Computers and Chemical Engineering 34 (2010) 19992006

Contents lists available at ScienceDirect

Computers and Chemical Engineering

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

Real time optimization (RTO) with model predictive control (MPC)

Glauce De Souza a , Darci Odloak a, , Antnio C. Zanin b
a
LSCP/CESQ, Department of Chemical Engineering, University of So Paulo, Av. Prof. Luciano Gualberto, n. 380, trav. 3, 05508-900, So Paulo, SP, Brazil
b
PETROBRAS, Av. Chile, 65 Sala 2102, 20035-900, Rio de janeiro, Brazil

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

Article history: This paper studies a simplied methodology to integrate the real time optimization (RTO) of a continuous
Received 4 December 2009 system into the model predictive controller in the one layer strategy. The gradient of the economic
Received in revised form 30 June 2010 objective function is included in the cost function of the controller. Optimal conditions of the process at
Accepted 5 July 2010
steady state are searched through the use of a rigorous non-linear process model, while the trajectory to be
Available online 13 July 2010
followed is predicted with the use of a linear dynamic model, obtained through a plant step test. The main
advantage of the proposed strategy is that the resulting control/optimization problem can still be solved
Keywords:
with a quadratic programming routine at each sampling step. Simulation results show that the approach
Real time optimization
Model predictive control
proposed may be comparable to the strategy that solves the full economic optimization problem inside
Fluid catalytic converter the MPC controller where the resulting control problem becomes a non-linear programming problem
Integration of control and optimization with a much higher computer load.
2010 Elsevier Ltd. All rights reserved.

1. Introduction degrees of freedom of the system in order to maximize throughputs

at the bottlenecks of the system.
Real time optimization (RTO) deals with methods of maximizing Economic optimization studies for oil rening and chemicals
an economic objective related to the operation of a continu- production systems have proven to be very benecial and the
ous process. Most RTO systems are based on non-linear steady resulting tools and algorithms have been accepted by the industry
state models of the process system, combined with data rec- (Georgiou, Taylor, Galloway, Casey, & Sapre, 1997; Rotava & Zanin,
onciliation or parameter estimation to update key-parameters, 2005). Similar algorithms that have been used in the petrochemical
such as feed compositions and efciencies (Marlin & Hrymak, industry were applied by Mercangoz and Doyle (2008) to the pulp
1996). Typically, the RTO application optimizes the process oper- and paper manufacturing system.
ating conditions and updates the set-points to local MPCs, which In a continuous process, maximizing the economic benet,
are based on linear dynamic models. A steady state RTO may while maintaining the desired product quality is considerable
not be sufcient if, for instance, the set of active constraints harder in a dynamic environment than at steady state conditions
with signicant economic importance changes frequently. For (Engell, 2007). In this work, we focus on the integration of real time
example, this could be the case of a plant in which the through- optimization and control of continuous processes through the strat-
put bottleneck moves frequently. At least in theory, for this egy called one-layer approach (Zanin, Tvrzsk de Gouva, & Odloak,
type of process, it is more suitable to use dynamic optimiza- 2002) or full optimizing control (Rolandi & Romagnoli, 2005).
tion with a non-linear model, which may be achieved using In process plants, the economical optimal operation is usually
dynamic RTO (DRTO) or non-linear MPC with economic objective addressed by a two-layer structure (Marlin & Hrymak, 1996).
(Kadam et al., 2003; Tosukhowong, Lee, Lee, & Lu, 2004). How- In the upper layer, the operating point or steady state of the
ever, the dynamic optimization of a complex process may still plant is optimized based upon a rigorous non-linear stationary
not be achievable with the available computing resources (Lu, plant model. The optimal operating point is characterized by
2003). targets to a set of controlled and manipulated variables. These
Aske, Strand, and Skogestad (2008) propose a simple solution to targets are passed to a lower level layer, where MPC controllers
the RTO problem for the case where the economic optimal opera- keep the chosen process variables as close to these set-points as
tion of the plant is the same as maximizing plant throughput. These possible by manipulating the available degrees of freedom of the
authors propose the use of a coordinator MPC, which explores the process within certain bounds. This two layer structure has some
drawbacks, as optimization is only performed intermittently at a
low sampling rate and inconsistencies may arise from the use of
Corresponding author. Tel.: +55 11 3091 2261; fax: +55 11 3091 2246. different process models at the two layers. These issues are partly
E-mail address: [email protected] (D. Odloak). addressed by schemes in which the economic optimization is

0098-1354/$ see front matter 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compchemeng.2010.07.001
2000 G. De Souza et al. / Computers and Chemical Engineering 34 (2010) 19992006

integrated within a linear MPC controller on the lower level. In the subject to
two layer structure, process data is rst tested for steady state and,    
if a stationary situation is conrmed, the model is reconciled and h u(k + m k), y(k + k) =0 (2)
key operational parameters are calculated. Then, an optimization 
case is solved and new values for critical state variables of the umin u(k + j k ) umax j = 0, 1, ..., m 1 (3)
plant are computed which optimize an economic cost function 
while meeting the constraints imposed by the equipment, product ymin y(k + k ) ymax (4)
specications, safety and environmental regulations. Finally, these 
umax u(k + j k ) umax j = 0, 1, ..., m 1 (5)
values are forwarded to the process control layer, which uses
these values as targets and implements appropriate moves to the 
y(k + j k) = gj (x(k), uk ) , j = 1, ... , p (6)
manipulated variables. For a discussion of the implementation of
the optimal steady state with linear model predictive controllers, where p and  m are the output and input horizons, respec-
see Rao and Rawlings (1999). In this approach, the time interval tively; u(k + j k) is the control input computed at time k to
between successive RTO implementations must be large enough 
be applied at time step k + j; y(k + j k) is the output predic-
for the plant to reach a steady state after each optimizing moves.
Thus, the sampling period of the RTO must be larger than the time- 
tion at time step k + j; r is the desired
 value of the output;
u(k + j k) = u(k + j k) u(k + j 1 k) ; W1 , W2 and W3 are
constant of the controlled process. Typical sampling times for the
positive weights; feco is the economic objective computed at the
RTO layer in a two layer approach are of the order of several hours
predicted steady state of the process corresponding to u(k + m k) ;
(Engell, 2007) while the control layer provides tracking and distur- 
bance rejection on shorter time-scales from seconds to minutes. In y(k + k) is the output steady state obtained through the
this structure, the control algorithm is based upon linear models, non-linear model dened in Eq. (2).
which can be determined from plant data. This implies that the The constraint dened in (6) corresponds to the output
prediction performed with the linear dynamic model of the sys-
models on the optimization layer and on the control layer will in     T
general not be fully consistent (Backx, Bosgra, & Marquardt, 2000). tem, and u = u(k k) u(k + 1 k) u(k + m 1 k)
Zanin et al. (2002) describe an experiment where the non-linear is the control sequence. The constraints dened in (3)(5) corre-
steady state optimization was successfully integrated into the spond to the input and output constraints that are usually included
linear MPC of an industrial uidized bed catalytic cracker, FCC. In in an MPC controller.
their approach the economic function computed at the predicted In the RTO/MPC algorithm dened above, the outputs are con-
steady state is included as an additional term of the cost function of trolled by zones, which means that the desired values r of the
the controller. Therefore, the MPC problem becomes a non-linear controlled outputs do not have xed values. These zones reect the
programming (NLP) problem with a non-linear objective and the process requirements as product specications, equipment restric-
non-linear steady state model is included as additional constraints tions and adequate operating practices.
of the control problem. Although, their approach could be imple- Note that, the economic objective feco is computed using a non-
mented in the industrial system with a small sampling period, it is linear steady state process model. The different components of the
not clear whether this approach can be implemented in a system cost function are weighted such that the economic criterion and the
with a more complex model. dynamic compensation of the output error have a similar inuence
This work presents a simplied strategy, which can be consid- on the values of the overall cost.
ered as an alternative to integrate the economic optimization into Although the integration of RTO and MPC through the solution of
the linear model predictive controller in the one layer approach. the problem dened in (1)(6) seems promising for practical appli-
Instead of including the economic objective at the predicted steady cations, there are some problems related to this approach. Since in
state in the control objective, an approximation of the reduced the general case, feco and h are non-linear functions of the inputs
gradient of the economic function with an appropriate weight is and outputs of the system, the problem solved by the controller is
iteratively included in the control cost. Simulation results obtained a NLP problem and requires a more sophisticated solver than the
with the non-linear model of Moro and Odloak (1995) are used to one used in the conventional MPC, which requires the solution of
compare the proposed approach to the one of Zanin et al. (2002), a QP. Usually a SQP algorithm is used in the solution of the prob-
and show that the results of the algorithm proposed here may lem of the MPC with an economic objective. Although, the recent
be comparable to the results of their approach, not only in terms progress in the algorithms for the numerical solution of NLP prob-
of the economic benets, but also in terms of the stability of the lems and the increasing capacity and power of computing systems
closed loop system. It is also discussed the conditions in which the enable the solution of the one layer RTO/MPC problem for several
proposed approach can be used as a substitute to the rigorous real practical systems, one can expect some difculties in solving the
time optimization. RTO/MPC problem for large dimension systems. This means that
the computer power may not be sufcient to solve the one layer
2. The one-layer RTO/MPC controller problem within the sampling period of the control system, which
is typically of the order of 1 min. Also, as the most efcient SQP
Zanin, Tvrzsk de Gouva, and Odloak (2000, 2002) propose a solvers take advantage of techniques that deal with infeasible path
new method to integrate the real time optimization of continuous algorithms, an intermediary solution to the RTO/MPC problem may
processes into the model predictive controller by solving at each be infeasible and cannot be implemented in practice. Consequently,
sampling step k an optimization problem that can be represented if the solver of the one layer RTO/MPC problem fails to produce a
as follows: converged solution within the sampling period of the controller,

p
   no control action will be available to be implemented and the sys-
min W1 (y(k + j k ) r)2 tem will remain in open loop, which is highly undesirable. Thus,
u(k+i|k );i=0,...,m1 2
there is interest in developing an alternative solution to the one
j=1
layer RTO/MPC problem that preserves some of the advantages of

m1
       the integration strategy proposed above, but without excessively
+ W2 u(k + j k )2 + W3 feco u(k + m k ), y(k + k ) (1) delaying the computation of the optimizing control action or jeop-
2
j=0 ardizing the stability of the closed loop system.
 G. De Souza et al. / Computers and Chemical Engineering 34 (2010) 19992006 2001

3. The simplied one-layer RTO/MPC to take into account that a given input or controlled output will
reach its bound, the calculation of the gradient of the economic
It is well known that to nd the unconstrained maximum of a objective is performed as follows.
continuous concave function or the unconstrained minimum of a At each sampling time k, vector d and matrix G are computed
convex function is equivalent to searching for a point where the using the predicted steady state corresponding to the control input
gradient of the function becomes equal to zero. This is true, even implemented at time step k 1. Then, the problem dened in
when the optimal point lies outside the set dened by the con- (10) is initially solved assuming that all controlled variables and
straints. In this case the reduced gradient that considers all the inputs will not be constrained, and then the corresponding con-
active constraints would be equal to zero. Now, consider a mul- trol sequence is computed. Then, using this control sequence the
tivariable system with ny outputs and nu inputs, and at any time resulting predicted values of the controlled variables are checked
step k, suppose that the stationary prediction of the controlled out- in order to conrm that there are no violations of the constraints.
put related to the present value of the system input u is y. Also, Now, depending on the existence or not of violations of the output
consider that the economic function associated with the operation bounds, the following actions are taken:
of the system is known to be a concave function whose maximum
has to be searched (or a convex function whose minimum has to 1) If the predicted values of the system outputs and inputs do not
be searched) and this economic function can be represented as a violate any constraint, the solution obtained through the solu-
function of the predicted steady state as follows: tion to the problem represented in (10) is implemented in the
F = feco (u, y) (7) real system.
2) If the prediction of one of the outputs is such that yi ymax,i or
Then, assuming that the vector that represents the control action yi ymin,i . Then, the problem dened in (10) is solved again at
is changed to u + u, the rst order approximation to the gradient the same time step with the following additional constraint.
of the economic function can be represented as follows:
   
nu

dF  2 
dF  + d F  u
yi = Kpi,j uj = 0 (11)
u+u =  = (8)
du u+u du u du2  j=1
u

If the non-linear steady state model dened in (2) is available, then Also, if the computed value of an input indicates that it will
y/u and 2 y/u2 can be calculated using this model. Moreover, reach a bound, this input is not considered in the computation
since y/u corresponds to the process gain, it is here represented of the gradient .
by Kp . Now, using (7) to calculate the derivatives of F, Eq. (8) 3) The output and input predictions corresponding to the solu-
becomes: tion dened in the previous step should be checked to verify
if any other output or input will break their bounds. If this
u+u = d + Gu (9)
happens, additional constraints of the type represented in (11)
where should be included in the control problem or inputs should be
removed from the calculation of the gradient. This procedure
F F
d= Kp + should be repeated until no other output constraint is required
y u
to be included in the control problem.

2 F 2 F F 2 y 2 F F The MPC with economic objective uses a linear dynamic model

G = KpT K +
2 p
Kp + + Kp +
(y) uy y u2 yu u2 to predict the output trajectory along the prediction horizon. To
produce an offset free controller, a state space model in the incre-
and
mental form is used as follows:
u = u(k + m 1) u(k 1) is the total move of the input vec-
tor. x(k + 1) = Ax(k) + Bu(k) (12)
Remark
In the unconstrained case, a point where = 0 corresponds to y(k) = Cx(k) (13)
an extremum of the economic function, which is a maximum if
G < 0 or a minimum if G > 0. In the constrained case, the extremum Using the model dened in (12) and (13) to represent the output
point corresponds to r = 0, where r is the projection of on the prediction in terms the control move and using (9) to express the
tangent space to the active constraints. This point will be a maxi- economic term in (10) as a function of the control move, the control
mum if T G < 0 for any lying on the tangent space to the active objective of the SRTO/MPC dened in (10) can be represented as
constraints. Analogously the extremum point will be a minimum if follows:
T G > 0.
J = uT Hu + 2Ct u + c (14)
The simplied real time optimization integrated into the lin-
ear MPC controller (SRTO/MPC) proposed here corresponds to the where
following optimization problem:
H = BT W1,p B + W2,m + GT W3 G;

p
  
min J= W1 (y(k + j k ) r)2
u(k+i|k );i=0,...,m1 2
T
j=1 Ct = (Ax(k) Ir) W1,p B + 2dT W3 Gu


m1
    2
+ W2 u(k + j k )2 + W3 T  (10)  T  
2 u+u 2 c = Ax(k) Ir W1,p Ax(k) Ir + dT W3 d
j=0

subject to: (3) to (6) and (9).

W1,p = diag(W1 .......W1 ), W2,m = diag(W2 ........W2 ), G = [ G G ]
The integration of the economic function into the controller can         
push one or more outputs towards their min/max bounds. In order p m m
2002 G. De Souza et al. / Computers and Chemical Engineering 34 (2010) 19992006

Fig. 1. Schematic representation of the FCC process system.

In the simulations presented here, it is considered a simpli-

CA CB 0 0 Iny ed version of the industrial control system. The most important
CA2 CAB CB 0 Iny manipulated inputs are: the total air ow rate to the two-stage cat-
A = . , B = , I = .
.. ... ..
.
.. alyst regenerator u1 (ton/h), the valve opening of the regenerated
catalyst u2 (%), the gasoil feed ow rate u3 (m3 /h) and the tem-
CAp CAp1 B CAp2 B CApm B Iny perature of the feed u4 (C). Adopting the zone control approach,
the controlled outputs considered here are: the temperature in
Iny is the identity matrix with dimension ny.
the dilute phase of the regenerator y1 (C), the temperature in the
Observe that the control cost function of the MPC with economic
dense phase of the regenerator y2 (C), the conversion of the cracking
objective represented in (14) is a quadratic function of the vector of
reaction y3 (%) and the riser temperature y4 (C).
control moves. As the constraints represented in (3) to (6) and (9)
A typical economic objective of the FCC unit considered
are linear, the optimization problem that produces the proposed
here is to maximize the production of liqueed petroleum
MPC with economic objective is a QP, or, it is a problem with a
gas (LPG), which is a non-linear function of the feed prop-
computer demand similar to the demand of the conventional MPC.
erties and process operating conditions (Zanin et al., 2000).
It should also be noticed that the control objective dened in (14)
Another possible objective is to maximize the production of
depends on the knowledge of the model state at the present time
the light cycle oil (LCO), which may be part of the diesel pool.
that is not usually measured. Then, the proposed controller assumes
In the simulations performed here, the initial steady state  of
the presence of a state estimator that is not discussed in this work.
the FCC is given by: uss = 220.7 85 404.0 234.9 and
In the next section the controller dened above is applied to  
a typical industrial system where real time optimization is usu- yss = 697.3 699.4 78.2 544.4 . The tuning parameters
ally justied because of the large economic gain obtained with the adopted in the RTO/MPC of section 2 are the following: m = 2,
optimal operation of the system. p = 70, W1 = diag(0.2 0.1 0.1 1), W2 = diag(5 5 5 5) and W3 = 1.5. The

constraints 
on the manipulated  considered here are: umax =
inputs
228 98 406 235 , umin = 200 50 400 234.9 and
4. The FCC (uid catalytic cracking) unit  
umax = 2 1 2 0.4 . The output zones are dened as follows:
   
To compare the performance of the SRTO/MPC algorithm pro- ymax = 705 725 95 547 and ymin = 695 695 60 540 .
posed in Section 3 with the RTO/MPC of Zanin et al. (2002) The sampling period is equal to 1 min.
summarized in Section 2, it is considered here the same process The proposed SRTO/MPC controller adopts the same tuning
system studied by Zanin et al. (2002). This example represents the parameters and constraints as the RTO/MPC controller except the
typical industrial problem, where the multivariable system has to weight of the economic component, which in this case is W3 = 0.3.
be optimized on-line subject to constraints and dynamic objectives. We initially consider the two controllers with the objective of
Furthermore, as the evaluation of the proposed approach is going to maximizing the volumetric production of LPG. Fig. 2 shows the
be performed through simulation, the models (rigorous stationary behavior of the LPG production for the two approaches when the
and linear dynamic) have to be well known and tested. The system system starts from the steady state and the optimizer is turned on
considered in this work is the FCC converter, which is represented at time step 50 min. We observe that the economic results of the
in Fig. 1 (Moro & Odloak, 1995). two controllers are almost coincident. Figs. 3 and 4 show the con-
 G. De Souza et al. / Computers and Chemical Engineering 34 (2010) 19992006 2003

benet will respond to a change in the operating conditions or dis-

turbances. However, as the constraints on the system outputs are
included in the control/optimization problem as control zones and
through the reduced gradient, the result of a large economic weight
W3 may be undesirable. If the optimal unconstrained operating
point of the controlled system lies outside the convex set dened
by the constraints, the economic term in the control cost function
dened in (9) tends to push some of the system outputs to outside
the control zones. In the example simulated here, Fig. 3 shows that
y1 tends to stay slightly below the minimum bound for quite long
time, while y4 tends to stay slightly above the maximum bound for
some time. This means that in the SRTO/MPC, the proper balance
between weighing factors W3 and W1 is a critical point in the tuning
procedure of the controller. This also shows that for this particular
optimal operating objective the constraints related to the minimum
bound of the regenerator dense phase temperature and maximum
bound of the riser temperature will be active and consequently the
degrees of freedom of the system will be reduced. Fig. 3 also shows
that the two other controlled outputs remain inside their control
Fig. 2. LPG production for RTO/MPC () and SRTO/MPC (). zones and do not affect the performance of the controller. Also, Fig. 4
shows that the gasoil feed ow, which is input u3 is pushed by the
controller to its maximum bound. In this simulated case the gasoil
trolled outputs and the manipulated inputs, respectively. One may feed temperature u4 was kept inside a very narrow range to rep-
observe that the simplied algorithm (SRTO/MPC) to integrate real resent the case where the gasoil heating furnace is shut off, which
time optimization and model predictive control in a single layer is a very common condition in this FCC unit. This indicates that all
can capture practically the same economic benet as the one-layer the four degrees of freedom of the system have been consumed by
(RTO/MPC) algorithm proposed by Zanin et al. (2002). The increase satisfying the system constraints and the optimal operating point
in the economic term is slightly faster in the one layer RTO/MPC, lies at the intersection of four constraints and the reduced gradi-
but the nal value of the economic benet is the same in the two ent of the economic objective function becomes equal to zero. A
approaches. Obviously, in both methods the speed of response of straightforward way to reduce the period of time in which the out-
the economic benet will depend on the tuning weight W3 . The puts tend to remain outside the control zones, is by increasing the
larger the absolute value of this parameter, the faster the economic corresponding elements of weight W1 . However, as the dynamic

Fig. 3. Controlled outputs for RTO/MPC () and SRTO/MPC ().

2004 G. De Souza et al. / Computers and Chemical Engineering 34 (2010) 19992006

Fig. 4. Manipulated inputs for RTO/MPC () and SRTO/MPC ().

model used in the controller is a linear model that only approxi- It is also interesting to compare the performances of the two
mates the real non-linear model, there is a limit on the maximum controllers when the economic objective is to maximize the pro-
value of this weight before the closed loop system becomes unsta- duction of light cycle oil (LCO), which may be important when the
ble. Thus, the integration of real time optimization with linear MPC issue is the optimization of the diesel pool of the renery. Initially,
in a single layer, still poses some open questions as the develop- observe the behavior of the LCO production shown in Fig. 5 when
ment of tuning rules related to the selection of weights W1 and the controller maximizes the production of LPG. It is clear that in
W3 , which directly affect the performance of the controller as well this case the two controllers minimize the LCO production. This
as the stability of the closed loop system when uncertainty of the means that the function that relates the production of LCO with the
linear dynamic model is signicant. variables of the FCC system has a minimum at the neighborhood
of operating region dened by the input and output constraints

Fig. 5. LCO production when the economic objective is to maximize LPG, RTO/MPC Fig. 6. LCO production when the economic objective is to maximize LCO, RTO/MPC
() and SRTO/MPC (). () and SRTO/MPC ().
G. De Souza et al. / Computers and Chemical Engineering 34 (2010) 19992006 2005

Fig. 7. Controlled outputs for RTO/MPC () and SRTO/MPC ().

Fig. 8. Manipulated inputs for RTO/MPC () and SRTO/MPC ().

2006 G. De Souza et al. / Computers and Chemical Engineering 34 (2010) 19992006

(3)(6). Then, consider the behavior of the two optimizing con- mized is a concave function of the process variables in the operating
trollers when the objective is to maximize the LCO. Fig. 6 shows region dened by the inputs and outputs constraints. Analogously,
the LCO proles when the simulation is repeated starting from the the method also works if the objective is to minimize a function that
initial steady state and with this new objective. It is clear that the has a local minimum in the operating region of the system. Tuning
SRTO/MPC fails completely as it reduces the LCO ow instead of of the optimizing controller shows an augmented level of complex-
maximizing it, while the RTO/MPC behaves quite nicely increasing ity when compared to the conventional MPC as the new controller
signicantly the related product ow rate. We observe from Fig. 7 has an additional tuning parameter which is the weight of the gra-
that the operating point corresponding to the maximal LCO pro- dient term. This term plays the same role as the economic function
duction corresponds to the temperature of the regenerator dense term that is included in the control objective of the conventional
phase at its upper bound and the riser temperature at its lower RTO/MPC.
bound, which is the opposite of what happens to the SRTO/MPC
that tends to minimize the LCO production pushing the FCC system References
to the same constraints as in the case of the maximization of the
LPG production. Fig. 8 shows that the manipulated inputs tend to Aske, E. M. B., Strand, S., & Skogestad, S. (2008). Coordinator MPC for maximizing
plant throughput. Computers and Chemical Engineering, 32(12), 195204.
stabilize at completely different points for the two controllers. This Backx, T., Bosgra, O., & Marquardt, W. (2000). Integration of model predictive control
clearly shows that the simplied version of the MPC controller that and optimization of processes. In Proceedings of IFAC Symposium ADCHEM Pisa,
implements the economic objective cannot be used in the maxi- (pp. 249260).
Engell, S. (2007). Feedback control for optimal process operation. Journal of Process
mization of the LCO ow rate. This is so because the operating point Control, 17, 203219.
where the reduced gradient of the economic function is zeroed cor- Georgiou, A., Taylor, P., Galloway, R., Casey, L., & Sapre, A. (1997). Plantwide closed
responds to a local minimum of this function as shown in Fig. 6. So, loop real time optimization and advanced control of ethylene plant (CLRTO)
improves plant protability and operability. In Proceedings of NPRA Comp. Conf.,
in the FCC system considered here, the SRTO/MPC algorithm pro-
Nat. Petr. Ass., paper CC 97-139, New Orleans.
posed here can be applied either to the maximization of LPG or to Kadam, J., Marquardt, W., Schlegel, M., Backx, T., Bosgra, O., & Brouwer, P. J. (2003).
the minimization of LCO. The applicability of the proposed method Towards integrated dynamic real-time optimization and control of indus-
to maximize or to minimize the economic objective can be eas- trial processes. In Proceedings Foundations of Computer-Aided Process Operations
(FOCAPO2003) (pp. 593596).
ily tested by verifying if the Hessian matrix G dened in Eq. (9) is Lu, J. (2003). Challenging control problems and emerging technologies in enterprise
negative or positive denite respectively. optimization. Control Engineering Practice, 11, 847858.
Marlin, T. E., & Hrymak, A. N. (1996). Real-time operations optimization of contin-
uous processes. In Proceedings of CPC V, AIChE Symposium Series, Vol. 93, 1997,
5. Conclusion pp. 156164.
Mercangoz, M., & Doyle, F. J., III. (2008). Real-time optimization of the pulp mil
Here, it is studied a simplied version of the one layer RTO/MPC benchmark problem. Computers and Chemical Engineering, 32(45), 789804.
Moro, L. F. L., & Odloak, D. (1995). Constrained multivariable control of uid catalytic
controller presented in previous works that integrates the two cracking converters. Journal of Process Control, 5, 111.
functions, real time optimization and multivariable constrained Rao, C. V., & Rawlings, J. B. (1999). Steady states and constraints in model predictive
control, in a single layer. In this strategy, control and optimiza- control. AIChE Journal, 45, 12661278.
Rolandi, P. A., & Romagnoli, J. A. (2005). An integrated environment for support
tion of the plant are solved simultaneously in the same problem. In of process operations. In AIChE Annual Meeting, Conference Proceedings (pp.
the approach proposed here, the reduced gradient of the economic 60266027).
objective function is included as an additional term of the cost Rotava, O., & Zanin, A. C. (2005). Multivariable control and real-time
optimizationAn industrial practical view. Hydrocarbon Processing, 84(6),
function of the MPC controller. The control objective is to zero the 6171.
reduced gradient of the economic objective while maintaining the Tosukhowong, T., Lee, J., Lee, J., & Lu, J. (2004). An introduction to a dynamic plant-
system outputs inside their control zones. The main advantage of wide optimization strategy for an integrated plant. Computers and Chemical
Engineering, 29, 199208.
the proposed approach is that the integrated control/optimization
Zanin, A. C., Tvrzsk de Gouva, M., & Odloak, D. (2000). Industrial implementation
problem becomes a QP, instead of a NLP as in the previous approach. of a real-time optimization strategy for maximizing LPG in a FCC unit. Computers
Simulation results with an FCC system have shown that almost and Chemical Engineering, 24, 525531.
the same economic benet as the in the conventional RTO/MPC Zanin, A. C., Tvrzsk de Gouva, M., & Odloak, D. (2002). Integrating real-time
optimization into the model predictive controller of the FCC system. Control
algorithm can be obtained when the economic benet to be maxi- Engineering Practice, 10, 919931.