Control Engineering Practice 19 (2011) 1137–1146

Contents lists available at ScienceDirect

Control Engineering Practice

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

Optimizing model predictive control of an industrial distillation column

C.R. Porfı́rio a, D. Odloak b,n
a
Petrobras, Refinery of Cubata~ o, Pc- a Stênio Caio de Albuquerque Lima, 1, 11555-900 Cubatao,
~ Brazil
b ~ Paulo, PO Box 61548, 05424-970 Sa~ o Paulo, Brazil
Department of Chemical Engineering, University of Sao

a r t i c l e i n f o abstract

Article history: The main scope of this work is the implementation of an MPC that integrates the control and the
Received 10 December 2010 economic optimization of the system. The two problems are solved simultaneously through the
Accepted 2 June 2011 modification of the control cost function that includes an additional term related to the economic
Available online 25 June 2011
objective. The optimizing MPC is based on a quadratic program (QP) as the conventional MPC and can
Keywords: be solved with the available QP solvers. The method was implemented in an industrial distillation
Real-time optimization system, and the results show that the approach is efficient and can be used, in several practical cases.
Model predictive control & 2011 Elsevier Ltd. All rights reserved.
Distillation control
Integration of control and optimization
Process optimization

1. Introduction controller. Rao and Rawlings (1999) discuss the implementation

of the optimal steady-state with linear model predictive
In systems subject to frequent disturbances such as changes in controllers.
the feed composition, feed flow rate, amount of available energy The multilayer approach may not be satisfactory if frequent
and changes in the market values of feed and products, real-time disturbances drive the system to optimum operating points
optimization (RTO) may produce significant economic benefits as corresponding to different sets of active constraints as in the case
RTO deals with methods of how to best operate a process system in which the throughput bottleneck is frequently changed. For
under frequent disturbances (Adetola & Guay, 2010). RTO appli- this kind of system, it could be more suitable to use dynamic real-
cations require nonlinear models to represent the process system, time optimization (DRTO) with a nonlinear dynamic model or
and, in the conventional approach, RTO also involves data recon- nonlinear MPC with economic objective (Adetola & Guay, 2010;
ciliation and estimation of suitable parameters of the process Hosen, Hussain, & Mjalli, 2011; Kadam et al., 2003; Kawathekar &
system, such as efficiencies of equilibrium stages or activity of Riggs, 2007; Tosukhowong, Lee, Lee, & Lu, 2004). However, the
chemical reaction catalysts (Marlin & Hrymak, 1997). Particularly, online dynamic optimization of complex processes seems not yet
in the oil refining and petrochemical systems, real-time economic achievable with the available computing power (Lu, 2003).
optimization has proven to be very effective and the resulting Another possible disadvantage of the multilayer approach is
tools and algorithms have been widely accepted in industry that the time interval of the RTO cycle must be large enough for
(Georgiou, Taylor, Galloway, Casey, & Sapre, 1997; Rotava & the plant to reach the new optimum steady-state. The conse-
Zanin, 2005). Usually, a three-layer structure is adopted, where quence is that typical sampling times for the RTO layer in a
in the upper layer RTO optimizes the process operating conditions multilayer approach may be of several hours (Engell, 2007) while
at steady-state and the optimum operating point is characterized the MPC layer needs a much shorter sampling time of a few
through set-points to the linear MPC (Ying & Joseph, 1999). The minutes to perform set-point tracking or disturbance rejection.
problem that is solved in the RTO layer considers the equipment To overcome the possible disadvantages of the multilayer
constraints and product specifications. Optimum values for the structure of real-time optimization and control, De Gouvêa and
inputs and/or outputs are computed and passed to an intermedi- Odloak (1998) proposed an approach where steady-state RTO and
ary layer where, based on the estimation of the disturbances, linear MPC are performed simultaneously in a single layer. The
feasible targets for the inputs and outputs are re-computed and strategy was implemented in a real Fluid Catalytic Cracking (FCC)
fed to the dynamic layer that contains the model predictive plant by Zanin, De Gouvêa, and Odloak (2002). In their approach,
the optimizing MPC is formulated to perform both functions by
adding an economic objective term to the standard MPC objective
n
Corresponding author. Tel.: þ551130912261; fax: þ 55113813 2383. function. It was observed that the one-layer approach could
E-mail address: [email protected] (D. Odloak). respond to frequent disturbances that push the system to new

0967-0661/$ - see front matter & 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.conengprac.2011.06.003
1138 C.R. Porfı́rio, D. Odloak / Control Engineering Practice 19 (2011) 1137–1146

plant operating conditions faster than the multilayer approach with To simplify this problem, suppose that the economic objective
a possible advantage in terms of the economic benefit. However, a function is designated F(us,ys) where us and ys define the system
disadvantage of this method is that the inclusion of the economic steady-state, and this economic function is a concave function
objective turns the optimization problem, which is solved in the whose maximum defines the optimum operating point of the
MPC into a Nonlinear Programming (NLP) problem, where the process system. In this case, the optimum unconstrained point
objective function is nonlinear and there are nonlinear constraints corresponds to a steady-state where Bus ¼0 with Bus ¼ dF=dus . In
corresponding to the steady-state model of the process system. this case, De Souza et al. (2010) represent the economic term in
Consequently, the expected computational effort required to com- the cost function as feco ¼ Bus T Bus
pute the control sequence can be much higher than in the conven- Observe that if the vector of control actions is changed from us
tional MPC. In order to preserve the structure of the conventional to us þ Dus, the first order approximation to the gradient of the
MPC, which is usually translated into a Quadratic Programming (QP) economic function can be written as
problem, De Souza, Odloak, and Zanin (2010) propose a simplified
Bus þ Dus ¼ d þGDus ð2Þ
version of the one-layer optimizing MPC. In their approach, the
objective function of the MPC controller is also modified to include a The calculation of d and G involves the rigorous steady-state
term related to the economic objective, but the economic informa- model as shown in De Souza et al. (2010). In Eq. (2),
tion is restricted to an estimation of the gradient of the economic Dus ¼u(k þm 19k) u(k 1) is the total move of the input vector.
objective. For those process systems, where the economic objective As a result of these assumptions feco can be approximated by a
is a convex function to be minimized or a concave function to quadratic function of the control move.
be maximized, searching for the optimum operating point is equi- To calculate the output prediction, the state space model
valent to searching for an operating point in which the gradient of adopted here is defined as follows:
the objective function is equal to zero. If the optimum point is
xðkþ 1Þ ¼ AxðkÞ þ BDuðkÞ
constrained, then the reduced gradient should be zeroed. Following
this strategy, the optimizing controller is obtained from the solution yðkÞ ¼ CxðkÞ ð3Þ
to a QP, where the process outputs are controlled by zones and the The model defined in (3) is written in the incremental form in
direction of the gradient of the economic function is dynamically the input as this is a suitable form to remove offset in the output
corrected considering the predicted steady-state and the active predictions (Maciejowski, 2002). Consequently, the state matrix A
constraints in the system outputs. has some eigenvalues equal to 1. This means that the infinite sum
In this work, the approach of De Souza et al. (2010) is extended of the cost defined in (1) is not bounded. To circumvent this
by considering the infinite horizon MPC, which has nominal problem, consider the eigenvalue-eigenvector expansion of the
stability and the new controller is implemented in a real indus- state matrix: A¼V D V  1. When the eigenvalues, l1, y, lnA, of the
trial distillation column in an oil refinery in Brazil. In the next process system are non-repeated, matrix D has the following form
section, the optimizing MPC considered here is developed. Then, D ¼diag(l1, y, lnA, 1, y, 1).
the distillation system is presented and the rigorous model that is Then, if one defines N ¼ ½0 Iny  A RnyðnA þ nyÞ and matrix V is
used to represent the multi-component distillation system is decomposed as V¼[Vs Vi] where the number of columns of Vs
introduced. Next, the results of the method obtained in the real and Vi correspond to the number of stable and integrating poles,
plant are presented and discussed. Finally, the paper is concluded. respectively, it can be shown that the infinite sum defined in (1)
will be bounded if

2. Model predictive control with economic objective

CVi NV 1 xðk þm9kÞysp,k dy,k ¼ 0 ð4Þ
Real-time optimization and model predictive control may be If (4) is satisfied by the control sequence, then the control cost
implemented in one single layer as proposed in Zanin et al. defined in (1) can be written as follows:
(2002). In this case, the cost function of MPC is extended with
an economic term and, for instance, if an infinite horizon cost is X
m
considered, then the controller cost may be written as follows: Jk ¼ ðyðk þj9kÞysp,k dy,k ÞT Qy ðyðk þj9kÞysp,k dy,k Þ
j¼0
X
1
Vk ¼ ðyðk þ j9kÞysp,k dy,k ÞT Qy ðyðkþ j9kÞysp,k dy,k Þ T
þ xðk þ m9kÞT Pxðk þm9kÞ þ W ðd þGDus Þ ðd þ GDus Þ
j¼0
X
m 1
X
m1
þ Duðk þj9kÞT RDuðkþ j9kÞ þ dTy,k Sy dy,k
þ Wfeco þ Duðkþ j9kÞT RDuðkþ j9kÞ þ dTy,k Sy dy,k ð1Þ
j¼0
j¼0

where y(kþj9k) and Du(kþj9k)¼u(kþj9k)  u(kþj  19k) are the where P is the terminal weight that can be computed as described
predicted output and the control move, respectively, computed at in Odloak (2004).
time k and corresponding to time step kþj, m is the control or Then, the economic optimizing MPC proposed here results
input horizon, Qy, R and Sy are positive weighting matrices of from the solution to the following problem
appropriate dimensions, ysp,k is the output set-point, which can be Problem P
an additional decision variable of the zone control strategy and dy min Jk
is a slack variable that guarantees that the control problem is Duðk9kÞ,. . ., Duðk þ m19kÞ
always feasible. The economic term feco, in general, is a nonlinear ysp,k , dy,k
function associated with the predicted steady-state of the system
and W is the tuning weight of this term. Then, because of the subject to (4) and
presence of feco, the control objective of the economic MPC as
Dumax r Duðk þj9kÞ r Dumax , j ¼ 0,. . .,m1
defined in (1) is not a quadratic function of the decision variables
of the optimization problem that defines the controller. Conse- umin r uðk þj9kÞ r umax , j ¼ 0,. . .,m1
quently, the control problem becomes a NLP, which may result in
an unacceptable computer demand. ymin r ysp,k rymax
 C.R. Porfı́rio, D. Odloak / Control Engineering Practice 19 (2011) 1137–1146 1139

In the unconstrained economic optimization, the operating 3. The toluene distillation system
point where the gradient B is equal to zero corresponds to a local
maximum (when G o0) or a local minimum (when G 40) of the The distillation system considered in this work is schematically
economic function. However, when constraints of the control represented in Fig. 1. This system is located at the Petrobras refinery
problem are active, the optimum corresponds to the point where of Cubata~ o, Brazil, where a column with 44 real stages produces at
the reduced gradient of the economic function is equal to zero. the top an aromatic rich stream with at least 99.7% (vol) of toluene.
The reduced gradient is obtained through the projection of the The bottom stream is a mixture of xylene and heavier components
gradient on the tangent space of the active constraints. For with a maximum of 1% (vol) of toluene. The feed stream is
instance, if the prediction of the output yi at a future time instant introduced at stage 23 and, typically, it is a mixture of hydrocarbon
t is such that yi ðk þ t9kÞ Zymax,i or yi ðkþ t9kÞ rymin,i , then, Problem components as represented in Table 1. The control loops of the
P should include the following additional constraint: regulatory level of this industrial system are also represented in
Fig. 1, where there are two control loops related to the composition
X
nu
Dyi ¼ Kpi,j Duj ¼ 0 of the top and bottom streams. The TDC (Temperature Difference)
j¼1 control loop, which controls the difference between the tempera-
tures at stages 4 and 14 and the FFC (Flow Ratio) control loop, which
where controls the liquid/vapor flow ratio at the top stage of the column.
Duj ¼ uj ðkþ m19kÞuj ðk1Þ

Problem P can be solved with any available quadratic program- 3.1. Steady-state rigorous simulation of the distillation column
ming (QP) solver as in a conventional MPC where the computer
demand, to calculate the sequence of control actions, is usually In the optimizing MPC, a rigorous steady-state simulation
small. The main difference between the proposed controller and model is included. This model is necessary to compute vector d
the conventional MPC is that in the new controller, at each time and matrix G in Eq. (2) that defines the gradient of the economic
step, one needs to know parameters d and G that are associated objective. Fig. 2 depicts all the mass and energy streams involved
with the economic objective. The calculation of these parameters in the rigorous distillation model, which is used in the controller
should be done in a separate module that includes the rigorous implemented here. For any equilibrium stage j, Fj is the molar feed
steady-state model of the process. The computer effort to obtain flow rate, Uj and Wj are the molar flow rates of the liquid and
these parameters may be high mainly for complex systems. But, vapor streams that are withdrawn from stage j, Vj is the vapor
eventually, if the calculation of d and G within the same sampling flow rate that leaves stage j and is fed into stage j  1, Lj is the
period as the controller shows to be impossible because of liquid flow rate that leaves stage j and is fed into stage j þ1, Qj is
computer limitation, a different time period can be used to update the heat duty introduced in stage j. N is the number of ideal
the economic parameters. For instance the economic parameters equilibrium stages of the column.
can be updated with a time period that is a multiple of the The rigorous model of the distillation column studied here is
sampling time of the controller. composed of the following equations.

Fig. 1. Schematic representation of the distillation system.

1140 C.R. Porfı́rio, D. Odloak / Control Engineering Practice 19 (2011) 1137–1146

Table 1
Composition of the feed stream.

Component Mass fraction

Cyclopentane 2.07  10–7

Methylcyclopentane 3.09  10–7
Cyclohexane 5.25  10–8
3-Methylhexane 2.68  10–7
2–3-3-Trimethyl-hexane 1.81  10–4
Methylcyclohexane 1.17  10–4
1-Octene 1.18  10–4
3-Methyloctane 1.48  10–4
Cyclodecane 2.08  10–4
n-Decane 1.19  10–4
Benzene 6.21  10–5
Toluene 0.6217
m-Xilene 0.1773
o-Xilene 8.29  10–2
p-Xilene 4.64  10–2
Ethyl-benzene 3.17  10–2
n-Propyl-benzene 1.57  10–3
1-Methyl-3-ethylbenzene 1.06  10–2
1–2-4-Trimethylbenzene 2.26  10–2
o-Cymene 9.72  10–4
n-Buthylbenzene 3.09  10–4
o-Diethylbenzene 3.07  10–3

3.2. Molar balance of component i at stage j

Lj1 xi,j1 þ Vj þ 1 yi,j1 þFj zi,j ðLj þ Uj Þxi,j ðVj þWj Þyi,j ¼ 0 ð5Þ
where xi,j, yi,j and zi,j are the molar composition of component i at the
liquid stream, vapor stream and feed stream of stage j, respectively.

3.3. Vapor–liquid equilibrium

yi,j ¼ ki,j xi,j ð6Þ

where ki,j is the equibrium constant of component i at the stage j

3.4. Stoichiometric relations

Nc
X Nc
X
yi,j ¼ 1, xi,j ¼ 1 ð7Þ
Fig. 2. Variables of the distillation model.
i i

where Nc is the number of pure chemical components in the With the above specifications, considering all the variables of
distillation column. the distillation system and Eqs. (5)–(8) written for all the
equilibrium stages and components, it can be shown that two
3.5. Energy balance at stage j degrees of freedom are still remaining to be defined in order to
completely characterize the column steady-state. So, for instance,
Lj1 hj1 þVj þ 1 Hj þ 1 þFj HFj þQj ðLj þUj Þhj ðVj þ Wj ÞHj ¼ 0 ð8Þ
one can assume that the reflux flow rate and the temperature of
where hj, Hj and HFj are the molar enthalpy of the liquid, vapor one of the equilibrium stages are fixed to compute the operating
and feed streams, respectively, at stage j. point of the distillation column through the solution of
Eqs. (5)–(8) should be written for each of the ideal stages of the Eqs. (5)–(8).
distillation column and for each of the 22 components represented From the solution to the model described above, the following
in Table 1 that corresponds to the typical hydrocarbon mixture, process variables, which are important to the operation and
which is fed to the distillation column. This model was fitted to optimization of the distillation system, can be obtained:
empirical data collected from the true plant and the optimum fitting
was obtained with N¼39 and NF ¼23 (feed stage). (a) The heat duties of the condenser and reboiler.
The model defined through Eqs. (5)–(8) for the system con- (b) The temperature profile along the column.
sidered here has nearly 1800 nonlinear algebraic equations. (c) Composition of the liquid and vapor streams at each equili-
To solve this model, several variables must be specified. For the brium stage of the column, as well as the compositions of the
distillation column considered here, the following variables are bottom and top product streams.
assumed to be known: (d) Molar flow rate of the liquid and vapor streams at each
equilibrium stage.
(a) The composition of the feed stream, which is provided by an Other variables that are important to the control strategy
online analyzer. adopted in this work are also computed: the liquid to vapor flow
(b) The flow rate of the feed stream, which is read from the plant. ratio (L/V) at stage 1 and the difference between the temperatures
(c) The column pressure that is controlled at the regulatory level at stages 4 and 14.
and the set-point to the pressure loop is available in the DCS The available degrees of freedom of the distillation column can
system. be used to define fixed values, for instance, for the temperature
 C.R. Porfı́rio, D. Odloak / Control Engineering Practice 19 (2011) 1137–1146 1141

difference between stages 4 and 14 and the liquid/vapor ratio at conventional MPC, but the cost function is extended through the
stage 1. However, in this case, an iterative procedure must be inclusion of an economic term as defined in Eq. (2). Thus, the
adopted as these two variables do not appear explicitly as controller implemented here results from the solution to Problem
variables of the distillation column model. One can start, for P where the outputs are controlled inside zones, the output
instance, with arbitrary values for the reflux flow rate and the prediction horizon is infinite and the matrix weight S of the
reboiler heat duty and iterate until the desired values of the vector of slack variables is large enough to eliminate offset in the
temperature difference and liquid to vapor ratio are achieved. controlled outputs.

4.2. Experimental results

4. Implementation of the proposed MPC at the industrial
distillation column To illustrate the performance of the implemented version of
the optimizing MPC, a series of experiments were performed in
The industrial system where the controller with economic the industrial system. In the first experiment reported here, the
objective was implemented has been operating with a conventional adopted economic objective is to minimize the flow rate of steam
MPC for several years. The controlled outputs of this controller are: that is introduced in the reboiler. Then, following the notation of
Section 2, in this case, the economic objective will be defined as
(a) (y1) The contents of aromatics with more than eight atoms of follows:
carbon (C8 þ ) in the top stream of the column, which is the
commercial toluene produced in the system. There is a QN
Fðus ,ys Þ ¼ ð9Þ
specification on the maximum value of this concentration ls
that should be kept below 0.3%. where QN is the reboiler heat duty computed through the solution
(b) (y2) The contents of toluene in the bottom stream of the to the model defined in Eqs. (5)–(8) and ls is the heat of
column should be controlled below a maximum value to vaporization of water.
prevent the loss of toluene to the heavy aromatics stream that In the control structure considered for this distillation column,
is produced in the bottom of the column and that has a low the economic objective function defined in (9) will depend only
market value. on the manipulated inputs (u1 and u2). This is so, because, once
the difference of temperature along the distillation column and
The existing MPC controller of the distillation column manip- the ratio between the liquid and vapor flow rates in the column
ulates the set points to two control loops of the regulatory level is fixed, all the remaining variables of the distillation model
that lies in the DCS system: (including the reboiler heat duty) can be uniquely obtained. This
means that in Eq. (2), which shows how the gradient of this
(a) (u1) The set point of the TDC that controls the difference of economic function is computed, one can consider d ¼ @F=@us and
temperatures between stages 4 and 14. As shown in Fig. 1, the G ¼ @2 F=@u2s that simplifies the computation of the gradient of the
TDC cascades the controller of the reflux flow rate. economic objective as follows:
(b) (u2) The set point of the FFC that controls the ratio between   " 2 #
the liquid flow rate and the vapor flow rate at stage 1 of the @F @ F
zu þ Du ¼ þ Du
distillation column. As shown in Fig. 1, the FFC cascades the @u @u2
controller of the flow rate of steam to the reboiler. For the first experiment reported here, the tuning parameters
and constraints of the optimizing controller are initially the
The existing MPC controller also considers in a feedforward following: control horizon m¼1; output weight Qy ¼ diagð3 2Þ;
scheme the feed flow rate as a measured disturbance (d). For a
input weight R ¼ diagð1 0:95Þ; Sy ¼ diagð1 1Þ  103 ; input con-
sampling period equal to 1 min, the linear model considered in      
0:005 4:50 3:00
the existing conventional MPC is shown in Table 2. This controller straints Dumax ¼ , umax ¼ , umin ¼ ; weight
was designed mainly for regulatory purposes and did not have 0:001 0:69 0:66
 
any real-time optimization facility. The model represented in 0:2
of the economic term W¼ 0.01; and output zones ymax ¼ ,
Table 2 was obtained through step tests performed several years 0:5
 
ago at the real plant. Since the existing conventional MPC has a 0:1
ymin ¼ .
good performance record, this model will also be adopted to 0:0
represent the system dynamics in the controller developed here. The tuning of the proposed MPC with economic objective should
It is interesting to note that all the transfer function models follow the same rules as the conventional MPC, mainly for the
represented in Table 2 have time constants equal to 18 min, control horizon m and weights R and Qy. The slack variable weight Sy
which is rather small for an industrial distillation column. should be high enough to satisfy the condition presented in Odloak
(2004) and, as long as this condition is satisfied, S has little effect on
4.1. The optimizing MPC the controller performance. There is no explicit rule to define the
economic weight W that has to be obtained experimentally by
The optimizing controller of the industrial distillation system trying increasing values of W until an acceptable economic perfor-
has the same controlled and manipulated variables as the existing mance is obtained while the closed loop stability is preserved.
Although, one can find in the control literature methods to
Table 2 access the performance of the conventional MPC with known
Transfer function model of the distillation column. output set points (Schafer & Cinar, 2004; Yu & Qin, 2008a, 2008b,
2009) or optimum output targets defined by the RTO layer in
u1 u2 D
a multilayer control structure (Xu, Huang, & Akande, 2007;
y1 0:005404 2 0:06485 2 0:0001481 2 Zhao, Zhao, Su, & Huang 2009), those methods cannot be applied
z z z
z0:946 z0:946 z0:946 here because, as the RTO and the MPC are merged into a single
y2 0:005945 2 0:03242 2 5:096  105 2
z0:946
z
z0:946
z z algorithm, the optimum set points or targets are not known a
z0:946
priori and it becomes not trivial for the proposed controller to
1142 C.R. Porfı́rio, D. Odloak / Control Engineering Practice 19 (2011) 1137–1146

define a benchmark controller as the minimum variance (MVC) or of maintaining the output y1 on the minimum bound while
the linear quadratic Gaussian (LQG) for the conventional MPC. So, output y2 is controlled inside its control zone defined through
here it is tried to evaluate the economic performance of the the dashed lines in Fig. 3. From Fig. 4, one can observe that the
proposed controller by simply comparing the performance of controller tends to reduce the difference of temperatures along
the controller with the economic term to the performance of the column while the liquid/vapor ratio is kept constant as it was
the conventional MPC without the economic term. already at its minimum bound. Fig. 4 shows the true process
Fig. 3 shows the behavior of the controlled outputs, the variable of the TDC loop instead of the set point to this controller,
concentration of C8 þ in the distillate and the concentration of which is the variable that is manipulated by the MPC controller. It
toluene in the bottom stream with the optimizing MPC in closed also shows the set-point to the FFC that controls the liquid/vapor
loop along a time period of 5 h. These data were collected from ratio at stage 1. It is clear that the optimizing controller tends to
the process data base of the toluene distillation system. The solid push both inputs to their minimum bounds. But, the controller
lines represent the process variables, while the dashed lines has to take some control actions to compensate for the changes in
define the bounds or control zones of these variables. Fig. 4 shows the feed flow rate that is represented in Fig. 4, where one can see
the behavior of the two inputs, the temperature difference that, although the mean of this important disturbance is kept
between stages 4 and 14 and the liquid to vapor flow ratio, almost constant there are some significant oscillations, that is
respectively. Again the solid lines represent the process variables reflected in the controlled outputs. Fig. 5 shows that the con-
and the dashed lines are their bounds. Fig. 5 shows the economic troller is quite efficient in achieving a better economic goal, by
objective that in this case is the steam flow rate to the reboiler. reducing smoothly and continuously the steam to the reboiler
At time t ¼0.5 h, the economic term of the controller is switched until the optimum point is reached after about 4 h. One can
on and the minimum bound of output y1 is increased from 0.1 to observe that, as only the minimum bound of y1 is reached, the
0.15. This was done to illustrate the activation of this bound and optimum operating point corresponds to a point where the
to show how the controller performs in the zone control strategy. projection of the gradient into the plane orthogonal to constraint
As shown in Fig. 3, this new bound tends to be reached when the y1 ¼0.15 is zeroed. A more aggressive controller could be obtained
controller tries to improve the economic objective (minimize the by increasing the economic weight W, but the plant operators
steam consumption in the reboiler), but the controller is capable prefer a more cautious tuning.

Fig. 3. Minimizing the steam to reboiler, controlled outputs: C8 þ (mass %) in the distillate flow (y1) and toluene (mass %) in the bottom stream (y2).
 C.R. Porfı́rio, D. Odloak / Control Engineering Practice 19 (2011) 1137–1146 1143

Fig. 5. Economic objective of the controller (steam flow rate to the reboiler, kg/h).

Fig. 4. Minimizing steam, inputs: temperature difference (C) between stages

17 and 4 (u1), liquid/vapor ratio in the column (u2) and disturbance d (feed flow Fig. 6. Maximizing toluene, controlled outputs: C8 þ (mass %) in the distillate flow (y1)
rate—m3/day). and toluene (mass %) in the bottom stream (y2).

In the second experiment performed in the industrial system,

the economic function is the flow rate of toluene (that is the top different of the operating condition of the first experiment. Fig. 6
stream of the distillation column and represented as U1 in Fig. 2), shows the controlled outputs of the distillation system, which are
which has to be maximized. The tuning parameters of the kept inside their control zones except for output y1 that reaches
optimizing MPC are the same as in the first case, except the input its maximum bound for a short period of time. Fig. 7 shows the
constraints that are as follows: behavior of the system inputs when the controller is switched on.
   
6:0 4:00 It is shown that input u2 tends to saturate at its minimum bound
umax ¼ , umin ¼ and the output zones that are as
0:69 0:65 while input u1 remains within the selected bounds. Fig. 7 also
    shows the behavior of the feed flow rate, which is the measured
0:3 0:1
follows: ymax ¼ , ymin ¼ . disturbance. It is clear that in this experiment this disturbance is
1:0 0:0
almost constant and has little effect on the control and optimiza-
Figs. 6–8 show the responses of the system for a period of tion of the distillation column. Fig. 8 shows the economic function
4.78 h. The optimizing MPC is switched on at time 1.3 h with the that is significantly increased and reaches a new steady-state
objective of maximizing the flow rate of the toluene stream. The about 1 h after the controller is started. In this case, the optimal
system starts from an initial operating condition that is quite economic point corresponds to an operating point where the
1144 C.R. Porfı́rio, D. Odloak / Control Engineering Practice 19 (2011) 1137–1146

Fig. 8. Economic objective (flow rate of the toluene stream, m3/day).

Fig. 7. Minimizing toluene, inputs: temperature difference (C) between stages

17 and 4 (u1), liquid/vapor ratio in the column (u2) and disturbance d (feed flow
rate—m3/day).

projection of the gradient of the economic function into the plane Fig. 9. Maximizing profit, controlled outputs: C8 þ (mass %) in the distillate flow
orthogonal to constraint u1 ¼0.65 is zeroed. (y1) and toluene (mass %) in the bottom stream (y2).
In the third experiment performed in the industrial system,
the economic objective function of the controller is represented
by the following function: of the distillation column, one should also include the cost of the
feed stream, which is not a commercial product and does not have
Fðus ,ys Þ ¼ pT U1 þpX UN pS QN ð10Þ
a known commercial value. However, as the feed flow rate was
where U1 is the flow rate (m3/day) of the product stream at the kept constant along the experiment, one can consider that any
top of the column, which is sold as toluene with a commercial increase observed in the function defined in (10) will correspond
value pT ¼300 US$/m3. UN is the flow rate (m3/day) of the product to the increase in the profit resulting from the implementation of
stream at the bottom of the column, which is sold as xylene the optimizing controller.
with a commercial value pTX ¼200 US$/m3. QN is the flow rate The tuning parameters of the optimizing MPC are the same as
(ton/day) of the steam to the reboiler, with cost pS ¼15 US$/ton. in the two cases before, except the maximum output bound that
 
The function defined in Eq. (10) corresponds to the distillation 0:27
is now ymax ¼ .
column income in US$/day. To compute the profit of the operation 1:0
 C.R. Porfı́rio, D. Odloak / Control Engineering Practice 19 (2011) 1137–1146 1145

Fig. 11. Economic objective function (profit, US$/day).

forces the system to a new steady-state with an increase in the

operating profit of about US$ 250/day, which corresponds to an
integrated profit near to US$ 100 thousand/year that is quite
substantial for a system of a relatively small capacity.

5. Conclusion

This paper presents a predictive controller that integrates the

control and the economic tasks in a single optimization problem. To
improve the nominal stability of the closed loop system, the
controller is developed assuming an infinite output horizon. The
method was implemented in an industrial toluene distillation column
in the Petrobras refinery at Cubata~ o (Brazil). In the case presented
here, a rigorous distillation model is included in the controller and it
is used in the computation of the gradient of the economic objective.
An advantage of the proposed approach is that the optimizing
predictive control problem is still a quadratic program that can be
easily solved using the available QP solvers. Although the method is
restricted to the case were the economic objective function to be
maximized is a concave function of the manipulated inputs, practical
results reported here, show that the approach is efficient and robust
for several economic objectives of the distillation system. The
controller remains in continuous operation since its implementation
without any reported operation problem.

References

Fig. 10. Maximizing profit, inputs: temperature difference (C) between stages 17 Adetola, V., & Guay, M. (2010). Integration of real time optimization and model
and 4 (u1), liquid/vapor ratio in the column (u2) and disturbance d (feed flow predictive control. Journal of Process Control, 20, 125–133.
rate—m3/day). Hosen, M. A., Hussain, M. A., & Mjalli, F. S. (2011). Control of polystyrene batch
reactors using neural network based model predictive control (NNMPC): An
experimental investigation. Control Engineering Practice, 19, 454–467.
Figs. 9–11 show the behavior of the distillation column along a De Gouvêa, M. T., & Odloak, D. (1998). One-layer real time optimization of
time period of 5 h when the optimizing MPC is switched on at LPG production in the FCC unit: Procedure, advantages and disadvantages.
time t ¼0.5 h. Figs. 9 and 10 show that the controller tends to Computers & Chemical Engineering, 22, S191–S198.
De Souza, G., Odloak, D., & Zanin, A. C. (2010). Real time optimization (RTO) with
drive output y1 towards its maximum bound and to drive input u2 model predictive control (MPC). Computers & Chemical Engineering, 34,
to its minimum bound, and these two constraints remain active 1999–2006.
along all the experiment. This means that, in this particular case, Engell, S. (2007). Feedback control for optimal process operation. Journal of Process
Control, 17, 203–219.
the reduced gradient of the economic objective is zeroed at the Georgiou, A., Taylor, P., Galloway, R., Casey, L., Sapre, A. (1997). Plantwide closed
optimum point because of the presence of two active constraints loop real time optimization and advanced control of ethylene plant (CLRTO)
(y1 ¼y1,max, u2 ¼u2,min) that consume the available degrees of improves plant profitability and operability. In Proceedings of NPRA computer
conference, National Petroleum Refiners Association, paper CC 97-139, New
freedom of the system. Figs. 9 and 10 also show that output y2
Orleans.
and input u1, respectively, are kept within their bounds and do Kadam, J., Marquardt, W., Schlegel, M., Backx, T., Bosgra, O., & Brouwer, P. J. (2003).
not tend to activate any other constraint. Fig. 10 shows that in Towards integrated dynamic real-time optimization and control of industrial
this experiment the distillation column was very little disturbed processes. In Proceedings foundations of computer-aided process operations
(FOCAPO2003), pp. 593–596.
by the changes in the feed flow rate that was kept almost fixed. Kawathekar, R., & Riggs, J. B. (2007). Nonlinear model predictive control of a
One can see from Fig. 11 that after about 0.5 h, the controller reactive distillation column. Control Engineering Practice, 15, 231–239.
1146 C.R. Porfı́rio, D. Odloak / Control Engineering Practice 19 (2011) 1137–1146

Lu, J. (2003). Challenging control problems and emerging technologies in enter- Xu, F., Huang, B., & Akande, S. (2007). Performance assessment of model predictive
prise optimization. Control Engineering Practice, 11, 847–858. control for variability and constraint tuning. Industrial & Engineering Chemistry
Maciejowski, J. M. (2002). Predictive control with constraints. New Jersey: Prentice-Hall. Research, 46, 1208–1219.
Marlin, T. E., & Hrymak, A. N. (1997). Real-time operations optimization Ying, C. M., & Joseph, B. (1999). Performance and stability analysis of LP-MPC and
of continuous processes. In Proceedings of CPC V, AIChE symposium series, Vol. QP-MPC cascade control systems. AIChE Journal, 45, 1521–1534.
93, pp. 156–164. Yu, J., & Qin, S. J. (2008a). Statistical MIMO controller performance monitoring. Part I:
Odloak, D. (2004). Extended robust model predictive control. AIChE Journal, 50(8), Data-driven covariance benchmark. Journal of Process Control, 18, 277–296.
1824–1836.
Yu, J., & Qin, S. J. (2008b). Statistical MIMO controller performance monitoring.
Rao, C. V., & Rawlings, J. B. (1999). Steady states and constraints in model
Part II: Performance diagnosis. Journal of Process Control, 18, 297–319.
predictive control. AIChE Journal, 45, 1266–1278.
Yu, J., & Qin, S. J. (2009). MIMO control performance monitoring using left/right
Rotava, O., & Zanin, A. C. (2005). Multivariable control and real-time
optimization—An industrial practical view. Hydrocarbon Processing, 84(6), diagonal interactors. Journal of Process Control, 19, 1267–1276.
61–71. Zanin, A. C., De Gouvêa, M. T., & Odloak, D. (2002). Integrating real-time
Schafer, J., & Cinar, A. (2004). Multivariable MPC system performance assessment, optimization into the model predictive controller of the FCC system. Control
monitoring, and diagnosis. Journal of Process Control, 14, 113–129. Engineering Practice, 10, 819–831.
Tosukhowong, T., Lee, J., Lee, J., & Lu, J. (2004). An introduction to a dynamic plant- Zhao, C., Zhao, Y., Su, H., & Huang, B. (2009). Economic performance assessment of
wide optimization strategy for an integrated plant. Computers and Chemical advanced process control with LQG benchmarking. Journal of Process Control,
Engineering, 29, 199–208. 19(4), 557–569.