Control Engineering Practice 52 (2016) 103–113

Contents lists available at ScienceDirect

Control Engineering Practice

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

Model predictive control of reactive distillation for benzene

hydrogenation
Vishal Mahindrakar a, Juergen Hahn a,b,n
a
Department of Chemical and Biological Engineering, Rensselaer Polytechnic Institute, Troy, NY 12180, USA
b
Department of Biomedical Engineering, Rensselaer Polytechnic Institute, Troy, NY 12180, USA

art ic l e i nf o a b s t r a c t

Article history: Benzene hydrogenation via reactive distillation is a process that has been widely adopted in the process
Received 23 April 2015 industry. However, studies in the open literature on control of this process are rare and seem to indicate
Received in revised form that conventional decentralized PI control results in sluggish responses when the reactive distillation
13 March 2016
column is subjected to disturbances in the feed concentration. In order to overcome this performance
Accepted 9 April 2016
limitation, this work investigates model predictive control (MPC) strategies of a reactive distillation
column model, which has been implemented in gPROMS. Several MPCs based upon different sets of
Keywords: manipulated and controlled variables are investigated where the remaining variables remain under
Model predictive control regular feedback control. Further, MPC controllers with output disturbance correction and, separately,
Reactive distillation
with input disturbance correction have been investigated. The results show that the settling time of the
Packed column
column can be reduced and the closed loop dynamics significantly improved for the system under MPC
Dynamic modeling
Feedback control control compared to a decentralized PI control structure.
& 2016 Elsevier Ltd. All rights reserved.

1. Introduction However, this reaction is not selective and toluene, which is a

desirable component of the reformate stream and present in sig-
Benzene is carcinogenic and has been classified by the En- nificant amounts, is also reduced. As toluene has a high octane
vironmental Protection Agency (EPA) as a Mobil Source Air Toxic rating (RON), it is desirable to maintain it in the final product; see
(MSAT) due to its presence in gasoline (Environmental Protection Eqs. (1) and (2) for detail of the reactions.
Agency, 2000). The EPA regulates automobile emissions of ben-
benzene (100RON )+3H2 →cyclohexane (83RON ) (1)
zene by limiting the amount of benzene in gasoline to 0.62 vol%
(Environmental Protection Agency, 2007). The gasoline pool in a
refinery consists of a blend of streams from several different
toluene (120RON )+3H2 →methylcyclohexane (75RON ) (2)
sources which also have different benzene concentrations. While
the reformate stream is one of the most important sources for The issue of selectivity can be overcome by using a reactive
blending, as it enhances the octane number of the pool, it is also distillation (RD) column as reactive distillation can make use of the
the main contributor of benzene (Environmental Protection different volatilities of the components. Fig. 1 shows a schematic of
Agency, 2006). As such, removal strategies for benzene from re- a reactive distillation column used for benzene hydrogenation.
formate streams are essential for gasoline production. A variety of Reactive distillation can enable selective reactions as separation
different approaches exist, ranging from removing compounds and reaction are simultaneously occurring in the same vessel, e.g.,
that form benzene in the reformate feed, to removal of benzene the catalyst zone can be located in a part of the column where one
via solvent extraction or hydrotreating (Palmer, Kao, Tong & or more of the components of the undesirable side-reaction are
Shipman, 2008). The latter approach will be investigated in this not present in significant concentrations. In addition to enhancing
work and it involves hydrogenation of benzene in the presence of the selectivity of the reactions, using reactive distillation can result
a catalyst. in savings in capital expenditure and also energy utilization for
some processes (Harmsen, 2007).
n
While reactive distillation for benzene hydrogenation can have
Corresponding author at: Center for Biotechnology and Interdisciplinary Stu-
dies, Rensselaer Polytechnic Institute, Troy, NY 12180, USA. benefits over a conventional process, combining separation and
E-mail address: [email protected] (J. Hahn). reaction in a single vessel can lead to operational challenges as

http://dx.doi.org/10.1016/j.conengprac.2016.04.008
0967-0661/& 2016 Elsevier Ltd. All rights reserved.
104 V. Mahindrakar, J. Hahn / Control Engineering Practice 52 (2016) 103–113

Notation rxn reaction rate, mol s
1 kg
1

s Laplace variable
a geometric surface area of packing per unit volume, T0 reaction reference temperature, K
m2 m
3 Tcondin temperature of cooling water entering the condenser,
A cross section area of column K
CA concentration, mol m
3 Tcondout temperature of cooling water leaving the condenser, K
dp packing particle diameter, m Tj temperature on stage j, K
ds column diameter, m u specific liquid load, m s
1
D distillate flow rate, mol s
1 V vapor flow rate, mol s 1
Ea reaction activation energy, J mol
1 v measurement noise
F feed flow rate, mol s
1 w process noise
hi total liquid holdup based on empty column, m3 m
3 x liquid mole fraction
Hl j molar enthalpy of liquid stream on stage j, J mol
1 y vapor mole fraction
Hv j molar enthalpy of vapor stream on stage j, J mol
1 y* equilibrium vapor mole fraction
HETP height equivalent to a theoretical stage, m z feed mole fraction
k ereaction rate constant, mol s
1 kg
1
K wall factor Greek letters
KA reaction adsorption coefficient, m3 mol
1
KH reaction adsorption coefficient, m3 mol
1 ε packing void fraction
L liquid flow rate, mol s
1 ∅l j, i liquid fugacity coefficient of component i on stage j
m number of manipulated variables ∅ v j, i vapor fugacity coefficient of component i on stage j
M mass holdup, kg ρcat catalyst density, kg m
3
Ml j liquid molar holdup on stage j, mol η Murphree efficiency
Mv j vapor molar holdup on stage j, mol λ relative gain
N number of stages Λ relative gain array
P pressure, Pa τ transfer function time constant, s
∆P0, j dry column pressure drop across stage j, Pa τc controller design parameter, s
∆Pj irrigated column pressure drop across stage j, Pa θ transfer function time delay, s
Q external heat energy input, J ψ resistance coefficient
 process noise covariance matrix
r number of controlled variables Subscripts
R reflux ratio
 measurement noise covariance matrix
i component index
Rgas gas constant, J mol
1 K
1
j stage index
Rev vapor Reynolds number

Vent (C4 and lights) (Environmental Protection Agency, 2006). This makes it challen-
ging to develop a control system for the process as the control
system needs to effectively maintain all the controlled variables at
their set points and at the same time rapidly reject disturbance
C5 to C6 effects.
It is essential for any control system to be tested in an industrial
setting or, in a first step, in detailed simulations. In this regard, this
work makes use of a rigorous first principles-based dynamic
Hydrogen
model of the benzene hydrogenation reactive distillation column.
Reformate This model has been implemented in gPROMS as part of a recent
study and consists of over 2400 differential and 5000 algebraic
equations (Mahindrakar & Hahn, 2014). This prior study developed
a decentralized control scheme in conjunction with a feedforward
controller for this process. While the investigated control config-
uration worked reasonably well, it had the drawback that the
feedforward controller was only beneficial if the delay associated
with the feed composition measurement was small, which in turn
C7+ requires a composition analyzer which can be expensive to pur-
chase and maintain (Luyben, 2006). In order to address these
Fig. 1. Schematic of a reactive distillation column. drawbacks, this work investigates several MPC control structures
in order to improve the closed loop dynamics of the RD column for
there are fewer controlled and manipulated variables than if se- disturbance rejection without requiring continuous and near in-
parate reaction and separation processes would be used. This can stantaneous measurements of the feed concentrations.
be especially challenging for benzene hydrogenation because this The outline of this paper is as follows. Section 2 presents pre-
process is known to be subjected to significant feed concentration liminary information about the dynamic model of the reactive dis-
fluctuations. For example, the reformate stream benzene con- tillation column and model predictive control relevant to this work.
centration can vary between 3 vol% and 11 vol% on a regular basis The details of the model predictive control scheme are provided in
 V. Mahindrakar, J. Hahn / Control Engineering Practice 52 (2016) 103–113 105

Section 3. Section 4 compares the responses of various control methods

discussed in this paper. Conclusions are provided in Section 5.

2. Preliminaries

This section reviews preliminary information needed for the

remainder of the paper. Section 2.1 reviews the dynamic model
used for the reactive distillation column (Mahindrakar & Hahn,
2014). Existing model predictive control strategies relevant to this
study are reviewed in Sections 2.2 and 2.3 presents estimator/filter
methods that will be used in this paper.

2.1. Dynamic model of benzene hydrogenation column

The model used in this work is based upon the model from a
previous study which was implemented in the dynamic simulator
gPROMS (Process Systems Enterprise, 1997). The column is a
packed column where part of the column is filled with catalyst to
facilitate the reaction. The packed column has been modeled as a
staged column with 70 theoretical stages and the entering re-
formate stream consists of 15 components. The column has 10
catalyst stages and the reformate feed enters the column at stage
30. Hydrogen feed enters the column at stage 29. The column has a
partial condenser as a significant amount of hydrogen remains
unreacted and it is economically infeasible to condense hydrogen.
This hydrogen is recycled to the hydrogen feed. Fig. 2 shows a
schematic of the reactive distillation column.
Equilibrium stage modeling (Al-Arfaj & Luyben, 2000; Alejski &
Duprat, 1996; Georgiadis, Schenk, Pistikopoulos & Gani, 2002; Ja-
cobs & Krishna, 1993; Peng, Lextrait, Edgar & Eldridge, 2002;
Sneesby, Tadé & Smith, 1998; Sneesby, Tadé, Datta & Smith, 1997)
was used for developing a dynamic model for the benzene hy-
drogenation process. Furthermore, the column model includes Fig. 3. Plots of predicted controlled variable trajectory, ` , and manipulated variable
trajectory, \ , for an MPC controller.
dynamic liquid and dynamic vapor holdups due to the operating
pressure of 8 atm (Choe & Luyben, 1987). Equations for liquid
hold-up and pressure drop have been adopted from Maćkowiak’s
(Hall & Mackowiak, 2009) study on packed columns. The model appendix.
equations for the reactive distillation column can be found in the
2.2. Model predictive control

Model predictive control is a popular control method that uti-

lizes a process model to predict the controlled variables and cal-
culate manipulated variable moves. The controlled variables pre-
diction is optimized over a prediction horizon 7 by estimating
changes in one or more manipulated variables over a control
horizon 4 (see Fig. 3). MPC solves the optimization problem,
implements the first computed value(s) of the manipulated vari-
able(s) and after the first time sample, corrects for plant-model
mismatch. This procedure is repeated for each time step. The MPC
controller used in this work is based upon a linear model of the
process while the “plant” is given by the rigorous dynamic model
described in the previous subsection.
Most MPC algorithms can account for constraints and some
algorithms can deal with process nonlinearities. As a result, MPC-
based controllers have been found to perform better than con-
ventional PID based controllers in a variety of instances (Qin &
Badgwell, 1997). Also, many vendors have developed MPC specific
algorithms for industrial implementations enabling wide-spread
adoption of MPC in the process industries (Qin & Badgwell, 1997;
Morari & Lee, 1999; Froisy, 2006).

xk + 1 = Φxk + Γuk

Fig. 2. Schematic of benzene hydrogenation RD column with control loops.

yk +1 = Cxk + 1 (3)
106 V. Mahindrakar, J. Hahn / Control Engineering Practice 52 (2016) 103–113

7
T
4− 1 selected based on the operating range of the manipulated and
( ) (
J = ∑ rk + i k − `^k + i k W y rk + i k−`^k + i k + ) ∑ ∆\Tk + i W \△\k + i controlled variables as given in Eqs. (11) and (12).
i=1 i=0 (4)
2.3. State estimation
^T ^
J = E W Y E +∆\Tf W U∆\f (5) Model-based control and optimization techniques often require
knowledge about states which cannot be directly measured (Qin &
where, Badgwell, 1997). These states can be estimated using an observer
⎡Wu … 0 ⎤ ⎡Wy … 0 ⎤ (Schei, 2008).
⎢ ⎥ ⎢ ⎥
WU = ⎢ ⋮ ⋱ ⋮ ⎥ ; WY = ⎢ ⋮ ⋱ ⋮ ⎥ xk + 1 = Φxk + Γuk + wk
⎣ 0 ⋯ W ⎦
u ⎣ 0 ⋯ W ⎦
y
(6) yk = Cxk + vk (13)

^ ^ (7)
E = r −Y ^ k + 1 = Φ_
_ (
^ k +Γ\k +Lk ` −C _
^k )
k (14)
⎡ `^ ⎤
⎡ rk + 1 ⎤ ⎢ k+1 k ⎥ ⎡ Δ\k ⎤ −1
⎢r ⎥
k+2
⎢ ^
` ⎥ ⎢
⎥; Y^ = ⎢ k + 2 k ⎥; ∆\f = ⎢
Δ\k + 1 ⎥
⎥ Lk = A Σ (
^C T CΣ
^ C T +
) (15)
r =⎢
⎢ ⋮ ⎥ ⎢ ⋮ ⎥ ⎢ ⋮ ⎥
⎢⎣ rk + 7 ⎥⎦ ⎢ `^ ⎥ ⎢⎣ Δ\k + 4 − 1⎥⎦ −1
⎣ k+7 k⎦ (8)
^ = AΣ
Σ ^ AT + − AΣ
^C T CΣ
(
^ C T +
) ^ AT
CΣ (16)

Consider a state space system as described in Eq. (13) which

∆\k + i = \k + i +1−\k + i (9)
includes process ( w ) and measurement noise ( v ). Often, the states
of the system are unknown and need to be determined based on
⎡ `^ ⎤ ⎡ ⎤
⎢ k+1 k ⎥ ⎡ CΦ ⎤ CΓ the measurements. A Kalman filter can be used to estimate states
⎢ ⎥ ⎢ ⎥
⎢ ⎥ C ΦΓ + C Γ of the system. The states are estimated by applying a filter gain to
^k k + ⎢ ⎥
^
` 2
Y = ⎢ k + 2 k ⎥ = ⎢ CΦ ⎥ _
^
⎢ ⋮ ⎥\k−1 the difference between the estimated outputs and the process
⎢ ⋮ ⎥ ⎢ ⋮ ⎥ ⎢ 7 ⎥
⎥ ⎢⎣ CΦ 7 ⎥⎦ measurements, where _ ^ k + 1 is the estimated state, and L k is the
⎢ `^ i−1
⎢⎣ ∑ i = 1 CΦ Γ ⎥⎦
⎣ k+7 k⎦ filter gain at time sample k . For time invariant systems, the gain L k
⎡ CΓ 0 … 0 ⎤ can be calculated from Eq. (15). Σ ^ is the steady state error covar-
⎢ ⎥
⎢ ΦΓ + CΓ CΓ … 0 ⎥ iance which can be computed from the algebraic Riccati equation
+⎢ ⋮ ⋮ … ⋮ ⎥ assuming that appropriate choices for the process noise covariance
⎢ 7 i−1
7− 1 7 − 4 +1 ⎥ matrix  and measurement noise covariance matrix  : Eq. (16).
⎢⎣ ∑ i = 1 CΦ Γ ∑i=1 CΦi − 1Γ … ∑i=1 CΦi − 1Γ ⎥⎦
⎡ Δ\k ⎤
⎢ ⎥
⎢ Δ\k + 1 ⎥ 3. Controller development
⎢ ⋮ ⎥
⎢⎣ Δ\k +4 − 1 ⎥⎦ (10) Reactive distillation columns can be viewed as an extension of
conventional distillation columns. As such, many of the existing
⎡ w1\ ⎤ control strategies for reactive distillation columns are in some
⎢ 2
⋯ 0 ⎥ form based upon distillation column control techniques. Skogestad
u
⎢ 1range ⎥
and co-workers have provided guidelines for selecting manipu-
W\ = ⎢ ⋮ ⋱ ⋮ ⎥
⎢ ⎥ lated and controlled variables for conventional distillation
⎢ 0 wm\ ⎥
⋯ columns (Skogestad & Morari, 1987; Skogestad, Lundström &
⎢⎣ umrange ⎥⎦
2
(11) Jacobsen, 1990; Skogestad, 1997).
⎛ D ⎞ ⎛ h1 ⎞
⎡ w` ⎤ ⎜ ⎟ ⎜ ⎟
⎢ 1
2
⋯ 0 ⎥ ⎜ V1 ⎟ ⎜ P1 ⎟
⎢ y1range ⎥ ⎜ LN ⎟ ⎜h ⎟
W` = ⎢ ⋮ ⋱ ⋮
⎥ ⎜ ⎟ ⎜ N⎟
⎢ ⎥ \ = ⎜ F H2 ⎟; ` = ⎜ V1 ⎟
`
⎢ wr ⎥ ⎜ Tcond ⎟ ⎜T ⎟
⎢ 0 ⋯
yrrange2 ⎥⎦ ⎜ 1⎟
out
⎜ ⎟
⎣ (12) ⎜ QN ⎟ ⎜ T19 ⎟
⎜ ⎟ ⎜ ⎟
⎝ R ⎠ ⎝ T55⎠ (17)

One implementation of MPC uses a state space representation ⎛ D ↔ h1 ⎞

(Froisy, 2006; Li, Lim & Fisher, 1989) described by Eq. (3), where ⎜ ⎟
xk ϵ9n are the states, uk ϵ9m are the manipulated variables, and ⎜ V1 ↔ P1 ⎟
⎜ ⎟
yk ϵ9r are the controlled variables. Eq. (4) describes the objective ⎝ LN ↔ hN ⎠ (18)
function for MPC that penalizes the prediction error over the
Based on the guidelines and other criteria, a set of 7 controlled
prediction horizon 7 and changes in the manipulated variables
variables, ` , and 7 manipulated variables, \ , can be selected for the
over the control horizon 4 . ^`k + i k represents the predicted con-
column under investigation (see Eq. (17)). Three pairs of ma-
trolled variable at time sample k + i , when the current time
nipulated variables and controlled variables given in Eq. (18) can
^
sample is k . Eq. (4) can be written in the form of predicted error E be paired by heuristics as the outflow of a vessel is directly related
^
as given in (Eqs. (5)–10). Y is the predicted controlled variable over to the level height (liquid) or pressure (vapor) in that vessel.
the prediction horizon. The weight matrices W y and W \ can be Several different control schemes are investigated in this section
 V. Mahindrakar, J. Hahn / Control Engineering Practice 52 (2016) 103–113 107

for controlling the remaining 4 sets of manipulated and controlled Table 2

variables. While the focus of this paper is on evaluating the per- RGA for RD column at operating conditions resulting from a feed benzene con-
centration of 6 vol%.
formance of model predictive control techniques, a base reference
needs to be established for comparing the performance with Tcond QN R F H2
conventional control strategies such as decentralized PI control.
Therefore decentralized PI control has been evaluated along with T1 1.01 0.00
0.01 0.00
different MPC controllers. Specifically, the following control T19
0.01 1.65
0.34
0.30
T55 0.02
0.64 1.96
0.34
methods have been developed, each of which will be discussed in
V1
0.02
0.01
0.62 1.64
subsequent subsections.

i) Decentralized PI control Table 3

ii) Decentralized PI plus feedforward control (based upon dis- Tuning parameters for decentralized PI controllers.
turbance measurement)
Manipulated variable Controlled variable Kc τi (min)
iii) SISO MPC with output disturbance
iv) SISO MPC with input disturbance Tcondout T1 21.82 20.0
v) MIMO MPC with output disturbances QN T19 3.96x105 97.0
vi) MIMO MPC with input and output disturbances R T55 −0.561 60.0
FH2 V1 0.074 11.3
All the control methods are derived for discrete time and a
sampling time of 1 minute is considered for the MPC controllers.
However, it should be noted that controllers are based upon ⎛ 1⎞
Gc (s ) = Kc ⎜ 1 + ⎟
continuous time transfer function determined from simulations of ⎝ τi s ⎠ (20)
the rigorous model and are then discretized for controller design.
In order to develop the controllers, transfer function relations The pairing of the variables is consistent with previous studies
between the manipulated variables and controlled variables were (Mahindrakar & Hahn, 2014). Internal model control (IMC) tuning
obtained by subjecting the partially-open gPROMS model to step relations are used to obtain the PI controller (Eq. (20)) tuning
changes in the manipulated variables and the disturbance variable parameters are shown in Table 3. The decentralized PI control
(Table 1). The three inventory control loops (liquid level in reflux scheme shown in this subsection serves as a benchmark for
drum, liquid level in reboiler drum, pressure at top of the column) comparison with the more advanced control schemes introduced
were closed for evaluation of the transfer function relations. next.

3.1. Decentralized PI control

3.2. Decentralized PI plus feedforward control
The benzene hydrogenation RD column control is a multi-input
It is known from previous studies (Mahindrakar & Hahn, 2014)
multi-output (MIMO) control problem. One of the simplest
that the feed composition disturbances affect the reactive dis-
methods of treating MIMO control problems is to use a decen-
tralized or multi-loop control strategy. Most conventional columns tillation column adversely when a decentralized PI feedback-only
use decentralized PI feedback-only control (Skogestad, 1997). De- control scheme is used. Adding feedforward control can provide
centralized PI control systems are convenient to tune and also significant benefit over a decentralized PI feedback-only scheme.
relatively easy to troubleshoot which explains their popularity in One of the main reasons for this is that feedback control does not
industry (Skogestad & Postlethwaite, 2007). The approach for es- take corrective action until after deviations in the controlled
tablishing a benchmark is to design a decentralized PI feedback- variables occur. However, as the effects of feed disturbances will
only control structure. only be detected after a while, this lack of predictive control can
The relative gain array (RGA) (Bristol, 1966) is evaluated (Ta- limit the overall column performance, especially for a column,
ble 2) for the set of manipulated and controlled variables based on such as the one investigated here, that includes large time con-
the transfer function relations from Table 1. RGA analysis indicates stants or time delays.
that the variable pairing should be done as given in Eq. (19).
Gd
⎛ FH Gf = −
↔ V1 ⎞ G p G mf (21)
⎜ 2
⎟
⎜ Tcondout ↔ T1 ⎟
⎜ ⎟ V1
⎜⎜ Q N ↔ T19 ⎟
⎟ Gp =
⎝ R ↔ T55⎠ (19) F H2 (22)

Table 1
Transfer functions for partially open benzene hydrogenation RD column estimated at 6.0 vol% feed benzene concentration. The units for all time constants are minutes.

Tcond QN R F H2 zC6 H6

T1 0 . 50077 2 . 2718x10−7 −0.20442 (1 − 96.03s) 0.1732 (1 + 110.45s) −116.5

1 + 20 . 20s (1 + 34 . 02s) (1 + 169.28s) (1 + 188s)(1 + 17.4s) 1 + 94.69s
T19 0 . 03025 (1 − 45 . 03s) 4 . 2662x10−6 (1 + 100.95s) −5.1542 (1 + 310.85s) e−8.33s −5.7528 (1 + 104.69s) 4200.2
1 + 2 (0 . 595)(11 . 35) s + (11 . 35s)2 (1 + 168 . 3s) (1 + 112.12s) (1 + 142.87s)(1 + 21s) 1 + 54.93s

T55 0 . 07445 1.1267x10−6 (−36.43s) −7.82 1.7219 −1305.6

1 + 161 . 93s e 1 + 149.47s 1 + 178.65s 1 + 186.03s
1 + 217 . 73s
V1 1.6616e−25s −2.244x10−7 (1 − 660.93s) −124.17 (1 − 281.32s) 161.54 (1 + 107.28s) −1.2103x105 (1 + 108.80s)
1 + 191.82s (1 + 134.65s) (1 + 171.35s) (1 + 186.9s)(1 + 24.32s) (1 + 187.25s)(1 + 30.51s)
108 V. Mahindrakar, J. Hahn / Control Engineering Practice 52 (2016) 103–113

V1 1 min has been used as before.

Gd =
zC6 H6 (23)
i) Output disturbance: It is assumed that the disturbances occur
749. 2 (1 + 108. 80s )(1 + 24. 32s ) at the output of the process. Even though this assumption is
Gf = not accurate for most processes, it simplifies the model and
(1 + 30. 51s )(1 + 107. 28s ) (24)
does not require an explicit disturbance model.
The feedforward control system investigated here assumes that the
xk + 1 = Φxk + Γuk
mole fraction of benzene in the feed is measured and it adjusts the
hydrogen flowrate based upon this measurement. Feed benzene con- `k + 1 = Cxk + pk (27)
centration, zC6 H6 is the measured disturbance for the feedforward
⎡ CΦ ⎤ ⎡ S1 ⎤ ⎡ S1 ⎤ ⎡ I⎤
controller. Vapor flowrate at the top of the column, V1 and hydrogen ⎢ ⎥ ⎢ ⎥ ⎢
0 … 0
⎥ ⎢ ⎥
2 ⎢ S2 ⎥ ⎢S S1 0 ⎥Δ + ⎢ I⎥ ^
feed to the column, FH2 are the controlled variable (CV) and manipu- ^
Y = ⎢ C Φ ⎥ ^
xk k + ⎢ \ k −1 + ⎢ 2
…
\f p
⎢ ⋮ ⎥ ⋮ ⎥ ⋮ ⋮ … ⋮ ⎥ ⎢ ⋮⎥ k k
lated variable (MV), respectively. The resulting feedforward controller ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢⎣ I ⎥⎦
⎣ C Φ7 ⎦ ⎣ S7 ⎦ ⎣ S 7 S 7−1 … S 7 − 4 + 1⎦ (28)
transfer function Gf for a system can be given by Eq. (21). Gd is the
disturbance transfer function, Gp is the process transfer function, and p^k k = `k −`^k k −1 (29)
Gmf is the disturbance sensor/transmitter transfer function which is set
to one here. Gp and Gd for the benzene hydrogenation RD column are −1 T
given by Eqs. (22) and (23). These transfer function relations can be (
Δ\f = STf W Y Sf + W \ ) Sf W Y E (30)
obtained from Table 1 and result in Eq. (24). It should be noted that
where,
this is the only control method in this study that utilizes a composition
analyzer for a predictive control action. This control method has been ⎡ CΦ ⎤ ⎡ S1 ⎤ ⎡ I⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
investigated to set a benchmark of the control performance that can be C Φ 2
⎢ S2 ⎥ I
E= ⎢ ⎥ ^
_k k + \ + ⎢ ⎥ p^
achieved if the key disturbance is measured accurately. ⎢ ⋮ ⎥ ⎢ ⋮ ⎥ k −1 ⎢ ⋮⎥ k k
⎢⎣ C Φ 7 ⎥⎦ ⎢ ⎥ ⎢⎣ I ⎥⎦
⎣ S7 ⎦ (31)
3.3. Model predictive control
The discrete state space model with output disturbances can be
Given the performance concerns resulting from feedback-only described by Eq. (27). The predicted controlled variable is given
control and the requirements for a composition measurement for a by Eqs. (28) and (29). It is assumed that the output disturbances
feedforward-feedback control scheme, model predictive control are constant throughout the prediction horizon. The solution
offers an alternative to these traditional control schemes. Since for the unconstrained MPC problem for the objective function
several of the controlled variables can result in good performance given in Eq. (5) is shown in Eq. (30). It was observed that with
for a feedback-only scheme, only the variables which will benefit appropriate tuning of MPC controllers none of the controlled
and manipulated variables exceeded their limits. As such it is
from a predictive, multivariable control scheme will be placed
sufficient for these investigations to focus on unconstrained
under MPC control. For this, different sets of controlled and ma-
MPC in this work.
nipulated variables were selected and a discrete state space model
ii) Input disturbance: Eq. (27) represents a process model with an
was developed for each set of CVs and MVs.
output disturbance. As disturbances rarely occur at the output in a
⎡ CΦ ⎤ ⎡ S1⎤ ⎡ S1 0 … 0 ⎤ real system, the performance of a MPC controller with output
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
2 S S S1 0 ⎥ disturbance can be poor if there are significant dynamics between
x k k + ⎢ 2 ⎥ uk − 1 + ⎢ 2
^ CΦ ⎥ ^ …
Y =⎢ Δuf
⎢ ⋮ ⎥ ⎢ ⋮⎥ ⎢⋮ ⋮ … ⋮ ⎥ the place where the disturbance actually occurs and the assumed
⎢⎣ C Φ P ⎥⎦ ⎢ ⎥ ⎢ ⎥ output (Shinskey, 1994). However, the closed-loop performance is
⎣ SP ⎦ ⎣ SP SP −1 … SP − M +1⎦ (25)
better if the disturbance can be incorporated as an input along
where, with a disturbance model. Pannocchia (Pannocchia, 2003) has
j
pointed out that input disturbance model for MPC algorithms can
Sj = ∑ C Φi −1Γ improve the performance of distillation processes by describing
i=1 (26) unmeasured disturbances in a more realistic way.

The state space model along with the output prediction Eq. (10) xk + 1 = Φxk + Γuk + Γ ddk
can be utilized to develop the MPC controller. Eq. (10) can be
yk = Cxk (32)
written in the form of step response coefficients Sj : Eq. (25). It
should be noted that Eq. (25) does not account for plant-model
⎡ ⎤
mismatch, however, the ability to deal with plant-model mismatch ⎢ C Γd ⎥
⎡ CΦ ⎤ ⎡ S1 ⎤ ⎡ S1 0 … 0 ⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ C ΦΓ d + CΓ d ⎥
is crucial for any control system. One way to account for the plant- ^ ⎢ C Φ 2
⎥ ^ ⎢ S2 ⎥ ⎢ S2 S1 … 0 ⎥ \ ⎢ ⋮ ⎥^
Y = _ + \ + ⎥Δ f + ⎢ 7 ⎥ dk k
model mismatch is to include a disturbance correction term in the ⎢ ⋮ ⎥ k k ⎢ ⋮ ⎥ k −1 ⎢ ⋮ ⋮ … ⋮
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣ C Φ7 ⎦ ⎣ S7 ⎦ ⎣ S 7 S 7−1 … S 7 − 4 +1⎦ ⎢ ∑ C Φi −1Γ d ⎥
controlled variable prediction. The disturbance correction term ⎢⎣ ⎥⎦
i =1 (33)
enables the controllerþpredictor system to maintain the controlled
variables at their set-points and avoid offsets (Muske & Badgwell,
⎡ C Γd ⎤
2002). There are different approaches to account for the dis- ⎡ CΦ ⎤ ⎡ S1 ⎤ ⎢ ⎥
turbance correction term. The two methods investigated here are ⎢ ⎥ ⎢ ⎥ d
⎢ C ΦΓ +CΓ ⎥ d
C Φ 2 S
^ k k + ⎢ 2 ⎥ \k −1 + ⎢ ⎥ d^
output disturbance correction and input disturbance correction and E =⎢ ⎥_ ⋮
⎢ ⋮ ⎥ ⎢ ⋮⎥ ⎢ 7 ⎥ kk
both will be utilized for developing MPC controllers. The prediction ⎢⎣ C Φ 7 ⎥⎦ ⎢ ⎥ ⎢ i 1 d ⎥
⎣ S7 ⎦ ⎢ ∑ CΦ Γ ⎥
−
horizon for all the MPC methods is 40 min and the control horizon ⎣ i=1 ⎦ (34)
is 4 min. Larger control horizons have also been investigated but the
results were only minimally affected by larger horizons and only the The state space model can be modified to include input dis-
results for the 4 min horizon are presented here. A sampling time of turbances (Eq. (32)). The predicted controlled variable is given by
 V. Mahindrakar, J. Hahn / Control Engineering Practice 52 (2016) 103–113 109

^ disturbance (Eqs. (35) and (36)).

Eq. (33), where dk k are the estimated disturbances of the plant.
The manipulated variable moves are calculated using Eqs. (30) ⎡ 0. 0001 0 ⎤ ⎡ 10 0 ⎤
=⎢ ⎥; =⎢
⎣ 0 10⎥⎦
^
and (34). dk k can be estimated by using an observer or a filter. MPC ⎣ 0 0. 0001⎦ (37)
controllers with output disturbances and input disturbances for
different sets of manipulated variables and controlled variables The values used for the process noise covariance matrix , and
will be developed in the following subsections. measurement noise covariance matrix  are given in Eq. (37). The
choice of these matrices was determined by (a) taking into account
3.3.1. Single-input single-output MPC controller that the process noise is much larger due to the plant-model
MPC is a multivariable control technique and one of the main mismatch, and (b) a trial and error procedure with the goal of
reasons of implementing MPC is to account for all the interactions achieving good closed-loop performance. The low value of  is
between the manipulated and controlled variables unlike a de- representative of the high model uncertainty, as the linear model
centralized control technique. However, the performance of the used for estimation is only accurate near the nominal conditions,
control mechanism for the benzene hydrogenation reactive dis- while the disturbance under investigations is significant in mag-
tillation column is significantly affected by how effectively the nitude and will have a profound impact on the operating condi-
controller is able to change the hydrogen feed to the RD column tions. This last point is discussed in more detail in Section 4.
when it is subjected to a feed benzene concentration disturbance.
Therefore, it is hypothesized that implementing MPC technique to 3.3.2. Multi-input multi-output MPC controller
just the single control loop of FH2−V1 may improve the perfor- As the process under investigation is inherently multivariable
mance of the column. An MPC controller implemented on the in nature, MIMO MPC is also investigated. In this section, MPC
single control loop FH2−V1 will be henceforth referred to as SISO controllers are developed for the manipulated variables, u and
MPC controller. The remaining control loops are closed with de- controlled variables, y :
centralized PI controllers as given in Table 3. Multivariable MPC
⎛ FH ⎞ ⎛ V1 ⎞
controllers will be discussed and evaluated in later subsections. ⎜ 2
⎟ ⎜ ⎟
\ = ⎜⎜
Tcondout ⎟
; ` = ⎜ T1 ⎟
3.3.1.1. SISO MPC with output disturbance. A SISO MPC controller is ⎟ ⎜ T19 ⎟
⎜ QN ⎟ ⎜ ⎟
developed based on the transfer function V1
from Table 1 and Eqs. ⎝ R ⎠ ⎝ T55⎠ (38)
F H2

(30) and (31). This controller uses a constant step output dis- The controlled variables and manipulated variables in Eq. (18)
turbance for offset free control. are still paired and controlled by proportional only controllers.
1 This requires the use of all the transfer function relations from
W\ = Table 1. However, it is observed that the controlled variable V1,
10 (35)
primarily depends on the hydrogen feed flow rate FH2 and on the
1 benzene feed concentration zC6 H6 . Incorporating the effect of other
W` =
1000 (36) manipulated variables ( Tcondout , Q N , and R ) on the controlled
variable V1 and the effect of benzene feed concentration zC6 H6 on
The weight matrices are selected using the criteria described by
the other controlled variables ( T1, T19 , and T55) leads to abnormal
Eqs. (11) and (12) and are given by Eqs. (35) and (36).
control behavior. This occurs because the transfer function rela-
3.3.1.2. SISO MPC controller with input disturbance. In the previous tions are sensitive to operating conditions and operating condi-
subsection a constant output step disturbance was used to ensure tions are affected significantly by feed concentration. As the feed
offset free control. However, in reality the disturbance (benzene composition changes, the transfer functions relating benzene
feed concentration) is occurring upstream of the process, i.e., there concentration with temperature controlled variables change sig-
are non-negligible dynamics between the place where the dis- nificantly. This is also true for the effect of the manipulated vari-
turbance occurs and when they affect the measurement. This ables ( Tcondout , Q N , and R ) on the controlled variable V1. While it
implies that there is potential room for improvement for the would be possible to address these points via a nonlinear MPC
closed loop performance of the process by considering the dis- formulation, this is beyond the scope of this work. In order to
turbance to be an input disturbance. MPC can utilize an input avoid these issues, an MIMO MPC controller is based on the
disturbance model to compensate for the disturbance effects. transfer function relations given in Table 4. The transfer function
However, just like feedforward control, using an input disturbance relations that correspond to the effect of disturbance variable, zC6 H6
requires knowledge about the effect of the actual disturbance on on the temperature controlled variables ( T1, T19 , and T55) and the
the output, i.e., a disturbance model. If a disturbance model is effect of the manipulated variables Tcondout , Q N , and R are found to
used, it is then possible to estimate the disturbance value. The be most sensitive to the operating conditions. These transfer
state space model with input disturbance (Eq. (32)) can be mod- function relations have not been included in Table 4 as they are not
ified to represent an augmented state space system. A Kalman
part of the MPC controller. MIMO MPC controllers with output
filter, which was tuned by the procedure described previously, was
disturbances and input disturbance are discussed and developed
applied to the augmented state space system for disturbance es-
in the following sections.
timation (Eqs. (A30)–(A33)).
V1 V1
The transfer functions and from Table 1 are used as a 3.3.2.1. MIMO MPC controller with output disturbances. An MIMO
F H2 z C6 H6

basis to develop the expressions of the discrete state space model. MPC controller with output disturbances is developed in this
The MATLAB functions “ss” and “c2d” are used for the conversion of section. A discrete state space model is derived from the transfer
transfer functions to state space. Based on the discrete state space function relations for the 4 pairs of controlled variables and ma-
model, and Eqs. (30)–(34) and (A30) through (A33), an MPC con- nipulated variables shown in Table 4. Eqs. (30) and (31) are used to
troller with an input disturbance model is designed for the RD develop the MIMO MPC controller with output disturbances.
column. The weights for the SISO MPC controller with the dis- Weights for the MIMO MPC controller are selected based on the
turbance model are the same as for the SISO MPC with output expected range of the manipulated and controlled variables.
110 V. Mahindrakar, J. Hahn / Control Engineering Practice 52 (2016) 103–113

Table 4
Transfer function relations used for developing MIMO MPC controllers

Tcond out QN R F H2 zC6 H6

T1 0 . 50077 2 . 2718x10−7 −0.20442 (1 − 96.03s) 0.1732 (1 + 110.45s) 0

1 + 20 . 20s (1 + 34 . 02s) (1 + 169.28s) (1 + 188s)(1 + 17.4s)
T19 0 . 03025 (1 − 45 . 03s) 4 . 2662x10−6 (1 + 100.95s) −5.1542 (1 + 310.85s) e−8.33s −5.7528 (1 + 104.69s) 0
1 + 2 (0 . 595)(11 . 35) s + (11 . 35s)2 (1 + 168 . 3s) (1 + 112.12s) (1 + 142.87s)(1 + 21s)

T55 0 . 07445 1.1267x10−6 −36.43s −7.82 1.7219 0

1 + 161 . 93s e 1 + 149.47s 1 + 178.65s
1 + 217 . 73s
V1 0 0 0 161.54 (1 + 107.28s) −1.2103x105 (1 + 108.80s)
(1 + 186.9s)(1 + 24.32s) (1 + 187.25s)(1 + 30.51s)

⎡1 0 0 0 ⎤ ⎡ T1 ⎤ ⎡ p1k ⎤
⎢ ⎥ ⎢ k⎥ ⎢ ⎥
⎢ 0 100 0 0 ⎥ ⎢ T19k ⎥ ⎢ p2k ⎥
⎢ ⎥ yk = ⎢ ⎥ = Cx + D p⎢
1 k
p ⎥
W\ = ⎢ 0 0
1000
0 ⎥ ⎢ T55k ⎥ ⎢ 3k ⎥
⎢ ⎥ ⎢⎣ V1 ⎥⎦ ⎢⎣ p4k ⎥⎦
⎢0 0 1 ⎥ k

⎢⎣ 0
100 ⎥⎦ (39) where,
⎡1 0 0 0⎤
⎡1 ⎤ ⎢ ⎥
⎢ 0 0 0 ⎥ 0 1 0 0⎥
⎢2 ⎥ Dp = ⎢
⎢0 0 1 0⎥
⎢ 1 ⎥ ⎢⎣ 0
⎢0 0 0 ⎥ 0 0 0⎥⎦ (42)
4
W =⎢
` ⎥
⎢0 1
0 0 ⎥ x˜ k + 1 = Φ̃x˜ k + Γ̃uk
⎢ 4 ⎥
⎢ ⎥ ˜ ˜k
⎢0 1 ⎥ y = Cx
k (43)
0 0
⎣ 104 ⎦ (40)
⎡ xk ⎤ ⎡ Φ Γd 0⎤ ⎡ Γ⎤
The weights selected for the MIMO MPC controller are given by ⎢ ⎥ ⎢ ⎥
_˜ k = ⎢ dk ⎥; Φ̃ = ⎢ 0 I 0⎥; Γ̃ = ⎢ 0 ⎥; C˜ = ⎡⎣ C 0 Dp ⎤⎦
Eq. (30) and (31). The range for the manipulated variables was ⎢ ⎥
determined by considering the steady state values for these vari- ⎢⎣ pk ⎥⎦ ⎢⎣ 0 0 I ⎦⎥ ⎣ 0⎦ (44)
ables at the extreme operating conditions with respect to the feed
benzene concentration (i.e. 3 vol% and 11 vol%). These weights (
_˜ k +1 = Φ̃_˜ k + Γ̃\k + L `k −C˜ _˜ k ) (45)
selected for the manipulated variables Q B and R are chosen to be
relatively large since it is observed that these MVs need to be The discrete state space model for the MPC controller can then
varied gradually for a stable response. A relatively small weight is be modified to include both the input and output disturbances as
selected for the recycle stream flow rate V1, in accordance with its given by Eq. (41). An augmented state space system can be defined
operating range. for this system, which can be used to estimate the input dis-
turbance and states using an observer ((Eqs. (43)–45)). Note that
3.3.2.2. MIMO MPC controller with input disturbance. This section Dp (4,4)=0, i.e no output disturbance is being considered for the
extends the MIMO MPC by using an input disturbance and an controlled variable V1 as it is assumed to be an input disturbance.
output disturbance model. The previous subsection made use of The state space matrices in Eq. (41) are estimated from the transfer
output disturbances for all the controlled variables. The logical functions given in Table 4. Since the effect of the feed composition
progression to this method would be to include input disturbance disturbance on the temperature controlled variables is eliminated
terms for all of the controlled variables for developing a MIMO in Table 4, the temperature controlled variables are not corrected
MPC controller with input disturbance. However, the linear by the disturbance variable. The augmented state space system in
transfer function relations between the disturbance variables and Eq. (44) changes the representation of the variable dk from an
the temperature controlled variables ( T1, T19 , and T55) will not been input variable to a state variable. This enables estimation of the dk
considered here as they are very sensitive to changes in the op- using an observer.
erating conditions. One approach to deal with this situation is to
use a combination of input and output disturbances as part of the
MPC controller. Muske and Badgwell (Muske & Badgwell, 2002) 4. Comparison of responses for different control schemes
have discussed the rotational factor modeling technique for this
type of problem. This technique can be implemented for the Previous simulation studies (Mahindrakar & Hahn, 2014) have
benzene hydrogenation column to develop a MIMO MPC con- indicated that decentralized PI feedback-only control, is adequate
troller that utilizes the input disturbance for prediction of vapor for disturbances in feed flow rate and feed temperature. However,
flow rate, V1 and output disturbances for prediction of the tem- significant settling times were observed when the reactive dis-
perature controlled variables. tillation column was subjected to disturbances in feed con-
⎡ uk ⎤ centration. The objective of this section is to evaluate and compare
xk + 1 = Φxk + ⎡⎣ Γ Γd ⎤⎦ ⎢ ⎥ the simulated responses of the reactive distillation column aug-
⎣ dk ⎦ (41)
mented with controllers developed in Section 3, when subjected
to feed concentration disturbances.
 V. Mahindrakar, J. Hahn / Control Engineering Practice 52 (2016) 103–113 111

The controllers developed in Section 3 are implemented on the Table 5

rigorous RD column model. The reactive distillation column with Comparison of deviations as measured by Eq. (46) for different control methods
when the RD column is subjected to a step change in feed benzene concentration
the controllers is subjected to step changes in the set-points of the
from 6 vol% to 11 vol%. %reduct. refers to reduction obtained in deviation when
controlled variables, a feed flow rate disturbance and feed tem- compared to the PI decentralized control only method.
perature disturbances. As expected from the previous simulation
studies that involved only traditional control schemes (Mahin- Control structure Dev % Reduct. Max Bz Settling Time
vol% (min)
drakar & Hahn, 2014), it is observed that all the controlled vari-
ables attain their set-points within reasonable time and without Decentralized PI feedback only 938 0.82 219
significant upset conditions. These responses have not been pre- Decentralized PI plus feedfor- 272 71 0.65 130
sented here in the interest of space. ward control
SISO MPC with output 688 27 0.73 172
∞
disturbance
Dev = ∫0 / (Bz. vol% − 0. 62)*(Bz. vol% − 0. 62) dt
(46) SISO MPC with input 324 65 0.69 127
disturbance
⎧ 0,x<0 MIMO MPC with output 828 12 0.80 198
/ (x) = ⎨ disturbance
⎩ 1,x≥0 (47) MIMO MPC with input 458 51 0.68 190
disturbance
The column model is then subjected to a step change in feed
benzene concentration from 6.0 vol% to 11.0 vol% as can be rea-
listically found in refineries (Environmental Protection Agency,
2006). Responses of the outlet benzene concentration for the a)
different control methods are presented in Fig. 4. The perfor- 350
mances of the various control methods are compared by using Eqs.
300
(46)–(47) to quantify the deviation of the outlet benzene con-
centration in excess of the regulatory limit. The deviations along 250
with the maximum benzene concentrations are given in Table 5. V1 (mol/sec) Decentralized PI control
200 Decentralized PI plus feedforward control
It is evident from Fig. 4 that decentralized PI feedback-only SISO MPC with output disturbance
control has the largest deviation and results in the worst response 150 SISO MPC with input disturbance
among all the control methods discussed. It can be further ob- MIMO MPC with output disturbance
100 MIMO MPC with input and output disturbances
served from Table 5 that decentralized feedback plus feedforward
control has the smallest deviation. However, this controller re- 50
2 4 6 8
quires the use of a composition analyzer, i.e., it works under the
assumption that accurate knowledge of the feed composition is Time (hours)

available almost instantaneously and a previous study with this b)

130
type of control scheme has shown that the performance strongly
degenerates if measurement time delays are present (Mahindrakar 120
& Hahn, 2014). The low deviation can be explained by the hy- 110
FH2 (mol/sec)

drogen feed flowrate FH2 and vapor flowrate V1 trajectories shown

100 Decentralized PI control
in Fig. 5. There is a spike in the hydrogen feed to the column based Decentralized PI plus feedforward control
on the disturbance information available to the controller, which 90 SISO MPC with output disturbance
SISO MPC with input disturbance
reduces the deviation significantly. While this may not be a very 80 MIMO MPC with output disturbance
realistic scenario, due to the expenses involved in continuous real- MIMO MPC with input and output disturbances
70
time measurement of the feed concentration, it can serve as a
good reference for the obtainable benefits. The settling times for 60
2 4 6 8
the benzene concentration to reach 95% of the final steady state
Time (hours)
have also been presented in Table 5.
All the MPC methods investigated in this work performed Fig. 5. Responses of (a) vapor flow rate at the top of the column and (b) hydrogen
feed to the column when the RD column with different control structures is sub-
better than decentralized PI feedback-only controller using any of
jected to step change in feed benzene concentration from 6 vol% to 11 vol%.
the measures presented in Table 5. This can be attributed to the
predictive nature of MPC methods that provide an advantage over
obtained depends on the type of MPC method used.
the traditional PI controllers. However, the extent of benefit
Benzene hydrogenation reactive distillation is a multivariable
control system and in general it is expected that multivariable
0.8
control methods will perform better than decentralized control
Benzene outlet conc (vol%)

0.7 methods. However, for this particular case, it is observed that

MIMO MPC methods do not provide additional benefit over SISO
0.6
Decentralized PI control
MPC methods. This can be explained from the RGA for the ma-
0.5 Decentralized PI plus feedforward control nipulated and controlled variable shown in Table 2. All the RGA
SISO MPC with output disturbance
0.4 SISO MPC with input disturbance
elements along the diagonal (corresponding to paired variables)
MIMO MPC with output disturbance are close to one and non-diagonal elements are close to zero. This
0.3 MIMO MPC with input and output disturbances
indicates that the interactions between the control loops are not
0.2
2 4 6 8
significant (Bequette, 2003). Furthermore, the performance is
Time (hours) mainly affected by how effectively the control system changes the
hydrogen feed to the column. Since SISO MPC methods specifically
Fig. 4. Comparison of benzene outlet concentration over time for different control
schemes when the reactive distillation column model is subjected to a step change focus on this control loop, there may not be an additional benefit
in feed benzene concentration from 6.0 vol% to 11.0 vol%. from using MIMO MPC. At the same time, MIMO MPC requires that
112 V. Mahindrakar, J. Hahn / Control Engineering Practice 52 (2016) 103–113

the interactions between the controlled and manipulated variables Appendix

are reasonably well known; however, the transfer functions re-
presenting the input-output behavior are found to be very sensi- Model
tive to the operating conditions which are strongly affected by the
hydrogen recycle flow rate. As such, this example produced the Condenser and reflux drum
somewhat surprising result that MIMO MPC did not return better, Mass balance
and even marginally worse, results, than SISO MPC coupled with
d ( Ml1 + Mv1)
decentralized PI control. = V2 − L1 − D − V1
dt (A1)
Finally, one other trend that can be observed from the devia-
tion values in Table 5 is that MPC methods with input disturbance Component balances
correction performed better than MPC methods with output dis-
turbance correction. This is because the feed benzene concentra- (
d Ml1x1, i + Mv1y1, i ) = V2 y2, i − L1x1, i − V1y1, i ∀ i : 1 to n−1
tion is a disturbance that occurs upstream of the process and it is dt (A2)
not surprising that MPC with an input disturbance predicts the
n n
controlled variables better than output disturbance-based correc-
tion. It should be noted that the extent of the performance in-
∑ x1, i = 1; ∑ y1, i = 1
i=1 i=1 (A3)
crease is dependent on the accuracy of the input disturbance
model used. As mentioned earlier, due to the nonlinear nature of Vapor–liquid equilibrium
the process, disturbance estimation using linear disturbance x1, i ∅l1, i = y1, i ∅v1, i ∀ i : 1ton (A4)
models cannot be perfectly accurate. As such it is expected that
nonlinear MPC can provide an even better disturbance rejection Energy balance
response. If this potential benefit is substantial can only be de-
0 = V2 Hv2 − V1H0 + Q 1 (A5)
termined by additional future investigations.

d ( Ml1Hl1 + Mv1Hv1)
= V2 H0 − (L1+D) Hl1 − V1Hv1
5. Conclusions dt (A6)

Benzene hydrogenation via reactive distillation is an operation of Q 1 = U *A*LMTD (A7)

significant interest in the process industries. However, control of this
process is non-trivial, especially for feed composition disturbances, as ( T1 − Tcondin ) − ( T2 − Tcondout )
decentralized PID control schemes result in sluggish responses. This LMTD =
paper investigated several alternate control schemes in simulations
ln ( T1 − Tcondin
T2 − Tcondout ) (A8)
for a reactive distillation column used for the hydrogenation of
Packed section
benzene. A combination of feedforward and feedback control, where
Plate j
the feedforward control seeks to minimize the effect of feed com-
Mass balance
position disturbances, can result in acceptable performance, how-
n
ever, this comes at the cost of having no time delay in the dis- d ( Ml j + Mv j )
= Vj +1 + L j −1 − Vj − L j + Ahj (HETP )ρcat ∑ rxnj, i
turbance measurements which requires a feed composition analyzer. dt (A9)
i=1
An alternative to this approach is to use MPC for control of this
process. SISO MPC and MIMO MPC methods with input and output Component balances
disturbances have been implemented. All the MPC methods provide
responses which are better than the decentralized PI feedback-only
(
d Ml j xj, i + Mv j yj, i ) = Fj zj, i + Vj +1yj +1, i − L j −1xj −1, i −Vj yj, i − L j xj, i
control scheme. Surprisingly, MIMO MPC methods did not provide dt
benefits over SISO MPC methods when the column is subjected to + Ahj (HETP )ρcat rxnj, i ∀ i : 1 to n−1 (A10)
feed benzene concentration disturbances as the interaction between
the control loops is not significant. At the same time, MPC methods n n

with input disturbances provide a significant reduction in the upset ∑ x j, i = 1; ∑ yj, i = 1

i=1 i=1 (A11)
condition over MPC methods using only an additive output dis-
turbance correction. It can be concluded from these simulation stu- Vapor–liquid equilibrium and Murphree efficiency
dies that a decentralized control structure with a model predictive
xj, i ∅l j, i = y* j, i ∅l j, i ∀ i : 1 to n − 1 (A12)
controller for manipulating the hydrogen feed to the column war-
rants further consideration for this process. Furthermore, using an
input disturbance model for controlled variable prediction is prefer- yj, i = yj +1, i (1 − η) + y* j, i η (A13)
able over using an output disturbance. An observer can be used to
Energy balance
estimate the feed concentration disturbance and thus eliminates the
need for a composition analyzer. Extensive simulation studies have d ( Ml jHl j + Mv jHv j )
= Vj +1Hj +1 + L j −1Hl j −1 − L j Hl j − Vj Hv j − Hrxn, j
been performed on a detailed fundamental model to evaluate the dt (A14)
different control schemes which led to the stated conclusions.
Flow rate and holdups
Liquid:
Acknowledgment Mj voll j = hj (HETP ) A (A15)

The authors gratefully acknowledge partial financial support hj = 0. 34a1/3uj2/3 (A16)

from the American Chemical Society - Petroleum Research Fund
(Grant PRF# 50978-ND9). Vapor:
 V. Mahindrakar, J. Hahn / Control Engineering Practice 52 (2016) 103–113 113

Pj A (HETP ) ( ε−hj ) = Mv j Rgas Tj (A17) a a

x^k +1 = Φax^k + Γ auk + La ( yk −C axka ) (A33)
∆P0, j (1−ε) uv2ρv
=ψ 3
HETP ε dp K (A18)

150 References
ψ = +1. 75
Rev (A19)
Al-Arfaj, M., & Luyben, W. L. (2000). Comparison of alternative control structures for an
⎡ 2hj ⎤
−5 ideal two-product reactive distillation column. Industrial and Engineering Chemistry
∆Pj = ∆P0, j ⎢ 1− ⎥ Research, 39(9), 3298–3307.
⎣ dp a ⎦ (A20) Alejski, K., & Duprat, F. (1996). Dynamic simulation of the multicomponent reactive
distillation. Chemical Engineering Science, 51(18), 4237–4252.
Bequette, B. W. (2003). Process control: modeling, design, and simulation (1st ed.). Upper
Pj +1 = Pj+∆Pj (A21) Saddle River, NJ: Prentice Hall Professional.
Bristol, E. (1966). On a new measure of interaction for multivariable process control. IEEE
Reaction rate Transactions on Automatic Control, 11(1), 133–134.
Choe, Y. S., & Luyben, W. L. (1987). Rigorous dynamic models of distillation columns.
kj,1K A1KH1C A j,1CH j Industrial and Engineering Chemistry Research, 26(10), 2158–2161.
rxnj, Benzene = − Environmental Protection Agency (2000). Control of emissions of hazardous air pollutants
⎛ 1 ⎞3 from mobile sources. Available: 〈http://www.epa.gov/otaq/regs/toxics/f00055.pdf〉.
⎜ 3K A1C A
j,1 + ( K H1C H j ) 2 +1⎟
⎝ ⎠ (A22) Environmental Protection Agency (2006). Control of hazardous air pollutants from mobile
sources: 40 CFR Parts, 59, 50, 85 and 85. Available: 〈http://www.epa.gov/otaq/regs/
toxics/proposedrule.pdf〉.
kj,2K A2KH2C A j,2 CH j Environmental Protection Agency (2007). Control of hazardous air pollutants from mobile
rxnj, Toluene = − sources : final rule to reduce mobile source air toxics. Available: 〈http://www.epa.gov/
⎛ 1 ⎞3 oms/regs/toxics/420f07017.pdf〉.
⎜ 3K A2C A
j,2 + ( K H2C H j ) 2 +1⎟
⎝ ⎠ (A23) Froisy, J. B. (2006). Model predictive control—Building a bridge between theory and
practice. Computers and Chemical Engineering, 30(10–12), 1426–1435.
Georgiadis, M. C., Schenk, M., Pistikopoulos, E. N., & Gani, R. (2002). The interactions of
⎡ E ⎛ 1 1 ⎞⎤ design control and operability in reactive distillation systems. Computers and Che-
ki, j = ki,0 exp ⎢ − a ⎜ − ⎟ ⎥ mical Engineering, 26(4–5), 735–746.
⎢⎣ Rgas ⎝ Tj T0 ⎠ ⎥⎦ (A24) Hall, C., & Mackowiak, J. (2009). Fluid dynamics of packed columns: Principles of the fluid
dynamic design of columns for gas/liquid and liquid/liquid systems (1st ed.). New York,
Reboiler NY: Springer.
Harmsen, G. J. (2007). Reactive distillation: The front-runner of industrial process in-
Mass balance tensification: A full review of commercial applications, research, scale-up, design and
operation. Chemical Engineering, 46(9), 774–780.
d ( MlN ) Jacobs, R., & Krishna, R. (1993). Multiple solutions in reactive distillation for methyl tert-
= LN −1 − LN − VN
dt (A25) butyl ether synthesis. Industrial and Engineering Chemistry Research, 32(8),
1706–1709.
Component balances Li, S., Lim, K. Y., & Fisher, D. G. (1989). A state space formulation for model predictive
control. AIChE Journal, 35(2), 241–249.
Luyben, W. L. (2006). Evaluation of criteria for selecting temperature control trays in
d ( MlN xN, i )
= LN −1xN −1, i − LN xN, i − VN yN, i ∀ i : 1 to n−1 distillation columns. Journal of Process Control, 16(2), 115–134.
dt (A26) Mahindrakar, V., & Hahn, J. (2014). Dynamics and control of benzene hydrogenation via
reactive distillation. Journal of Process Control, 24(3), 113–124.
n Morari, M., & Lee, J. H. (1999). Model predictive control: past, present and future. Com-
puters and Chemical Engineering, 23(4–5), 667–682.
∑ xN, i =1 Muske, K. R., & Badgwell, T. A. (2002). Disturbance modeling for offset-free linear model
i=1 (A27) predictive control. Journal of Process Control, 12(5), 617–632.
Palmer, R. E., Kao, S. H. , Tong, C., Shipman D. R. (2008). Consider options to lower benzene
Energy balance levels in gasoline. Available: 〈http://www.woodgroup.com/SiteCollectionDocuments/
news-tech-articles/2008-06_GasBenzeneLevels_Mustang.pdf〉.
d ( MlNHlN ) Pannocchia, G. (2003). Robust disturbance modeling for model predictive control with
= LN −1HlN −1 − (LN +VN ) HlN application to multivariable ill-conditioned processes. Journal of Process Control, 13
dt (A28)
(8), 693–701.
Peng, J., Lextrait, S., Edgar, T. F., & Eldridge, R. B. (2002). A comparison of steady-state
VN HlN − VN HvN + Q N =0 (A29) equilibrium and rate-based models for packed reactive distillation columns. In-
dustrial and Engineering Chemistry Research, 41(11), 2735–2744.
Process Systems Enterprise (1997–2015). gPROMs ModelBuilder V3.7. Available: 〈www.
Augmented state space system for disturbance estimation.
psenterprise.com/gproms〉.
⎡ xk + 1⎤ ⎡ Φ Γ d ⎤ ⎡ xk ⎤ ⎡ Γ⎤ Qin, S. J., & Badgwell, T. A. (1997). An overview of industrial model predictive control
⎢ ⎥=⎢ ⎥ ⎢ ⎥ + ⎢ ⎥ uk
⎣0⎦
technology. AIChE Symposium Series, 93, 232–256.
⎣ dk + 1⎦ ⎣ 0 I ⎦ ⎣ dk ⎦ Schei, T. S. (2008). On-line estimation for process control and optimization applications.
Journal of Process Control, 18(9), 821–828.
⎡ xk ⎤ Shinskey, F. G. (1994). Feedback controllers for the process industries (1st ed.). New York,
yk = ⎡⎣ C 0⎤⎦ ⎢ ⎥
⎣ dk ⎦ (A30)
NY: McGraw-Hill Professional.
Skogestad, S. (1997). Dynamics and control of distillation columns – A critical survey.
Modeling, Identification and Control, 18(3), 177–217.
Skogestad, S., & Morari, M. (1987). Control configuration selection for distillation col-
umns. AIChE Journal, 33(10), 1620–1635.
xka+ 1 = Φaxka + Γ auk Skogestad, S., & Postlethwaite, I. (2007). Multivariable feedback control: analysis and de-
sign (2nd ed.). New York, NY: John Wiley & Sons.
yk = C axka (A31) Skogestad, S., Lundström, P., & Jacobsen, E. W. (1990). Selecting the best distillation
control configuration. AIChE Journal, 36(5), 753–764.
where, Sneesby, M. G., Tadé, M. O., & Smith, T. N. (1998). Steady-state transitions in the reactive
distillation of MTBE. Computers and Chemical Engineering, 22(7–8), 879–892.
⎡ xk ⎤ ⎡ d⎤ ⎡ Γ⎤ Sneesby, M. G., Tadé, M. O., Datta, R., & Smith, T. N. (1997). ETBE synthesis via reactive
xka = ⎢ ⎥; Φa = ⎢ Φ Γ ⎥; Γ a = ⎢ ⎥
⎣ dk ⎦ ⎣0 I ⎦ ⎣ 0⎦ (A32)
distillation. 2. Dynamic simulation and control aspects. Industrial and Engineering
Chemistry Research, 36(5), 1870–1881.