Dynamic Considerations in the Synthesis

of Self-Optimizing Control Structures

Michael Baldea
Praxair Technology Center, Praxair Inc., Tonawanda, NY 14150

Antonio Araujo and Sigurd Skogestad

Dept. of Chemical Engineering, Norwegian University of Science and Technology, N-7491 Trondheim, Norway

Prodromos Daoutidis
Dept. of Chemical Engineering and Materials Science, University of Minnesota, Minneapolis, MN 55455

DOI 10.1002/aic.11470
Published online May 8, 2008 in Wiley InterScience (www.interscience.wiley.com).

This work builds on our prior results to develop novel control structure design prin-
ciples for integrated plants featuring multiple time scale dynamics. Specifically, the
concept of self-optimizing control can be used to identify the variables that must be
controlled to achieve acceptable economic performance during plant operation. This
approach does not, however, provide guidelines on control structure design and
control loop tuning; a detailed controllability and dynamic analysis is generally
needed to this end. In this work, we employ a singular perturbation-based framework,
which accounts for the time scale separation present in the open loop dynamics of
integrated plants, to identify the available controlled and manipulated variables in
each time scale. The resulting controller design procedure thus accounts for both
economic optimality and dynamic performance. The developed concepts are sub-
sequently successfully applied on a reactor-separator process with recycle and purge.
Ó 2008 American Institute of Chemical Engineers AIChE J, 54: 1830–1841, 2008
Keywords: self-optimizing control, singular perturbations, dynamic analysis

Introduction authors6,7,8,9,10 have indirectly assumed this time scale multi-

Modern chemical plant designs increasingly rely on tight plicity to propose tiered control structures, featuring at least
integration between process units, using heat and material two levels of control action: a primary layer addressing
recycle streams, to reduce capital and operating costs. In inventory and temperature control at the unit level and pro-
integrated plants, economic gains come, however, at the price viding stability in operation, and a supervisory layer, acting
of an increased dynamic complexity and control challenges. over a slower time scale, that targets the control objectives at
The complex dynamic behavior of integrated processes has the plant level, such as product purity and production rate.
been characterized in several works.3,4,5 In particular, inte- In our previous work,2 we relied on singular perturbation
grated processes have long been recognized to exhibit a arguments to rigorously characterize the nonlinear dynamic
dynamic behavior that spans multiple time scales. Many behavior of integrated processes with large recycle streams
and purge streams, demonstrating that it features three time
scales, associated, respectively, with the evolution of the
Correspondence concerning this article should be addressed to S. Skogestad at states of the individual units, with the evolution of the total
[email protected].
material holdup of the network, and with the impurity levels
Ó 2008 American Institute of Chemical Engineers in the network. We derived reduced order nonlinear models

1830 July 2008 Vol. 54, No. 7 AIChE Journal

 which accounts for the time scale separation present in
the open loop dynamics of integrated plants, to identify
the available controlled and manipulated variables in each
time scale. The resulting controller design procedure,
thus, accounts for both economic optimality and dynamic
performance.
This article is structured as follows: a brief description of
self-optimizing control is provided in the next section, suc-
ceeded by an account of singular-perturbation based model
Figure 1. Generic reactor-separator process network reduction and controller design. A motivating case study is
with large recycle and purge. introduced, and the proposed controller synthesis approach is
then presented. Finally, the newly developed framework is
for the dynamics of the process in the three time scales, and demonstrated via simulations.
our analysis aided in delineating a multitiered controller
design framework, using three layers of control action to
address objectives, both at the unit and at the network level, Self-Optimizing Control
using the manipulated inputs identified to be available in Self-optimizing control1 is when one can achieve an
each time scale. acceptable loss with constant setpoint values for the con-
Our previous work also addressed the economic issues trolled variables without the need to reoptimize when distur-
encountered in the operation of process networks by propos- bances occur (real time optimization).
ing the concept of self-optimizing control.1 Specifically, To quantify this more precisely, we define the (economic)
modern plants tend to include optimization and scheduling loss L as the difference between the actual value of a given
layers atop the supervisory control system, in order to ensure cost function and the truly optimal value, that is
economic optimality. With this approach, however, economic
Lðu; dÞ ¼ Jðu; dÞ
Jopt ðdÞ (1)
performance is obtained at the price of computationally ex-
pensive real-time optimization calculations. Self-optimizing Truly optimal operation corresponds to L 5 0, but in general
control1 aims to alleviate this issued by identifying a set of L [ 0. A small value of the loss function L is desired as it
controlled outputs which, when maintained at their setpoints, implies that the plant is operating close to its optimum. The
ensure that the economic losses affecting the operation of the central issue to self-optimizing control is not finding optimal
plant in the presence of disturbances remain at an acceptable set points, but rather finding the right variables to keep con-
level. stant. The precise value of an ‘‘acceptable’’ loss varies from
This contribution draws on our aforementioned work, uti- case to case, and the selection is made on the basis of engi-
lizing our ideas in1 to identify the controlled outputs that neering and economic considerations.
ensure the near-optimal operation of an integrated process In1 we recommended that a controlled variable c suitable
that features multiple time scale dynamics. Self-optimizing for constant set point control (self-optimizing control) should
control does not, however, provide information concerning meet the following requirements:
the selection of the manipulated inputs to be used to control R1. The optimal value of c should be insensitive to dis-
the desired outputs. In this work, we rely on the analysis in,2 turbances, i.e., copt(d) depends only weakly on d.

Figure 2. Reactor-separator process.

AIChE Journal July 2008 Vol. 54, No. 7 Published on behalf of the AIChE DOI 10.1002/aic 1831
 Table 2. Selected Candidate Controlled Variables

Table 1. Dynamic Model of the Reactor-Separator with Candidate Notation

Recycle Network Reactor holdup (Reactor pressure) Preactor
Vapor mole fraction of A in the reactor yA,R
Differential equations
Vapor mole fraction of I in the reactor yI,R
dMR Vapor mole fraction of A in the separator yA
¼ Fo þ R
F Vapor mole fraction of I in the separator yI
dt
Liquid mole fraction of A in the separator xA
dyA;R 1 Liquid mole fraction of I in the separator xI
¼ ½Fo ðyA;o
yA;R Þ þ RðyA
yA;R Þ
k1 MR yA;R 
dt MR Liquid mole fraction of B in the separator xB
Separator pressure Pseparator
dyI;R 1 Flow out of the reactor (Valve opening) zF
¼ ½Fo ðyI;o
yI;R Þ þ RðyI
yI;R Þ
dt MR Liquid flow out of the separator (Valve opening) zL
Purge flow (Valve opening) zP
dMV Recycle flow R
¼F
R
N
P
dt Compressor power WS
dyA 1
¼ ½FðyA;R
yA Þ
NA þ yA N
dt MV
Table 3. Prices for the Components of the Objective
dyI 1
¼ ½FðyI;R
yI Þ
NI þ yI N Function in (6)
dt MV
Price Unit Value
dML
¼N
L pL $/mole 2.55
dt
pP $/mole 0.50
dxA 1 pFo $/mole 1.50
¼ ½NA
xA N
dt ML pW $/kW 0.08

dxI 1
¼ ½NI
xI N
dt ML
R2. The value of c should be sensitive to changes in the
Algebraic equations manipulated variable u, i.e., the gain from u to y should be
MR Rgas Treactor large.
Preactor ¼ R3. For cases with two or more controlled variables, the
Vreactor
selected variables in c should not be closely correlated.
MV Rgas Tseparator
Pseparator ¼ R4. The variable c should be easy to measure and con-
ðVseparator
MqLL Þ trol.

M During optimization some constraints are found to be
PSA L
NA ¼ KA a yA
xA active in which case the variables they are related to must be
Pseparator qL
selected as controlled outputs, since it is optimal to keep

PSI M them constant at their setpoints (active constraint control).
L
NI ¼ KI a yI
xI
Pseparator qL The remaining unconstrained degrees of freedom must be
h iM
fulfilled by selecting the variables (or combination thereof)
PSB L which yield the smallest loss L with the active constraints
NB ¼ KB a ð1
yA
yI Þ
ð1
xA
xI Þ
Pseparator qL implemented.
N 5 NA 1 NB 1 NI
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Multiple Time Scale Dynamics of Integrated
F ¼ Cvf zf Preactor
Pseparator Process Networks
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P ¼ Cvp zp Pseparator
Pdownstream In our previous work,2 we have demonstrated that the
dynamic model of process networks, such as that in Figure 1,
Ws
R¼ h i featuring a reaction and a separation section that contain a
1 cRgas Tseparator 3Preactor;max c
1
e c
1 ð Pseparator Þ
c
1

Where:
– MR, MV, and ML denote the molar holdups in the reactor and separator Table 4. Disturbances to the Process
vapor and liquid phases, respectively.
Nominal Disturbance (D)
– Rgas is the universal gas constant.
CP

c¼ is assumed constant: D1 Feed rate (Fo) [mole/min] 100 120 (120%)
CV D2 Feed rate (Fo) [mole/min] 100 210 (210%)
– Cvf and Cvp are the valve constants.
– Pdownstream is the pressure downstream the system (assumed constant). D3 Composition of inerts 0.02 10.004 (120%)
– e is the compressor efficiency. in the feed (yI,o)
– Preactor,max is the maximum allowed pressure in the reactor. D4 Product purity (xB) 0.8711 20.0436 (25%)
– The compressor and valves are modeled as first order systems, with time D5 Product purity (xB) 0.8711 10.0436 (15%)
constants scompressor 5 10 min and svalve 5 1 min. D6 Composition of product B 0 10.02†
– The flowrate L of the liquid product is used to control the separator liquid in the feed (yB,o)
level.
†
Reduction of yA,o by the same amount.

1832 DOI 10.1002/aic Published on behalf of the AIChE July 2008 Vol. 54, No. 7 AIChE Journal
 Table 5. Optimization Subject to the Disturbances Considered in Table 4
Unit Nominal D1 D2 D3 D4 D5 D6
Profit $/min 98.29 116.80 88.84 97.01 98.87 97.18 98.39
MR mole 6449 6449 6449 6449 6449 6449 6449
yA,R mole/mole 0.2625 0.3140 0.2366 0.2610 0.2496 0.2751 0.2564
yI,R mole/mole 0.5542 0.5023 0.5803 0.5647 0.5803 0.4970 0.5620
MV mole 12.36 12.01 12.32 12.32 12.31 12.30 12.32
yA mole/mole 0.2792 0.3323 0.2515 0.2758 0.2620 0.2934 0.2724
yI mole/mole 0.6234 0.5518 0.6608 0.6280 0.6730 0.5447 0.6326
ML mole 74150 74017 74236 74162 74406 73656 74173
xB mole/mole 0.8711 0.8711 0.8711 0.8711 0.8275 0.9147 0.8711
xA mole/mole 0.1288 0.1288 0.1288 0.1288 0.1724 0.0853 0.1288
xI mole/mole 0.0000557 0.0000527 0.0000555 0.0000555 0.0000554 0.0000551 0.0000554
Preactor Pa 2000000 2000000 2000000 2000000 2000000 2000000 2000000
Pseparator Pa 540094 453815 599573 546729 770258 340234 553578
NA mole/min 12.48 14.91 11.25 12.40 16.74 8.22 12.48
NI mole/min 0.03 0.03 0.03 0.03 0.03 0.03 0.03
NB mole/min 84.33 100.77 76.04 83.80 80.31 88.14 84.37
N mole/min 96.84 115.70 87.33 96.23 97.08 96.38 96.89
F mole/min 871.21 1289.62 716.16 954.74 704.46 1100.46 867.69
L mole/min 96.84 115.70 87.33 96.23 97.08 96.38 96.89
P mole/min 3.16 4.30 2.67 3.77 2.92 3.62 3.11
R mole/min 771.21 1169.62 626.16 854.74 604.46 1000.46 767.69
zF – 0.25 0.36 0.20 0.27 0.21 0.30 0.24
zP – 0.19 0.31 0.15 0.23 0.12 0.36 0.19
WS kW 171.96 291.20 140.37 189.08 105.92 296.32 168.45

recycle loop with the recycle flow rate R, and relying on a such that, at steady state, we have e1 5 Fo,s/Rs  1, and
purge stream of flow rate P to eliminate any impurities pres- that, conversely, the flow rate of the purge stream is signifi-
ent in small quantities, is captured by a stiff system of equa- cantly lower than the network throughput, i.e., e2 5 Ps/Fo,s
tions of the form  1.
Using nested singular perturbation arguments, we demon-
d 1 strated that the dynamic behavior of the process network in
x ¼ fðx; us Þ þ Gl ðxÞul þ e2 gP ðxÞup (2)
dt e1 Figure 1 features three components, that evolve over three
distinct time scales. Specifically:
 a fast component, evolving in the fast time scale s 5 t/
l
In Eq. 2, x [ Rn is the state vector, with ul [ Rm being the
vector of scaled input variables corresponding to the flows e1, described by an equation system of the form
rates of the internal streams within the recycle loop, us [ Rm
being the vector of scaled input variables corresponding to d
the flow rates of the streams outside the recycle loop x ¼ Gl ðxÞul (3)
ds1
(excluding the purge stream), and up being a scaled input
variable corresponding to the flow rate of the purge stream;
f(x,us), gP(x) are n-dimensional vector functions, and Gl(x) is The ‘‘stretched’’ time scale s1 is in the order of magnitude of
a n 3 ml- dimensional matrix. the time constants of the individual process units with large
Equation 2 is developed based on the assumption that the material throughput that are part of the recycle loop, and
flow rates of the recycle loop streams are are of comparable thus the model in Eq. 3 effectively captures the dynamics of
magnitude, and much higher than the network throughput, thes individual process units.

Table 6. Loss Evaluation ($/min) for Selected Candidate Variables Based on Table 2†
Candidate D1 D2 D3 D4 D5 D6 Avg.
yA,R Inf 0.36661 0.02890 0.17707 Inf 0.09963 Inf
yI,R Inf 0.01265 0.00456 0.00939 Inf 0.00199 Inf
yA Inf 0.22442 0.01111 0.12012 Inf 0.02979 Inf
yI Inf 0.22490 0.01116 0.37894 Inf 0.02979 Inf
xA 0.00003 0.00000 0.01111 0.57275 Inf 0.02981 Inf
xI 0.00003 0.00000 0.00000 0.00000 0.00009 0.00004 0.00003
Pseparator Inf 0.13732 0.01110 0.56009 Inf 0.02906 Inf
R 0.07558 0.00730 0.00263 0.00813 0.02706 0.00005 0.02013
zF 0.07967 0.00740 0.00266 0.00512 0.02170 0.00007 0.01944
zP Inf 0.30234 Inf 0.54822 Inf 0.02980 Inf
WS 0.13227 0.01391 0.00241 0.03465 0.13660 0.00006 0.05332
†
Inf means infeasible operation.

AIChE Journal July 2008 Vol. 54, No. 7 Published on behalf of the AIChE DOI 10.1002/aic 1833
 Table 7. Control Structure Selection Based on the Singular of manipulated inputs (respectively, ul, us and up), that act
Perturbation Analysis. upon and can be used to address control objectives in the
Time scale Controlled output Manipulation respective time scale.
By way of consequence, process networks featuring signif-
Fast MR (Preactor) F (zf) icant material recycling, as well as a purge stream for elimi-
Fast MV (Pseparator) R (zp)
Intermediate ML L nating impurities, lend themselves naturally to a hierarchical
Intermediate xb MR,setpoint (Preactor,setpoint) control structure, featuring three layers of control action:
Slow yI,R P  Control objectives at the unit level should be addressed
in the fast time scale, using the large flow rates of the inter-
nal material streams ul, as manipulated inputs.
 an intermediate component, evolving in the time scale t  The control of the network wide objectives (such as
d product purity and production rate), should be undertaken in
x ¼ ~fðx; us Þ þ e2 ~
gP ðxÞup the intermediate time scale, using the flow rates us of the ma-
dt (4)
0¼G ~ l ðxÞul ðxÞ terial streams outside the recycle loop as manipulated inputs.
 The impurity levels in the network should be regulated
with 0 5 G̃l(x)ul(x) being the linearly independent con- using the flow rate up of the purge stream, over a long time
straints that denote the quasi-steady state of the fast dynam- horizon (slow time scale).
ics in the time scale t. Equation 4 is a description of the core In the following section, we demonstrate how these results,
dynamics of the process network, that is due to the presence along with the concept of self-optimizing control reviewed
of the recycle loop with large recycle flow rate. earlier, can be successfully fused in a control design proce-
 a slow component, evolving in the compressed time dure that accounts for both economic optimality and dynamic
scale y 5 e2t, of the general form performance.
d ^
x¼^ gP ðxÞup þ BðxÞ
dh Case Study on Reactor-Separator with
^
0 ¼ GðxÞu s
ðxÞ (5) Recycle Process
~
0 ¼ G ðxÞu ðxÞ
l l In this section, we present a case study that considers a
reactor-separator network, interconnected via a recycle
Note that, in Eq. 5, not only the fast dynamics, but also the stream with a large flow rate (compared to the network
intermediate dynamics of the network are considered to be at throughput), and the inert impurities present in the feed are
a quasi-steady state. The slow component captures the dy- eliminated by purging. The generic process was studied in,2
namics associated with the presence of small amounts of and for this article the pressure-flow relations for the flows
feed impurity that are removed by the small purge stream. F, P, and R and economic data were added.
Note that description of each of the models of the dynam- We rely on this representative example to develop and
ics in the fast, intermediate, and slow time scales described illustrate a hierarchical controller synthesis procedure, that
previously (respectively, Eqs. 3,4,5), features a distinct group accounts for both the time scale separation in process dynam-

Figure 3. Original configuration based on singular perturbation with control of xB, Pseparator, and yI,R.

1834 DOI 10.1002/aic Published on behalf of the AIChE July 2008 Vol. 54, No. 7 AIChE Journal
 Figure 4. Simplest self-optimizing configuration with control of xB, Preactor, and R.

ics, and for economic criteria, in order to ensure dynamic per- The objective is to ensure a stable operation while control-
formance and economic optimality of the closed-loop system. ling the purity of the product xB.
k1
A first-order reaction takes place in the reactor, i.e., A ! B.
In the condenser-separator, the interphase mole transfer rates
Process description and modeling
for the components A, B, and I are governed by rate expres-
The process consists of a gas-phase reactor and a con- PS
sions of the form Nj ¼ Kj aðyj
Pj xj Þ Mq L , where Kj a represents
denser-separator that are part of a recycle loop (Figure 2). A L

low single-pass conversion requires that a large (with respect the mass transfer coefficient, yj the mole fraction in the gas
to the feed flow rate) recycle flow rate R be used in order to phase, xj the mole fraction in the liquid phase, PSj the satura-
achieve the desired purity of the product B. The feed stream tion vapor pressure of the component j, P the pressure in the
contains a small amount of an inert, volatile impurity yI,o condenser, and qL the liquid density in the separator. A com-
which is removed via a purge stream of small flow rate P. pressor drives the flow from the separator (lower pressure) to

Figure 5. Modification of Figure 3: Constant pressure in the reactor instead of in the separator.

AIChE Journal July 2008 Vol. 54, No. 7 Published on behalf of the AIChE DOI 10.1002/aic 1835
Figure 6. Final structure from modification of Figure 5: Set recycle flow rate (R) constant instead of the inert
composition (yI,R).

the reactor. Moreover, valves with openings zf and zp allow where pL, pP, pFo, and pW are the prices of the liquid product
the flow through F and P, respectively. Assuming isothermal L, purge P (here assumed to be sold as fuel), feed Fo, and
operation (or, equivalently, perfect temperature control), the compressor power Ws, respectively (see also Table 3 for
dynamic model of the system has the form given in Table 1. cost-related information).

Economic approach to the selection of controlled Identification of important disturbances

variables: Self-optimizing control computations We will consider the disturbances and process changes listed
Degree of Freedom Analysis. The open loop system has in Table 4. Specifically, we account for the possibility of varia-
three degrees of freedom at steady state, namely the position tions in the feed flow rate and composition (including the possi-
of the valve at the outlet of the reactor (zF), the position of bility of having a small quantity of product present in the feed),
the purge valve (zP), and the compressor power (Ws). as well as for possible changes in the product purity requirement.
Table 2 lists the candidate controlled variables considered
in this example. With three degreesof freedom and 14 candi- Optimization
date controlled outputs, there are 14 3 ¼ 3!11! ¼ 364 possible
14!
Two constraints are active at the optimum throughout the
ways of selecting the control configuration, which constitutes calculations (each of which corresponds to a different dis-
a rather large number if we consider the dimension of the turbance), namely the reactor pressure Preactor, at its upper
problem. Therefore, in order to avoid evaluation of each one bound, and the product purity xB, at its lower bound (Table
of these possible configurations, we determine whether there 5). These consume two degrees of freedom, since it is opti-
are active constraints during operation. mal to control them at their setpoint,11 leaving one uncon-
strained degree of freedom.
Definition of optimal operation
Unconstrained variables: Evaluation of the loss
The following profit is to be maximized
In order to identify the remaining controlled variable, we
ð
JÞ ¼ pL L þ pP P
pFo Fo
pW Ws (6) evaluate the steady-state economic loss incurred in the pres-

subject to Table 8. Final Control Structure Based on Dynamic Analysis

and Optimality
Preactor  2000 kPa
Time scale Controlled output Manipulation
Pseparator  1000 kPa
Fast MR (Preactor) F (zf)
xB  0:8711 (7) Fast MV (Pseparator) R (Ws)
Intermediate ML L
WS  300 kW Intermediate xb MV,setpoint (Pcondenser,setpoint)
zF ; zP 2 ½0; 1 Slow R (WS) zp

1836 DOI 10.1002/aic Published on behalf of the AIChE July 2008 Vol. 54, No. 7 AIChE Journal
 Table 9. Controller Tuning Parameters for Each Control Configuration in Figures 3–6
Feedback loop Figure 3 Figure 4 Figure 5 Figure 6
ML 3 L Kc 5 0.001 Kc 5 0.001 Kc 5 0.001 Kc 5 0.001
Preactor 3 zF Kc 5 0.001 Kc 5 0.001 Kc 5 8 Kc 5 8
sI 5 10 sI 5 16 sI 5 16
Pseparator 3 WS Kc 5 0.0013 Kc 5 0.0013 Kc 5 1.31025
R 3 zP Kc 5 0.005
sI 5 1000
R 3 WS Kc 5 0.01
sI 5 2
yI,R 3 zP Kc 5 10 Kc 5 10
sI 5 500 sI 5 410
xB 3 zP Kc 5 100
sI 5 1000
xB 3 Preactor,sp Kc 5 8
sI 5 100
xB 3 Pseparator,sp Kc 5 4 Kc 5 4
sI 5 100 sI 5 100

Time constants are in minutes.

ence of disturbances, when the candidate controlled variable, out the optimizations shown in Table 5. However, composi-
in addition to the two active constraints, is perfectly con- tion measurements have large dead times and are unreliable
trolled (i.e., it is kept constant). and, we, therefore, disregard this candidate as the potential
Table 6 shows the results of the loss evaluation. We can self-optimizing variable.
see that the smallest average loss was found for the liquid Two other candidates which show smaller average losses
mole fraction of inert in the separator (xI). This was some- are the recycle flow rate R and valve opening zF, with aver-
how expected since its value is essentially constant through- age losses of 0.02013 and 0.01944 $/min, respectively. As

Figure 7. Closed-loop responses for configuration in Figure 3.

AIChE Journal July 2008 Vol. 54, No. 7 Published on behalf of the AIChE DOI 10.1002/aic 1837
 Figure 8. Closed-loop responses for configuration in Figure 4.

the valve opening zF is typically required to address the con- sure, while maintaining the reactor pressure at its setpoint.
trol of the reactor pressure, we select the recycle flow rate R, This will be discussed later in this article.
which has an acceptable loss, as the unconstrained (self-opti-
mizing) controlled variable.
In summary, by the self-optimizing approach, the primary Control Configuration for Optimality and
Dynamic Performance
variables to be controlled are then y 5 [Preactor xB R] with
the manipulations u 5 [zF zP WS]. In addition, secondary The objective of this study is to explore how the configu-
controlled variables may be introduced to improve the rations suggested by the two different approaches can be
dynamic behavior of the process. With these variables, a merged to produce an effective control structure for the sys-
number of control configurations can be assigned and some tem. Thus, as a starting point, we employ the following two
of them will be assessed later in this article. ‘‘original’’ configurations: Figure 3 presents the original con-
figuration from the singular perturbation approach.2 Figure 4
depicts the simplest self-optimizing control configuration
Singular Perturbation Approach for the with control of the active constraints (Preactor and xB) and
Selection of Controlled Variables self-optimizing variable R.
According to the hierarchical control structure design pro- The configuration in Figure 4 does not account for opti-
posed in,2 based on the time scale separation of the system, mality and could give rise to infeasibility with respect to
the variables to be controlled, and their respective manipula- operating constraints on pressure. On the other hand, the
tions are given in Table 7. structure outlined in Figure 4 does not directly control the
In the prior work,2 economics were not a consideration. impurity levels in the network, and employs the flow rate of
Moreover, in the aforementioned configuration the reactor the purge streams to control the product purity, a solution
pressure is employed to control the purity of the product, and which, according to our previous results, could lead to poor
is evidently required and allowed to vary, which could lead, dynamic performance. These observations will be confirmed
in some cases, to the violation of the operating constraints by the simulation results presented later in this article.
included in the present problem formulation. A simple modi- Since one usually starts by designing the regulatory control
fication that allows the pressure constraint in the reactor to system, the most natural starting point is the configuration in
be satisfied entails controlling xB using the separator pres- Figure 3. The first evolution of this configuration is to

1838 DOI 10.1002/aic Published on behalf of the AIChE July 2008 Vol. 54, No. 7 AIChE Journal
 Figure 9. Closed-loop responses for configuration in Figure 5.

change the pressure control from the separator to the reactor Figure 3, the reactor pressure rises over the 2MPa bound
(Figure 5). In this case, both active constraints (Preactor and (Figure 7) when a setpoint increase for xB occurs. The
xB) are controlled in addition to impurity level in the reactor dynamic response in terms of the product purity—a key per-
(yI,R). formance indicator—is, however, very good. Moreover, the
The final modification towards buiding a self-optimizing aforementioned behavior is to be expected since the original
control structure is to change the primary controlled vari- configuration was based on varying the reactor pressure to
able from yI,R to the recycle flow rate R (Figure 6). The control the purity of the product.
latter evolution also ensures that the compressor power is With Preactor controlled with a controller with integral
controlled around its steady-state optimal value over a action (configuration of Figure 5), and manipulating the con-
longer time scale, using the flow rate of the purge stream denser pressure to control the product purity, a similar
to this end. The final control configuration is summarized dynamic response in xB is obtained (Figure 9). This is again
in Table 8. to be expected, since, as explained earlier, the structures
depicted in Figures 3 and 5 are dynamically similar. Note,
however, that in this case, tracking the purity setpoint as it
Results and Discussion increases at t 5 50 h requires a significant increase in the
Simulations were carried out to assess the dynamic per- energy consumption of the compressor (WS exceeds, in
formance of the control configurations proposed earlier. The effect, the 300 kW bound imposed in the problem formula-
tuning parameters for the controllers in each configuration tion), intuitively leading to a less than optimal profit.
are shown in Table 9. The simulation study considered two The proposed self-optimizing configuration of Figure 4,
major disturbances: a 10% drop in the feed flow rate (Fo whereby the controlled variables are selected based on eco-
from 100 to 90 mol/min) at t 5 10 h followed by a 5% nomics, results in a rather poor dynamic performance for the
increase in the setpoint for the product purity (xB from controlled variable xB as seen in Figures 8 and 11. The ex-
0.8711 to 0.9147) at t 5 50 h. planation lies in the fact that xB is controlled by the small
The results are found in Figures 7 through 10. flow rate P (using valve position zP), which leads to a slug-
Based on the aforementioned simulation results presented gish response. Note also that obtaining a dynamic perform-
above, notice that, in the case of the original system in ance in terms of xB comparable to that of the aforementioned

AIChE Journal July 2008 Vol. 54, No. 7 Published on behalf of the AIChE DOI 10.1002/aic 1839
 Figure 10. Closed-loop responses for configuration in Figure 6.

configurations entails using a high gain controller. Consider- scaled variables) in the configurations discussed above are,
ing the data in Table 9, the gain of the purity controller in respectively, Kc 5 8 and Kc 5 4. As a consequence, in the
the basic self-optimizing configuration is Kc 5 100, while response of Figure 8, the purge flow P is significantly
the gains of the purity controllers (expressed in terms of increased for an extended period of time.
Finally, the configuration in Figure 6 gives feasible opera-
tion with a good transient behavior and low compressor
energy consumption (Figure 10).
The developments above demonstrate that the approaches
proposed in our previous work1 and2 are complementary in
developing a control configuration at the network level,
that is both economically optimal and has good dynamic
performance.
The controlled outputs in the configuration illustrated in
Figure 6 were selected based on economic considerations,
that is, (1) the active constraints in the optimization calcula-
tions, and (2) the variables that ensure a minimum loss in the
presence of disturbances. The input-output pairings are based
on the time scale that each controlled output evolves in, and
on the manipulated inputs available in the respective time
scale.
Specifically, the pressure changers (valve and compressor)
that determine the large internal flowrates are used for the
fast regulation of the pressure/holdup in the reactor and con-
Figure 11. Closed-loop responses for the product pu- denser vapor phase. Subsequently, the setpoint of the con-
rity xB for the configurations in Figures 3 denser vapor phase pressure controller is used to control the
(solid), 4 (dash), 5 (dot) and 6 (dash-dot). product purity in a slower time scale. Finally, the economic

1840 DOI 10.1002/aic Published on behalf of the AIChE July 2008 Vol. 54, No. 7 AIChE Journal
analysis recommends that the original structure be modified controller design procedure that merges the aforementioned
in the slow time scale, using the small purge flowrate to reg- concepts, thereby accounting for both economic optimality
ulate the recycle flow at a set value, thereby keeping the and dynamic performance. Numerical simulation results indi-
compressor power consumption constant. cated that the resulting control system exhibited very good
Notice that the final control structure in Table 8 is not transient response characteristics, while maintaining the
only optimal, but also in complete agreement with the con- parameters of the system within the desired economic per-
trol structure design framework based on singular perturba- formance envelope.
tion analysis (Table 7). Specifically, the control configura-
tions are identical in the fast time scale. Moreover, both the
reactor and the condenser pressure setpoints are manipulated
inputs acting in the intermediate time scale, with the former
Literature Cited
being selected in the original configuration (Figure 3) and
the latter selected in the final configuration based on optimal- 1. Skogestad S. Plantwide control: The search for the self-optimizing
control structure. J Proc Contr. 2000;10:487–507.
ity considerations. Last, since the flow rate of the recycle 2. Baldea M, Daoutidis P. Control of integrated process networks–A
stream varies in the intermediate time scale, the purge stream multi-time scale perspective. Comp Chem Eng. 2007;31:426–444.
must be used in the slow time scale to reset the recycle flow 3. Mizsey P, Kalmar I. Effects of recycle on control of chemical pro-
rate. cesses. Comp Chem Eng. 1996;20:S883–S888.
Note also that final control configuration proposed earlier 4. Contou-Carrére MN, Baldea M, Daoutidis P. Dynamic precompensa-
tion and output feedback control of integrated process networks. Ind
is in agreement with Luyben’s rule of flow-controlling one Eng Chem Res. 2004;43:3528–3538.
of the streams in the recycle loop.12 However, this rule 5. Kiss AA, Bildea CS, Dimian AC, Iedema PD. Design of recycle sys-
should be applied with caution.13 tems with parallel and consecutive reactions by nonlinear analysis.
Ind Eng Chem Res. 2005;44:576–587.
6. Findeisen W, Bailey FN, Brdyś M, Malinowski K, Tatjewski P,
Conclusion Woźniak A. Control and Coordination in Hierarchical Systems.
John Wiley and Sons; 1980.
This work utilized our prior results to develop novel con- 7. Price RM, Georgakis C. Plantwide regulatory control design proce-
trol structure design principles for integrated plants featuring dure using a tiered framework. Ind Eng Chem Res. 1993;32:2693.
multiple time scale dynamics. Specifically, our concept of 8. Luyben ML, Tyreus BD, Luyben WL. Plantwide control design pro-
cedure. AIChE J. 1997;43:3161–3174.
self-optimizing control was employed to identify the varia- 9. Kothare MV, Shinnar R, Rinard I, Morari M. On defining the partial
bles that must be controlled in order to achieve acceptable control problem: concepts and examples. AIChE J. 2000;46:2456–
economic performance during plant operation. This approach 2474.
does not, however, provide guidelines on control structure 10. Stephanopoulos G, Ng C. Perspectives on the synthesis of plant-
design and control loop tuning. We, therefore, relied on our wide control structures. J Proc Contr. 2000;10:97–111.
11. Maarleveld A, Rijnsdorp JE. Constraint control on distillation col-
previously introduced singular perturbation-based analysis umns. Automatica. 1970;6:51–58.
and control framework, which accounts for the time scale 12. Luyben WL, Tyréus BD, Luyben ML. Plantwide process control.
separation present in the open loop dynamics of integrated McGraw-Hill; 1998.
plants, to identify the available controlled and manipulated 13. Larsson T, Govatsmark MS, Skogestad S, Yu CC. Control structure
selection for reactor, separator, and recycle processes. Ind Eng Chem
variables in each time scale. Res. 2003;42:1225–1234.
Using a prototype reactor-separator process, we success-
fully demonstrated the development and implementation of a Manuscript received Jan. 18, 2007, and revision received Jan. 31, 2008.

AIChE Journal July 2008 Vol. 54, No. 7 Published on behalf of the AIChE DOI 10.1002/aic 1841