00
Printedin GreatBritain.All rightsreserved Copyright0 1987Pergamon Journals Ltd
L. B. EVANS
Massachusetts Institute of Technology, Cambridge, MA 02139, U.S.A.
(Received 17 February 1987; final revision received 8 June 1987; received for publication 23 June 1987)
AbstractThe sequential modular and twotier simultaneous modular approaches to solving the process
flowsheet optimization problem are described. Results from implementation of both of these strategies
in an existing sequential modular simulator are presented and an analysis of the benefits and deficiencies
of each method is given.
For the sequential modular approach, a new method, called infeasible path hybrid (IPH), is shown to
be superior to existing infeasible path and feasible path methods.
For the twotier simultaneous modular approach, the requirements for good reduced models are
discussed in detail and a new reduced model for plug flow reactors is presented. It is shown that, with
good reduced models, the twotier simultaneous modular approach should outperform the sequential
modular approach.
ScopeThis paper discusses new developments in both the sequential modular and twotier simultaneous
modular approaches to process flowsheet optimization using sequential modular simulators. Results from
implementation of both of these strategies in an existing sequential modular simulator, ASPEN PLUS,
are presented and an analysis of the benefits and deficiencies of each method is given.
For the sequential modular approach, a new method, the infeasible path hybrid (IPH) method is
presented. The IPH method combines the robustness of feasible path methods with the speed of infeasible
path methods.
The twotier simultaneous modular methodology, in which equationoriented optimization techniques
are applied iteratively to reduced (i.e. simplified) models of the flowsheet, is described in detail. Particular
attention is paid to the requirements reduced models must meet in order to ensure superior performance
of the method. A new reduced model for plug Row reactors is presented. This model demonstrates the
applicability of the twotier approach to unit operations described by differential equations. A technique
for ensuring a true optimal solution using the nonlinear simultaneous modular approach is presented.
The relationship between sequential modular optimization and linear simultaneous modular opti
mization is also considered.
Conclusions and SignIScenc~The results of the test problem show that the twotier simultaneous modular
method can be much more efficient than the sequential modular methods tested. Less time was needed
to optimize the flowsheet using the simultaneous modular approach than was required to simulate it using
the sequential modular approach. TrevinoLozano [l] shows that, as the complexity of the simulation
problem increases, the relative efficiency of the twotier simultaneous modular method increases. However,
in addition to requiring changes to the architecture of an existing sequential modular simulator, there are
other factors which will slow the development of a generalpurpose twotier simultaneous modular
simulator/optimizer. Further research is needed to determine how equality constraints affect the behavior
of the algorithm. Good reduced models are needed to take full advantage of the efficiency of the approach.
If linear models are used for some unit models, the efficiency of the algorithm will be decreased.
Results for the sequential modular optimization methods indicate that the infeasible path methods are
generally faster than the feasible path method tested. This conclusion differs from that reported in earlier
work by Biegler [2]. Although the infeasible path method (IP) required more iterations than the feasible
path method (CFV), IP was faster because it did not converge the simulation problem at each optimization
iteration. The new infeasible path hybrid (IPH) method was faster than either IP or CFV, generally taking
the same number of iterations as CFV but requiring less computation per iteration. These trends are
typical of our experience with sequential modular optimization in ASPEN PLUS.
Although the twotier simultaneous modular method promises to be the most efficient, more research
is required to fully develop the method. The sequential modular approach is important since it requires
no changes to the architecture of an existing simulator and since reduced models are not required. The
benefits of flowsheet optimization can be realized today using the sequential modular approach and
existing flowsheet simulators.
INTRODUCTION CO, Z) z 0
Computer simulation of chemical processes is an
L,<ZdU, (3)
important and widely used tool in the operation of
existing plants and in the design of new ones. Math where Zrepresents all of the decision variables; R = 0
ematical optimization of process simulations is re is the collection of flowsheet describing equations,
ceiving increasing attention because of the significant including any additional equations necessary to
economic benefits that can be attained. Millions of evaluate the objective function; H = 0 is the col
dollars may be saved in yearly operating expenses by lection of the design specification constraints; G 2 0
identifying and implementing optimal operating con is the collection of inequality constraints; and L and
ditions. U are the bounds on the decision variables.
A chemical process plant consists of a series of unit Given the complexity of the problem and the
operations connected by process streams. Each pro computer time needed to solve it, there is a strong
cess unit may be modeled by a set of describing incentive to develop efficient and robust solution
equations, which include material and energy bal methods.
ances, phase and chemical equilibrium relations, rate
equations and physical property correlations. These
equations ultimately relate the outlet stream variables APPROACHES TO THE PROBLEM
to the inlet stream variables, for a given set of
There are two broad classes of methods for solving
equipment parameters.
flowsheet optimization problems: feasible path meth
The process simulation problem is to solve:
ods and infeasible path methods. In feasible path
R(X) = 0, methods, the equality constraints of problem (3)
(1)
are satisfied for every intermediate estimate of the
where X represents all the variables in the process and decision variables along the path towards the optimal
R = 0 is the collection of the describing equations of solution.
all the units plus the stream connectivity relations. In infeasible path methods, the equality constraints
In many applications additional design specifi are required to be satisfied only at the optimal
cations are imposed on the process simulation prob solution. These methods solve the process opti
lem (1). For example, the purity of a product or the mization problem and the simulation problem associ
temperature of an internal stream may be specified. ated with it simultaneously. There are three types of
The resulting constrained simulation problem is to infeasible path optimization algorithms: sequential
solve the system of equations: modular, equationoriented and twotier simul
taneous modular methods.
R(X,Z)=O Infeasible path sequential modular methods use the
modular architecture of existing simulators. Tear
H(X, Z) = 0, (2) streams and the optimization problem are converged
simultaneously by a convergence block [24].
where Z represents the decision variables freed to
Infeasible path equationoriented methods solve
satisfy the design specifications and H = 0 is the
the optimization problem (3) directly as one large
collection of design specification equations.
nonlinear programming problem. These methods
In a typical process simulation, systems of one
promise to provide the greatest computational
thousand to hundreds of thousands of nonlinear
efficiency and are the most elegant. However, due to
equations are solved simultaneously. These equations
the numerical problems encountered in chemical en
may be very poorly behaved and difficult to solve for
gineering applications, more research is needed be
real systems.
fore a reliable and robust generalpurpose equation
When a flowsheet is optimized, a similar system of
oriented optimizer is developed [5, 61.
equations is solved with some process parameters The twotier simultaneous modular methodology
freed and determined so as to maximize a given may be viewed as the iterative application of the
objective function, F. Additionally, inequality con equationoriented methodology, where simplified
straints may be imposed on the optimization prob flowsheet optimization problems are solved at each
lem. For example, a lower limit may be placed on the step. These algorithms use existing sequantial modu
purity of a product stream. It may also be necessary lar simulator unit operation models to generate
to add equations to evaluate the objective function, simplified flowsheet describing equations which are
such as cost correlations, to the system of equations.
solved with an equationoriented methodology
The optimization problem is to:
[712, 11. Some recent work [13] that has been called
simultaneous modular actually falls into the category
maximize F(X, Z)
which we have called infeasible path sequential
subject to R(X, Z) = 0 modular. As will be shown later, the two algorithms
can be related. Nevertheless, this nomenclature is
H(X, Z) = 0 confusing, so we prefer to reserve the name simul
Sequential and simultaneous modular strategies 569
taneous modular for the class of twotiered algorithms is essentially infinite; no attempt is made to converge
described in this paper. the tear streams. We can define the new hybrid
In this study we have focused on the methods method, IPH (which is more precisely a class of
which may be implemented in existing sequential methods lying on a continuum between IP and CFV),
modular simulators: feasible path and infeasible path by setting the tolerance to some value between these
sequential modular, and the twotier simultaneous extremes, or by adjusting the tolerance during the
modular algorithms. These methods will be presented course of the optimization.
in more detail in the following sections, along Alternatively, the hybrid method can be im
with performance comparisons obtained through plemented by adjusting the number of passes through
implementation in the ASPEN PLUS simulator. the sequence of calculation modules taken by the IP
algorithm. At each step of the SQP algorithm, two
or more passes through the sequence of modules
SEQUENTIAL MODULAR OJTIMIZATION
are performed, updating the tear stream variables.
In a sequential modular simulator the recycle This approach was used in implementing IPH by
problem can be stated as in (l), except that X Kisala [4].
represents only the variables in the torn streams and The algorithm for IPH can be summarized as
R = 0 is the collection of tear stream equations: follows.
R(X)=g(X)X, (4) 1. Provide initial guesses for the decision variables.
2. Take several sequential modular flowsheet iter
where X is the vector of assumed values of the tear
ations to partially converge the tear streams. The
stream variables and g(X) is the vector of values
number of iterations can be based on a tear stream
calculated by executing one pass through all the
convergence tolerance. If only one iteration is taken,
modules.
the method becomes the same as IP. Note that, unlike
Feasible path methods nest the tear stream con
some previous feasible path methods [14161, the
vergence inside the optimization problem. Infeasible
design specifications are not converged.
path methods converge the tear stream and opti
3. Compute the derivatives of the objective func
mization problem simultaneously.
tion, F, the tear stream equations, R = 0, the design
Currently, the most efficient methods for solving
specification contraints, H = 0 and the inequality
the process optimization problem in a sequential
constraints, G 2 0, with respect to the tear stream
modular simulator are the methods originally pro
variables, X and the decision variables, 2. This is
posed by Beigler [2,3]. Biegler presented two feasible
done by perturbing each variable independently and
path methods (RFV and CFV) and one infeasible
performing a sequential modular pass through the
path method (IP). One of the purposes of this work
sequence of modules. Note that the simulation prob
was to further develop these methods and to study
lem (2) is not converged for each perturbation. In
their application to realistic flowsheeting problems.
general, not all the blocks in the calculation sequence
Based on results for small hypothetical problems,
need to be executed for each derivative perturbation,
Biegler concluded that the feasible path methods
only the blocks which are downstream of the
performed better than the infeasible path method,
perturbed variable. Care must be taken to order the
although his results show that the differences between
derivatives such that the base point of the per
the methods is usually small. Kisala [4] showed that
turbation is preserved.
the infeasible path method was more efficient than the
4. Generate new values for the decision variables.
feasible path methods when applied to realistic flow
The successive quadratic programming (SQP) algor
sheeting problems. The behavior of the algorithms
ithm [17] is used to generate new values for the
is greatly influenced by variable scaling and other
decision and tear stream variables. The details of
implementation details.
the implementation of SQP used for this work are
described in Appendix A.
HYBRID METHODS 5. Check for optimality. If converged, stop.
6. Go to Step 2.
Kisala [4] proposed and tested a third sequential
modular method. This method, which we call the IPH has several advantages. Since the tear stream
infeasible path hybrid (IPH), combines properties of constraints are not tightly converged at each iter
the CFV and IP methods. The IP and CFV methods ation, much of the computation time spent by CFV
were shown in Kisala [4] to differ only in the con converging the constraints can be eliminated. An
vergence path: IP follows an infeasible path while advantage CFV can have over IP is that it sometimes
CFV converges the tear stream equations at each converges the optimization problem in fewer iter
iteration. Otherwise the algorithms are identical. The ations due to line search difficulties in IP; thus it can
two methods may actually be thought of as the same require less computation time than IP. IPH shares
algorithm, differing only in one parameter: the tear this characteristic with CFV and thus can converge in
stream convergence tolerance in the CFV algorithm. fewer iterations than IP. The hybrid method can,
For CFV, the tolerance is a small number. For IP, it therefore, converge faster than either CFV or IP.
570 T. P. KISALA et al.
1
of the constraints. An alternative solution to the
problem is to insure that the tear stream equations
are converged to a tight tolerance at each point of a Inside Loop
lems are stated as nonlinear programming problems return to Step 2 to initiate the correction sequence
of the form: using linear reduced models.
terms in the firstorder approximations to the inlet than the sequential modular approaches analyzed in
outlet relations of the blocks. this paper.
This problem may be mathematically decomposed
in many different ways. In the particular implemen
tation used in this work, the problem is decomposed REDUCED MODEL CHARACTERISTICS
automatically via the successive quadratic pro
The previous two sections present arguments which
gramming algorithm proposed by Locke et al. [20].
lead to the conclusion that the efficacy of the twotier
The variables in this decomposition are chosen dur
simultaneous modular approach depends upon the
ing the matrix manipulation operations based on
use of suitable nonlinear reduced models of individ
numerical stability criteria. Nevertheless, other sets of
ual unit operations. In our work we have established
variables could be used to decompose the problem.
a set of desirable characteristics of reduced models
Let us consider another possible set of variables:
which we have found to impart superior performance
the decision variables 2 combined with the tear
characteristics to the approach. In this section
stream variables X. As the Jacobian matrix of
we present these characteristics and discuss their
the reduced flowsheet describing equations is de
significance.
composed, the chain rule would insure that the
The principal subset of desired characteristics is
decomposed problem is equivalent to a linear
listed below.
approximation of the tear stream equations
R(X, Z) = 0 shown in (4). If the optimization algor 1. The number of equations in a reduced model of
ithm uses linear approximations to the nonlinear a unit operation should be much smaller than the
equality constraints at each iteration (as successive number of equations solved by the rigorous model.
quadratic programming does), the solution of prob 2. The equations in the reduced model should be
lem (6) as the reduced problem of the twotier well behaved.
simultaneous modular approach would be equivalent 3. The number of degrees of freedom of a reduced
to the solution of problem (3) in an infeasible path model should be the same as that of the correspond
sequential modular approach. ing rigorous model.
The equivalence described above provides an in 4. The variables included in a reduced model
centive for a detailed comparison between the steps are: inlet and outlet stream variables, equipment
involved in a linear twotier simultaneous modular parameters, computed properties and the internal
optimization method and an infeasible path sequen variables that may be unique to the reduced model.
tial modular one. The two algorithms are almost 5. The parameters of a reduced model are com
equivalent if successive quadratic programming is puted from a base solution of the rigorous model in
used to solve the optimization problems and the such a way that the reduced model equations are
derivatives of the tear stream equations in the latter satisfied exactly at the base point.
method are computed by chain ruling. The main 6. The simple model equations and the Jacobian
differences between the two approaches are: associated with them should be evaluated analyti
cally, without other auxiliary calculations, such as
(1) a linear twotier simultaneous modular thermophysical properties and without requiring
approach will usually use a better choice of iterative calculations.
decomposition variables than the sequential modular
Characteristics 1, 2 and 6 guarantee that the calcu
approach which depends on tear stream variables;
lations needed to solve the reduced problem are
(2) the derivatives of the tear stream equations are
relatively simple, making it particularly attractive to
updated at each iteration of the successive quadratic
use equation oriented methods. Characteristics 3, 4
programming algorithm in a sequential modular ap
and 5 insure that the reduced and rigorous models
proach. The twotier simultaneous modular method
give consistent results when the solution is reached.
keeps the Jacobian matrix of the linearized inlet
The list of characteristics for reduced models
outlet relations constant until the reduced problem
presented above does not impose any constraints on
is converged. Thus, the computed Jacobian of the
the actual form of the model being used for a unit.
tear stream equations remains unchanged until the
That is, these characteristics would apply to both
optimization problem is converged.
linear and nonlinear models. The following set of
characteristics have to do with the performance of the
Given the similarities between the two approaches,
reduced models.
it is reasonable to expect both methods to have about
the same performance for the solution of a given 7. The equations in a nonlinear reduced model
flowsheet optimization problem. should be representative of the type of unit being
The work performed by Jirapongphan shows that modeled, rather than just empirical relationships.
a nonlinear twotier simultaneous modular approach 8. The material and energy balances around each
performs much better than a linear one [8,9]. There unit should always be satisfied, even away from the
fore, it is possible to infer that a nonlinear twotier base point.
simultaneous modular approach will also be faster 9. The parameters of a reduced model should be
Sequential and simultaneous modular strategies 573
obtainable from the results of the rigorous model by reactor. If this is the case, a differential energy
analytical expressions. These expressions may involve balance for each phase is necessary to compute the
any variables of the rigorous model (internal vari different temperature profiles). The inlet flow rates of
ables as well as inlet and outlet stream variables). the components in the system are represented by the
vector fi. A series of reactions take place inside the
Characteristics 7 and 8 indicate that a reduced
reactor and heat is added or removed in order to
model should make use of the knowledge available of
control the reactor temperature.
the physical processes taking place in the unit oper
The component flow rates and the stream tempera
ation. The reduced model would then be a suitable
ture and pressure inside the reactor are described by
engineering approximation of the rigorous model and
a system of nonlinear differential equations:
away from the base point at which the model par
ameters were computed, would have a wider range
of extrapolation than a reduced model based on g = H/U T, PI
empirical relationships alone.
Characteristic 9 is intended to rule out the use of dT
models that require the computation of numerical dz = HAL T, f)
derivatives across each unit. Such calculations require
excessive computer time and should be avoided dP
whenever possible. Analytical expressions for the dz = H,(f, T, P),
reduced model parameters result in increased
where z is the axial coordinate. At the outlet of the
efficiency of the parameter generation calculations.
reactor, z = L, the stream variables have values f,,
Finally, the following requirement was introduced
to guarantee convergence to the true optimal i0, P,.
The purpose of the reduced reactor model is to
solution.
approximate the values of the outlet stream variables,
10. At convergence the inputoutput Jacobian of
the reduced model must be the same as that of the
f,, To, PO, given the inlet conditions A, Ti, Pi and
other relevant exogenous data such as heat transfer
rigorous model of the unit.
rates. This approximation must be made through a
Gradient type linear models always satisfy this system of algebraic equations and the equations in
requirement, because the model parameters are the the model should take into account the coupling
derivatives of the outlet variables with respect to the among the original differential equations.
inlet variables. While nonlinear reduced models hav One way to account for coupling in the original
ing characteristics l9 will not in general satisfy this differential equations is to approximate the complete
requirement, they should have enough physical composition, temperature and pressure profiles
meaning to give a good approximation to the through the reactor. In the proposed reduced model
inputoutput Jacobian, resulting in a suboptimal this is done using polynomials of the form:
solution which in general would be close to the true
optimal solution. As discussed in a previous section,
f =A,+(A,,+rZ,)z +A,fz*+..
linear models are then used in one or more additional T=A,,+(A,,+l,)z+A,,z*+~~~
iterations to reach the true optimal solution, using the
P=A,,+(A,,+rl,)z+A,,z*+.... (8)
suboptimal solution obtained with the nonlinear
models as a starting point when the linear models are The main feature of this proposed reduced model
introduced. is that the polynominal coefficients A,, A,, A2, etc.
are not reduced model parameters. Instead, these
REDUCED MODEL EXAMPLE
coefficients are reduced model variables to be deter
mined for any given set of inlet conditions. The
Examples of reduced models possessing the charac coefficients 1 are reduced model parameters which are
teristics just presented have been described previously determined from the rigorous solution of the differ
by Lee and Boston [21] and TrevinoLozano et al. ential equations for the basepoint inlet conditions.
[22] for multicomponent, multistage separation oper The function of these coefficients will be discussed
ations. In this section we show the derivation of a later in this section.
reduced model for a general plug flow reactor to The equations needed to determine the poly
illustrate how a unit operation described by differ nominal coefficients A,, . . , are obtained from the
ential as well as algebraic equations can be handled. application of collocation techniques to linearized
It is necessary to describe the plugflow reactor sets of differential equations. Given a set of M
problem before the reduced model is presented. Let collocation points z,, z2, . . . : z,, it is possible to
us consider a system of n components including define a system of m(n + 2) algebraic equations of the
reactants and all possible products. A material stream form:
enters the reactor at inlet temperature T, and inlet
pressure Pi. (In some problems there may be multiple H,f+ i dH (A,,+A,Zj+ . . . fF)
phases at different temperatures flowing through the ,=, af, zj
574 T. P. KISALAet al.
3Hk
+F z,(A,,+A,,zj+P*)
=A,,+2Azzj+.. (9)
fork = 1,2,. . . , n + 2 (n = number of components)
i=l,2,... , rn (m = degree of polynomials)
The left hand side of this equation corresponds to
a firstorder expansion of the derivative function Hk
at the collocation point zj. The value of the stream
variable whose profile is described (in differential REACTKIN A  Px
form) by Hk is approximated by the corresponding
polynomial without the linear correction term &z. MAXMUE P  300 in Product
Table 1. Results for plug flow reactor problem simulation problem. The details of the imple
Algorithm Number of iterations STE mentation are given by Kisala [4]. For the twotier
Simultaneous modular Sb 0.92 simultaneous modular approach, changes to the
Squential modular architecture of the simulator were necessary. Trevino
IP 4c 2.66
CFV T 3.50 Lozano [l] describes, in general, the changes neces
lPHd 2 2.46 sary to convert an existing sequential modular
STE = Simulation time equivalents (time for optimization/time for simulator to a twotier simultaneous modular
sequential modular simulation at optimum). architecture.
bNumber of outside loop iterations.
Number of SQP iterations. The first problem is an adaptation of a flowsheet
3 passes through the sequence of modules. used by Evans as an example to demonstrate the
. MIXER 
Stream FEED
24.0% NiImgen I
74.3% Hydrogen STAGE 1 INTERCOOLER STAGE 2
0.6% Argon COMPRESSOR COMPRESSOR
1.l% Methane
 SPLITTER e
BLEED
t
COMPRESSOR
HIGH
PRESSURE
FLASH
LOW
PRESSURE
FLASH
JPRODUCT
Equality constraints
1. Reactor is adiabatic
2. STAGE2 pressure = recompression pressure
Inequality constraints
1. Ammonia in purge stream is no more than 2.0 lbmol hl.
2. Vapor fraction entering STAGE2 is at least 0.99.
3. High pressure flash pressure is at least 4 atm (58.8 psia) leas than outlet pressure for feed compressor.
4. Product stream is at least 99.9% ammonia.
576 T. P. KISALA et al.
principles of sequential modular simulation [24]. The Table 3. Ammonia synthesis process results
flowsheet is shown in Fig. 2. This purely hypothetical Number
process consists of five unit operations. The feed Number of SQP of flowsheet
Algorithm iterations evaluations STE*
stream is fed to a mixer where it is mixed with a
IP 6 117 4.88
recycle stream rich in unreacted raw material, A. The CFV 4 121 5.60
liquid is fed to a reactor where it forms a product, P, IPHb 4 88 3.79
which further reacts to form an undesired by STE = Simulation time equivalents (time for optimization/time for
product, G. The outlet from the reactor is fed to a simulation at optimum).
b3 passes through ihe sequence of modules.
flash drum where the more volatile product, P, is
separated from unreacted A and the byproduct G.
The liquid stream from the flash is recycled after a optimization problem in less computation time than
small fraction of the stream is purged from the was required for one sequential modular simulation.
process. The infeasible path sequential modular method was
The feed stream consists of 0.0126 kmol si of pure the fastest of the sequential modular methods, al
A at 298 K and 1 atm (1.013 x 105Nm2). The though the differences among the sequential modular
reactor is a plugflow reactor operating at a methods were not great.
temperature of 449.82 K and a pressure of The second problem is an ammonia synthesis loop,
1.0342 x lo6 Nmm2. The diameter of the reactor is first presented by Parker [15]. The details of the
0.1533 m. The rates of the two reactions taking place process are given by Kisala [4]. The flowsheet for this
in the reactor are given by Arrhenius type kinetic process is given in Fig. 4. The objective function for
expressions, which are functions of reactant concen this problem was to maximize the net present value
tration and temperature: of the project. The economic analysis and net present
value were calculated rigorously using the ASPEN
A+P+G
PLUS costing and economic evaluation system.
rl = (3.53 x lO)exp(  35570/RT)C,, (12) Eleven variables were freed for the optimization
problem. The variables include operating parameters
r2 = (53.346)exp(  23670/RT)Ci, (13) for every unit in the process. There were two design
specifications and four inequality constraints im
where C, and C, are the concentrations of A and P,
posed on he optimization problem. The variables and
respectively, T is the temperature in K, and R is the
the constraints are summarized in Table 2.
gas constant. The flash drum operates at atmospheric
Table 3 gives a comparison of the sequential
pressure.
modular optimization algorithms for this problem.
The physical properties were computed assuming
Although IP required more iterations than CFV, IP
the following compounds in the system:
was somewhat faster, due to the amount of com
A = isobutyric acid, putation time necessary to solve the simulation prob
B = ethyl acetate, lem at each iteration for CFV. IPH converged in the
C = nbutyric acid. same number of iterations as CFV and was
significantly faster than either IP of CFV. These
Ideal physical properties were assumed for enthalpy
trends are typical of our experience with sequential
and equilibrium calculations.
modular optimization in ASPEN PLUS.
For this problem, the length of the plug flow
reactor was varied to maximize the price of the
product stream given by the expression: REFERETCES
product stream and g is the molar flow rate of by 2. L. T. Biegler, Optimization methods fo; sequential
product G in the same stream. The value of the modular simulators. Ph.D. Thesis, University of
Wisconsin, Madison (198 1).
objective function is plotted for different reactor 3. L. T. Biegler and R. R. Hughes, Infeasible path optim
lengths in Fig. 3. The optimal value is obtained with a ization with sequential modular simulators. AZChE JI
reactor length of 12.77 m which corresponds to a value 28, 994 (1982).
of the objective function of 8.89 x 10m3kmol sr. 4. T. P. Kisala, Successive quadratic programming in
The results for the twotier simultaneous modular sequential modular process flowsheet simulation and
optimization. Sc.D. Thesis, Department of Chemical
method and the sequential modular methods Engineering, Massachusetts Institute of Technology
described earlier are given in Table 1 in terms (1985).
of the number of simulation time equivalents 5. P. K: Gupta, R. C. Lavoie and R. R. Radcliffe, An
(STE = amount of time to solve the optimization industrial evaluation of SPEEDUP. Paper presented at
the 1984 AZCHE And Mtg, San Francisco (1984).
problem/amount of time to perform a sequential 6. M. H. Locke, A CAD tool which accommodates an
modular simulation at the optimum). The twotier evolutionary strategy in engineering design calculations.
simultaneous modular method was able to solve the Ph.D. Thesis, CarnegieMellon University (1981).
Sequential and simultaneous modular strategies 511
to [ to a large negative number ( IOk, for example). The routine itself allows bounds on the variables, the bounds
elements in the extra column and row of B for r are set to were passed to the QPP directly, instead of entering them as
zero. If no solution to the QPP can be found with 5 > 0, inequality constraints.
then the algorithm fails.
4. Check for convergence. Calculate the KuhnTucker Step limiting
error, K: One difficulty encountered with SQP is that on the first
few iterations, the solution of the QPP can predict very large
K=IVFTdl+lJ.CI+IpTGI (21)
changes in the variables, often leaving the variables at their
If K 4 c is the convergence tolerance, stop. upper and lower bounds. This can lead to failure of the
5. Find a value of the line search step size, (I, to satisfy algorithm, especially if the line search does not reduce the
the inequality predicted step sufficiently.
A heuristic which was used to aid convergence from poor
Pr(x+ ad, Ik, rk) C Pr(x, lk, #) + O.la/i, (22) initial points was to artificially to limit the change in the
Pr is a penaltytype line search objective function: variables on a given iteration. In the solution of the QPP
(23), bounds are placed on the change in the variables:
Pr(x,I,C)=F+C1:ICiI+CF:Imin(O,Gj)I, (23) l<d<u
, i (27)
A+ and p* are constraint weighting factors which are d is the vector representing the change in X. I and Y are the
functions of the Lagrange and KuhnTucker multipliers at lower and upper bounds on d. Thus, we can easily limit the
the solution of the QPP. On the first iteration, they are set change in any variable to, say, 20% of the step necessary to
equal to the Lagrange and KuhnTucker multipliers: reach one of the bounds by modifying the upper and lower
limits on d:
1*=111,
(28)
P*=IPI. (24)
where:
After the first iteration,
I+ = 0.2 *i,
1* = max { Ilk], l/2(1*, + ]1])},
u* =0.2 *u. (29)
p*=max{I~I, lP(~*,+l~~l)), (25)
Evaluation of derivatives
iWr
p is an approximation to  VFOZAD requires that the gradient of the objective
aa .=O function and the constraints be calculated on every call,
r9 = aFT6+ F  Pr(xk, Ak,flk), (26) even during the line search. Since this information is only
needed on the last line search point (the new point), a
B would be equal to
modification was made so that the gradients were not always
aPr required.

aa m=0 Convergence test
if all the constraints were linear.
Instead of using the absolute convergence test in equa
6. Set xk+=xk+ad, tions (2&6), the tolerance used in the convergence test was
Setk=k+l, the specified tolerance multiplied by the value of the objec
Go to Step 1. tive function at that point (6* F), unless F was less than 1.O,
in which case the specified tolerance was used. This elimi
Powells implementation of SQP is available as subroutine nated the need to change the tolerance based on the
VFOZAD in the Harwell Subroutine Library [25]. This
magnitude of the objective function.
implementation was used as the basis for the imple
mentation by Kisala [4]. Several modifications were made to Line search procedure
the routines which will be discussed below.
A change was made to the line search procedure so that
Quadratic programming package the algorithm is not allowed to terminate if a line search fails
to converge after five line search points. This is important
Perhaps the most significant change to Powells imple
since the line search may have trouble converging on early
mentation was a change in the algorithm for solving
iterations if the starting point is far from the optimum. Thus
the quadratic programming problem (QPP) (19). Powells
it is usually helpful to allow the algorithm to continue if this
implementation used a quadratic programming package
occurs on an early iteration, but to terminate if it happens
developed by Fletcher 1251.As pointed out by Biegler 121,
repeatedly or on later iterations.
this method is susceptible to roundoff error. Powell noted
Another change to the line search procedure was to
that a faster and more accurate quadratic programming
change the maximum step size reduction factor. In Powells
package was desirable. For the present implementation, the
implementation, the step size, a, would not be reduced by
SOL/QPSOL package [26] was used. QPSOL is an imple
less than 10% per line search point. This is actually quite
mentation of the active set nullspace method of Gill et al.
restrictive, allowing too much of a reduction, since, at the
[26]. Among the advantages of this package is that con
end of five line search noints, a could be as small as 0.00001.
straints believed to be satisfied at the solution of the QPP
In the present implementation, value of 30% was used,
can be specified. This warm start capability can save
makine the minimum value of a be 0.00243. This usually led
computer time in SQP since a good guess of the active
the algorithm to take larger steps and to converge in fewer
constraints is available from the solution of the QPP on the
iterations.
last iteration. The storage requirements for QPSOL are also
significantly less than for Fletchers routine.
APPENDIX B
Bounds on the variables LockeEdahl Westerberg Algorithm
In Powells implementation of SQP, bounds on the vari The LockeEdahlWesterberg (LEW) [20] algorithm is a
ables had to be entered as inequality constraints. This is modification to Powells SQP algorithm. In LEW, the
inefficient because of the storage space required for the optimization problem is decomposed into dependent and
larger Jacobian matrix of the constraints. Since the QPP independent (decision) variables. The equality constraints
Sequential and simultaneous modular strategies 579
are used to eliminate the dependent variables and the savings in storage requirements for the approximate Hessian
quadratic programming problem is solved in the decision matrix and in the computation time for updating the
variable space only. Thus the algorithm is similar to the approximate Hessian. Thus, this algorithm has been useful
generalized reduced gradient (GRG) algorithm [28], which for equationoriented [6] and simultaneous modular ap
also eliminates the dependent variables and the equality proaches to the process optimization problem where the
constraints from the optimization problem. Unlike the ratio of the number constraints to the number of decision
GRG method, LEW does not converge the constraints variables is very high (on the order of 100 or 1000 to 1).
before eliminating them. Rather it eliminates the constraints TrevinoLozano [1] observed that the LEW algorithm will
based on a linear approximation of the constraints, thus sometimes fail when the initial values of the variables make
performing one NewtonRaphson iteration toward con the quadratic programming problem on the first iteration or
verging the constraints. other early iterations infeasible. He recommends taking
Although this algorithm usually requires more iterations several NewtonRaphson iterations toward converging the
to converge the optimization problem than the original SQP constraints when this is likely to occur, thus approaching a
algorithm, the decrease in the number of variables in the feasible path method and gives a criterion for how many
quadratic programming problem results in a substantial NewtonRaphson iterations are necessary.