. Information and Control, 8(3), 338–353.
Emerging Trends in Computing, Communication and Nanotechnology (pp. 660–665). Zakeri, H., & Sadeghi, M. S. (2010). A 2 – DoF multi-objective H∞ /L1 robust controller
Tirunelveli, India. design for a high purification distillation column using LMIs. In IEEE international
Moghadam, A. A., Aksikas, I., Dubljevic, S., & Forbes, J. F. (2011). Distributed optimal conference on computer applications and industrial electronics (pp. 683–687). Kuala
control of a dimethyl ether (DME) catalytic distillation column. In IEEE conference Lumpur.
on decision and control and European control conference (pp. 2770–2775). Orlando,
USA.