Você está na página 1de 9

ARTICLE IN PRESS

Control Engineering Practice 15 (2007) 231239


www.elsevier.com/locate/conengprac

Nonlinear model predictive control of a reactive distillation column


Rohit Kawathekar, James B. Riggs
Chemical Engineering, Texas Tech University, Lubbock, TX 79409-3121, USA
Received 5 July 2005; accepted 10 July 2006
Available online 7 September 2006

Abstract

This paper considers the application of nonlinear model predictive control (NLMPC) to a highly nonlinear reactive distillation
column. NLMPC was applied as a nonlinear programming problem using orthogonal collocation on nite elements to approximate the
ODEs that constitute the model equations for the reactive distillation column. Diagonal PI controls were used to identify that the
[L/D,V] and the [L/D,V/B] congurations performed best. NLMPC was applied using the [L/D,V] conguration and found to provide a
factor of 23 better performance than the corresponding PI controller. The effect of process/model mismatch on the performance of the
NLMPC controller was also evaluated.
r 2006 Elsevier Ltd. All rights reserved.

Keywords: Reactive distillation; Nonlinear model predictive control; Conguration selection; Process/model mismatch

1. Introduction distillation column using a nonlinear rst-principles model


combined with an articial neural network model. Reactive
Reactive distillation combines both separation and distillation is a challenge for control due to process
reaction in one unit and has been applied industrially for nonlinearity and complex interactions between vaporli-
a number of years. Reactive distillation can offer sig- quid equilibrium and chemical reactions.
nicant economic advantages for certain cases, particularly Al-Arfaj and Luyben (2000), Sneesby, Tade, Datta, and
for systems that involve reversible reactions. Examples of Smith (1997) and Kumar and Daoutidis (1999) have
commercial successes of reactive distillation include nylon considered decentralized PI control structures for reactive
6,6 processes, methyl acetate processes, ethyl acetate distillation columns. Khaledi and Young (2005) applied PI
processes and methyl tert-butyl ether processes (Doherty controls and linear model predictive control to a dynamic
& Buzad, 1992). simulation of an ETBE reactive distillation column using
Most of the literature available on reactive distillation is the (L,V) conguration. The linear model predictive
based on steady-state conditions including process design controller used rst-order plus deadtime models for each
(e.g., Chang & Seader, 1988; Krishnamurthy & Taylor, of its input/output models. Kumar and Daoutidis (1999)
1985) and the analysis of multiple steady states (e.g., have discussed the superior performance of nonlinear
Abufares & Douglas, 1995; Jacobs & Krishna, 1993; controllers compared to linear controllers for reactive
Sneesby, Tade, & Smith, 1998). Dynamic modeling and distillation systems. Volker, Sonntag, and Engell (2006)
simulation (e.g., Grosser, Doherty, & Malone, 1987; Ruiz, applied a linear reduced-order MPC controller to a semi-
Basualdo, & Scenna, 1995) have also been studied, but a batch reactive distillation column. Gruner et al. (2003)
relatively small amount of work has been reported on the developed a general controller for reactive distillation
control of reactive distillation columns. Chen, Hontoir, systems based on asymptotically exact input/output
Huang, Zhang, and Julian Morris (2004) performed a linearization and applied their controller to a simulation
nonlinear dynamic modeling case study of a reactive of an industrial reactive distillation column with improved
performance compared to a well-tuned linear controller.
Corresponding author. Tel.: +1 806 742 1765; fax: +1 806 742 3552. This paper considers the application of nonlinear model
E-mail address: jim.riggs@ttu.edu (J.B. Riggs). predictive control (NLMPC) to an ethyl acetate reactive

0967-0661/$ - see front matter r 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.conengprac.2006.07.004
ARTICLE IN PRESS
232 R. Kawathekar, J.B. Riggs / Control Engineering Practice 15 (2007) 231239

Nomenclature xi,j the mole fraction of the jth component in the


liquid on the ith stage
B bottom product ow rate yi,j the mole fraction of the jth component in the
D distillate product ow rate vapor on the ith stage
Fi the external feed rate to the ith stage. ysp the setpoint for the controlled variable
Hi the enthalpy of the vapor leaving the ith stage
hi the enthalpy of the liquid leaving the ith stage Greek Symbols
ki the reaction equilibrium rate constant for the
ith stage gi,j the liquid phase activity coefcient for the jth
kfi the reaction rate constant for the ith stage component on the ith stage
L the reux rate mj the stochiometric coefcient for the jth compo-
L/D the reux ratio nent in the primary reaction
Mi l the molar holdup of liquid on the ith stage ri the liquid density for the ith stage
nc the number of components
Pi the pressure on the ith stage Subscripts
Psi,j the vapor pressure of component j on the ith
stage A acetic acid
QC the condenser duty B ethyl alcohol
QR the reboiler duty C ethyl acetate
QMS the mover suppression factor for the NLMPC c condenser
controller D water
ri the reaction rate on the ith stage i stage number
u the manipulate variable f feed
V the boilup rate j component number
V/B the boilup ratio n the top tray number
Voli the volume of the liquid holdup on the ith stage

distillation column. In Section 2, the case study is described for the reactive distillation column. A reactive column with
in detail. Application of diagonal PI controls is presented countercurrent ow of reactants was used. Pure ethanol
in Section 3. Development and implementation of NLMPC and acetic acid were fed separately into a column that
algorithm is discussed in Section 4. In Section 5, the operates at atmospheric pressure. Under these conditions,
comparison between PI and NLMPC control results is acetic acid is the heaviest of the components and moves
presented. The effect of process-model mismatch on the toward the bottom of the column. Ethyl acetate is the
closed-loop performance of NLMPC controller is analyzed lightest and moves toward the top of the column. It is
in Section 6. expected that the middle portion of the distillation column
is the primary reaction zone. The rectifying section
2. Model development fractionates the ethyl acetate out of acetic acid, and the
stripping section removes alcohol from water. This
This study is based on the production of ethyl acetate by quaternary system consisting of ethanol, acetic acid, water
a reactive distillation using the esterication reaction and ethyl acetate is highly nonideal. It can form four
between acetic acid and ethanol. The achievable conversion binary azeotrope mixtures and one ternary azeotrope. Over
in this reversible reaction is limited by the equilibrium a wide range of composition, ethanol and water do not
conversion. The reaction is slightly endothermic and takes differ greatly in volatility, making it difcult to produce
place in the liquid phase. Though the esterication reaction only water as the bottom product. The reactant ethanol has
is self-catalyzed, sulfuric acid can act as external catalyst to a relatively high volatility in the reaction zone. This leads
enhance the reaction rate: to a low composition of ethanol in the liquid phase,
reducing the production rate of ethyl acetate.
CH3 COOH C2 H5 OH Under these conditions, the rate of the esterication
Sulfuric acid reaction between acetic acid and ethanol is generally low,
! CH3 COOC2 H5 H2 O: 1
which implies that it is favored by long residence times in
The reaction kinetics for this reaction were proposed by each stage. It is evident from all the previous studies on
Aljeski and Duprat (1996) and include reaction equilibrium ethyl acetate reactive distillation columns that an unfavor-
limitations. Bock and Wozny (1997) reported a detailed able physical equilibrium makes the production of high-
analysis of ethyl acetate reactive distillation column. The purity ethyl acetate impossible from a single distillation
slow reaction rate for this system leads to lower conversion column.
ARTICLE IN PRESS
R. Kawathekar, J.B. Riggs / Control Engineering Practice 15 (2007) 231239 233

Reactive Column Recovery Column reported by Seferlis and Grievink (2001). This dynamic
model served as the process that was controlled by the PI
and NLMPC controllers in this study. An equilibrium
stage dynamic model of the reactive distillation column
was implemented with the following assumptions: (1)
AA feed each tray was assumed to be an ideal stage, (2) nonequal
101.3 350.3
molar overow was considered, (3) the liquid activity
kPa kPa
Purge
coefcient was modeled using the approach of Suzuki,
EtOH feed RX Yagi, Komatsu, and Hirata (1971) [15], (4) the SRK
RC method (Soave, 1992) was used to model the enthalpy
departure functions, (5) a hydraulic time constant was
used to model the liquid ow from each tray, (6) tray
temperatures in the rectifying and stripping sections were
used to infer the impurity composition of the overhead and
bottom products, respectively, (7) the heat of dilution of
acetic acid in water was neglected and (8) equilibrium
Ethyl acetate product limited reaction kinetics (Aljeski & Duprat, 1996) are
assumed.
Recycle
For simulation purposes, the process is modeled using a
Fig. 1. Two-column sequence for the production of high-purity ethyl detailed tray-to-tray model. For a standard reactive
acetate. distillation tray where vapor holdup is considered negli-
gible and reaction takes place in liquid phase, the modeling
equations can be written as follows:
However, Seferlis and Grievink (2001) proposed adding Xnc
a recovery column (Fig. 1), which operates at a higher dM li
F i Li1 V i1  Li  V i Vol i m j ri , (2)
pressure to produce a high-purity ethyl acetate product. dt j1
The recovery column was combined with the reactive
distillation column and designed to operate at a higher dM li xi;j
pressure (350 kPa) to break the azeotropes, which increases F i xi;f Li1 xi1;j V i1 yi1;j
dt
the overall conversion and produces a high-purity ethyl
 Li xi;j  V i yi;j Vol i mj ri , 3
acetate product. The distillate stream from the reactive
column is fed to the recovery column. At a relatively high
pressure, ethyl acetate becomes heavier than ethanol and dM li hi
F i hf i Li1 hi1 V i1 H i1  Li hi  V i H i , (4)
water so that it appears as the bottom product. The target dt
purity level of the ethyl acetate was set at 99.5%. Reaction
in the recovery column is negligible because the column dhi
 0, (5)
operates without any sulfuric acid catalyst and the liquid dt
holdup on the trays is small. Virtually all of the acetic acid  
1
in the overhead of the reactive column ends up in the ri r2i kfi xA;i xB;i  xC;i xD;i , (6)
bottom product of the recovery column, and therefore, ki
directly affects the purity of nal ethyl acetate product.
Hence, a specication is imposed on the maximum yi;j Pi gi;j xi;j Psi;j . (7)
allowable concentration of acetic acid in the distillate of For the partial condenser, the material balance and
the reactive column. To assist in the reduction of the acetic energy balance equations are as follows:
acid in the distillate of the reactive column, the reactive
column was designed with a reduced liquid phase holdup in dM lc Xnc

the upper section of trays of the reactive column to V n  Lc  V D Vol c m j rc , (8)


dt j1
suppress the reverse reaction of ethyl acetate.
Seferlis and Grievink (2001) developed an economically
optimized steady-state design for a reactive distillation dM lc xc;j
V n yn;j  Lc xc;j  V D yD;j Vol c mj rc , (9)
process based on the two-column conguration (Fig. 1). dt
The exact design and steady-state operating conditions of
the reactive distillation column from Seferlis and Grievink dM lc hc
V n H n  Lc hc  V D H D  Qc , (10)
(2001) were used as the basis of this study while the dt
recovery column was not directly modeled. An equilibrium
stage dynamic model of the reactive distillation column dhlc
 0, (11)
was developed and benchmarked against the results dt
ARTICLE IN PRESS
234 R. Kawathekar, J.B. Riggs / Control Engineering Practice 15 (2007) 231239

 
1 The overhead level and pressure control structures are
ri r2i kfi xA;i xB;i  xC;i xD;i , (12)
ki shown in Fig. 2 based on producing a vapor distillate
product. The overhead pressure control loop sets the
yi;j Pi gi;j xi;j Psi;j , (13) distillate vapor ow rate. The overhead pressure was
modeled considering the total number of moles of vapor in
where subscript c represents condenser stage, subscript n
the overhead and the vapor volume in the overhead
represents the top tray of the column and subscript D
line and the overhead condenser. A ooded condenser is
represents the distillate. VD represents the vapor distillate
used and the level of liquid in the condenser is used to
from the partial condenser stage.
vary the surface area in the condenser available for
The reux accumulator stage is modeled as a liquid
condensing the overhead vapor. An accumulator level
holdup tank. The terms relating to the vapor phase, i.e.,
controller determines the setpoint for the condenser level
vapor ow, vapor composition, vapor enthalpy corre-
controller.
sponding to reux accumulator are absent in the material
Because all the acetic acid that leaves in the overhead of
and energy balances.
the reactive distillation column ends up in the ethyl acetate
The material and energy balance equations for the
product produced by the recovery column, the primary
reboiler are written as follows:
composition control point is the acetic acid concentration
dM lB Xnc in the overhead product from the reactive distillation
Li  B  V B Vol B mj rB , (14) column (i.e., the design setpoint is 0.04% acetic acid in the
dt j1
overhead). The composition setpoint for the bottoms is
0.08% ethyl acetate, which was determined by Seferlis and
dM lB xB;j
Li xi;j  BxB;j  V B yB;j Vol B mj rB , (15) Grievink (2001) using a steady-state economic analysis of
dt the two-column system.
The temperature of the second tray from the top was
dM lB hB
Li hi  BhB  V B H B QR , (16) used to infer the impurity level in the overhead product and
dt the fth tray from the bottom was used to infer the
impurity level in the bottom product. These tray tempera-
dhlB
 0, (17) tures were identied using a steady-state model of the
dt reactive distillation column to determine the tray tempera-
  tures that exhibited the strongest correlation between
1
ri r2i kfi xA;i xB;i  xC;i xD;i , (18) changes in the product purities and changes in the tray
ki
temperatures. The correlations between tray temperature
yi;j Pi gi;j xi;j Psi;j . (19) and product composition were updated using new compo-
sition analyzer information when it became available. It
was determined that this inferential tray temperature
scheme provided excellent estimates of the product
D
composition by comparing the inferential estimate of the
product impurity levels with the corresponding values
generated by the process simulator. Therefore, an observer
was not considered necessary in this case.
Table 1 lists the steady-state design conditions for the
reactive distillation column studied here. The model
PC LT
equations for the reactive distillation column (i.e., the
liquid dynamics equation, the component material bal-
ances and the energy balance for each tray, the component
RSP CW
LC LC material balances and the overall material balances for the
accumulator, reboiler and condenser, and the PI control
equations for the pressure controller and the level
LT controllers) were integrated numerically using the Euler
PT method with a step size of 0.5 s. To improve the
computational efciency of the dynamic reactive distilla-
tion model, the inside-out algorithm (Boston & Sullivan,
1974) was used to calculate the component K-values. If a
tray temperature changed by more than 1 1C, the para-
L
meters of the K-value correlation were updated. In
addition, all K-value correlations were updated each 30 s
Fig. 2. Schematic of the column pressure controls and the accumulator of simulated time. The accumulator and reboiler level
level controls. controllers were tuned for critically damped behavior.
ARTICLE IN PRESS
R. Kawathekar, J.B. Riggs / Control Engineering Practice 15 (2007) 231239 235

Table 1 4.6E-04
Steady-state design conditions for the reactive distillation column
L/D,V

Overhead impurity
4.4E-04

(mole fraction)
Number of stages 31 L/D,V/B
Rectifying section 253a L/D,B
Reactive section 724 4.2E-04
Stripping section 16
Stage holdups 4.0E-04
Rectifying section (m3) 0.151
Reactive section (m3) 1.75
3.8E-04
Stripping section (m3) 1.65
0 500 1000 1500 2000
Distillate ow (kmol/s) 4.88  103 Time (min)
Bottoms ow (kmol/s) 9.09  104
Reux ratio 0.8211 Fig. 3. Overhead composition control using PI controls for an unmea-
sured feed composition upset.
Feed flow rate
Acetic acid feed (kmol/s) 8.92  104
Ethanol feed (kmol/s) 8.74  104
Recycle feed (kmol/s) 4.03  103
Distillate composition (mol frac) 3. Diagonal PI control results
Ethanol 0.2383
Acetic acid 0.0004
Water 0.2093
Diagonal PI composition controllers were used to
Ethyl acetate 0.5520 compare the control performance of various control
congurations and as a benchmark for the NLMPC
Bottoms composition (mol frac)
Ethanol 0.0249
control results. Because the primary control objective for
Acetic acid 0.0876 the reactive distillation column is control of the impurity in
Water 0.8867 the overhead product, the bottom composition control
Ethyl acetate 0.0008 loop was tuned rst for a closed-loop overdamped
Reboiler duty (kW) 171.0 response. Then the overhead PI composition controller
Condenser duty (kW) 137.7 was tuned using setpoint changes. The overhead tuning
a
procedure was based on using an ATV test using a
Low holdup was used to reduce reaction in the rectifying section.
single tuning factor (Riggs, 2001) combined with Tyreus
Luyben settings (Tyreus & Luyben, 1992) and nally
adjusted (Riggs, 2001) to minimize the IAE from setpoint
Table 2
Steady-state changes for a 10% change in a manipulated variable
for a series of setpoint changes. The control performance
of the overhead and bottoms composition controllers
Conguration % change in the process gain was evaluated for several unmeasured disturbances in feed
rate and feed composition. Because fresh acetic acid and
Overhead Bottom
ethyl alcohol are each fed to the reactive column as well as
[L/D,V] 200 300 the recycle stream from the recovery column, feed
[L/D,V/B] 150 300 composition upsets were implemented by changing the
[L,V] 150 600
ratio of fresh acetic acid and ethyl alcohol and feed
[L,V/B] 33 300
rate changes were implemented by changing the recycle
feed rate.
The [L,B], [L,V], [L,V/B], [L/D,B], [L/D,V] and [L/D,V/
The steady-state gains for the ethyl acetate reactive B] congurations were compared for control performance
distillation column are listed in Table 2. Note that for based on dual composition PI control (note that for a [A,B]
approximately a 10% change in each manipulated variable conguration A is manipulated to control the overhead
of each conguration, the process gain mostly changed by composition and B to control the bottom composition).
150600%, indicating severe steady-state process nonli- Congurations using the distillate rate, D, as a manipu-
nearity. In addition, several of the off-diagonal gains lated variable were not considered because D is used to
demonstrated bi-directionality. As is usually the case, control the column pressure (Fig. 2).
process integration (i.e., adding the recovery column with Figs. 3 and 4 show the overhead and bottoms composi-
recycle in this case) offers signicant steady-state economic tion control performance for the [L/D,B], [L/D,V] and [L/
benets but at the expense of a much more challenging D,V/B] congurations for an unmeasured feed composition
control problem. The dynamic model of the reactive upset. Table 3 lists the IAEs for each conguration for a
distillation column described in this section was used as feed rate upset. Based on these results, the [L/D,V] and [L/
the process, which was controlled by PI and nonlinear D,V/B] congurations provided the best overall control
MPC controllers. performance.
ARTICLE IN PRESS
236 R. Kawathekar, J.B. Riggs / Control Engineering Practice 15 (2007) 231239

8.3E-04 differences observed between the process output and the


8.2E-04 model prediction are due to additive step disturbances in
impurity (mole fraction)

the output. These disturbance terms are also assumed to


Bottom ethyl acetate

8.1E-04
8.0E-04 remain constant over the prediction horizon.
7.9E-04
7.8E-04
4.1. Solution algorithm
7.7E-04 L/D,V/B
L/D,V
NLMPC was implemented by setting up the control
7.6E-04
problem as a nonlinear programming (NLP) problem and
7.5E-04 L/D,B
solving it over the prediction horizon. It is necessary to
7.4E-04
simultaneously solve an optimization problem (the tradeoff
0 2000 4000 6000
between control to setpoint and changes in the manipu-
Time (min)
lated variables) and the system model equations. These two
Fig. 4. Bottom composition control performance using PI control for an procedures may be implemented either sequentially or
unmeasured feed composition upset. simultaneously.

4.1.1. Sequential solution and optimization algorithm


Table 3 The sequential algorithm employs separate algorithms to
Dual PI composition control performance indices for unmeasured feed solve the differential equations and to carry out the
rate disturbance
optimization. First, a manipulated variable prole is
Conguration IAE selected and the differential equations are solved numeri-
cally to obtain the controlled variable prole. The objective
Overhead Bottoms function is then determined. The gradient of the objective
[L,V] 0.143 1.278 function with respect to the manipulated variables is
[L,B] 0.177 0.830 determined either by numerical perturbation or by using
[L,V/B] 0.142 1.239 analytical derivatives. The control prole is then updated
[L/D,V] 0.115 1.270 using an optimization algorithm. The process is repeated
[L/D,B] 0.393 1.215
until the optimal prole is obtained. This is referred as a
[L/D,V/B] 0.114 0.988
sequential solution and optimization algorithm.
The implementation of the sequential solution and
optimization algorithm is relatively simple to apply.
4. Implementation of NLMPC However, there are some drawbacks associated with this
approach. The sequential solution and optimization
NLMPC uses nonlinear models in a model predictive requires the solution of differential equations at each
control framework to choose control action. The NLMPC iteration of the optimization. Jones and Finch (1984) found
controller chooses the future control moves that minimize that such methods spend about 85% of the time integrating
the following objective function: the model equations in order to obtain gradient informa-
tion. This can make the implementation of this algorithm
X
N 1 X
N 1
computationally expensive for cases involving a large
min ysp  yj 2 QMS Duj 2 , (20)
j0 j0
number of model equations. The gradient information
required for the optimization procedure is often obtained
subject to bounds on manipulated, output and state through numerical differentiation, as the analytical deri-
variables. vatives are not always available for highly nonlinear model
The rst term represents the error from setpoint and the equations involving complicated thermodynamic relations.
second term represents the weighted variation in the To obtain the gradients using nite difference typically
manipulated variables. Once the optimum set of future involves differencing the output of an integration routine
moves is determined, the initial control move is imple- with adaptive step sizes. Gill, Murray, and Saunders (1988)
mented. In the next control cycle, the next set of optimum points out that the integration error is unpredictable and
control moves is determined. Even though the full set of hence differencing output of an integration routine greatly
optimum control moves into the future are calculated at degrades the quality of the nite difference derivatives. It is
each control interval, only the rst move is actually also difcult to incorporate the constraints on state
implemented at each control interval. variables with the use of the sequential solution and
The formation of the NLMPC objective function will optimization approach (Meadows & Rawlings, 1997).
involve a disturbance term, which is added to the output
prediction over the entire prediction horizon to match the 4.1.2. Simultaneous solution and optimization algorithm
model prediction to the current process measurement. The The simultaneous solution and optimization algorithm
NLMPC procedure in this work assumes that the involves the model equations appended to the optimization
ARTICLE IN PRESS
R. Kawathekar, J.B. Riggs / Control Engineering Practice 15 (2007) 231239 237

problem as equality constraints. Then the NLP problem is Table 4


posed to optimize the objective function such that the Ethyl acetate NLMPC control performance indices for overhead impurity
setpoint tracking
algebraic form of the model differential equations are satised,
and the upper and lower constraints on states, controlled Conguration Overhead loop IAE Bottoms loop IAE
variables and manipulated variables are met. This can greatly
increase the size of optimization problem, leading to a tradeoff [L/D,V] NLMPC 0.91 2.14
[L/D,V] PI 2.41 6.84
between the two approaches. Meadows and Rawlings (1997)
reported that for small problems with few states and a short
prediction horizon, the sequential solution and optimization
algorithm is more computationally efcient. For larger Table 5
problems, the simultaneous solution and optimization NLMPC control performance indices for unmeasured feed composition
approach is more computationally efcient and more reliable. disturbance rejection
In the simultaneous solution and optimization approach,
Conguration IAE
the model equations and the process constraints (e.g., a
manipulated variable cannot be less than zero) become Overhead Bottoms
constraints for the optimization problem to determine the
[L/D,V] NLMPC 0.0609 0.483
future control moves. Since the model equations largely
[L/D,V] PI 0.115 1.270
appear as ODEs, orthogonal collocation on nite elements
(Finlayson, 1980) was used to convert the ODEs in time into
a set of algebraic equality constraints for the optimization the simulation of the ethyl acetate reactive distillation
problem. Due to the size of the resulting NLMPC problem column. The previous PI dual-ended composition control
considered here, the simultaneous solution approach was results indicated that [L/D, V] conguration provided good
chosen for this work. The process model and the NLMPC control performance for setpoint tracking as well as the
controller were implemented in Fortran code. The NLP rejection of unmeasured disturbances. NLMPC was
problem resulting from the implementation of NLMPC was applied using the [L/D, V] conguration to determine the
solved using the SNOPT software (Gill et al., 1988). Typical benets compared to conventional PI controls.
computational times for 30 h of simulation of the NLMPC Table 4 summarizes the comparisons between diagonal
controller using an 1100 MHz in an AMD PC with PI control and NLMPC using the [L/D,V] conguration
Windows XP required 10 h CPU. for a series of setpoint changes in the overhead impurity.
Table 5 summarizes the comparisons between diagonal PI
4.2. Selection of tuning parameters control and NLMPC for an unmeasured feed composition
upset. These results are based on an NLMPC controller
The tuning parameters that have a signicant effect on with an almost perfect process model. Based on the IAEs of
NLMPC performance are the prediction horizon, control these tests, the NLMPC controller outperformed the PI
horizon, sampling interval, penalty weight matrices and controller by a factor of 23. Due to the severe nonlinearity
move suppression factors. A set of heuristics based on the of this process, one would expect a larger disparity between
linear systems and numerical simulations were used to PI and NLMPC. On the other hand, this column was
select the nal tuning parameters operated largely as a single-ended column control problem
Following is a discussion of the tuning parameters used for because the bottom composition control was detuned
the tuning of each NLMPC controller considered: (1) The compared to the more important overhead product, which
control interval was set at 20 min, which was shown to be signicantly simplied the control problem.
equivalent to continuous control using a correlation by
Marlin (1995). (2) The prediction horizon was set at 80 by 6. Process/model mismatch
using different levels and comparing control performance. (3)
A control horizon of 15 was found to provide good control Process/model mismatch for the NLMPC controller was
performance. (4) The equal concern errors were selected to represented by introducing a 5% and a 25% error between
provide a 10 times greater weighting of the overhead errors the reaction equilibrium constant used in the dynamic
from setpoint than the bottoms because impurities in the process simulator and the model equations used by the
overhead directly affect the purity of the ethyl acetate nal NLMPC controller to calculate control action. Figs. 59
product. (5) The move suppression factors were adjusted to show the effect of process model/mismatch for NLMPC
minimize the IAE from setpoint for the overhead product for for an unmeasured feed composition upset. The PI control
a series of overhead impurity setpoint changes. results are also shown for this case.
To more completely analyze the sensitivity of PI and
5. Comparison between PI and NLMPC NLMPC control performance to unmeasured disturbances,
a sinusoidal disturbance in the recycle feed rate with
The NLMPC control algorithm described in the previous constant amplitude and varying frequency was applied to
sections was applied for dual-ended composition control of the process with a PI controller and an NLMPC controller
ARTICLE IN PRESS
238 R. Kawathekar, J.B. Riggs / Control Engineering Practice 15 (2007) 231239

4.2E-04 1.68E+02
0% mismatch

Reboiler duty (btu/sec)


1.66E+02
Overhead impurity

4.0E-04 5% mismatch
(mole fraction)

1.64E+02 25% mismatch


3.8E-04 1.62E+02 PI
0% mismatch
1.60E+02
5% mismatch
3.6E-04
25% mismatch 1.58E+02
PI
3.4E-04 1.56E+02
0 200 400 600
1.54E+02
Time (min) 0 500 1000 1500
Fig. 5. Control results for the overhead product for an unmeasured feed Time (min)
composition upset showing the effect of process/model mismatch for Fig. 8. The manipulated variable prole for bottoms for an unmeasured
NLMPC and compared with PI control for the [L/D,V] conguration. feed composition upset for the [L/D,V] conguration.

8.3E-04 0.07

0.06 PI
impurity (mole fraction)
Bottoms ethyl acetate

NLMPC
8.1E-04 Overhead impurity
amplitude ratio
0.05
NLMPC
0.04 5% mismatch
7.9E-04 0.03

0% mismatch 0.02
7.7E-04 5% mismatch
0.01
25% mismatch
PI 0
7.5E-04 0 500 1000 1500
0 500 1000 Time period for sinusoidal feed
Time (min) disturbance (min)

Fig. 6. Control results for the bottom product for an unmeasured feed Fig. 9. A comparison of closed-loop amplitude ratios for PI and NLMPC
composition upset showing the effect of process/model mismatch for for the [L/D,V] conguration.
NLMPC and compared with PI control for the [L/D,V] conguration.

sinusoidal disturbance to the amplitude of the recycle feed


1.3 disturbance. The closed-loop performances of the PI
control structure, NLMPC controller as well as NLMPC
1.2
with 5% process/model mismatch were compared using the
1.1 unmeasured sinusoidal recycle feed rate disturbance. A plot
Reflux ratio

of the closed-loop amplitude ratio versus the period of the


1.0 sinusoidal feed disturbance (Fig. 9) shows that for the PI
0% mismatch controller as well as the NLMPC controller exhibit
0.9
5% mismatch sensitivity to the feed disturbance in a specic range of
0.8 25% mismatch frequencies (120 min period to a 720 min period). The
PI closed-loop amplitude ratio for the diagonal PI control
0.7 structure is shown to be 23 times larger than that for
0 200 400 600 NLMPC over the sensitive range. At low periods (high
Time (min)
frequencies), the inertia of the process damps out the effect
Fig. 7. The manipulated variable prole for overhead for an unmeasured of the sinusoidal disturbance. At large periods (low
feed composition upset for the [L/D,V] conguration. frequencies), the feedback controllers have ample time to
absorb the disturbances by feedback. Fig. 9 also indicates
that the 5% process/model mismatch does not signicantly
in use. A sinusoidal disturbance to the recycle feed rate was affect the control performance for the NLMPC controller.
applied with an amplitude of 10% of the base case recycle These results are consistent with the previous results for
feed rate. The closed-loop amplitude ratio for the overhead the comparisons between PI and NLMPC control perfor-
impurity was calculated by taking the ratio of the mance for the setpoint tracking as well as unmeasured
amplitude of overhead impurity resulting from the disturbance rejection, but more quantitatively dene the
ARTICLE IN PRESS
R. Kawathekar, J.B. Riggs / Control Engineering Practice 15 (2007) 231239 239

control performance differences between PI and NLMPC Finlayson, B. A. (1980). Nonlinear analysis in chemical engineering. New
for this case. York, NY: McGraw-Hill.
Gill, P. E., Murray, W., & Saunders, M. A. (1988). Users Guide for
SNOPT 5.3/6.0: A Fortran Package for Large Scale Nonlinear
7. Conclusions Programming, Systems Optimization Laboratory, Stanford University,
CA.
As expected, NLMPC was found to provide signicantly Grosser, J. H., Doherty, M. F., & Malone, M. F. (1987). Modeling of
better control performance than PI controls, but this work reactive distillation systems. Industrial and Engineering Chemistry
quanties the advantage as a factor of 23 reduction in Research, 26, 983989.
Gruner, S., Mohl, K.-D., Kienle, A., Gilles, E. D., Fernholz, G., &
variability. The advantage of the NLMPC controller comes Friedrich, M. (2003). Nonlinear control of a reactive distillation
from faster closed-loop dynamic performance compared to column. Control Engineering Practice, 11(8), 915925.
the PI controller resulting from using a nonlinear dynamic Jacobs, R., & Krishna, R. (1993). Multiple solutions in reactive distillation
model of the reactive distillation column considered here. for methyl tert-butyl ether synthesis. Industrial and Engineering
In addition, NLMPC was found not to be particularly Chemistry Research, 32, 17061709.
Jones, D. I., & Finch, J. W. (1984). Comparison of optimization
sensitive to process/model mismatch. Even though the algorithms. International Journal of Control, 40, 747.
reactive distillation column considered in this study was Khaledi, R., & Young, B. R. (2005). Modeling and model predictive
shown to be extremely nonlinear, the PI control was able to control of composition and conversion in an ETBE reactive
control the process reasonably well. distillation column. Industrial and Engineering Chemistry Research,
44, 31343145.
Krishnamurthy, R., & Taylor, R. (1985). A nonequilibrium stage model of
Acknowledgements multicomponent separation processes. AIChE Journal, 32, 449465.
Kumar, A., & Daoutidis, P. (1999). Modeling, analysis and control of
This work was supported by the member companies of ethylene glycol reactive distillation column. AIChE Journal, 32,
the Texas Tech Process Control and Optimization Con- 449465.
sortium. The authors would also like to acknowledge Marlin, T. E. (1995). Process control: Design processes and control systems
for dynamic performance. New York: McGraw Hill.
assistance from Charles R. Cutler, Cutler-Technologies. Meadows, E. S., & Rawlings, J. B. (1997). Model predictive control. In M.
Henson, & D. Seaborg (Eds.), Nonlinear process control. Prentice-Hall
References PTR.
Riggs, J. B. (2001). Chemical process control (2nd ed.). Lubbock, TX:
Abufares, A. A., & Douglas, P. L. (1995). Mathematical modeling and Ferret Publishing.
simulation of an MTBE catalytic distillation process using SPEEDUP Ruiz, C. A., Basualdo, M. S., & Scenna, N. J. (1995). Reactive distillation
and aspen plus. Chemical Engineering Research and Design, Transac- dynamic simulation. Chemical Engineering Research and Design,
tions of the Institute of Chemical Engineers Part A, 73, 312. Transaction of the Institution of Chemical Engineering, Part A, 73,
Al-Arfaj, M. A., & Luyben, W. L. (2000). Comparative control of ideal 363378.
and methyl acetate reactive distillation. Chemical Engineering Science, Seferlis, P., & Grievink, J. (2001). Optimal design and sensitivity analysis
57, 50395050. of reactive distillation units using collocation models. Industrial and
Aljeski, K., & Duprat, F. (1996). Dynamic simulation of the multi- Engineering Chemistry Research, 40, 16731685.
component reactive distillation. Chemical Engineering Science, 51, Sneesby, M. G., Tade, M. O., Datta, R., & Smith, T. N. (1997). ETBE
42374252. synthesis via reactive distillation. 2. Dynamic simulation and control
Bock, H., & Wozny, G. (1997). Analysis of distillation and reaction rate in aspects. Industrial and Engineering Chemistry Research, 36, 18701881.
reactive distillation. Distillation and absorption 97. Institute of Sneesby, M. G., Tade, M. O., & Smith, T. N. (1998). Multiplicity and
Chemical Engineering, Symposium Series, 142, 553564. pseudo-multiplicity in MTBE and ETBE reactive distillation. Chemical
Boston, J. F., & Sullivan, S. L. (1974). A new class of solution methods for Engineering and Research Design, Transactions of the Institution of
multicomponent, multistage separation processes. CanadianJournal of Chemical Engineers, Part A, 76, 525531.
Chemical Engineering, 52, 5263. Soave, G. (1992). Chemical and Engineering Science, 27, 1197.
Chang, Y. A., & Seader, J. D. (1988). Simulation of continuous reactive Suzuki, I., Yagi, H., Komatsu, H., & Hirata, M. (1971). Calculation of
distillation by homotopyContinuation method. Computers and multicomponent distillation accompanied by chemical reaction.
Chemical Engineering, 12(12), 12431255. Journal of Chemical Engineering of Japan, 4, 2633.
Chen, L., Hontoir, Y., Huang, D., Zhang, J., & Julian Morris, A. J. A. Tyreus, B. D., & Luyben, W. L. (1992). Tuning PI controllers for
(2004). Combining rst principles with black-box techniques for integrator/deadtime processes. I&ECR, 31, 26252635.
reaction systems. Control Engineering Practice, 12(7), 819826. Volker, M., Sonntag, C., & Engell, S. (2006). Control of integrated
Doherty, M. F., & Buzad, G. (1992). Reactive distillation by design. processes: A case study on reactive distillation in a medium-scale pilot
Chemical Engineering Research and Design, Transactions of the plant, Control Engineering Practice, in press, Corrected Proof,
Institute of Chemical Engineers, Part A, 70, 448458. Available online 5 June 2006, doi:10.1016/j.conengprac.2006.03.00.

Você também pode gostar