Professional Documents
Culture Documents
This article introduces a framework for including different uncertainties at the chemi-
cal plant design stage. Through an integrated robust optimization approach and prob-
lem formulation, equipment, operating, control, and quality costs are simultaneously
taken into account, leading to system, parameter, and tolerance design. Rather than
using single pointwise solutions in the decision space, operating windows leading to
overall best performance are identified and defined. Such windows and their width al-
low us to point out control needs and goals at a very early stage of plant design. Two
small-scale case studies (for a CSTR and a batch distillation column) provide enough
evidence to support the practicality of the optimization framework: the robust solutions
found are digerent and much better than the corresponding solutions obtained with the
fully deterministic optimization paradigms.
Introduction
Process control, parameter, and tolerance design issues are sion variables (concentrations, pressures, flows, etc.) will al-
not explicitly taken into account at the design stage for most ways occur, eventually bounded within a narrow, but not null,
chemical plants. Traditionally, design and control are treated operation window.
in a sequential way with the following two steps: As a consequence of not taking into account this process-
(1) Design stage-definition and sizing of equipment and inherent uncertainty, the final solutions found by the conven-
determination of the operating nominal point (usually based tional approach may be suboptimal: the decision-space zone
on steady-state models). that surrounds the best deterministic pointwise solution does
( 2 ) Control stage-choice and design of the control system not correspond in general to the zone of the decision space
based on the operating nominal point determined in the first where best average performance can be achieved.
stage (taking into account dynamic issues). Besides randomness of operating variables, the determinis-
Furthermore, design stage calculations are traditionally tic approach also ignores other uncertainties (Pistikopoulos,
made under deterministic optimization paradigms: an eco- 1995) such as model-inherent uncertainty (kinetic constants,
nomic objective function is optimized and leads to a point- physical properties, thermodynamic parameters) or external
wise solution in the decision space, without taking into con- uncertainty (feedstream availability, product demands, eco-
sideration different sorts of uncertainty. Designing a control nomic indexes, environmental regulations).
system based on this solution may be difficult or even impos- Several researchers have recently attempted to integrate
sible since the approach does not consider operability as- process design and control. Luyben and Floudas (1994) com-
pects, such as controllability and flexibility. bine economic and control objectives through a multi-
As Saraiva and Stephanopoulos (1992) and Saraiva (1993, objective optimization framework, incorporating open-loop
1996) have shown in previous studies, by considering this tra- controllability measures in the mathematical formulation of
ditional approach one ignores the fact that operating deci- process synthesis. Bahri et al. (1996, 1997) present a method-
sion variables behave as random variables and are always as- ology for considering flexibility and controllability in process
sociated with some variability. No matter how good control design, where both steady-state and dynamic cases are con-
systems happen to be, in reality ranges of values for the deci- sidered. Mohideen et al. (1996) approach the problem of op-
timal design for dynamic systems under uncertainty consider-
ing both flexibility and control aspects, with the best control
Correspondence concerning this article should be addressed to P. M. Saraiva structure being proposed through multiloop controllers.
I
I I
0
I I
I
I
I I
I I
I I
I
I dz I
I
I
I
I
I
I
I Subroutine O W N
Computes Cd,C, and
I
I C=Cd+Cz+E(C,)+E(Co)
I
I
, I I
Subroutine Model
Computes C,, ,CQ1,Y,,~
several kinds of input PDFs (independent or correlated, con- and standard deviation for the relevant outputs with enough
tinuous or discrete). accuracy (a maximum error of 1% is considered). Since N,
The generic optimization approach presented so far allows depends on the process model used, sampling accuracy must
almost any sort of process design problem subject to several be evaluated for each plant.
kinds of uncertainty to be addressed. 3) Computational effort depends directly on the value of
Some issues related with the above optimization methodol- N,, since the total number for deterministic model evalua-
ogy deserve further discussion. tions that are needed is N, x N, + D, where N,. represents
1) The external algorithm progression does not take into the number of iterations of the external optimization proce-
account the deterministic model as an explicit restriction. In dure and D the number of evaluations needed to estimate
fact, the deterministic model works as an implicit restriction partial derivatives. The HSS technique significantly reduces
because the objective function evaluation needs the output the magnitude of N, when compared with other techniques,
values, y , , , , to be computed for all observations. If the deter- such as MCS (Monte Carlo Sampling) or LHS (Latin Hyper-
ministic model doesnt converge for a specific sample point, cube Sampling).
then the region Z that includes that point is considered in- 4) During the external algorithm progression we kept N,
feasible and the objective function (annual cost) is penalized constant, although Chaudhuri and Diwekar (1996, 1997) have
with an infinite value, in order to guarantee that presented a stochastic annealing algorithm that uses in-
P M d , z , y , 0 ) = 0) = 1. creasing values of N, when the solution seems to approach
2 ) Before starting the stochastic Optimization procedure, optimality. When applied to process synthesis problems, this
the accuracy of the sampling technique used in the internal alternative has contributed to additional computational effort
cycle must be evaluated. This is done by computing the num- reduction and may also be explored within our optimization
ber of observations, N,, that are needed to predict the mean framework.
CSTR application
(18)
In this section we will consider the CSTR already briefly
introduced (Figure 1) with the nomenclature presented in
Table 1. where (Y and P are constants, p, is the mean value of z , and
Two first-order reactions in series ( A B -+ C ) with Ar-
--$
uz its standard deviation; quality costs are expressed by a
rhenius kinetics are assumed to occur. The stationary process symmetric Taguchi loss function:
model is obtained through an energy balance around the re-
actor and two material balances, referring to components A
and B:
F(C,, -C,)+ k j exp(- EA/RT)CAV Table 2. CSTR Parameter and Initial Conditions
Parameters Initial Operating Cond.
EA 3.64 X lo4 0.391
En 3.46 x 104 1.004 (1 +0.10)
The model variables can be classified according to our 8.4 x 105 2 . 5 4 107
~ (i+o.ioj
problem mathematical formulation (P): k: 1.6 X lo4 3,118 (1 +0.05)
d = [ V]-design variable AH, -2.12~10~ +
342.5 (1 0.05)
z = [ F , Q, CAo,C,,, T,,IT-operating variables AH2 - 6.36 X lo4 314 (1 k0.05)
O=[E,, E,, k j , k i , AHl, AH,, p , process model P 1,180
3.2 x 103
2,416
940.2
parameters cP
R 8.314 311.6
y , = [ PB]-quality variable
0.8
Table 3. Comparison Between the Deterministic and Robust
~
Solution
Deterministic Robust y 0.4
I/ (m) 0.3918 0.3463 0.2 ,
F (m3/min) 0.6621 0.5023 (1 f0.099) 0
Q (kJ/min) 1,619 +
146.7 (1 0.080)
280 300 320 340
C,qo (moi/m3) 3,449 3,140 (1 f0.050)
CBo (moi/m3) 34.25 510.7 (1 f0.050) To 00
To (K) 305.7 313.8 (1 f0.005)
Figure 4. Optimal robust CSTR operating region.
External uncertainties are also accounted for through the 0.15 0.03
consideration of the mixture average molar weight as a pa-
0.02 -
rameter with a given probability density function. N
w
v
As stated before, the accuracy of the sampling technique 0.01
must be evaluated for each process model, with 500 observa-
tions needed in this case to reliably predict the xT, mean and 0 -~--
standard deviation. 85 105 125 60 80 100 120 140
V (mofi) F (mol)
The results obtained by the two distinct approaches are
compared in Table 7. In the deterministic formulation the
decision variables are the number of theoretical stages, N ; 6
30- 9
with F expressed in mymin. Equation A3 results from a where c3 = 0.00935 $/kg and t , = 0.1 h is the stop time be-
polynomial regression based on values presented by Douglas tween batches (Diwekar and Kalagnanam, 1997a,b);
(1988); Eq. A4 is based on an installation cost graphic corre-
lation for centrifugal pumps (Peters, 1991).
Control costs are estimated by Eq. 18. The residual cost,
+
a , is considered as 6% of C, Cpumpfor the initial operating
conditions (Table 21, which give LY = $143.2. The p parame-
ter is computed considering the control cost for a single oper- where a and p have the same values as for the CSTR exam-
+
ating variable as 7.5% of C, Cpump (for the initial operating ple: a = $143.2 and p = 1.736.
conditions), when the variable tolerance is E, = 0.15, leading The nominal profit is based on the nominal sales value for
to p = 1.736. The total control cost is the sum of the contri- the product relative to the cost of feed, P/ = 0.1 $/mol (Di-
butions corresponding to the five operating variables: wekar and Kalagnanam, 1997a,b):
Finally, quality costs are expressed by the symmetric When the average product composition differs from the
Taguchi loss function Eq. 19, where the constant k is com- desired value (0.95), the product sales price, Pr, is penalized
puted assuming that a 10% deviation on the quality-related through an asymmetric quadratic Taguchi loss function. For
variable PB leads to an increase of 50% in the cost C, + C, a given observation i, corresponding to an average product
(for the initial operating conditions). This assumption results composition x&, the penalty function is:
in a k value of 6.536.
P: - kl(xE,8-0.95)2 x;,, I 0.95
Batch distillation column application pr,, = (A101
P;"- k,(x;,, -0.95)* > 0.95,
Our objective function is the plant annual profit given by
Eq. 21.
Equipment, operating costs, and control costs are given by where k , = 20 and k , = 4. The associated quality cost is given
the following expressions: by:
c,vMP c,v
C,,$/yr = -+-, iA6)
G* Gb