Beruflich Dokumente
Kultur Dokumente
This work addresses the application of the Robust Innite Horizon Model Predictive Control (RIHMPC) to a heat integrated propylene distillation
system at a Petrobras renery. The approach proposed here is tested on the rigorous dynamic simulation software (Dynsim1) that reproduces the
system as a virtual plant and is able to communicate with the MPC algorithms developed in Matlab, through an Open Platform Communication (OPC)
interface. The controller is based on a minimal order statespace model that is equivalent to the system step response and considers the zone control of
the outputs and optimizing targets for the inputs. The optimizing targets are obtained through the steadystate economic optimization using the real
time optimization package (ROMeo11). The proposed integration approach provides convergence and stability to the closedloop system. The
propylene distillation system is simulated with the proposed control and optimization strategies and the results show that, from the economic
performance and robustness viewpoint, for this particular system, the proposed robust MPC is signicantly better than the nominal IHMPC based on a
single linear model obtained at the most probable operating point.
Keywords: advanced control of distillation, real time optimization, dynamic simulation, propylene production unit
INTRODUCTION several practical features such as the zone control and the
consideration of optimizing the targets for the inputs and
O
ne of the key challenges in the process industry is how to outputs. The method can be applied to stable, integrating and
best control and stabilize the process plant while timedelayed systems and provides guaranteed nominal
searching for the economic optimum operating point. stability.
Advanced control based on MPC, which is a control standard in The lack of robust stability is also a weakness of available model
the oil rening industry, is frequently implemented as one of the predictive controllers that are usually implemented in industry.
layers of the control structure where real time optimization A robust controller is able to maintain the stability of the closed
(RTO) lies in an upper layer and denes optimum targets for loop at different operating conditions. Typically, at each operating
some of the inputs and outputs.[1] Examples of successful point the process system can be represented by a different linear
implementations of MPC have been reported in the literature model. This case is called the multiplant uncertainty,[9] and a
in the last 30 years.[24] Most of the MPC applications in industry robust control strategy (RIHMPC) can be proposed in which a set of
are still based on the stepresponse model of the process system as models is used to represent the uncertain system and the controller
in the seminal application of dynamic matrix controller (DMC) is stable for any single model of that set.[10,11] In a similar strategy,
presented in Cutler and Ramaker.[5] Despite the good perfor- polytopic uncertainty can also be considered where the true
mance of the MPC based on the step response model, some model is assumed to be the convex combination of a nite set of
limitations of this approach can be mentioned, as the lack of models that represent the vertices of a polytope.[12,13] These ideas
nominal stability related with the nite output prediction can also be included in the innite horizon MPC and result in
horizon. Also, the state of the equivalent state space model is various strategies for implementing robust controllers that can be
nonminimal, which means that a state with a smaller dimension integrated with the RTO layer.[14,15]
could be obtained. The state estimation strategy is also not In practice, the design, tuning and implementation of IHMPC or
optimal and adhoc solutions need to be adopted for the unstable RIHMPC integrated with RTO would be costly and time consuming
or integrating system. mainly because the approach would need the identication of the
As a consequence, system representations based on the state process model at different operating points, which would demand
space approaches that allow the use of innite prediction horizon extensive plant testing. The alternative is to develop a study based
have been considered in MPC implementations. First, Rodrigues on simulation scenario. Steadystate and dynamic simulation
and Odloak[6] developed a minimal statespace representation based on rst principles are now mature technologies and have
for stable and integrating systems that is equivalent to the step
response of systems that can also be represented by transfer
functions. The model was extended by Carrapio and Odloak[7]
for time delayed systems, but it resulted in a nonobservable *Author to whom correspondence may be addressed.
model for systems with multiple delays. More recently, Santoro Email address: odloak@usp.br
and Odloak[8] developed a spacestate representation that solves Can. J. Chem. Eng. 92:12131224, 2014
the observability problem of the delayed systems. The new state 2014 Canadian Society for Chemical Engineering
DOI 10.1002/cjce.21980
space system representation approach allows the implementa- Published online 26 May 2014 in Wiley Online Library
tion of innite horizon model predictive control (IHMPC) with (wileyonlinelibrary.com).
VOLUME 92, JULY 2014 THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING 1213
been extensively applied in industry. As plant designs are about the implementation of the advanced control strategy at the
becoming more complex, integrated and interactive, the control propylene/propane splitter are presented.
of their dynamic behaviour tends to represent challenges of
increasing difculty.[16] Then, commercial process simulators tend THE CONTROL PROBLEM AND OPENLOOP SIMULATION OF
to be used for the understanding of the process interacting
dynamics as well as the evaluation and tuning of control strategies THE PROPYLENE/PROPANE SPLITTER
before implementation. Control practitioners have adopted the use The industrial propylene process system considered here consists
of dynamic simulation as an alternative to plant testing so that the of four sections: the depropanizer column (T01), the deethanizer
required information about the process dynamics can be obtained column (T02), the propylene/propane splitter (T03) and the
without disturbing the process operation. The advantages of caustic treatment section. The system was designed to produce
performing a step test on a dynamic simulation instead of on the highpurity propylene (99.5% molar) from liqueed petroleum gas
real plant are obvious. As no plant test is required, the effort can be (LPG), which is fed at column T01, where the bottom product
minimized especially for processes with many variables and/or (butane) is sent to storage while the top product is sent to the
long settling times.[17] Once all the required data for model caustic treatment so that its sulphur content is reduced to 4 ppm.
identication is available, the implementation of the advanced The desulphurized stream is then fed to column T02 where ethane
control becomes much simpler.[18] Here, this approach will be and the lighter ends are separated and incorporated in the renery
extended to the study of the implementation of IHMPC and fuel gas, while the bottom product is sent to column T03, which is
RIHMPC integrated with RTO in a real propylene/propane splitter schematically represented in Figure 1. In this column, propylene is
of an oil renery. separated from propane, which also carries other hydrocarbons
The main scope of this work is to study the implementation of the with four atoms of carbon. A typical composition of the feed of
IHMPC and the RIHMPC, based on the statespace representation of column T03 is shown in Table 1. The propylene stream is
the process step response as will be explained in the paper, in an produced as the top stream of the splitter and is sold to a nearby
industrial propylene/propane splitter without performing any petrochemical plant, while the propane stream obtained as the
plant test. Also, the closedloop performance of these controllers bottom product is stored in propane spheres.
will be compared using the dynamic simulation in order to evaluate The distillation system studied here includes an energy recovery
the possible advantage of the RIHMPC over the conventional system through the integration of the top cooling system into
IHMPC. The control scheme to be tested assumes that a RTO layer is the bottom heating system. As it can be observed from Figure 1,
present in the control structure and provides the optimum targets there is a vapour recompression system where energy savings of
for the manipulated inputs. about 50% have been reported.[19] In this way, compressor V01
The paper is organized as follows: The control problem of the increases the pressure and temperature of the vapour leaving
propylene/propane splitter and the dynamic simulation is briey the top of the column at about 16 kgf/cm2 and 50 8C, and the
discussed in the next section. Then, the statespace system recompressed vapour is condensed in the reboilers M01 A/B and
representation developed in Santoro and Odloak[8] and adopted M02 at the bottom section of the column. This column presents
in the MPC implementation is highlighted. Next, the IHMPC and three bottom reboilers that work in parallel. Reboilers (M01 A/B)
robust IHMPC are formulated to be tested in the propylene/ are vertical, while reboiler (M02), which was included in
propane system. Results of simulation scenarios are presented and the system in a revamp project of the system, is horizontal.
discussed in the simulation section. Finally, some conclusions All the reboilers have variable heat transfer areas that depend on
1214 THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING VOLUME 92, JULY 2014
Table 1. Typical feed composition of the propylene/propane splitter The high concentration of propylene in the product stream and
the variable heat transfer area of the bottom reboilers justify the use
Component % Molar fraction
of a multivariable advanced controller to maximize the economic
Ethane 0.0102 benet and to attain the product specications. As a consequence,
Propylene 64.41 the implementation of MPC in conjunction with RTO needs to be
Propane 34.77 evaluated. Possible manipulated inputs for the advanced controller
iButene 0.337 are the feed ow rate (FC1), the reux ow rate (FC2) and the heat
1Butene 0.061 pump ow rate (FC3), and the controlled outputs are the molar
cis2Butene 0.0334 concentration of propane in the propylene stream (AC1), the molar
trans2Butene 0.0334 concentration of propylene in the propane stream (AC2) and the
1,3Butadiene 0.012 level of liquid in separator O03 (LC5), which affects the heat
iButane 0.298
transfer area in the bottom reboilers M01 A/B. The controller to be
Butane 0.0334
implemented should drive the system inputs to the optimum
targets while keeping the system outputs inside control zones that
are dened by the operators as will be seen in the next section.
Besides, the propylene splitter has a highly nonlinear behaviour,
the level of the condensed liquid inside drums O03 and O04. which is a common characteristic of highpurity distillation
The existing PID control loops of the system are presented in processes. Then, stability is an additional issue in the control
Table 2. problem to be studied here and justies the consideration of new
The operation of the main regulatory control loops described in control strategies as the ones that will be described in the following
Table 2 can be described as follows: FC5 and LC1 are cascaded sections.
such that the liquid level at the bottom of T03 is maintained at The dynamic simulation of the propylene distillation column
1.9 m. In the same way, LC4 cascades FC7 such that liquid level in will be developed in the simulation package Dynsim. The idea is to
drum O04 is kept at 65 % of its range. Also, PC1 cascades FC6 in use the rigorous dynamic model as a virtual plant so that the design
order to keep the top pressure at 9.0 kgf/cm2. There is an override of the advanced control, the controller tuning and the model
control strategy involving LC3, LC3A and FC4. The resulting identication will not result in any cost. All the regulatory PID
signal of the high selector of the outputs of LC3 and FC4 is control loops will be included in the simulation in order to test the
compared in a low selector with the output of LC3A so that the implementation of the advanced control in a simulated system that
level of liquid propylene in separator O02 is held inside its upper closely represents the real system. The dynamic simulation will
and lower limits. Since the percentage of propylene in the also help the identication of the linear models that represent
propylene product is the most important controlled variable, there the process at different operating points and will be included in the
is a composition analyser (AC1) that cascades FC2 so that the RIHMPC. To characterize the dynamics of the propylene/propane
reux ow rate setpoint will be driven by the AC1 output signal. splitter, Figure 2 shows the step responses corresponding to three
Finally, in order to minimize the loss of propylene to the bottom different operating points. From these responses two main
product, there is a composition analyser (AC2) that cascades FC3 conclusion can be drawn: the model gain can change signicantly
and the resulting signal is compared with the output of LC05 in a from one operating point to the other and the settling time of the
low selector so that the liquid level in drum O03 is maintained system is very large (about 30 h), which can make the model
above a minimum value. identication experiments in the real system very time consuming.
To study the implementation of the advanced control strategy,
which was developed in MATLAB, and the real time optimization,
which was based on the rigorous steadystate optimization
Table 2. Main PID controllers used in the propylene/propane splitter package ROMeo integrated with the dynamic simulation that
PID
represents the true plant, it was necessary to integrate the various
controller Function software into a realtime data communication. Therefore, there is a
communication interface based on OPC technology that allows the
FC1 Feed flow rate to column T03 realtime data transfer between Dynsim, MATLAB and ROMeo.
PC1 Pressure at the top of T03 The OPC drive was designed to provide a common bridge for
LC1 Liquid level at the bottom of T03 Windows based software applications and process control
FC2 Reflux flow rate of column T03 hardware. To obtain a successful communication, there must be
PC2 Pressure at the outlet of compressor V01 (Relief) at least one OPC server and one or various OPC clients. In this case,
LC2 Liquid level in the knockout drum O01
the OPC server is the OPC Gateway which lies in Dynsim and
FC3 Total flow rate through the heat pump
the OPC clients are the OPC DA which is part of the OPC toolbox
LC3 Upperliquid level in separator O02
LC3A Lowerliquid level in separator O02
of MATLAB, and the OPC external data interface (EDI) of
FC4 Propylene flow rate to storage ROMeo. A representative scheme of this structure is shown in
LC4 Liquid level in drum O04 and reboiler M02 Figure 3.
FC5 Propane flow rate to storage
LC5 Liquid level in drum O03 and reboilers M01 A/B STATESPACE FORMULATION OF THE STEP RESPONSE
FC6 Flow rate through water cooler M03
FC7 Outlet flow rate of drum O04 MODEL
AC1 Propane molar composition in the propylene Consider a system with nu inputs and ny outputs, and assume that
product stream there are transfer function models relating the inputs and outputs.
AC2 Propylene molar composition in the propane Then, a state space realisation of the step response model of the
product stream system originally presented in Rodrigues and Odloak[6] and
VOLUME 92, JULY 2014 THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING 1215
O03 level x Heat pump (%/ton/h) O03 level x Feed (%/ton/h) O03 level x Reflux (%/ton/h)
0 0.18 0.2
Model 1
Model 2 0.16 0.18
Model 3
0.14 0.16
0.14
0.5 0.12
0.12
0.1
0.1
0.08
0.08
1 0.06
0.06
Model 1 Model 1
0.04
Model 2 0.04 Model 2
Model 3 Model 3
0.02 0.02
1.5 0 0
0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000
minutes minutes minutes
Propane in Propylene x Heat pump (molar %/ton/h) Propane in Propylene x Feed (molar %/ton/h) Propane in Propylene x Reflux (molar %/ton/h)
0.06 0.025 0
Model 1
Model 2
0.05 Model 3
0.01
0.02
0.04
0.02
0.015
0.03
0.03
0.02
0.01
0.04
0.01
Model 1 Model 1
0.005
Model 2 Model 2 0.05
0 Model 3
Model 3
0.01 0 0.06
0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000
minutes minutes minutes
Propylene in Propane x Heat pump (molar %/ton/h) Propylene in Propane x Feed (molar %/ton/h) Propylene in Propane x Reflux (molar %/ton/h)
0.1 0.4 0.7
Model 1 Model 1
0 Model 2 Model 2
0.35 0.6
Model 3 Model 3
0.1
0.3
0.5
0.2
0.25
0.3 0.4
0.2
0.4 0.3
0.15
0.5
Model 1 0.2
0.1
0.6 Model 2
Model 3
0.05 0.1
0.7
0.8 0 0
0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000
minutes minutes minutes
1
2 3
x k s
6 xd k 7
6 7
6 7
6 z1 k 7
6 7
yk I ny C 0 0 0 6 7
6 z2 k 7 2
6 7
6 .. 7
6 7
4 . 5
zumax k
where,
1216 THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING VOLUME 92, JULY 2014
component xs corresponds to the predicted output steady state and where ddi;j;k is the coefcient of the partial fraction expansion
xd corresponds to the stable dynamic modes that tend to zero when corresponding to the kth pole of the step response of the transfer
the system approaches steady state. For the case of nonrepeated function Gi, j(s).
poles F is a diagonal matrix with components of the form er i Dt Alternatively, if l 6 0, then each matrix Bdl would have the same
where ri is a pole of the system and Dt is the sampling period. The dimension as DdFN where those elements corresponding to
upper right block of matrix A is included to account for the time transfer functions with dead time different from l are replaced
delay of the system. with zeros.
Matrices Bsl , with l 1; ; umax can be computed as follows: Finally, matrix c is dened as follows:
VOLUME 92, JULY 2014 THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING 1217
a zone. This cost explicitly incorporates an input deviation and (4b) are satised it can be shown that the cost dened in
penalty that tries to accommodate the system at an optimal Equation (3) can be developed as follows:
economic stationary point. The slack variables dy,k and du,k are
included to guarantee that the innite horizon objective function is
Xmax
mu
bounded and to eliminate any infeasibility of the control problem. Vk yk jjk ysp;k dy;k T Qy yk jjk ysp;k dy;k
It can be shown that the cost dened in Equation (3) will be j0
bounded only if the following constraints are included in the X
1
The IHMPC also includes the usual constraints related with the 7
input bounds and input moves:
Finally, the optimization problem solved by the IHMPC can be
dened as follows:
Dumin Duk jjk Dumax j 0; 1; ; m 1
5
umin uk jjk umax ; j 0; 1; ; m 1 min Vk 8
Duk ; dy;k ; du;k ; di;k ; ysp;k
1218 THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING VOLUME 92, JULY 2014
plant is one of the components of this set. Therefore, we can dene subject to Equations (4b), (10) and
the set of possible plant models as V fQ1 ; ; QL g where each Qn
corresponds to a particular model. This approach is particularly Dumin Duk jjk Dumax ; j 0; 1; ; m 1
suitable to a nonlinear process system that can operate at different umin uk jjk umax ; j 0; 1; ; m 1
operating points that correspond to different product specica-
tions, market economic conditions, unknown disturbances, etc. In ymin ysp;k Qn ymax ; n 1; ; L
this case, each model Qn only represents the true system locally
around an operation point. ~ k Du
V k Duk ; ysp;k Qn ; dy;k Qn ; du;k ; Qn V ~ k ; ~ysp;k Qn ;
Badgwell [10] developed a robust linear quadratic regulator ~dy;k Qn ; ~du;k ; Qn n 1; ; L 12
for stable systems with the multiplant uncertainty. Later,
Odloak[21] extended the method to the output tracking of stable
systems considering the same kind of model uncertainty. These where, QN corresponds to the most probable model and assuming
strategies include a new constraint that prevents the plant cost that Duk1 ; ysp;k1 Qn ; dy;k1 Qn ; du;k1 is the optimum solu-
function to increase at successive time steps. More recently, tion to the control problem at time step k1, we dene:
Gonzlez and Odloak[15] presented an extension of the method
to time delayed systems and included the idea of considering the h
iT
output set point as a new restricted decision variable to develop ~ k Du kjk 1T Du k m 2jk 1T
Du 0 ; ~ysp;k Qn
a robust MPC for systems where the control objective is to ysp;k1 Qn
maintain the output variables into their corresponding feasible
zone. In this case the system can be driven to the desired optimal and ~dy;k Qn , ~du;k are such that
economic operation point given by the input target of the
supervisory stationary optimization (RTO). Therefore, in the ~ k I ny ~ysp;k I ny~dy;k 0
N s Ap Qn xk Aumax Qn CoQn Du
control of the propylene splitter, we consider the multimodel 13
uncertainty and assume that each model is characterized
by a
set of parameters dened as Qn Bsn ; Bdn ; F n ; un ; n 1; ; L ~ k udes I nu~du;k 0
uk 1 I Tnu Du 14
Also, we dene p m maxun, which is used in the model
construction to guarantee that the state vector has the same
dimension for all models belonging to set V. Then, for stable Stability of the propylene distillation column with the proposed
systems, we can dene, for each model Qn, the following cost robust MPC is achieved through the inclusion of the non
function: increasing cost constraints dened in Equation (12), which
prevent the cost corresponding to the true plant to increase. The
inclusion of these constraints, which are nonlinear, turns the
X
p
control problem a nonlinear program and a NLP solver is required
V k Qn yn k jjk ysp;k Qn dy;k Qn T Qy yn k jjk ysp;k Qn dy;k Qn
j0 for its solution.
X
1
yn k jjk ysp;k Qn dy;k Qn T Qy yn k jjk ysp;k Qn dy;k Qn
jp1
Economic Objective of the Propylene/Propane Splitter
X
m1
The input optimizing targets udes,k are dened by the RTO layer,
uk jj udes;k du;k T Qu uk jjk udes;k du;k
j0 which is based on a rigorous steadystate simulation of the
X
1
distillation process and computes the optimum operation point of
uk jjk udes;k du;k T Qu uk jjk udes;k du;k
jm the plant that maximizes the economical function:
X
m1
Duk jjkT RDuk jjk dy;k T Qn Sy dy;k Qn du;k T Su du;k
j0 X
product streams feedX
streams X
utilities
15
Following the same steps as in the case of the nominal system,
we can conclude that the cost dened in Equation (9) will be where PPS is the price of the propylene product [$/ton], PFR is the
bounded only if the control actions, set points and slack variables product ow rate [ton/h], PFS: is the price of the feedstock [$/ton],
are such that Equation (5) and Equation (4b) are satised, and FFR is the feed ow rate [ton/h], PU is the price of electricity
Equation (4a) can be written as follows: [$/kWh], UC is the electricity consumption [kWh/h].
The economic function dened in Equation (15) is maximized
xn s k m umax jkQn ysp;k Qn dy;k Qn 0 n 1; ; L producing the optimum input and/or output targets subject to the
following constraints:
10
The rigorous steadystate model that relates the system inputs
and measured disturbances to the outputs.
where xn s k m umax jk Qn is the output prediction at steady Lower and upper bounds to the input targets.
state corresponding to model Qn. Lower and upper bounds to the controlled outputs.
The robust controller of the propylene/propane splitter results
The usual hierarchical structure that integrates the RTO and
from the solution to the following optimization problem:
MPC layers through the optimizing targets to the MPC layer and,
subsequently, to the distributed control system (DCS) layer
min V k QN 11 through the set points to the regulatory loops is shown in Figure 4.
Duk ; ysp;k QN ; dy;k QN ; du;k
VOLUME 92, JULY 2014 THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING 1219
Simulation Results
The implementation of the control and optimization structure
presented in the previous section on the propylene/propane splitter
described in the paper was tested through simulation. The system
studied is the propylene/propane splitter of the propylene
production unit of the Capuava Renery (RECAP/Petrobras)
located in So Paulo, Brazil. As described in a previous section,
the advanced control system considered has three manipulated
inputs and three controlled outputs. In order to simulate the control
structure based on the robust IHMPC, the linear models that locally
represent the system at the three different operating points
presented in Figure 2 were identied. Therefore, in this study
only three models constitute the multimodel set V on which the
robust controller is based. These points correspond to the following
operating conditions: Point 1 corresponds to the normal operating
condition, Point 2 corresponds to an operating condition near the
maximum production capacity and Point 3 corresponds to an
operating point where the ow rate to the heat pump system is near
the upper bound. The operating conditions corresponding to each
of these points are dened in Table 3. These operating points were
dened based on the process history database within the last
2 years. The three models considered (Table 4) show the transfer
Figure 4. Hierarchical control structure of the propylene system. functions, where the time constants are in minutes.
2 3
1:237 0:18exp3s 0:0928exp3s
6 32:2s 1 138:79s 1 126:54s 1 7
6 7
6 5 6 6 6 7
6 5:361 10 s 2:51 10 exp7s 1:252 10 exp4s 2:843 10 exp92s 7
6
GQ2 6 7
s 2 0:0225s 6:873 105
s 2 0:01439s 9:732 10 5 2 0:01465s 7:883 10 5 7
6 0:6s 7
6 5 5 7
4 0:0003005s 3:039 10 0:000122exp3s 2:11 10 exp7s 5
5 4 5
0:2s2 0:0129s 3:846 10 1:4s2 0:054s 6:75 10 s2 0:0141s 3:956 10
2 3
1:44 0:1684exp6s 0:1821exp10s
6 31:5s 1 159s 1 119:8s 1 7
6 7
6 7
6 7:657 105 s 3:69 106 exp7s 2:094 10 6
exp4s 3:501 10 6
exp98s 7
6
GQ3 6 7
s 2 0:017156s 6:439 10 5
s 2 0:01776s 9:857 10 5 2 0:013s 6:257 10 5 7
6 0:6s 7
6 5 5 7
4 0:0003423s 3:704 10 0:0001589 1:99 10 exp9s 5
0:2s2 0:0142s 4:986 105 1:4s2 0:0423s 4:194 104 s2 0:0184s 5:66 105
1220 THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING VOLUME 92, JULY 2014
340
Table 5. Output zones of the propylene/propane splitter
u1
320
Output ymin ymax
300
y1 (% level) 4 80 0 200 400 600 800 1000 1200 1400 1600 1800
y2 (% molar) 0 0.45 nT (min)
y3 (% molar) 0 2 35
u2
Table 6. Input constraints of the propylene/propane splitter 30
0 200 400 600 800 1000 1200 1400 1600 1800
Input ~umax umin umax nT (min)
300
u1 (ton/h) 0.15 220 350
u2 (ton/h)
u3
0.02 10 45 280
u3 (ton/h) 0.13 200 320
260
0 200 400 600 800 1000 1200 1400 1600 1800
nT (min)
In this system, the manipulated variables correspond to: u1 the
heat pump ow rate (FC03), u2 the feed ow rate (FC01) and u3 Figure 6. Manipulated inputs ( IHMPC), ( RIHMPC),
the reux ow rate (FC02). The controlled variables are: y1 the ( Optimizing targets).
liquid level in the bottom reboiler, y2 the propane molar % in the
propylene stream and y3 the propylene molar % in the propane
stream. nominal model. The closed loop simulation began at the steady
The output and input constraints, as well as the bounds in the state corresponding to Point 3 as dened in Table 3, which
input increments, are shown in Tables 5 and 6, respectively. corresponds to input u0 [302 30 268] and to output y0 [42 0.5 1]
The following tuning parameters are considered for the controllers that are read from the dynamic simulation at the beginning of the
simulated: Dt 1 min, m 3, Qy diag(6 25 2), R diag(0.5 3 0.5), test. The initial steadystate corresponds to feco 14 800 $/h. Then
Qu diag(0.1 10 1), Sy 107diag(1 10 1), Su 104diag(0.1 100 1). with the assumed market conditions and available feedstock to the
The reading and writing of data from and to the dynamic propylene distillation system, ROMeo computes a new optimum
simulation, is performed through the timer function of Matlab. The operating point and denes the optimum targets to the MPC. These
time period for the reading of data from Dynsim to Matlab was set input targets are udes [329.6 34 294.8], which corresponds to an
equal to 5 s, while the sampling time for the innite horizon MPC is increase of the feed ow rate while the heat pump and reux ow
1 min and the control action is computed based on the average of rates are minimized. At this operating point the value of the
the last 12 readings of the outputs. This is also the time period in economic function is feco 16 400 $/h.
which the input setpoints calculated by the MPC algorithms are The simulation of the dynamic behaviour of the propylene/
written on the PID controllers of the regulatory layer of the propane splitter starting from the initial state towards the optimum
dynamic simulation. In addition, the transfer of data from ROMeo steadystate dened by the RTO with the nominal IHMPC are
to Matlab and Dynsim is set to be performed at a much larger time shown in Figures 5 and 6 in blue lines. From these gures, it is easy
period of several hours. For this purpose, the export function of to realize that the controller is able to stabilize the plant and to
OPC EDI was used, and in the same way, the reading of data was maintain the controlled variables inside their respective zones,
done using the import and download functions. while the manipulated variables are driven to their desired targets.
Therefore, the implementation of the advanced control and The responses are quite slow and the system takes about 24h to
optimization in the propylene distillation column was tested reach the new steady state. In Figure 7 it is represented in blue
considering initially the innite horizon MPC using G(Q1) as the lines, the behaviour of the instantaneous and the accumulated
economic objective along the response time of the propylene
system with the IHMPC. In the computation of the economic
60
function, it is assumed that when the propylene product is out of
specication it can only be valued as propane, which has a lower
y1
40
market value.
20
0 200 400 600 800 1000 1200 1400 1600 1800 At this point, it should be interesting to note that the new
nT (min) operating point corresponding to the optimum economic point is
0.5 quite far from the initial operating point where the nominal model
was obtained. Certainly, model uncertainty is signicant at this
y2
5
other two models represented in Table 4 to account for the multi
0 model uncertainty. The system starts from the same steadystate as
0 200 400 600 800 1000 1200 1400 1600 1800
nT (min) in the previous simulation case with IHMPC and the optimum
targets dened by ROMeo are also the same as in case of the
Figure 5. Controlled outputs ( IHMPC), ( RIHMPC). IHMPC, as well as the input and output bounds.
VOLUME 92, JULY 2014 THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING 1221
5
x 10
20000 4.5
15000 3.5
10000 2.5
($/h)
($)
2
eco
eco
f
5000 1.5
f
0 0.5
5000 0.5
0 500 1000 1500 0 500 1000 1500
nT (min) nT (min)
30
Figure 5 compares the system outputs for RIHMPC (red line)
alongside the responses for IHMPC (blue line), and the output
y1
25
zones (dashed line) for the total simulation time (1800 min). It is
20
clear that the two controllers have acceptable performances, but 1800 2000 2200 2400 2600 2800 3000 3200 3400 3600
when comparing the IHMPC and RIHMPC performances, it is easy nT (min)
to see that the robust controller has a better performance than the 1
IHMPC. The robust controller drives the controlled variables to
y2
0.5
their control zones faster than the nominal IHMPC. As a result,
with the robust controller, the controlled output y2, which 0
corresponds to the most important process specication of the 1800 2000 2200 2400 2600 2800 3000 3200 3400 3600
nT (min)
propylene/propane splitter and started at a point outside its control
4
zone, was brought to inside the control zone very rapidly and
maintained almost all the simulation time in its control zone.
y3
2
However, with the IHMPC, y2 violates the upper bound of its
control zone for a signicant period of time. Figure 6 compares the 0
1800 2000 2200 2400 2600 2800 3000 3200 3400 3600
calculated inputs for the two controllers and also shows the nT (min)
optimizing input targets (green line) calculated by ROMeo. It is
clear, that both controllers are capable of driving the propylene/ Figure 8. Controlled outputs ( IHMPC), ( RIHMPC) for the feed
propane splitter to the optimum operating point, but with the disturbance.
RIHMPC, the inputs u1 (heat pump ow rate) and u3 (reux ow
rate) are driven faster to their targets, which corresponds to a better
transient economic performance (Figure 7). Observe that the
estimated accumulated economic benet for the strategy based on
the robust controller is nearly US$ 200 103 higher than the benet 330
for the strategy based on the nominal controller.
u1
325
It was clear from the simulation results presented above that, for
the propane/propylene splitter, the nominal IHMPC shows a 320
1800 2000 2200 2400 2600 2800 3000 3200 3400 3600
nT (min)
34.1
Table 7. Feed molar composition (disturbance)
u2
34
Component % Molar fraction
33.9
1800 2000 2200 2400 2600 2800 3000 3200 3400 3600
Ethane 0.0118
nT (min)
Propylene 50.37
298
Propane 48.69
iButene 0.3895
u3
296
1Butene 0.0705
cis2Butene 0.0386 294
1800 2000 2200 2400 2600 2800 3000 3200 3400 3600
trans2Butene 0.0386 nT (min)
1,3Butadiene 0.0134
iButane 0.3442 Figure 9. Manipulated inputs ( IHMPC), ( RIHMPC),
Butane 0.0386 ( optimizing targets) for the feed disturbance.
1222 THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING VOLUME 92, JULY 2014
5
x 10
20000 3
2.5
15000
2
10000
($/h)
($)
1.5
eco
eco
f
5000
f
0
0.5
5000 0
2000 2500 3000 3500 2000 2500 3000 3500
nT (min) nT (min)
Figure 10. Instantaneous and accumulated economic function ( IHMPC), ( RIHMPC) for the feed disturbance.
reasonable performance, but the highnonlinearity of the propyl- which represents the most valued product and the economic
ene distillation column may not be properly dealt with by this function is increased again (Figure 10). As in the rst simulation
controller if strict specications of product quality are enforced. case, the robust controller has a better economic performance,
This is evidenced by the long period of time that the controlled which corresponds to an accumulated benet nearly US$ 70 103
variables may remain outside of the control zones when a change higher than the nominal controller.
of the operating point is required. While the controlled outputs are
outside their control zones, the propylene stream will be out of
specication and cannot be valued as the propylene product to be CONCLUSION
commercialized. Meanwhile, the robust controller showed a better In this work, the implementation of two advanced control and
performance and does not seem to be signicantly affected by the optimization strategies based on the nominal IHMPC and on the
process nonlinearities since it considers plant models correspond- robust IHMPC in a propylene production unit were studied. The
ing to different operating points and is able to stabilize the plant study was based on the commercial dynamic simulation software
faster than the nominal controller. Dynsim integrated to the real time process optimizer ROMeo and
The second simulation experiment started when the plant was the real time facilities of Matlab. The robust control structure
already stabilized at time 1800 min and a disturbance, which was assumes that model uncertainty can be represented as a discrete
unknown by the controller, was introduced in the feed of the set of models (multimodel uncertainty), each one corresponding
propylene/propane splitter. The new feed molar composition to a possible operating point of the system. Representative
represented in Table 7 is signicantly different from the initial feed simulation examples were presented, leading to the conclusion
composition represented in Table 1. that, for the propane/propylene splitter, the robust controller
Assuming that this disturbance was known by ROMeo, a new shows a better performance when compared with the nominal
optimum steady state was calculated and corresponds to the one. Since this better performance was achieved with a reduced
following input targets udes [325.7 34 295.4]. Analogously to the set of only three models to represent the separation system, we
rst simulation case, the controllers should be able to reject the can conclude that the implementation of the control structure
disturbance and drive the distillation column to the new optimum based on the robust IHMPC in the real process is feasible in terms
operating point. The results of the simulation based on the IHMPC of the required computing time. The solution of the NLP problem
(blue lines) and RIHMPC (red lines) are shown in Figures 810. that produces the robust MPC considered in this work takes
When comparing the IHMPC and RIHMPC performances, it is about 1.44 s, which is only a small fraction of the controller
easy to realize that, as in the rst simulation case, the robust sampling time.
controller has a better performance than the nominal IHMPC,
because the controlled variables respond faster and were kept
ACKNOWLEDGEMENTS
inside the control zones more efciently. For instance, controlled
output y2, which started the simulation at a point at the upper Support for this work was provided by CNPq under grant 160465/
bound of its zone control, tends to leave the control zone for a 20125.
period of time, but it was brought back to the zone by the RIHMPC
faster than by the IHMPC. Figure 9 compares the inputs for the
REFERENCES
two controllers and also shows the optimum input targets (green
line) provided by ROMeo and corresponding to the feed column [1] D. E. Kassman, T. Badgwell, R. B. Hawkins, AIChE J. 2003,
composition (Table 7). As a consequence of the new feed 45, 1007.
composition, the ow rate of the propylene product has to be [2] C. R. Cutler, R. B. Hawkins, Constrained multivariable
decreased while the propane product ow rate is increased. The control of a hydrocracker reactor, in Proceedings of the
consequence is that the economic function tends to decrease, but American Control Conference, Minneapolis, USA, 1987,
the optimizer manages to maximize the propylene production, 1014.
VOLUME 92, JULY 2014 THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING 1223
[3] O. L. Carrapio, M. M. Santos, A. C. Zanin, D. Odloak,
Application of the IHMPC to an industrial process system,
in 7th IFAC International Symposium on Advanced Control
of Chemical Processes, 2009, 7, 851.
[4] C. Pinheiro, J. Fernandes, L. Domingues, A. Chambel, I.
Graa, N. Oliveira, H. Cerqueira, F. Ribeiro, Ind. Eng. Chem.
Res. 2012, 51, 1.
[5] C. R. Cutler, B. L. Ramaker, Dynamic matrix controla
computer control algorithm, in Joint Control Conference,
San Francisco, USA, 1980.
[6] M. A. Rodrigues, D. Odloak, Comp. Chem. Eng. 2003, 27,
1113.
[7] O. L. Carrapio, D. Odloak, Comp. Chem. Eng. 2005, 29, 1089.
[8] B. F. Santoro, D. Odloak, J. Process Control 2012, 22, 1209.
[9] S. J. Qin, T. A. Badgwell, Control Eng. Pract. 2003, 11, 733.
[10] T. A. Badgwell, Int. J. Control 1997, 68, 797.
[11] J. H. Lee, B. L. Cooley, Automatica 2000, 36, 463.
[12] M. V. Kothare, V. Balakrishnan, M. Morari, Automatica 1996,
32, 1361.
[13] Z. Wan, M. V. Kothare, J. Process Control 2002, 12, 763.
[14] L. A. Alvarez, D. Odloak, Comp. Chem. Eng. 2010, 34, 1937.
[15] A. H. Gonzles, D. Odloak, Robust Model Predictive Control
for Time Delayed Systems With Optimising Targets and Zone
Control, Robust Control, Theory and Applications, 2011.
ISBN 9789533072296.
[16] W. Y. Svrcek, D. P. Mahoney, B. R. Young, A RealTime
Approach to Process Control, John Wiley & Sons, 2000. ISBN
0471804525.
[17] N. Alsop, J. Ferrer, Chem. Eng. World 2004, 39, 42.
[18] N. Alsop, J. Ferrer, Step Test Free Advanced Process Control
Implementation Using dynamic simulation, in Proceedings
of AIChE annual meeting, Orlando, USA, 2006.
[19] D. Bruinsma, S. Spoelstra, Heat Pump in Distillation, in
Distillation and Absorption Conference, Eindhoven, The
Netherlands, 2010.
[20] A. H. Gonzles, D. Odloak, J. Process Control 2009, 19, 110.
[21] D. Odloak, AIChE J. 2004, 50, 1824.
1224 THE CANADIAN JOURNAL OF CHEMICAL ENGINEERING VOLUME 92, JULY 2014