Professional Documents
Culture Documents
Process Control Structure Selection Based On Economics: Jonathan A. Heath, Ioannis K. Kookos, and John D. Perkins
Process Control Structure Selection Based On Economics: Jonathan A. Heath, Ioannis K. Kookos, and John D. Perkins
on Economics
Jonathan A. Heath, Ioannis K. Kookos, and John D. Perkins
Centre for Process Systems Engineering, Dept. of Chemical Engineering and Chemical Technology,
Imperial College of Science, Technology and Medicine, London SW7 2BY, U.K.
In process systems, selecting suitable sets of manipulated and controlled ®ariables and
the design of their interconnection, known as the control structure selection problem, is
an important structural optimization problem. The operating performance of a plant
depends on the control structure selected, as well as the characteristics of the dis-
turbances acting on the plant. The economic penalty associated with the ®ariability of
main process ®ariables close to acti®e constraints was used in this work to de®elop a
quantitati®e measure for the ranking of alternati®e control structures. The problem is
formulated as a mixed integer nonlinear optimization problem of special structure, which
was used for an algorithm to sol®e this problem to global optimality. The final formula-
tion is a mixed integer linear problem for which effecti®e solution methods are currently
a®ailable. The ®alidity and usefulness of the method are demonstrated through a num-
ber of case studies.
Introduction
An important structural decision in terms of the operating negligible even for large-scale problems. The mathematical
performance of a process plant is related to the selection of programming-based methods, on the other hand, are rigor-
suitable sets of manipulated and controlled variables, and the ous and produce structures that are optimal within the limi-
design of their interconnection. This problem is known as the tations imposed by the mathematical methods and models
control structure selection problem. The last two decades used. An obvious limiting factor in the latter approaches is
have witnessed important theoretical developments in this the size and complexity of the models that can be tackled
area, and two major approaches have emerged. The first ap- within a mathematical programming framework. The method
proach is based on qualitative methods. Heuristic and logical presented in this article belongs to the mathematical pro-
rules are used in sequential decisions that lead to the inven- gramming-based methods and, as such, it consists of three
tion of feasible and plausible control structures. The work of major elements ŽNishida et al., 1981; Morari and Perkins,
Buckley Ž1964. had a profound impact on the development of 1995.: the problem representation, the problem e®aluation, and
these methods. Most of the available literature is reviewed in the solution strategy. The rationale behind the elements em-
Stephanopoulos Ž1983. and in Luyben et al. Ž1999.. The sec- ployed in this work is briefly discussed in what follows.
ond approach uses quantitative methods where mathematical Nishida et al. Ž1981. state that the main task of chemical
models are formulated and solved using mathematical pro- process control synthesis is to structure a dynamic system of
gramming methods. Georgiou and Floudas Ž1989., Moore measurements and manipulated variables so that certain con-
Ž1992., Luyben and Floudas Ž1992., Turkay et al. Ž1993., Nar- trol objectives are satisfied in the presence of disturbances.
raway and Perkins Ž1993, 1994., and Mohideen et al. Ž1996. They divide these control objectives into two main categories.
have presented methodologies that fall into this category. The The first set of control objectives stems from purely eco-
main advantage of the heuristic-based approaches stems from nomic considerations and can be related to the optimization
the fact that they do not involve any detailed evaluation. As a of an objective function that accounts for the profitability of
result, the engineering effort and computational time can be the plant. These control objectives are assigned to the
Žsteady-state . optimizing control system. The second set of
control objectives is related to maintaining certain process
Correspondence concerning this article should be addressed to J. D. Perkins. variables within the constraints that define the feasible re-
min J Ž x, z , p, u, d, Y . min J Ž x , z , p, u .
u, d, Y u
s.t. h D Ž x , z , p, u, d, Y . s 0 s.t. h D Ž ˙
x , x , z , p, u . s 0
h A Ž x, z , p, u, d, Y . s 0 h A Ž x , z , p, u . s 0
Ž P3 .
g Ž x , z , p, u, d, Y . q m F 0 g Ž x , z , p, u . F 0 Ž3.
x g X , z g Z, ugU x L F x F xU , z L F z F zU
nY u L F uF uU , ps pN , ˙
xs0
d g D, Y g 0, 1 4 , ps pN
The solution of this problem defines the optimum steady-state
This formulation ensures that its solution will give a new
operating point denoted by the subscript N Ž x N , z N , u N ..
nominal operating point for which the plant operation will
The dynamics of the process close to this nominal point can
remain feasible for the whole set of uncertainty. This is illus-
be described with adequate accuracy by the linearization of
trated in Figure 1. The solution of problem P2 defines the
the Eq. 3 at the nominal point
steady-state economics, while the solution of the problem P3
defines the dynamic economics of the process. hD hD hD
It should be noted that, since the feasible region of prob- d xq
ž ˙/ ˙ ž / ž d xq / dz
x x z
lem P3 is a subset of the feasible region of problem P2, the
following holds true hD hD
ž q d uq
/ ž / d ps 0
u p
d J s JŽ P 3. y JŽ P 2. ( Ý ll ml G 0 Ž2.
l g IA
hA hA hA hA
d xq
ž / ž / ž d zq d uq
/ ž / d ps 0 Ž4.
x z u p
that is, the dynamic economics is bounded from below by the
g g g g
steady-state economics. Furthermore, using information ob-
ž / ž / ž
gN q d xq d zq / ž /d uq d pF 0
tained from the optimality conditions of problem P2, namely, x z u p
n hD y1
hD hD hA y1
hA max < d yi Ž t . < F Ý max < g i j Ž j v . < ? < d pj < Ž 14.
v
Asy ž ˙/
x x
y
z ž / ž /
z x
Ž7. t j
where
I is the identity matrix and B, F, H, P, and S are defined
analogously to A. Finally, we define the vector s as the vec- d yi Ž s . s Ý g i j Ž s . d p j Ž s . , ; i Ž 15.
tor of deviations of the inequality constraints from their nom- j
inal value
It should be noted that Eq. 14 is also valid for systems with
n
s s g y gN Ž8. time delay since the magnitude of the time delay is one for
all frequencies.
Since for points close to the nominal point the following holds Before closing this section, it should be noted that the use
true of the inequality given by Eq. 14 restricts the applicability of
the proposed method to systems that exhibit stable open-loop
g f g N q H d x q P d uq S d pF 0 Ž9. dynamics. This is due to the fact that the step response of an
integrating or unstable system is not bounded. Thus, the as-
sumptions that need to be fulfilled for the application of the
Equation 8 can be written as
proposed methodology are the following: Ža. the open-loop
dynamics of the system is described by a DAE system of in-
y Is q H d x q P d uq S d ps 0 Ž 10. dex not exceeding one; Žb. the A matrix defined in Eq. 7 has
negative eigenvalues.
Comparing Eq. 8 with Eq. 1, we observe that the maximum
value of s l is a linear, local approximation to the exact value
Control Structure Selection Based on Perfect
of the maximum constraint deviation denoted by m l . Thus, Control
the following linear model can be used to estimate the neces-
sary open-loop Žwithout control. back-off In this section, the linear deviation model of the system
given by Eq. 11 will be combined with a mathematical de-
scription of the perfect controller in order to develop a
yId˙x q A d x q B d uq F d ps 0
Ž 11. method for estimating the closed-loop constraint back-off.
y Is q H d x q P d uq S d ps 0 Equation 11 describes the dynamics of the system and the
deviations of the inequality constraints from their nominal
by assuming that values, close to the steady-state optimal point, under the ef-
fect of uncertainty. However, this is an open-loop description
since no controller is included. Thus, a method has to be
m l f r l s max < s l Ž t . < , ;l, l s1, 2, . . . , n g
n
Ž 12. devised that would invent the optimal, regulatory controller
tgT
pŽ t .g P and then its state space description can be used together with
Eq. 11 in order to estimate the necessary constraint back-off.
This is a dynamic optimization problem that can be avoided In order to overcome the complicating nature of this prob-
by transforming the problem into the frequency domain. To lem, Narraway and Perkins Ž1993. proposed the use of the
this end, a bounding property connecting the maximum devi- perfect control assumption.
ation in the time domain with the maximum of the magni- Firstly, a set of potential controlled variables has to be de-
tude in the frequency domain is presented. fined. This set is a subset of the state vector and can be de-
AX R q v I x X I q BU R sy FP R
Equations 17, 18, 21, and 24 constitute a mixed-integer for-
y I S R q HX R q PU R sy SP R mulation of the perfectly controlled closed-loop system. For
IY R y CX R s 0 Ž 17. any feasible realization of the binary variables, this formula-
tion can be used in order to obtain the frequency response of
y v IX q AX q BU s 0
R I I
the closed-loop system that corresponds to the control struc-
y I S I q HX I q PU I s 0 ture defined by the binary variables. We are interested in
calculating the maximum of the magnitude of the variables
IY I y CX I s 0 S l Žs. over a frequency range
where X, U, P, S and Y are the Laplace transforms of the x,
u, p, s , and y vectors, superscript R Ž I . denotes the real max
ve V
'S R
l Ž v . 2q S lI Ž v . 2 , ;l s1, 2, . . . , n g Ž 25.
Žimaginary. part of a complex number, and v denotes the
frequency. The perfect control assumption implies that when
However, this nonlinear calculation would transform our for-
a potential controlled variable is selected at the regulatory
mulation to a MINLP problem. In order to preserve the lin-
control level, then the controller will keep this variable at its
earity of the formulation we use the following lower bounds
set point irrespective of any other factor. If we use the set of
on the magnitude
binary variables Z to denote the decisions associated with the
selection of controlled variables, then the following inequali-
ties can be used to impose the perfect control S lR Ž v p . F e l¶
y S lR Ž vp . F el
y yi , H Zi F Yi R F yi , H Zi
y yi , H Zi F Yi I F yi , H Zi 5 ; is1, 2, . . . , n y Ž 18. S lI •
Ž v p . F e l ;l s1, 2, . . . , n l Ž 26.
y S lI Ž vp . F el ps1, 2, . . . , nv
When Zi s 0, then Eq. 18 takes the form Žperfectly con-
trolled candidates .
'Ž 2 2
S lR Ž v p . . q Ž S lI Ž v p . . G e l ß
Yi R sYi I s 0 Ž 19 . Furthermore, in order for these lower bounds to approximate
tightly the maximum over the whole frequency range a multi-
and when Zi s1 then Žacceptable variation of the uncon- period Žor multifrequency. formulation was used. It is inter-
trolled candidates . esting to note that as the number of frequencies included in
the formulation tends to infinity, e approaches the maximum
y yi , H F Yi R F yi , H over the frequency domain with negligible error. Thus, at least
Ž 20. in principle, no accuracy is lost by this discretization of the
y yi , H F Yi I F yi , H frequency space. However, as the number of periods is in-
creased, the size and the complexity of the underlying prob-
We can associate a binary vector W with the potential manip- lem is also drastically increased.
ulated variables used in the regulatory control Having a method to calculate the back-off vector that cor-
responds to each control structure, then, using Eqs. 6 and 16,
y u j, H Wj FUjR F u j, H Wj we can calculate the economic penalty associated with this
min a T d x q b T d u
When Wj s 0 then Eq. 21 takes the form Žunused candidates . u
A d x q B d us 0
UjR sUjI s 0 Ž 22 . I d yy C d x s 0 Ž 27.
g N q H d x q P d uFy I r
and when Wj s1 then Žacceptable variation of the candidates
used for control. u L F uF uU
X
y I S I q HX I q PU I s 0 U
IY I y CX I s 0 s F Ž jv . P Ž 32.
Y
R
y yi , H Zi F Yi F yi , H Zi S
y yi , H Zi F Yi I F yi , H Zi 5 ;i
•; p where
y u j, H Wj FUjR F u j, H Wj
y u j, H Wj FUjI F u j, H Wj 5 ;j j v Ix yA
yC
yB
0 Iy
0 0
0
y1
F
0
S lR F e l ¶ ˜ Ž jv . s
F yH yP 0 Ig S Ž 33.
y S lR F e l 0 Eu 0 0 0
•
S lI F e l ;l 0 0 Ey 0 0
y S lI F e l
el F rl ß ß In order to calculate the upper bounds on the maxima of the
time domain deviations we define the following matrix
This formulation assumes that a model of the form given by ˜ s w˜i j s max < wi j Ž v . N
F Ž 34.
Eq. 3 is available. However, in many cases such a detailed vg V
model may not be available. In these cases we can use an
objective function of the form The frequency domain bounds can then be calculated as fol-
lows
D J s Ý wixx i q Ý wjyy j Ž 29.
i j x
˜P
y sF Ž 35.
where w are weights that reflect the economic penalty that is r
related to the maximum deviations of the state Ž x . and con-
trol Žy . variables. These maximum deviations can be calcu- where x , y , and r are the vectors of actual upper bounds
lated by applying the procedure used to calculate the maxi- on the maximum time domain deviations of the state, input,
mum deviations of s . and slack variables, respectively. Having calculated these
The formulation given by Eq. 28 is a mixed integer linear bounds, the actual value of the objective function can be cal-
programming problem which determines the regulatory con- culated using the LP formulation given by Eq. 27.
trol structure that minimizes the economic penalty associated In order to solve the control structure selection problem,
with the maximum time domain deviations of certain process the MILP formulation given by Eq. 28 is solved to the first
variables. However, due to the fact that the estimated max- integer solution. Then, the actual value of the objective func-
ima over the frequency domain are only lower bounds on the tion is calculated based on Eqs. 27 and 35. This value is used
real maxima, this formulation does not give Žfor each control as the entry cost at this integer solution and an integer cut is
structure defined by the binary variables. the actual objective added to exclude the current integer solution from further
function but only a lower bound on this objective function. consideration. In cases where no further integer solution can
This lower bound becomes increasingly tight as the number be found the algorithm has converged and the best entry cost
of discrete frequencies over which Eq. 28 is solved is in- corresponds to the global optimum. If no feasible integer so-
f jim
k ji is the controller gain and t ji is the integral Žor reset. time. Ý Ž y k ji . f jim cjim
We assume that these controller parameters can be calcu-
lated by considering each input-output pair and applying one
Uj s Ý
i m
ž 1y j
t ji v / Yi Ž 43.
of the well-known methods for SISO PI controller tuning We further define the following discrete controller gains and
ŽMarlin, 1995.. However, these parameters can be totally in- transformed measurements
appropriate in a multivariable closed-loop system. In order to
take the multivariable nature of the problem into considera- K jim sy k ji f jim Ž 44.
tion, Luyben Ž1990. suggested the use of a detuning factor f.
Using this factor, the parameters of the decentralized PI con- f jim
R
trollers can be related to the values obtained by the SISO Yjim sYi R q Yi I Ž 45.
t ji v
tuning procedures according to the following rules
I
f jim
k MIMO s k SISO f ji Ž 38. Yjim sYi I y Yi R Ž 46.
ji ji t ji v
t jiSISO
t jiMIMO s Ž 39. Using Eqs. 44]46, Eq. 43 becomes
f ji
f ji
ž
Uj sy k ji f ji 1y j
t ji v / Yi Ž 40. Equations 47 and 48 are nonlinear due to the multiplication
of continuous and binary variables. Fortunately, this is one of
the cases where efficient linearization techniques have been
In general, it is desirable to have continuous detuning fac- proposed. First, the following variables are introduced
tors and calculate their optimal values as part of the algo-
rithm. However, as can be observed from Eq. 40, using con- Gjim
R
s K jim c jimYjim
R
Ž 49.
tinuous detuning factors leads to the introduction of bilinear
terms involving the detuning factors and the measured vari- Gjim
I
s K jim c jimYjim
I
Ž 50.
ables. In order to formulate a linear problem, which can be
solved effectively to global optimality, the continuous detun-
and Eqs. 47 and 48 can be written in the following linear
ing factors are replaced by a discrete set of values. This is
form
done by introducing a discrete set of detuning factors f jim
and a set of binary variables c jim that are used to select the
particular discrete detuning factor according to: c jim s1 im-
UjR s Ý Ý Gjim
R
Ž 51.
i m
plies that f jim is the detuning factor used in j-i channel, c jim
s 0 implies that f jim detuning factor is not used in j-i chan- UjI s Ý Ý Gjim
I
Ž 52.
nel. i m
y Gjim
R
F K˜jimYjiR, H c ji m Ý Ý cjim F1, ;j
i m
Gjim
R
F K˜jimYjiR, H c ji m
Ž 54. Ý Ý cjim F1, ;i
y Gjim
R R
Fy K jimYjim q K jimYjiR, H ˜ Ž 1y cji m . j m
˜ AX 0R q BU0R sy FP R
Gjim
R R
F K jimYjim q K jimYji,R H Ž 1y cji m .
y I S R q HX 0R q PU0R sy SP R
where K˜jim denotes the absolute value of K jim and Yji,RH are IY0R y CX 0R s 0
upper bounds defined as follows
max Ž f ji .
y u j, H ž Ý Ý c / FU
i m
ji m
R
0 , j F u j, H ž Ý Ýc / ,
i m
ji m ;j
YjiR, H s 1q
ž ti j v / yiR, H Ž 55.
y yi , H ž 1y Ý Ý c / F Y
i m
ji m
R
0 , i F yi , H ž 1y Ý
j
Ý cji m
m
/ , ;i
Gjim
R
s0 Ž 56 . IY R y CX R s 0 •; p
y v p IX q AX I q BU I s 0
R
i
Ý Gjim
m
I
ß
property of perfect steady-state control that characterizes
controllers with integral action, we can neglect the structure S lR F e l ¶
of the controller at zero frequency and use the plant model
y S lR F e l
together with the perfect control constraints applied at that
frequency, that is S lI F e l ;l, • ;p
y S lI F e l
AX 0R q BU0R sy FP R el F rl ß
y I S R q HX 0R q PU0R sy SP R y Gjim
R
F K˜jimYjiR, H c ji m ¶
IY0R y CX 0R s 0 Ž 58. Gjim
R
F K˜jimYjiR, H c ji m
y Gjim
R R
Fy K jimYjim q K˜jimYjiR, H Ž 1y c ji m .
y u j, H ž Ý Ý c / FU
i m
ji m
R
0 , j F u j, H ž Ý Ýc /,
i m
ji m ;j
Gjim
R R
F K jimYjim q K˜jimYji,R H Ž 1y c ji m .
y Gjim
I
F K˜jimYjiI, H c ji m
y y j, H ž Ý Ý c / FY
i m
ji m
R
0 , i F yi , H ž 1y Ý
j
Ý cji m
m
/ , ;i
Gjim
I
F K˜jimYjiI, H c ji m •; i , j, m, p
y Gjim
I I
Fy K jimYjim q K jimYjiI, H ˜ Ž 1y cji m .
where the subscript 0 has been used in order to explicitly
denote that these equations apply only at steady state. It Gjim
I I
F K jimYjim q K jimYji,I H ˜ Ž 1y cji m .
should be noted that the imaginary parts of the variables have
f jim
been eliminated from these steady-state constraints since, as R
Yjim sYi R q Yi I
can be seen by examining Eq. 17, they are identically zero at t ji v p
zero frequency. In summary, the following MILP formulation f jim
I
sYi I y Yi R
for the solution of the control structure selection problem
with realistic, decentralized PI controllers is proposed
Yjim
t ji v p ß
AIChE Journal October 2000 Vol. 46, No. 10 2005
The algorithm presented for the case of perfect control is This set of equations is of the same form as the one pre-
also applied in the case of decentralized PI control since the sented for the case of perfect control given by Eq. 32. Thus,
calculated cost is a lower bound to the actual cost. In order the maximum constraint back-off can be calculated as indi-
to calculate the actual cost the worst case back-off is calcu- cated by Eq. 35.
lated first. Then the LP problem given by Eq. 27 is solved
and its solution is used as the entry cost that corresponds to Solution Algorithm
the current integer solution. However, Eqs. 30]35 were de-
rived for the case of perfect control and, thus, they do not In this section, a decomposition approach that can acceler-
apply in the case of decentralized PI controllers. In what fol- ate the solution of problem given by Eq. 59 is presented. The
lows the calculation of the worst case back-off for the case of necessity of this decomposition approach is due to the fact
decentralized controllers is presented. that the number of alternatives that has to be taken into con-
The system dynamics is described by Eqs. 11 and 16 sideration, even for small problems, is extremely large. This
number can be calculated using the following
y j v IX q AX q BUq FP s 0
N
y I Sq HX q PUq SP s 0 Ž 60. nu ny
IY y CX s 0
NS s Ý
k s1
ž /ž /
k k
k!n kt
N nu! n y!
We then define the following variables as the integrals of the
potential controlled variables s Ý k!n kt Ž 67.
k s1 k! Ž n u y k . ! k! Ž n y y k . !
q˙s y Ž 61.
where Ns min n u ,n y 4 and NS is the number of alternative
or solutions.
In order to apply the proposed algorithm more effectively,
j v QyY s 0 Ž 62. a decomposition is used that is based on the idea of separat-
ing the problem of selecting sets of controlled and manipu-
where Q is the Laplace transform of q. Furthermore, define lated variables from the pairing and tuning problem. This is
the n u by n y matrix J whose elements j ji are given by achieved by first solving the following steady state, perfect
control problem that is derived from Eq. 28 by eliminating
j ji s Ý f jim c jim Ž 63. the imaginary parts of variables
m
min a T d x q b T d u
that is, each element of the matrix J denotes whether or not Z, W , u
the pairing j-i is part of the chosen structure and, at the
A d x q B d us 0
same time, the value of the corresponding detuning factor.
Then, the description of the PI controller that corresponds to I d yy C d x s 0 Ž 68.
the current integer solution is given by
Ig N q H d x q P d uFy I g r
Uy EK Y y EI Qs 0 Ž 64. y u L F uF uU
ny nu
where
Ý Zi q Ý Wj s n y
EK s KmJ is 1 js 1
EI s KmTmJ Ž 65. AX R q BU R sy FP R
K s k ji , T s ty1
ji
y I S R q HX R q PU R sy SP R
y jv I y A yB 0 0 0 X F ¶
S lR F e l
yH yP 0 Ig 0 U 0 •;l
y S lR F e l
yC 0 Iy 0 0 Y s S P e F rß
l l
0 0 Iy 0 y jw I y S 0
0 Iu y EK 0 y EI Q 0 Then, for each structure chosen by the solution of the prob-
lem given by Eq. 68, the problem given by Eq. 59 is solved by
Ž 66. imposing the constraints
y1
L s l ji s G Ž 0 . m GT Ž 0 . Ž 71. dL2
rA s F1 y F4 y F2 Ž 73.
dt
and G Ž0. syCAy1 B is the steady-state gain matrix. Equa- dC2
tion 70 imposes the well-known necessary condition for inte- M s F1C1 y F2 C2 Ž 74.
dt
gral controllability with integrity ŽICI. ŽCampo and Morari,
1994.. Furthermore, systems that are not integral stabilizable dP2
ŽIS., that is, structures that do not satisfy the Žnecessary and C s F4 y F5 Ž 75.
dt
sufficient . condition for IS ŽCampo and Morari, 1994.
where
det w G Ž 0 . x / 0 Ž 72.
F4 s 0.16 Ž F1 q F3 .
never enter the variable pairing and tuning step, and are ex-
cluded from further consideration by using integer cuts. = Ž y0.3126C2 y0.5616 P2 q0.1538 P100 q41.57 .
This decomposition has proven to be efficient in practice
due to the fact that Eq. 68 can be solved easily to first integer y F1C p Ž 0.3126C2 q0.5616 P2 q48.43yT1 . 4 rl s1 Ž 76 .
solution even for large-scale problems, while the conditions
2UA 2 Ž 0.507P2 q55yT200 . C p F200
F5 s Ž 77.
l s2 Ž UA 2 q2C p F200 .
In order to find the optimum point of steady-state operation The results are summarized in Table 2 Žfor d p 1 ., Table 3 Žfor
d p 2 . and Table 4 Žfor d p 3 .. From these tables, we can con- tions and the definition of the model constraints, we can cal-
clude that: culate the exact value of the maximum deviation of the con-
v Decreasing the magnitude of the uncertainties generates straints from their nominal values Žsee the definition given by
more feasible structures, but the relative ranking of these Eq. 1.
structures is consistent.
v The calculated economic penalties can be predicted ex- 0.199
actly by the simplified calculation given by Eq. 2. 0.675
v For the model evaporation process, the full control m s 0.660 Ž 94.
structure is the best structure followed by the 1=1 structure 7.835
that controls perfectly the composition Ž y 1 .. 45.109
Finally, in Table 5 the calculated economic penalties for
the formulations given by Eq. 28 ŽMILP formulation with ap- Comparing m given by Eq. 94 with r given by Eq. 93, we can
proximate back-off. and Eqs. 27 and 35 ŽLP formulation with see that a key assumption of this work Žsee Eq. 12. is clearly
exact back-off. are compared. From this table, we can con- satisfied. The worst case combination of the maximum time
clude that, for the example under consideration, the pro- domain deviations of the nonlinear, closed-loop model Ž m . is
posed algorithm will have a satisfactory performance. This is well predicted by the worst case combination of the maxi-
due to the fact that the optimum control structure has an mum magnitudes over the frequency domain Ž r ..
economic penalty that is better than the approximate eco-
nomic penalties of almost all other structures. Only one Case study 2: Reacti©e distillation for ethylene glycol
structure Ž u1-y 1 structure . has a smaller approximate eco- synthesis.
nomic penalty.
In this case study the double feed, reactive distillation col-
The realistic controller-based MILP formulation was solved
umn for ethylene glycol ŽEG. production presented by Ciric
next only for case 2. The open-loop response was fitted to
and Gu Ž1994. is considered. EG is produced from ethylene
simple, first-order transfer functions, and the SISO PI con-
oxide ŽEO. and water ŽW.
troller parameters were calculated using the IMC tuning rules
with an open-loop to closed-loop time constant ratio of three EOqW™EG
ŽMorari and Zafiriou, 1989.. Five discrete detuning factors
equally spaced between 0.1 and 1 were used, as well as two EG reacts further with EO to produce the unwanted byprod-
discrete frequencies v 1 s10y8 and v 2 s10y1. The optimum uct diethylene glycol ŽDEG.
structure is given by the following J matrix
EOqEG ™ DEG
1 0
Js Ž 92. Ciric and Gu Ž1994. present an optimal flowsheet for a dou-
0 1
ble feed, reactive distillation column that consists of ten trays
that is, the best structure corresponds to the pairings C2-P100 shown in Figure 3. The lower part of the distillation column
and P2-F200 and both detuning factors equal to one. The Žtrays 1 to 4. is a simple separation zone, while the upper
back-off is given by the following vector part Žtrays 5 to 10. is a reaction zone. Pure water is fed on
tray 10, while pure ethylene oxide is fed on tray 5. A produc-
0.252 tion rate specification of 25 kmolrh of EG is assumed. The
0.660 modeling equations are presented in the following:
r s 0.660 Ž 93.
7.769
42.534 Table 5. Results for Case 3 (Evaporator Model-Perfect
Control)
and the economic penalty is 8.8 Ž$rh.. This economic penalty D J MI LP =10y3 D JLP =10y3
is considerably greater than the economic penalty of the per-
Open Loop 0.005614 0.023638
fectly controlled system, although the control structure se- Žu1, u2.yŽy1, y2. 0.000057 0.000109
lected in both cases is the same. Using the nonlinear model u1yy1 0.000100 0.000412
of the evaporator process and the best control structure, the u1yy2 0.024719 0.047399
response of the closed-loop system to step disturbances of u2yy1
u2yy2 0.005040 0.009664
magnitude d p 2 is calculated. From these closed-loop simula-
d
dt
Ž M j . s L jq1 y L j yVj Ž 101.
d
dt
Ž Mj x i j . s L jq1 x i jq1 y L j x i j yVj yi j Ž 102.
d
dt
Ž Mj . sVjy1 y L j Ž 103.
d
dt
Ž M j x i j . sVjy1 yi jy1 y L j x i j Ž 104.
Algebraic Equations
3r2
L j s c o r L j l W Ž h OW , j . , js1, 2, . . . , 10 Ž 105.
p DC2 rL j
M j s Ž hW q h OW , j .
ž /ž 4 MWL j / , js1, 2, . . . , 10
Ž 106.
NR
Ž 107.
Separation Section Ž js1, 2, 3, 4, is1, . . . , NC y1. r L j s r Ž Tj , Pj , x i j . , js1, 2, . . . , 10 Ž 108.
d Vj D H VAP y Q R s 0, js 0 Ž 109.
dt
Ž Mj . s L jq1 y L j yVj qVjy1 Ž 95.
Vj D H VAP y QC s 0, js10 Ž 110.
d
dt
Ž M j x i j . s L jq1 x i jq1 y L j x i j yVj yi j qVjy1 yi jy1 Ž96. MWL j E1
Ž 111.
NR
d MWL j E2
dt
Ž Mj . s L jq1 y L j yVj qVjy1 y M j Ý
k s1
R jk Ž 97. R j2 s
ž /
rL j ž
k 0 , 2 exp y
R g Tj / xEO j xEG j , js6, 7, 8, 9
d Ž 112.
dt
Ž M j x i j . s L jq1 x i jq1 y L j x i j yVj yi j qVjy1 yi jy1
K i j Pj y Pisat
j s 0, js1, 2, . . . , 10, is1, . . . , NC y1 Ž 113 .
NR
q Mj Ý n ik R jk Ž 98 . yi j y K i j x i j s 0, js1, 2, . . . , 10, is1, . . . , NC y1 Ž 114 .
k s1
Ý x i j y1s 0, Ý yi j y1s 0, js1, 2, . . . , 10 Ž 115.
Feed Trays Ž js 5,10, is1, . . . , NC y1. i i
FW y rFEO s 0 Ž 116.
where r s1.0488.
The perfect control formulation given by Eq. 29 is applied
Figure 4. Temperature profile in the reactive distillation in order to find the best temperature sensor location for the
column. minimization of the maximum time domain deviation of the
bottom product composition in EG. The four best solutions
are reported in Table 6. Controlling the temperature at tray
2 is the best solution followed by temperatures at trays 1 and
3. Then, the PI controller-based formulation is applied. The
Q R Ž QC . is the heat input in the reboiler Žcondenser., D H R open-loop responses of the tray temperatures were fitted to
is the heat of reaction, r L is the liquid density, and MWL is first-order transfer functions and single PI controllers were
the liquid molar weight. Subscript i refers to the set of com- designed based on the IMC tuning rules ŽMorari and Zafiriou,
ponents Ž1 to NC ., j refers to the set of trays Ž1 to NT q1., 1989.. Only one detuning factor was used Ž f s1. and one
and k to the set of reactions Ž1 to NR .. All required data are frequency v s 0.1. The results of the realistic control formu-
obtained from Ciric and Gu Ž1994.. In addition to the equa- lation were the same as the results of the perfect control case.
tions used by Ciric and Gu, we also use in this case study the This is due to the fact that Bode diagrams of all SISO loops
Francis weir formula to relate the liquid outflow from a tray were identical to that of a first-order system and, thus, their
with the liquid holdup. An 0.083 m weir is assumed in the maximum appears at zero frequency. Since the perfect con-
separation zone and a 0.166 m weir in the reaction zone. The trol and realistic PI control formulations are the same at zero
liquid holdups are restricted between 0.01 and 0.2 m3. A frequency, the same structures are chosen.
parametric optimization problem is solved in order to mini- In Figure 5, the closed-loop response of the bottom com-
mize the annual operating cost. The temperature profile is position is given. In these closed-loop simulations tray tem-
given in Figure 4. The feed flow rate of water and ethylene peratures are controlled using PI controllers. It is interesting
oxide are 26.314 kmolrh and 27.6 kmolrh, respectively. The to note that unstable closed-loop response is obtained by us-
heat input to the reboiler is 7.36 MW. Compared to the solu- ing any tray temperature above tray 3. Thus, the proposed
tion of Ciric and Gu, this solution has increased operating method was successful in predicting the tray temperatures
cost by 0.03=10 6 $ryr.
The structure of the regulatory control employed is shown
in Figure 3. The feed flow rate of water is used in order to
set the total production rate. The feed flow rate of ethylene
oxide is ratioed to the feed flow rate of water. A constant
excess of 4.887% in EO is used to achieve 95% EG in the
bottom product. The bottom product flow rate is adjusted to
control the bottom holdup, while heat input is adjusted in
order to control the product composition. The main distur-
bances acting on the system are the water feed flow rate
Žproduction rate level. and column pressure. The condensa-
tion rate at the top of the column is assumed to be used for
controlling the column pressure. However, in order to ac-
count for the effect of the inefficiencies of this control loop,
we consider the pressure variation as a major disturbance.
The production rate is assumed to vary by 50% and pressure
by 10% of nominal values. The problem investigated is the
selection of the best tray temperature sensor location for in-
ferential control of the EG composition at the bottom of the Figure 5. Product composition response under control
column. In order to mimic the ratio specification imposed on of different tray temperatures (reactive distil-
the feed flow rates of water and ethylene oxide we use the lation case study).
d
dt
Ž Mj x i j . s L jq1 x i jq1 y L j x i j yVj yi j qVjy1 yi jy1 q Fj x i R
Ž 117.
d
dt
Ž Mj . s L jq1 y L j yVj qVjy1 q Fj Ž 118.
ai
yi j s K i j x i j s xi j Ž 119.
Ý ak xk j
kgC
Vj sVjy1 q Ž 1y q V . Fj Ž 120.
3r2
L j s c o r L l W Ž h OW j . Ž 121.
p DC2 rL
M j s Ž hW q h OW j . Ž 122.
4 MW , L
F, js NF
Fj s ½ 0, j/ NF 5 Ž 123.
Figure 6. Reactorr
rseparator with recycle process. Column Base Ž js1, igC.
d
dt
Ž Mj x i j . s L jq1 x i jq1 y L j x i j yVj yi j Ž 124.
that should be controlled in order to minimize the variability
of the product composition using an inferential control d
scheme.
dt
Ž M j . s L jq1 y L j yVj Ž 125.
ai
Case study 3: The reactorr
rseparator process yi j s K i j x i j s xi j Ž 126.
Ý ak xk j
The reactorrseparator process is shown in Figure 6. The kgC
fresh feed consists of component A and an inert component QR
I. The first-order irreversible reaction A ™ B takes place in Vj s Ž 127.
the single-phase CSTR and transforms reactant ŽA. into D H VAP
product ŽB.. The output of the CSTR contains all three com-
ponents and is fed into a distillation column where the unre- Column Condenser and Reflux Drum Ž js NT q1, igC.
acted A is separated from the product and recycled back to
the reactor. In order to avoid accumulation of the inert com- d
dt
Ž M j x i j . sVjy1 yi jy1 y Ž L j q P q R . x i j Ž 128.
ponent in the process it is necessary to purge this material at
some point. This point is mainly determined by the relative d
volatility of the inert. In cases where the volatility of the inert Ž M j . sVjy1 y Ž L j q P q R . Ž 129.
is between the volatility of reactant and product then the most dt
plausible point is the rectifying section of the distillation col-
Reactor Ž igC.
umn.
In order to develop a relatively simple mathematical model
d
of the reactorrseparator process the following assumptions Ž MR x i R . s F0 x i0 q Rx i N Tq1 y Fx i R q MR x A R n i k Ž 130.
are used dt
Ž1. The liquid on trays and in the reactor is perfectly mixed d
and has uniform properties. Ž MR . s F0 q Ry F Ž 131.
dt
Ž2. Thermodynamic equilibrium on all trays is achieved.
Ž3. The pressure drop in the column is negligible. yE
k s k 0 exp
Ž4. Constant relative volatility and equimolar overflow.
Ž5. The vapor holdup is negligible.
ž R g Ž T qT0 . / Ž 132.
0 0 0 1 0 0 0 0
1 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
Js Ž 152.
0 0 0 0 1 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 1 0 0
Figure 7. Closed-loop response of the reactor r
This matrix corresponds to the control structure selected as separator system-manipulated variables.
the best solution when perfect control was used Žsee Table 20% change in production rate and 5% change in feed com-
7.. Heat input at the reboiler is used in order to control bot- position.
Greek letters
asgradient of the cost function with respect to the state vari-
ables
bsgradient of the cost function with respect to the input vari-
ables
dsdeviation variable
Lsrelative gain array
lsLagrange multipliers
rs linear estimate of the back-off vector
SsLaplace transform of the s vector
sslinear estimate of the deviation of the inequality con-
straints from their nominal values
vsfrequency
Subscripts
0s initial conditions
Asalgebraic variable
cls closed loop
Dsdifferential variable
Hs maximum allowable variation
Figure 8. Closed-loop response of the reactor r Lslower bound
separator system-controlled variables. Us upper bound
20% change in production rate and 5% change in feed com-
position.
Superscripts
Is imaginary part or integral variable
MIMOsmulti-input, multi-output system
to a subset of the potential measured variables, and a suit- Rsreal part
SISOs single-input, single-output system
able set of control variables is selected among the potential
manipulated variables so that the dynamic economics is as
favorable as possible. Two control algorithms, perfect control
and realistic, decentralized PI control, have been included in Literature Cited
the formulation for the examination of the closed-loop sys- Bahri, P. A., A. Bandoni, and J. Romagnoli, ‘‘Operability Assess-
tem. Integer variables are used to model these control algo- ment in Chemical Plants,’’ Comput. Chem. Eng., 20, S1787 Ž1996..
Belanger, P. W, and W. L. Luyben, ‘‘Plantwide Design and Control
rithms resulting in an MINLP formulation of special struc- of Processes with Inerts. 3. Intermediate Inerts,’’ Ind. Eng. Chem.
ture. By exploiting its structure, the problem is transformed Res., 37, 535 Ž1998..
into an MILP problem that can be solved effectively avoiding Brenan, K. E., S. L. Campbell, and L. R. Petzold, Numerical Solution
the need for complete enumeration of all alternatives. At the of Initial Value Problem in Differential-Algebraic Equations, North-
same time, global optimality is guaranteed. Application of this Holland, New York Ž1989..
Buckley, P. S., Techniques of Process Control, Wiley, New York Ž1964..
approach to a number of case studies demonstrates that the Campo, P. J., and M. Morari, ‘‘Achievable Closed-Loop Properties
method can effectively be used in order to answer important of Systems Under Decentralized Control: Conditions Involving the
questions arising in the control structure selection problem. Steady-State Gain,’’ IEEE Trans. Autom. Control., 39, 932 Ž1994..
Ciric, A. R., and D. Gu, ‘‘Synthesis of Nonequilibrium Reactive Dis-
tillation Processes by MINLP Optimization,’’ AIChE J., 40, 1479
Ž1994..
Notation
Dimitriadis, V. D., and E. N. Pistikopoulos, ‘‘Flexibility Analysis of
fsdiscrete detuning factor Dynamic Systems,’’ Ind. Eng. Chem. Res., 34, 4451 Ž1995..
f ssscaling factor used in Eq. 13 Figueroa, J. L., P. A. Bahri, A. Bandoni, and J. Romagnoli, ‘‘Eco-
hs vector of equality constraints nomic Impact of Disturbances and Uncertain Parameters in
Is identity matrix of dimension equal to the dimension of the Chemical Processes}a Dynamic Back-Off Analysis,’’ Comput.
vector given as subscript Chem. Eng., 20, 453 Ž1996..