You are on page 1of 16

Corn/x&chcm. Engng, Vol. 12, No. 12, pp. 1171-1186, 1988 0098-I 354188 $3.00 + 0.

00
Printedin Great Britain.All rights reserved Copyright0 1988 PergamonF’resspk

COMPUTATIONAL EXPERIMENTS IN
EQUATION-BASED CHEMICAL PROCESS FLOWSHEETING

S. E. ZITNEY and M. A. STADTHERR


Department of Chemical Engineering, University of Illinois, 1209 W. California St, Urbana,
IL 61801, U.S.A.

(Received 27 July 1987; J%UZI revision received 22 April 1988; received for publication 4 May 1988)

Abstract-The central computational problem in equation-based flowsheeting is the simultaneous solution


of a large sparse system of nonlinear algebraic equations. SEQUEL-II, our implementation of the
equation-based approach, provides a means for comparing nonlinear equation-solving techniques, in
particular the techniques used for computing correction steps, for generating initial guesses, and for
evaluating sparse Jacobian matrices. Comparisons are made on the basis of both efficiency and reliability.
On a set of fourteen process flowsheeting problems, Powell’s dogleg correction step is not only very
efficient, but also somewhat more reliable than the Newton step when good initial guesses are not
available. Schubert’s sparse Jacobian update performs very well, but Bogle and Perkins’ new least-
relative-change update gives improved performance,when used in connection with either the dogleg or
Newton correction step. The use of a hybrid update provides a better approximation of the Jacobian than
a sparse update alone, and so significantly improves, in terms of number of iterations, the efficiency of
the dogleg correction step.

INTRODUCFION reliability of the correction-step methods and the


quality of several automatic initialization schemes
The equation-based (EB) approach is an attractive
adapted from those described by Locke (1981), Stadt-
alternative to the usual sequential-modular approach
herr and Hilton (1982), and Gupta et al. (1984).
to process flowsheeting. Recent years have seen con-
3. Sparse Jacobian EvaluatioeA Jacobian matrix
siderable progress in EB flowsheeting (Stadtherr and
is needed to generate a new set of guesses at each
Vegeais, 1985), with significant accomplishments in
iteration of the EB approach. The need to evaluate
commercial software, as discussed in more detail
the Jacobian analytically or by finite difference can be
below. It appears, however, that further im-
avoided by updating the Jacobian approximation
provements in a number of areas regarding nonlinear
from the previous iteration. The sparse Jacobian
equation solving are needed before truly general-
update devised by Schubert has been found to be
purpose EB systems achieve widespread industrial
somewhat unreliable in an evaluation of the general-
use.
purpose EB package SPEEDUP (Gupta ef al., 1984).
In this paper we focus primarily on the following
This is consistent with the findings of other studies
areas:
(e.g. Mah and Lin, 1980; Perkins and Sargent, 1982)
1. Correction stepPublished work on general- based on chemical engineering applications. How-
purpose EB flowsheeting has concentrated on the use ever, Chen and Stadtherr (1984), using a small set of
of the Newton correction-step method for quite some mathematical test problems, report that Schubert’s
time. Chen and Stadtherr (1984) have recently pro- update can be used successfully, at least when incor-
posed an alternative method in which Powell’s dogleg porated in the dogleg method for computing a cor-
(Chen and Stadtherr, 1981) technique is used for rection step. We thus study here the efficiency and
computing a correction step. We apply here their reliability of Schubert’s update, as well as a new
modification of Powell’s method (MP) to a set of least-relative-change update devised by Bogle and
chemical process simulation and design problems. As Perkins (1987). We also make comparisons to the
a basis of comparison, the Newton correction-step sparse finite difference scheme of Curtis et al. (1974).
method is applied to the same EB flowsheeting 4. Hybrid Jacobian method--The use of a hybrid
problems. The comparisons are based on efficiency method which makes combined use of available
and reliability. analytical derivatives and a sparse Jacobian update
2. Automatic initializatio_Process variable initial- has frequently been suggested, often in the context of
ization has a strong effect on the performance of the a chemical process flowsheeting environment (e.g.
correction-step methods used in the EB approach Lucia and Macchietto, 1983; Lucia et al., 1984; Lucia
(Stadtherr and Hilton, 1982). And since initialization and Westman, 1984; Westman et al., 1984; Field er
schemes are largely empirical in nature, there are a al., 1985; Macchietto, 1985; Miller and Lucia, 1985;
variety of such schemes that are possible. We thus Pantelides, 1986). We thus study here, using the MP
study here the interaction between the efficiency and correction step, the effect of such a hybrid method on
1171
1172 S. E. ZITNEY and M. A. STADTHERR

the solution of EB flowsheeting problems. We also EB programs are in industrial use, primarily for
make comparisons to a hybrid implementation that utility system calculation (Gordon et al., 1978) and
uses the Newton correction step. In this method some systems of multistage separation units (Hegner and
partial derivatives are computed analytically, while Schoenmakers, 1985). It should also be emphasized
others are approximated by finite difference rather that this list of industrial applications of the EB
than by a sparse Jacobian update. approach is by no means exhaustive.
For these computational experiments, fourteen Today there are several flowsheeting systems that
benchmark problems, involving four chemical pro- implement the EB approach. As mentioned above,
cess flowsheets, are used. The flowsheets include a among those that have been used or tested on a
simple ethylene process (example 3) from the CHESS commerical level are SPEEDUP (Imperial College,
User’s Guide (Motard and Lee, 1971) the ammonia London) (e.g. Perkins and Sargent, 1982; Paloschi et
synthesis process studied by Stadtherr and Hilton al., 1983; Pantelides, 1986), TISFLO II (Dutch
(1982), the well-known four-flash-unit system used by Mines) (e.g. Van Meulebrouk et al., 1982), CHEOPE
Cavett (1963), and a light hydrocarbons recovery (Montedison) (e.g. Pagani and d’Arminio Monforte,
process adapted from exercise 25 in the FLOW- 1987) and a process simulator developed at Amoco
TRAN exercise book (Clark, 1977). (Barnard et al., 1986). Other systems, such as
ASCENDII(Camegie-MellonUniversity)(e.g. Locke,
1981), QUASILIN (University of Cambridge, U.K.)
BACKGROUND
(e.g. Hutchison et al., 1986a, b; Smith and Morton,
It should be noted at the outset that two compre- 1987), FLOWSIM (University of Connecticut/
hensive reviews of EB process flowsheeting (Shacham Control Data) (Shacham et al., 1982), SEQUEL
et al., 1982; Perkins, 1984) have appeared in recent (University of Illinois) (Stadtherr and Hilton, 1982)
years. The review by Perkins is particularly valuable and DIVA (University of Stuttgart) (e.g. Ho11 er al.,
and covers many aspects of EB flowsheeting in some 1987; Marquardt et al., 1987), have been used on an
detail. A more recent paper by Stadtherr and Vegeais academic level as experimental prototype systems.
(1985) discusses the current status of EB flowsheeting
and highlights recent developments in four areas
related to EB flowsheeting, including industrial ex- Nonlinear equation solving in EB process$owsheering
perience, nonlinear equation solving, inclusion of EB process flowsheeting requires the solution of a
procedures, and sparse matrix methods. large sparse system of nonlinear equations. While
The EB approach offers speed and flexibility in initial work in EB flowsheeting typically involved
making process flowsheeting calculations, especially solution by tearing to reduce the number of variables
when dealing with complex, highly-integrated pro- iterated on, today the solution is almost always by
cesses. It also is an excellent approach for performing simultaneous linearization. In this case all the equa-
process optimization and dynamic simulations, tions are linearized and all the variables iterated on
though these applications are not considered here. simultaneously using a Newton-type method.
These potential benefits are generally recognized to- When the Newton-Raphson (NR) method is used,
day, and there is increasing industrial interest in this the need at each iteration to evaluate the Jacobian,
approach. either analytically or by finite difference, is a com-
For instance, and EB package CHEOPE (Pagani putational burden. The use of complete analytical
and d’Arminio Monforte, 1987) was designed at derivative information is expensive, and for many
Montedison, and has been used in performing the flowsheeting applications analytical partial deriva-
simulation, design, and optimization of some indus- tives are not available when a functional relationship
trial chemical processes, including toluene oxidation, is defined by a procedure. In fact, Pantelides (1986)
methyl methacrylate synthesis, urea processing, phe- recently pointed out that this problem is accentuated
nol purification, and hydrogen cyanide recovery. An in SPEEDUP by the very flexibility of the package in
EB process simulator was developed at Amoco, and allowing the user to define new process models. The
has been used for the simulation of a process for gas partial derivatives can also be evaluated by finite
sweetening by amine absorption (Barnard et al., difference, but generally at the cost of a significant
1986). The EB program TISFLO-II was developed at decrease in computational efficiency. For instance, in
Dutch State Mines, and has been used successfully in their evaluation of SPEEDUP, Gupta et al. (1984)
industrial applications, including simulation of an found that the finite-difference NR method was al-
ammonia plant (Cronin et al., 1985) optimization of ways quite slow in convergence, and thus they used
utility generation and supply to a chemical complex this method very little in their testing.
(Van Meulebrouk et al., 1982), and olefin plant The fact that the NR method is frequently costly
optimization (De Leeuw den Bouter ef al., 1983; and time consuming has prompted considerable re-
Barendregt et al., 1987). The general-purpose EB cent interest in sparse quasi-Newton (QN) methods in
package, SPEEDUP, has been evaluated recently by which an approximation to the Jacobian is updated
Exxon (Gupta et al., 1984) and has been used else- at each iteration. One such update is the modification
where industrially as well. Also, some special-purpose of Broyden’s update (Broyden, 1965) proposed by
Computationalexperimentsin chemicalprocess flowsheeting 1173

Schubert (1970). Several studies have found Schu- formance of this new update is compared to Schu-
bert’s update to be somewhat unreliable. For in- bert’s update in this study.
stance, Gupta et al. (1984) found that a sparse QN Another approach to improving the reliability of
method based on an idea by Paloschi (1982) gave an EB system is to provide better automatic initial-
better results than Schubert’s method. The reason for ization schemes. In fact, the need for globally con-
these results may be that the implementation of vergent techniques for solving the nonlinear equation
Paloschi’s method in SPEEDUP uses partitioning to systems can be alleviated somewhat given effective
decompose the flowsheeting problems, while the im- initialization schemes. And since they are largely
plementation of Schubert’s method does not. empirical in nature, any number of such schemes may
Unfortunately, local solution strategies that use the be devised. For instance, the three schemes imple-
Newton correction step (NR and QN) may become mented in SEQUEL by Stadtherr and Hilton (1982)
unreliable for large flowsheeting problems when good use various preselected default and “average” values
initial estimates of the process variables are not with heuristics. The evolutionary strategy studied by
available. One approach to overcoming this problem Locke (1981) uses increasingly complex models to
is to try to devise nonlinear equation solving methods solve the flowsheet at each step. Another procedure
with good global convergence properties. described by Locke uses a tearing strategy to deter-
A class of methods with generally excellent global mine a set of tear variables, which if guessed, would
convergence properties is the class of so-called con- allow the remaining variables to be initialized by
tinuation (or homotopy) methods. The use of such sequencing through the flowsheet. The EB program,
methods in EB problems has been discussed recently QUASILIN (Hutchison, 1986a), uses still another
by Waybum and Seader (1984a, b), Seader et al. approach in which the variables are initialized by
(1984) Vickery and Taylor (1986), Burton and solving a simplified linear system approximately
Morton (1987) and Lin er al. (1987). These methods modeling the flowsheet. Another possible scheme,
solve the original problem by transforming it into a suggested most recently by Gupta et al. (1984) is to
sequence of subproblems, in which the solution of do a few sequential-modular iterations to initialize
one subproblem is used to generate an initial guess the variables. We adapt several of these initialization
for the next subproblem. In this sense the con- schemes for use in this study.
tinuation methods are similar to the evolutionary
strategy for problem initialization proposed by Locke
(1981). The efficiency of a continuation method es- TEST CODE: SEQUEL-II
sentially depends on the number of subproblems
needed to generate a good enough initial guess to In order to perform computational experiments
solve the original problem. Hlavacek (1984) has using the various nonlinear equation-solving tech-
reported that in his experience relatively few sub- niques described in this study, as well as to investigate
problems are needed in some chemical engineering other computational problems arising in EB
applications. While these methods do not guarantee flowsheeting, we have enhanced the EB system, SE-
convergence (e.g. Burton and Morton, 1987), except QUEL, developed by Stadtherr and Hilton (1982).
in special cases, they may converge to a solution when This new system, dubbed SEQEUL-II, is, like its
other methods fail and may find multiple feasible predecessor, perhaps best regarded as an experi-
solutions when they exist. There are apparently no mental prototype system. The recent developments
published comparisons of continuation methods to are discussed here. We first briefly outline the sparse
other initialization schemes in general-purpose EB nonlinear equation solver implemented in SEQUEL-
flowsheeting. II, and then discuss, in detail, several of the more
Chen and Stadtherr (1984) have shown recently important aspects of this implementation.
how the use of Powell’s dogleg (Chen and Stadtherr,
1981) technique for generating a correction step can Sparse nonlinear equation solver
be used to improve the reliability of Schubert’s
SEQUEL-II uses a modified version of the sparse
update. They found that the MP dogleg method, a
nonlinear equation solver, SPNEQ, developed by
combination of ideas from the QN method and the
Chen and Stadtherr (1984) and Sun and Stadtherr
steepest descent method, is much more reliable than
(1988a). Some results for the SPNEQ code on a
the QN method alone, thereby permitting a more
variety of chemical engineering problems can be
successful application of Schubert’s update than has
found in Sun and Stadtherr (1988a). A brief descrip-
been reported elsewhere (Mah and Lin, 1980; Perkins
tion of the modifications of SPNEQ used here fol-
and Sargent, 1982; Gupta et al., 1984). These results
lows:
and those of Sun and Stadtherr (1988a) show that
Given an initial estimate x, of the solution, an
Schubert’s update may be well suited to EB
initial step bound A, and a functionf(x), perform the
flowsheeting, at least when used as part of the MP
following calculations:
method. Bogle and Perkins (1987) have recently
produced a new sparse update based on the idea of 1. Evaluate the Jacobian, B. using a sparse finite
least relative change in the Jacobian. The per- difference scheme or a hybrid method.
1174 S. E. Z~TNEYand M. A. S~~~ruras

2. Calculate a diagonal function scaling matrix Dr For more details, see Chen and Stadtherr (1981,
and/or a diagonal variable scaling matrix D,, 1984).
and then scale the Jacobian. SEQUEL-II pro-
vides automatic scaling of functions and/or Correction-step variants
variables. Alternatively the user may provide As indicated in the outline above, two correction-
his own scaling or use no scaling at all. step variants within the SPNEQ code are available in
3. Perform the LU factorization of the scaled SEQUEL-II:
Jacobian. In order to maintain sparsity,
SEQUEL-II uses sparse matrix methods imple- MP-Implements the dogleg correction step in
mented in a two-pass approach. Step 4. This is the usual correction step used
4. For iteration k, calculate the correction step pk in SPNEQ by Chen and Stedtherr (1984).
and adjust pk as necessary to restrict variable DN-Implements the damped Newton correction
values to within physically feasible limits. The step in Step 4. This code variant is the same
scaled MP (dogleg) correction step is calculated as MP except that in Step 4 the calculation
here as follows: of the steepest descent direction and the
application of Powell’s dogleg formula are
DxPk = D,P: if A >O.Sl IID,pfj(, bypassed.
D,P, = aD,~f + (1 - alDx~;T Note that the DN and MP variants will yield the
same results when the size of the scaled Newton Step
if 0.81 IID,p,NII k A =- IID,pSII,
falls within the trust region defined using A. This is
DxPk = AD,&lllD,gk II if IlD,~fll 3 A, most likely to occur when A is large (indicating a
good linearization).
where )I- II indicates the Euclidian norm, p: is
It should be emphasized that these correction-step
the Newton correction step, p f is the step in the
variants are implemented in SEQUEL-II using the
steepest descent direction g,, and a is deter-
SPNEQ code as a framework. This means that if the
mined from IID,p, II = A as shown in Chen and
same Jacobian evaluation technique is used, the code
Stadtherr (1981). When the scaled Newton step
for solving a problem with MP and the code for
alone is used, it is damped using step-size
solving a problem with DN are exactly the same,
control, and the correction step pk is calculated
except that, as noted above, two sections of code are
using the following equations:
bypassed when DN is used. Thus, the MP and DN
D,P, = DA if A>O.SlIID,plyII, variants can be compared directly by simply using the
same Jacobian-evaluation variant. This is an attempt
DxPk = BD.xptv if O.SlllD,p~II 2 A,
to achieve a meaningful comparison of correction-
where /I is A/l/ D,pc 11. step methods by minimizing implementation effects.
5. Calculatef‘(x, + pk) and check for convergence. Sparse Jacobian -evaluation variants
The relative convergence test is:
For both the MP and DN correction-step vari-
llD,~LNll G 6 max[llD,x /I, 1.01 ants, SEQUEL-II provides five sparse Jacobian-
evaluation variants, including two Jacobian updates,
and
a finite difference scheme, and two hybrid methods:
SCH-Implements Schubert’s update.
where 6 is the user-provided number that BGL-Implements Bogle and Perkins’ least-
specifies the desired accuracy of the solution. relative-change update (1987).
6. Check for very slow convergence or non- CPR-Implements Curtis er al., finite difference
convergence. If there is little or no im- scheme (1974). A variable which occurs
provement, stop the algorithm and return error in a subset of the equations is perturbed
messages. at the same time as a variable appearing
7. Check whether it is necessary to calculate a new in another subset of equations provided
Jacobian. The Jacobian is re-evaluated occa- these two subsets have no equations in
sionally when the rate of convergence begins to common. By taking advantage of spar-
slow considerably. If a new Jacobian is re- sity in this manner, the total number of
quired, return to Step 1. Step 7 is ignored when function evaluations required to evaluate
the finite difference implementation is used. the Jacobian is fewer than the n + 1
8. Update X, and step bound A. Then return to function evaluations needed for a simple
Step 1 when the finite difference imple- finite difference approximation. In fact,
mentation is used; otherwise continue to Step 9. the number of function evaluations re-
9. Update the Jacobian using Schubert’s update, quired depends on the structure of the
Bogle and Perkins’ least-relative-change update occurrence matrix and is typically about
(1987), or a hybrid method. Return to Step 3. a few multiples of the number of com-
Computationalexperimentsin chemicalprocess flowsheeting 1I75

ponents for a process flowsheeting prob- been made here to determine which mix provides the
lem. Table 2 lists the number of function best combination of overall efficiency and reliability.
evaluations (NCPR) required by CPR to Also, as in the case of the correction-step variants, all
perform a finite difference Jacobian eval- of the Jacobian-evaluation variants have been impJe-
uation for the flowsheeting problems in mented within the framework of the SPNEQ code.
this study. This scheme is not always This means that if the same correction-step variant is
optimal, although the number of func- used, the only difference among the Jacobian-
tion evaluations required is usually close evaluation variants is that part of the code that
to the minimum. Coleman and More actually performs Jacobian evaluation or updating.
(1983) have recently proposed an alter-
native finite difference scheme based on
Scaring
the graph coloring problem. In a recent
comparative study using chemical en- It is not uncommon in EB flowsheeting problems
gineering problems, Sun and Stadtherr for the process variables and model equations to be
(1988b) have found that this new scheme very badly scaled, perhaps differing by several orders
performs only slightly better than the of magnitude. This can adversely affect the con-
CPR scheme while requiring signifi- vergence of the correction-step methods. To minimize
cantly more execution time. the sensitivity of the methods to scaling, SEQUEL-II
HCPR-Impiements hydrid version of the CPR provides Chen and Stadtherr’s automatic scheme
scheme. The partial derivatives of linear, (198 1) for the scaling of functions, variables, or both.
bilinear and trilinear terms are computed The SEQUEL-II user is also given the option of
analytically. This typically includes par- supplying his own scaling factors, or setting one or
tial derivatives from mass balance equa- both scaling factors to the identity matrix (no scal-
tions and some terms from equilibrium ing).
relationships, energy balance equations, An alternative approach to the scaling problem is
and linearized thermophysical property to use scale-invariant solution techniques. For in-
expressions. All other derivatives are stance, a QN method using BogJe and Perkins’
evaluated by the CPR scheme. The ele- least-relative-change update (1987) is scale invariant.
ments of the Jacobian to be computed However, since SPNEQ is designed to implement the
analytically are simply flagged so that dogleg approach as one of the correction-step alter-
they are not evaluated by finite differ- natives, and this approach is inherently scale de-
ence. Table 2 lists the number of func- pendent, no matter what Jacobian update is used, we
tion evaluations (NHCPR) needed to are not able to take advantage of the scale invariance
compute the finite difference portion of of Bogle and Perkins’ update. The fact that, as will
the hydrid Jacobian for the flowsheeting be seen below, this update still performs extremely
problems in this study. It should be well, suggests that the automatic scaling procedure
noted that since some partial derivatives used (function scaling only) is quite effective in
are computed analytically, NHCPR is in dealing with scaling problems.
general less than NCPR, and in some
cases much less. The percentage of non-
zero Jacobian elements calculated ana- Sparse matrix metho&
lytically for each problem is listed in SEQUEL-II uses a two-pass approach to solve the
Table 2. Note that some other strategy large sparse systems that arise after linearization in
for deciding which Jacobian elements are the EB approach. In the two-pass approach, the first
to be done analytically could be used, pass is a reordering phase, in which the coefficient
though the one used here is simple and matrix is reordered into a form desirable for pre-
yields a high percentage of analytically- serving sparsity. In the second pass, the numerical
determined elements. phase, the solution vector is calculated. If necessary,
HSCH-Implements hybrid version of SCH. The the column reordering determined in the first pass is
partial derivatives of linear, bilinear and changed to preserve numerical stability. A detailed
trilinear terms are computed analytically description of the reordering and solution algorithms,
(see HCPR above), while all other deriv- as well as the fundamental techniques common to
atives are approximated using Schubert’s most of them, is given by Stadtherr and Wood
update. The elements of the Jacobian to (1984a, b).
be computed analytically are simply
flagged so that they are not updated.
Variable constraint checking
It should be noted that any number of mixes of In many EB flowsheeting calculations, the process
exact and approximated derivative information can variables must be within a particular range to be
be used in such hybrid methods and no attempt has physically feasible. The values of the process vari-
1176 S. E. ZXNEY and M. A. STADTHERR

ables may vary considerably during the iterative specified, they are estimated, if possible, using the
solution process, and sometimes these values may be values for those components specified elsewhere in
outside the range of physical feasibility. In some cases the flowsheet. For an unspecified feed stream com-
this may cause the nonlinear equation solver to ponent for which estimation is possible, the feed
converge to an infeasible solution or to fail because flowrate is set to the average of the flows specified
of numerical difficulties in performing function and elsewhere for that component times the number of
derivative evaluations. Since the nonlinear equation process output streams. For components that do not
solver does not provide protection against constraint have a flowrate specified anywhere in the flowsheet,
violations, a routine for resetting the process vari- the feed stream flowrates are set to the average of all
ables must be used. This routine can implement either specified component flowrates.
local constraints on individual variables or global Scheme 2-This initialization scheme is adapted
constraints on all variables of a given type. from a scheme (type 3) given by Stadtherr and Hilton
In SEQUEL-II, the process variables are subject to (1982), and differs from Scheme 1 primarily in the
global constraints. Flowrates and pressures are con- method used to initialize the component flowrates. In
strained to be positive. The magnitudes of enthalpy this case a set of heuristics, one for each type of
variables are kept below some maximum value. Tem- process unit, is used to estimate the flowrates. An
peratures, K-values, and split fractions are con- initial estimate is generated by beginning with the
strained to be within minimum and maximum values. feed streams, which are initialized as described in
If a variable violates a constraint, we recompute or Scheme 1, and moving sequentially through the
simply reset the variable. For example, when a flowsheet until the component flowrates for all the
flowrate becomes negative at iteration k, its value is process streams have been generated. Only one se-
recomputed using the following procedure: quential pass is made through the flowsheet, so no
iterations are performed on the recycle streams. The
F *t t = -AFk if Fk=Q,
heuristics for the individual process units can be
otherwise found in Stadtherr and Hilton (1982). In using these
heuristics, values given by design specifications al-
F kt = b’F,,
ways override the guessed flows and assumed param-
,

where b is a user-provided number, AF, is the change eters. A variety of such heuristics could be devised for
in F predicted by iteration k, and r is the max this initialization scheme.
(- 2.0, AF,/F,). Note that the constraints used here Scheme L-This scheme is a simplified version of
are arbitrary and do not necessarily improve the sequential-modular initialization scheme sug-
efficiency, but they do extend the applicability of the gested by Gupta er al. (1984). In our scheme, two
SPNEQ code by making possible the solution of sequential passes are made through the process
problems with variables that must be within a partic- flowsheet. The passes are performed by beginning
ular range to be physically feasible. with the feed streams, which are initialized as de-
scribed in Scheme 1, and moving sequentially
Automatic initialization through the flowsheet using heuristics similar to those
used in Scheme 2. The heuristics for the process units
In EB flowsheeting, initial estimates of all the
are as follows:
process variables are needed to start the iterative
solution process. Since we regard SEQUEL-II as an
Stream mixer-for the first sequential pass,
experimental tool for testing different computational
guess the unknown input flowrates using the
methods, it is desirable to provide a number of
heuristic from Scheme 2. All stream flows into
initialization schemes, ranging from relatively poor to
the mixer are known in the second sequential
relatively good. SEQUEL-II systematically generates pass. Guess the output stream temperature by
an initial guess by using one of the following four
assuming that each input stream contributes an
automatic initialization schemes:
amount equal to its temperature times its per-
Scheme l-This initialization scheme is adapted centage of the total flow.
from a scheme (type 2) given by Stadtherr and Hilton Stream divider and split fraction separator-
(1982). In this scheme, temperatures and pressures guess the output stream flows using the corre-
are set to preselected reference values, enthalpies and sponding heuristics in Scheme 2. Set the tem-
heat loads are set to zero, and split fractions are set perature and pressure of the output streams
to be equal. These preselected values are overridden equal to the temperature and pressure of the
if other values are specified as design variables. The input stream.
K-values are initialized using Raoult’s Law, and Equilibrium flash-guess the output stream
vapor pressures are estimated using the Antoine flowrates by performing an isothermal flash
equation along with the preselected values (or design calculation. If the calculation does not converge,
variable values) for temperature and pressure. The use the heuristic from Scheme 2.
component flowrates for all streams are set to the feed Reactor-guess the output stream flowrates us-
stream values. If the feed stream values are not ing the corresponding heuristic in Scheme 2. Set
Computational experiments in chemical process flowsheeting 1177

the output stream pressure equal to the input The test problems are also classified based on the
stream pressure. type of flash specification used. The problems which
5. Heat exchanger-set the output stream have flash temperatures specified are referred to as
flowrates and pressure equal to their corre- isothermal flash problems (I). In other problems, the
sponding input values. heat loads are specified. These are referred to as
6. Pressure change--set the output stream adiabatic flash problems (A) even though the heat
flowrates and temperature equal to their corre- load may be specified to be nonzero.
sponding input values. The problems also differ with respect to which
variables are specified. The problems which have the
It should be noted that direct substitution is used
feed stream and equipment parameters specified are
to update the recycle stream flowrates and that no
categorized as simulation problems (S). The remain-
control loops are used to handle design specifications.
ing problems are design problems (D) with one or
As in Schemes 1 and 2, values given by design
more output flows specified.
specifications always override the guessed flows and
The test problems used in these computational
assumed parameters.
experiments are not particularly large by flowsheeting
Scheme &This scheme uses an evolutionary ap-
standards, but they are comparable in size to test
proach (Locke, 1981) to flowsheet initialization. The
problems used by others. Statistics on the sizes of the
solution of a mass balance problem is used as a seed problems are given in Table 2. This table lists for each
for the more difficult mass and energy balance prob- problem the number of variables (IV), the number of
lem. The initial guess for the mass balance problem
elements in the occurrence matrix (NZ), and the
is generated by one of the three schemes described
percentage density of the matrix.
above.
A brief description of each flowsheeting problem is
given here.
I. Simple ethylene plant. The first test problem is
COMPUTATIONAL EXPERIMENTS
the simple ethylene process (example 3) from the
In this section we briefly describe the flowsheeting CHESS manual (Motard and Lee, 1971). This prob-
problems used in the computational experiments, and lem only involves material balance equations with no
discuss the options used in SEQUEL-II. design constraints. Most equations are linear using
the equation-variable formulation in SFQUEL-II.
Flowsheeting problem details However, it should be noted that this problem can be
In our computational experiments we used four- formulated as a completely linear problem.
teen problems involving four process flowsheets, all 2. Ammonia synthesis process. The five ammonia
of which are familiar in the flowsheeting literature. As manufacture problems are listed as problems 2-6 in
shown in Table 1, these test problems include one Table 1. The process flowsheet, user input file, and
simple ethylene production problem, five ammonia problem specifications can be found in Stadtherr and
synthesis problems, five Cavett problems, and three Hilton (1982).
light hydrocarbon-recovery problems. We use 3. Cauett’s process. The third test problem is the
different versions of the same flowsheet to provide a well-known four-flash-unit process studied by Cavett
wide range of computational difficulty. For example, (1963). A sixteen-component system is used here. The
mass balance problems (M) are usually easier to solve five problems involving Cavett’s process are listed as
than mass and energy balance problems (M + E). problems 7-l 1 in Table 1. The flowsheet and

Table 1. Flowshectinn oroblcm classification

Problem BalallCX Flash Problem


Process flowsheet number type’ type2 type’
Ethylene Plant 1 M - s

Ammonia Synthesis Process 2 M I


3 M I :
4 M+E I s
5 M+E I D
6 M+E A s

Cavett’s Process I M I s
8 M I D
9 M+E I s
10 M+E I D
II M+E A s

Light hydrocarbon process 12 M I s


13 M+E I S
14 M+E A S

‘Balance type: M = material balance; M + E = material and energy balance.


‘Flash type: I = isothermal flash units; A = adiabaticflash units; - = no Rash units.
31’roblem Type: S = simulation problem; D = design problem.
1178 S. E. ZITNEYand M. A. STADTHERR

Table 2. Statistics of flowsheeting problems. Number of variables (N), number of nonzeros (NZ), percentage density, number of fun&on
evaluations (NCPR) required by CPR to perform a finite difference Jacobian evaluation, number of function evaluations (NHCPR) nqtid
by CPR to compute the finite difference portion of the Jacobian in HCPR, and percentage of n~nzer~ Jacobian elements calculated
analytically
_ . for HSCH and HCPR

Analytic
Problem Density derivatives
Process Rowsheet number N NZ (%) NCPR NHCPR for HSCH/HCPR (%)

Ethylene Plant 1 163 509 1.92 14 8 75.64

Ammonia synthesis process 2 133 399 2.26 10 6 87.47


3 133 399 2.26 10 6 87.47
4 155 518 2.16 10 10 75.68
5 155 518 2.16 10 10 75.68
6 155 518 2.16 10 10 75.68

Cawtt’s process 7 273 1054 1.41 19 1 87.86


8 273 1054 1.41 19 1 87.86
9 288 1273 1.53 18 75.02
10 288 1273 1.53 ;; 18 75.02
II 288 1273 1.53 19 18 75.02

Light hydrocarbon process 12 321 965 0.94 12 I 95.85


13 350 1217 0.99 14 12 81.18
14 350 1217 0.99 14 12 81.18

specifications for these problems are given in Stadt- 5. All runs using SEQUEL-II were done on a CDC
herr and Hilton (1982). Cyber 175 in single precision (60 bits) and under the
4. Light hydrocarbons recovery process. The last test Fortran extended (OPT = 1) compiler.
problem is a light hydrocarbons recovery process.
This problem is adapted from exercise 25 in the
FLOWTRAN exercise book (Clark, 1977). The three RESULTS

problems involving this recovery process are denoted The computational experiments comparing six
as problems 12-14 in Table 1. different combinations of correction-step methods
and Jacobian-evaluation methods are shown in
SEQUEL -II options
Tables 3-6. Each table gives the numerical results for
Throughout this study the following options are the test runs using a particular initialization scheme.
used in SEQUEL-II: For each run, we list the number of iterations (ITER),
1. All thermophysical properties are handled using function evaluations (NFCN) and Jacobian evalu-
ations (NJAC), and indicate failure by (F). The
thermodynamics case 2 described by Stadtherr and
Hilton (1982). In this case, linearized K-value equa- tabulated results represent the counts when the solu-
tions are included in the process matrix. In addition, tion is found or when the method fails. The function
a linearized enthalpy equation for each stream in the evaluation count includes those used for making
flowsheet is included in the matrix to allow the update finite difference evaluations of the Jacobian. At the
of each stream’s temperature. After a process matrix bottom of each table, we list the timing results. The
iteration, current values of temperature, pressure, total execution time includes setup time, nonlinear
and composition are provided to subroutines, which time and linear time. The setup time is essentially the
in turn update, perhaps iteratively, the thermo- time spent reading input data and generating the
physical property values required for function and initial guess. The nonlinear time is the time spent in
Jacobian evaluation. Linearized dew and bubble calculations such as function and derivative evalu-
point equations and phase determination equations ations, scaling the Jacobian, constraint checking, and
do not appear in the process matrix. These properties diagnostic printing. The linear time is the time spent
are evaluated in external subroutines. in the repeated elimination of the linearized matrix.
2. The automatic function scaling scheme of Chen All computation times are given in CPU s.
and Stadtherr (1981) is used to scale the Jacobian. To facilitate a quantitative comparison between
3. The P4 algorithm devised by Hellennan and the methods, we adapt several indices used by Bogle
Rarick (1972) is used in the reordering phase of the and Perkins (1987). As a basis for our measures, we
two-pass approach for solving the sparse linear sys- use NFCN, to denote the total number of function
tems. LUlSOL, a very efficient and reliable routine evaluations used by method i on problem j. If n,
described elsewhere (Stadtherr and Wood, 1984b; represents the total number of problems attempted
Chen and Stadtherr, 1984), is used in the subsequent with method i and si is the number successfully
solution phase. solved, then we define the following performance
4. The relative convergence tolerance is lo-‘. The factors:
perturbation factor in the Jacobian evaluation is EF-Efficiency factor. The EF for method i
10-5. The user-provided number, b, for resetting on problem j, denoted as EF,, is
variables that violate constraints is set to be 1.3. (min,NFCN,)/NFCN,), where I includes all
Table 3. Number of iterations (ITER), function evaluations (NFCN) and Jacobian evaluations (NJAC) required to solve tbe set of 14 flowsheeting problems using initialization Scheme I. An (F) indiites
failure to solve the problem. The overall timing results are listed in CPU sat the boltom of the table. MP and DN indicate correction-step methods and SCH, BGL and CPR indicate sparse Jacobian-evaluation
methods, as described in the text
Problem MP-SCH DN-SCH MP-BGL DN-BGL MPCPR DN-CPR
number (lTER/NFCN/NJAC) (lTER/NFCN/NJAC) (ITER/NFCN/NJAC) (ITER/NFCN/NJAC) (ITER/NFCN/?WAC) (ITER/NFCN/NJAC)

I 6/21/l 6/21/l 5/20/l 5/20/l 416114 416114


2 3116213 37/68/3 44/15/3 2814912 11/122/11 616716
3 4517613 2415513 3416513 3116213 9/100/9 9/100/9
4 28/59/3 5218313 (F) 2916013 4511613 (F) 13/144/13 17/188/17 (F)
5 59/lOO/4 41/82/4 50/81/3 4918013 12/133/12 717817
6 42/73/3 (F) 22/33/l (F) 4117213 (F) 26/37/l (F) 17/188/17 (F) l7/188/17 (F)
7 3417312 4019813 3217112 2916812 16/321/16 100/2081/100 (F)
8 66114314 12/32/l I: 64/141/4 10/30/l (F) 271541127 100/2001/100 (F)
9 2818613 3819613 (F) 2716612 2015912 12/241/12 s/la119
IO 44/W/4 (F) 78/174/5 (F) 33/9l/3 (F) 79/l75/5 (F) 100/2001/100 (I=) 100/2001/100 (F)
II 2416312 (F) 59/117/3 (F) 3417312 (F) 56/114/3 (F) 100/2001/100 (F) l00/2001/l00 (F)
I2 4219114 3415912 (F) 44l9314 55/92/3 (F) 30/391/30 100/1301/100 (F)
I3 59/102/3 (F) 3316212 (F) 3516412 (F) 30/59/2 (F) 100/1501/100 (F) Ioo/l5ol/loO (F)
14 39/82/3 (F) 86/157/5 (F) 37/80/3 (F) 72/129/4 (F) 100/1501/1w (F) lOD/l5OljlOO (F)
Timing results
Setup time (s) 2.398 2.439 2.412 2.408 2.410 2.401
Linear time (s) 67.839 78.318 64.337 74.29 I 101.219 121.616
Nonlinear time (s) 94.750 117.283 88.049 103.265 234.985 312.620
Total time (s) 164.987 198.040 154.798 179.964 338.614 436.637

Table 4. Number of iterations (ITER), function evaluations (NFCN) and Jacobian evaluations (NJAC) required to solve the set of I4 Rowsheetingproblems using initialization Scheme 2. The overall timing
results are listed in CPU s at the bottom of the table. MP and DN indicate correction-step methods, and SCH, BGL and CPR indicate sparse Jacobianevaluation methods, as described in the text
Problem MP-SCH DN-SCH MP-BGL DN-BGL MP-CPR DNCPR
number (ITER/NFCN/NJAC) (ITERJNFCNINJAC) (ITER/NFCN/NJAC (ITER/NFCN/NJAC) (ITER/NFCN/NJAC) (ITER/NFCN/NJAC)

I 4/19/l 4/19/l 4/19/l 4/19/l 314613 314613


2 I 9/40/2 2314412 14/35/2 1413512 515615 515615
3 30/51/2 29/20/2 17/38/2 1713812 7/78/7 l/18/7
4 3215312 29jsOj2 14/25/l 14/25/l 717817 8/89/S
5 I9/40/2 2114212 I E/39/2 18/39/2 10/l11/10 10/111/10
6 2415513 2614712 29/50/2 30/51/2 9/lcQ/9 10/111/10
7 21/60/2 2816112 2216112 2316212 6/121/6 6/121/6
a 21/60/2 17/M/2 I 715612 17/56/2 7/141/7 7/141/7
9 18/57/2 17/56/2 2O/S9/2 2216112 6/121/6 6/121/6
IO 22/80/3 M/88/3 25/83/3 31/89/3 7/141/7 Y/141/7
II 18/57/2 1815712 21/60/2 2616512 7/141/7 7/141/7
12 30/67/3 2214112 2515012 2014512 18/235/18 18/235/18
I3 39/82/3 40/97/4 3616512 2615512 17/256/17 17/256/17
I4 2sj54/2 2415312 3716612 3516412 7/W/7 P/106/7
Timing results
Setup time (s) 2.495 2.511 2.493 2.498 2.503 2.501
Linear time (s) 37.231 40.686 35.002 37.439 14.359 14.616
Nonlinear time (s) 60.785 61.417 64.260 64.347 112.207 I 12.620
Total time (s) 100.51 I 104.614 101.755 104.284 129.069 129.737

__ -. -- . - _. .- __..... ---- - - - _ _. _ - .. - _-_ . ._ . ..-..


Table 5. Number of iterations (ITER), function evaluations (NFCN) and Jacobian evaluations (NJAC) required to solve the set of 14 flowsheeting problems using initialixation Scheme 3. The overall
timing results are listed in CPU s at the bottom of the table. MP and DN indicate correction-step methods, and SCH, BGI and CPR indicate sparse Jacobian-evaluation methods, as described in the
text

Problem MP-SCH DN-SCH MP-BGL DN-BGL MP-CPR DN-CPR


number (ITER/NFCN/NJAC) (ITER/NFCN/NJAC) (lTER/NFCN/NJAC (ITER/NFCN/NJAC) (ITER/NFCN/NJAC) (ITER/NFCN/NJAC)
I 4/19/l 4/19/l 4/19/l 4/19/l 314613 314613
2 1313412 I 313412 10/21/l 10/21/l 414514 414514
3 1313412 1613712 10/21/l 10/21/l 6/67/6 616716
4 7/18/l 7/18/l 6/17/l 6/17/l 414514 414514
5 11/22/l 11/22/l 9/20/l 9/20/l 5/56/S 5/56/s
6 2214312 31/52/2 14/25/l 14/25/l 7/78/7 7/78/7
7 l2/51/2 I l/50/2 10/30/l 10/30/l 4/81/4 418114
8 3319113 2316212 2516412 21/60/2 9/l8l/9 ll/22l/ll
9 I l/50/2 I l/50/2 10/30/l 10/30/l 4/81/4 4/81/4
IO 2218013 26/84/3 2518313 3118913 15/301/15 I5/301/15
II I 615512 l6/55/2 I 315212 1515412 5/101/5 5/lOl/S
I2 4/17/l 4/17/l 3/16/l 3/16/l 3/40/3 3/40/3
I3 5/20/l 5/20/l 5/29/l 5/20/l 4/61/4 4/61/4
I4 I 6/45/Z I514412 1414312 l4/43/2 7/106/7 7/W/7

Timing results
Setup time (s) 4.335 4.3% 4.336 4.343 4.351 4.352
Linear time (s) 20.785 20.694 18.700 19.569 9.189 9.349
Nonlinear time (s) 46.374 47.548 43.665 44.149 IOo.059 loo.173
Total time (s) 71.494 72.590 66.701 68.061 113.519 113.874

Table 6. Number of iterations (ITER), function evaluations (NFCN), and Jacobian evaluations (.NJAC) required to solve the eight energy balance problems (problems 4-6,9-l I, I3 and 14) using initialization
Scheme 4. The overall timing results are listed in CPU s at the bottom of the table. MP and DN indicate correction-step methods, and SCH, BGL and CPR indicate sparse Jacobianevaluation methods,
as described in the text

Problem MP-SCH DN-SCH MP-BGL DN-BGL MPCPR DNCPR


number (ITER/NFCN/NJAC) (ITER/NFCN/NJAC) (ITER/NFCN/NJAC (lTER/NFCN/NJAC) (fTER/NFCN/NJAC) (ITER/NFCN/NJAC)
4 4/15/l 4/15/l 4/15/l 4/15/l 3/34/3 313413
5 4/15/l 4/15/l 4/15/l 4/15/l 3/34/3 313413
6 S/20/ I 9/20/l 7/18/l 7/18/l 6/67/6 a/67/6
9 6/26/I 6/26/l 5/25/l 5/25/l 5/lOl/5 5/101/5
IO 6/26/l 6/26/l 5/25/l I515412 5/lOl/S 5/101/5
II 14/53/2 I515412 1515412 I515412 5/101/5 5/101/5
I3 11/26/l 7/22/l 6/21/l 6/21/l 517615 5/76/5
I4 29/72/3 3517813 21/50/2 2215112 22/331/22 221331122
Timing results
Setup time (s) I.819 I.801 I .829 1.823 I .852 1.852
Linear time (s) 13.127 14.221 IO.991 II.131 8.630 8.630
Nonlinear time (s) 29.399 29:866 26.675 26.976 83.718 83.718
Total time (s) 44.345 45.888 39.495 39.930 94.200 94.2al

- - -. - - . - -..- - - . .. - - - - - - - - -. - - .. - _-_ . -.- .,=__


““-_“.~;__I . - ._ .,.
Computationalexperimentsin chemicalprocessflowsheeting 1181

,.oc

0.80

0.m

CAC

020

0.00
RF OPF

Fig. 1. Performancefactors for differentcorrection-step Methods


methods and Jacobian-evaluation methods using initial- Fig. 3. OPF for the different correction-step methods and
ization Scheme 1. Jacobian-evaluation methods using initialization Scheme 3.

successful methods tested. If method i failed In order to provide a similar quantitative com-
to solve problem j, EF, is assigned a value parison between initialization schemes, we simply
of zero. If method i is the most efficient of redefine NFCN,, to denote the total number of func-
all the methods on problem j, then EF,, has tion evaluations required by a method using scheme
a value of one. i for problem j. The definitions of the performance
OEF--Overall efficiency factor. The OEF for factors in this case follow naturally from those
method i is the summation of EF, over all defined previously for the solution methods. Figure 6
problems j divided by s,. presents the OEF, RF and OPF for the four initial-
RF-Reliability factor. The RF for method i is ization schemes on the entire flowsheeting problem
the fraction of problems successfully solved, set. There are 84 problems for each of the Schemes
si/ni. l-3 (14 problems with 6 different solution methods)
OPF-Overall performance factor. The OPF for a and 48 problems for Scheme 4 (8 problems with 6
method is the product of its OEF and its different solution methods).
RF.
The performance factors for the different combina-
tions of correction-step methods and Jacobian-
evaluation methods are shown in Figs l-4, with each 0.80

figure corresponding to a different initialization


scheme. When all problems are successfully solved, as 0.60

is the case using schemes 2-l, the RF of each method k


0
has a value of one. Thus, the OPFs are equal to the 0.40
OEFs for these methods, and we present only the
OPFs in Figs. 2-4. As a summary, Fig. 5 gives the 020

OEF, RF and OPF for all the methods on the entire


J
flowsheeting problem set. For each of the six meth- 0.00
Ma-sG-lct.l-SCH----
ods, the problem set includes 50 problems (14 prob-
Methods
lems with each of the Schemes 1-3 and 8 problems
with the Scheme 4). Fig. 4. OPF for the different correction-step methods and
Jacobian-evaluation methods using initialization Scheme 4.

.-_

oE.0 080

cm0 o.ao

b 0.40 0.40

020 020

0.00 0.00
*-se!- mtSCH bu-eGL -wzpR-
oEf= RF CR=

Methods Fig. 5. Performance factors for the different correction-step


Fig. 2. OPF for the different correction-step methods and methods and Jacobian-evaluation methods over all the
Jacobian-evaluation methods using initialization Scheme 2. initialization schemes.
1182 S. E. ZITNEY and M. A. STADTHERR

lb0 lutionary schemes may require a fair amount of


judgement on the part of the user in choosing the
cm0
m- sequence of problems to be solved.
-1-
0.60 - Correction step
-2-

0.40
CiE3-
-a-
1. Rest&s concerning the correction-step methods
EZZI- indicate that DN and MP are essentially equal in
terms of efficiency when using reasonably good start-
-4-
0.20
ing points (initialization Schemes 2-4). To under-
cl00 stand this it should first be recalled that the DN and
OEF RF OPF MP correction steps will yield the same results when
Fig. 6. Performance factors for automatic initialization the scaled Newton step is within the trust region.
Schemes 1-l on entire flowsheeting problem set. When schemes 2-4 are used, the Jacobian-evaluation
variants provide adequate linearizations at each iter-
DISCUSSION ation, thus A tends to be relatively large, and the
calculated Newton steps are usually within the trust
On the basis of the results shown in Tables 336 and region. Thus, the iteration sequence is essentially the
Figs l-6, we make the following observations: same for both the DN and MP methods. This is
especially evident when the CPR finite difference
Automatic initialization
scheme is used for Jacobian evaluation at each iter-
Of the three truly automatic initialization
1. ation. For example, when using Schemes 24, the
schemes tested (Schemes l-3), the best is the only problems for which the MP-CPR and DN-CPR
simplified sequential-modular strategy of Scheme 3 iteration sequences differ (because the scaled Newton
(RF = 1, OPF = OEF = 0.86). Other heuristics for step is outside the trust region) are problems 4 and
use in Scheme 3 can be developed for more complex 6 for initialization Scheme 2, and problem 8 for
process units and for systems with more complex Scheme 3. With the poor initializations provided by
equilibrium relationships. The use of an existing Scheme 1, however, the DN and MP iteration se-
sequential-modular flowsheeting system may provide quences are usually quite different, since the scaled
even better results. However, it should be noted that Newton step may often fall outside the trust region.
more setup time will generally be required for more In these cases, the MP correction step is somewhat
complex initialization schemes. The very simple less efficient than DN. This is primarily due to the
heuristic approach of Scheme 2 (RF= 1, fact that MP makes use in part of the steepest descent
OPF = OEF = 0.61) is not as efficient as scheme 3, direction, which is inefficient locally, though also
but performs comparably in terms of reliability. The more reliable.
development of a better set of heuristics may further 2. From the standpoint of reliability, the results for
enhance the performance. Scheme 1 (RF = 0.5, the correction-step methods are not surprising. When
OEF = 0.44, OPF = 0.22) is not competitive with the the more sophisticated initialization schemes
other schemes in terms of efficiency and reliability. (Schemes 24) are used, the MP and DN methods
The poor results for this primitive scheme are consis- both prove to be very reliable (RF = l), solving all
tent with those of Stadtherr and Hilton (1982), who the flowsheeting problems. When the most primitive
also found that using default and “average” values is scheme (Scheme 1) is used, the two methods become
ineffective for initializing large flowsheets. It is clear somewhat unreliable, but in this case MP is clearly
that the quality of the initialization scheme plays an the most dependable correction-step method tested.
important role in determining the ability of an EB The MP method solves 9 of 14 problems (RF = 0.64)
system to converge. with each of the three Jacobian-evaluation methods,
2. The evolutionary approach used in initialization while the DN method solves at most 6 problems
Scheme 4 (RF = 1, OPF = OEF = 0.92) is very (DN-BGL, RF = 0.43) and as few as 4 problems
effective for solving energy balance problems. This is (DN-SCH, RF = 0.29).
as expected since the mass balance solution is typi- 3. The best overall performer, based on the OPF
cally close enough to converge the energy balance results in Fig. 5, is MP-BGL (OPF = O-87), followed
problem quickly. It should be noted here that the closely by DN-BGL (OPF = 0.82) and MP-SCH
computational effort required to obtain the mass (OPF = 0.77). On the basis of these computational
balance solution used as the seed is not reflected in flowsheeting experiments, we recommend the use of
the analysis of Scheme 4. The total work required by the MP correction step, using either the BGL or SCH
the evolutionary approach includes not only the updates (see discussion below). However, DN should
computational effort neeeded to solve the energy also perform well, especially with the BGL update.
balance problem, but also the effort required to
Sparse Jacobian evaluation
obtain the original mass balance solution using any
of the three truly automatic initialization schemes 1. Overall results regarding efficiency (Fig. 5) indi-
(Schemes l-3). Other more sophisticated evo- cate that the best Jacobian-evaluation method is the
Computational experiments in chemical process flowsheeting 1183

BGL update (OEF = 0.97). Schubert’s update is also


very efficient (OEF = 0.87). As expected, the CPR
scheme is clearly the worst (OEF = 0.40) because
of the inefficiency of performing finite difference
Jacobian evaluations at each iteration. (Note here
that OEF is based on the number of function evalu-
ations, including those required for finite difference
Jacobian evaluations.) However, fewer iterations are
usually required with the CPR scheme because it
provides accurate Jacobian information. Thus, there
is a trade-off between accuracy and computational
efficiency when the Jacobian is evaluated by finite
difference
2. Schubert’s update is effective when used in
connection with both the MP and DN correction
steps, though MP-SCH is somewhat more reliable
than DN-SCH when good initial guesses are not
available. The dogleg correction step appears to
maintain numerical stability when SCH does not
provide an adequate approximation of the Jacobian.
The successful results here for SCH are noteworthy
because most previous experience (Mah and Lin,
1980; Perkins and Sargent, 1982; Gupta ef al., 1984)
with Schubert’s update has found it to perform
unreliably. This suggests that the poor reliability of
Schubert’s update reported in some previous studies
may not be due to the update itself, but due to the
implementation used. Note that we refer here not to
the coding of the update, but to other implementation
details of the code within which the update is used.
3. Overall the best of the sparse Jacobian updates
tested proves to be the BGL update. This new update
performs very well in connection with both MP and
DN, and is usually much more efficient than SCH. It
should be noted that Sun and Stadtherr (1988a), on
a variety of chemical engineering problems, also fmd
that BGL gives an improved performance over SCH,
but not as marked an improvement as in this study.
4. For quasi-Newton updates, such as BGL and
SCH, it is not uncommon for the Jacobian approxi-
mation to grow successively worse and inhibit
progress toward the solution. Thus, SEQUEL-II
monitors the rate of convergence, and uses the policy
devised by Chen and Stadtherr (1981) to decide
whether to re-evaluate the Jacobian by finite
difference. In order to directly compare the inherent
quality of the two Jacobian updates used here, we
can disable this Jacobian re-evaluation procedure.
Table 7 gives the numerical results for both updates
when used with the MP correction step and no
Jacobian re-evaluation. For each run, we list the
number of iterations (ITER) and function evalu-
ations (NFCN) and indicate failure by (F). Initial-
ization Scheme 4 is not applicable (NA) to the mass
balance problems. Table 7 clearly shows that
MP-BGL is much more reliable than MP-SCH. For
schemes 14, MP-BGL fails on ten fewer problems
than MP-SCH. In fact, MP-BGL does not fail on
any problem that MP-SCH solves. In terms of
efficiency, the number of function evaluations re-
1184 S. E. ZITNEYand M. A. ST,WTHERR

Table 8. Number of iterations (ITER), function evaluations (NFCN) and Jacobian evaluations (NJAC) required to solve the set of fourteea
flowsheeting problems when the MP correction-step method is used with rhe HSCH Jacobian update and the DN correction-step method
is used with the HCPR finite difference scheme. Problems are solved using initialization Schemes 2 and 3. The overall timing rcsubs are
listed in CPU s at the bottom of the table

MP-HSCH MP-HSCH DN-HCPR DN-HCPR


Problem Scheme 2 Scheme 3 Scheme 2 Scheme 3
NUltlbtX (ITER/NFCN/NJAC) (ITERINFCNJNJAC) (ITER/NFCN/NJAC) (ITER/NFCN/NJAC)

I 4/13/l 4/13/l S/28/3 312813


2 5/12/l 4/l l/l 5/36/5 412914
3 7/14/l 6/13/l 7/50/7 614316
4 7/18/l 7/18/l 7/87/7 414514
5 12/23/l 7/18/l 10/111/10 515615
6 12/23/l 10/21/l 12/133/12 9/100/9
7 6/8/l 4/6/l 6/13/6 41914
8 I l/13/1 10/12/l 7/l 517 11/23/11
9 ?/26/I 6/25/l 519615 4/77/4
10 9/28/I 1915612 7/134/7 15/286/l 5
II 9/28/I 10/29/l 6/l 1516 519615
12 18/20/I 3/5/l 18/37/18 31713
13 21/34/l 4/17/l 171222117 4/53/4
14 13/26/I 10/23/l 8/105/8 719217

Timine results

Setuptime (s) 2.525 4.499 2.535 4.514


Linear time (s) 18.339 12.940 14.133 10.586
Nonlinear time (s) 27.615 29.104 98.545 94.007
Total time (s) 48.479 46.543 115.213 109.107

quired by MP-BGL is less than or equal to the However, the use of HCPR does improve efficiency
number needed by MP-SCH for all problems solved by reducing the number of function evaluation
here. The results here indicate that BGL maintains a groups required by CPR for the finite difference
better Jacobian approximation than SCH. In fact, portion of the Jacobian.
many of the problems that required two Jacobian 3. In general, the improvements in efficiency when
evaluations for MP-BGL in Tables 34 are solved using hybrid Jacobian methods must be weighed
here using one Jacobian evaluation and only a few against the work necessary to compute Jacobian
more iterations. Thus, the re-evaluation policy used elements analytically. However, as long as the cost of
in SEQUEL-II for SCH may be too conservative for analytical derivative evaluations is not large, there is
use with BGL. The condition ensuring that con- no reason not to use hybrid methods of this sort.
vergence becomes quite slow before the Jacobian is In fact, the savings can be quite substational, as
re-evaluated can be changed to reflect the fact that demonstrated here.
the BGL maintains a better approximation.

Hybrid Jacobian method


CONCLUSIONS AND SIGNIFICANCE
1. The numerical results for MP-HSCH and
DN-HCPR are summarized in Table 8. Both combi- The computational flowsheeting experiments dem-
nations prove to be very reliable, solving all the onstrate that Powell’s dogleg method, as modified by
flowsheeting problems with initialization Schemes 2 Chen and Stadtherr (1981), is an effective correction-
and 3. step method for use in the EB approach. The cor-
2. Results regarding the efficiency of the hybrid rection step generated by the dogleg method is not
Jacobian methods are not surprising. The only efficient, but also somewhat more reliable than
MP-HSCH method is, in terms of number of iter- the Newton step when a poor initial estimate is
ations, much more efficient than MP-SCH because provided.
the hybrid update maintains a better approximation Of the three truly automatic initialization schemes
of the Jacobian at each iteration. This can be seen by tested, the best is the simplified sequential-modular
comparing the results for MP-HSCH in Table 8 with scheme. Results indicate that by using such a scheme,
those for MP-SCH in Tables 4 and 5. It should also perhaps in an evolutionary approach to problem
be noted that since a Jacobian factorization is re- initialization, larger and/or more complex flow-
quired at each iteration of the MP method, the use sheeting problems can be solved efficiently and re-
of HSCH with MP substantially reduces linear solu- liably using the EB approach.
tion time. On the other hand, a hybrid finite Overall the best of the sparse Jacobian-evaluation
difference scheme (HCPR) has little effect on the methods tested proves to be the new least-relative-
number of iterations and the amount of linear solu- change update recently devised by Bogle and Perkins
tion time required by the DN-CPR method, as can (1987). This new update performs very well in wn-
be seen by comparing the results for DN-HCPR in nection with both the dogleg and Newton correction
Table 7 with those for DNXPR in Tables 4 and 5. steps, and gives a marked improvement over the
Computational experiments in chemical process flowsheeting 1185

performance of Schubert’s sparse Jacobian update. Subscripts


Schubert’s update is substantially more effective than i = Method number
commonly believed, however. j = Problem number
k = Iteration number
Compared to a sparse update alone, the use of a
I = Successful methods
hybrid Jacobian update significantly improves, in
terms of number of iterations, the efficiency of the
dogleg correction step. This can result in a substantial
savings for large EB flowsheeting problems in which REFERENCES
the factorization of the Jacobian is a computational Barendregt S., H. M. Woerde, P. Boeren and C. Meotti, The
burden. application of equation based flowsheeting in advanced
plant management systems for olefin plants. Proc. Chem.
Acknowledgement-This work was supported in part by the Engng Fundam. ‘87, XVIII Cong.: The Use of Comput. in
National Science Foundation under Grant number Chem. Engng, European Federation of Chemical En-
DMC-8520663. gineering, Giardini Naxos, Italy, April 26-30 (1987).
Barnard W. L., D. R. Benjamin, D. L. Cummings, P. C.
Piela and J. L. Sills, Three issues in the design of an
equation-based process simulator. Paper No. 42d,
NOMENCLATURE Presented at AIChE Spring National Mtg, New Oreleans
(1986).
A = Adiabatic flash problem Bogle I. D. L. and J. D. Perkins, Improvements to sparse
E = Approximate Jacobian matrix quasi-Newton methods for algebraic systems. Proc.
b’ = Constraint checking parameter Chem. Engng Fun&m. ‘87, XVfZI Gong.: The Use of
BGL = Bogle and Perkins’ sparse Jacobian update Comput. in Chem. Engng. European Federation of Chem-
CPR = Curtis sparse finite difference scheme ical Engineering, Giardini Naxos, Italy, April 2630
D = Design problem (1987).
D, = Variable scaling matrix Broyden C. G., A class of methods for solving nonlinear
Df = Function scaling matrix simultaneous equations. Math Comput. 19, 577 (1965).
DN - Damped Newton correction-step method Burton P. J. and W. Morton, Differential arclength homo-
EF = Efficiency factor topy continuation in equation orientated simulation.
F = Solution method failure Proc. Chem. Engng Fun&m. ‘87, XVIll Cons.: The Use of
Fk = Flowrate variable Comput. in Chem. Engng, European Federation of Chem-
f = Function values ical Engineering, Giardini Naxos, Italy, April 2630
g = Steepest descent direction (1987).
HCPR = Hybrid version of CPR scheme Cavett R. H., Application of numerical methods to the
HSCH = Hydrid version of SCH update convergence of simulated processes involving recycle
I = Isothermal flash problem loops. Proc. Am. Petrol. Inst. 43, 57 (1963).
ITER = Number of iterations Chen H. S. and M. A. Stadtherr, A modification of Powell’s
L = Unit lower-triangular factor of B dogleg method for solving systems of nonlinear equa-
M = Mass balance problem tions. Comput. them. Engng 5, 143 (I 98 1).
M + E = Mass and energy problem Chen H. S. and M. A. Stadtherr, On solving large sparse
MP = Modified Powell’s correction-step method nonlinear equation systems. Comput. them. Engng 8, 1-7
N = Number of variables (1984).
n = Number of problems attempted Clark J. P., Exercises in Process Simulation Using FL.0 W-
NCPR = Number of function evaluations for CPR TRAN. CACHE Corp., Cambridge, Mass. (1977).
scheme Coleman T. F. and J. J. Mart, Estimation of sparse
NFCN = Number of function evaluations Jacobian matrices and graph coloring problems. SZAMJ.
NHCPR = Number of function evaluations for HCPR Numer. Analysis 20, 187 (I 983).
scheme Cronin P., S. Barendregt, R. Srinivasan and M. G. G. Van
NJAC = Number of Jacobian evaluations Meulebrouk, Optimization of an ammonia plant using an
NR = Newton-Raphson method equation based package. Paper No. 21d, Presented at
NZ = Number of nonzero elements AZChE Spring National, Houston, Texas (1985).
OEF = Overall efficiency factor Curtis A. R., M. J. D. Powell and J. K. Reid, On the
OPF = Overall performance factor estimation of sparse Jacobian matrices. J. Inst. Math.
p = Correction step Applic. 13, 117 (1974).
QN = Quasi-Newton method De Leeuw den Bouter J. A., M. G. G. Van Meulebrouk,
RF = Reliability factor A. G. Swenker and S. Barendregt, Olefin plant opti-
S= Simulation problem misation using SPYRO and TISFLO-II. Paper No. 4c,
s = Number of problems solved Presented at AIChE Spring National Mtg, Houston
SCH = Schubert’s sparse Jacobian update (1983).
U = Upper triangular factor of B Field A. J., R. P. Maddams and W. Morton, The incorpo-
x = Independent variables ration of procedures in an equation-oriented flowsheeting
environment. IChemE Symp. Ser. 92, 341 (1985).
Greek symbols Gordon E., M. H. Hashemi, R. D. Dodge and J. LaRosa,
01 = correction step parameter A versatile steam balance program. Chem. Engng Prog.
B = Step-size control parameter 74, 51 (1978).
A = Radius of trust region (step bound) Gupta P. K., R. C. Lavoie and R. R. Radcliffe, An industry
6 = Convergence parameter evaluation of SPEEDUP. Paper 63a, Presented at AIChE
Anni Mtg, San Francisco Calif. (1984).
Superscripts
Hegner B. and H. Schoenmakers, CHEMASIM-
N = Related to Newton method Experience with BASF’s simultaneous process simulator.
S = Related to steepest descent method IChemE Symp. Ser. 92, 365 (1985).
1186 S. E. ZITNEY and M. A. STADTHERR

Hellerman E. and D. Rarick, The partitioned preassigned simulation. Paper 4% presented at AIChE Ann! Mtg,
pivot procedure (P4) In Sparse Matrices and Their AppIi- Miami Beach, Florida (1986).
cations (D. .I. Rose and R. A. Willoughby, Eds). Plenum Perkins J. D., Equation-oriented flowsheeting. Proceedings
Press, New York (1972). of the Second fnternationai Conference on Four&lions of
Hlavacek V. H.. Invited discussion; nonsequential modular Computer -Aided Process Design (A. W. Westerherg and
flowsheeting, In Proceedings of the Second International H. H. Chien, Eds). CACHE, Ann Arbor (1984).
Conference on Foundations of Computer-Aided Process Perkins J. D. and R. W. H. Sargent, SPEEDUP: a computer
Deskn (A. W. Westerbera and H. H. Chien, Eds). oroaram for steady state and dynamic simulation of
CACHE; Ann Arbor (19845. chemical processes. in Selected Topb on Computer-Aided
Ho11 P., W. Marquardt and E. D. Gilles, DIVA-a powerful Process Design and Analysis (R. S. H. Mah and G. V.
tool for dynamic process simulation. Proc. Chem. Engng Reklaitis, Eds), AIChE Symp. Ser. 78, 1 (1982).
Fundam. ‘87, XVIII Cong.: The Use of Comput. Chem. Seader J. D.. R. Chavez and T. L. Wavburn. Multiple
Engng, European Federation of Chemical Engineering, solutions to systems of interlinked distillition cblumnsby
Giardini Naxos. Italy, April 26-30 (1987). differential homotopy continuation. Paper No. 63f,
Hutchison H. P., D. J. Jackson and W. Morton, The Presented at AIChE Annual Mtg, San Francisco, Califor-
develooment of an eauation-oriented flowsheet simu- nia (1984).
lation ‘and optimizati& package-I. The QUASILIN Schubert L. K., Modification of a quasi-Newton method for
program. Cornput. them. Engng 10, 19-29 (1986a). nonlinear equations with a sparse Jacobian. Math. Com-
H&cl&on H. p., D. J. Jackson and W. Morton, The pur. 25, 27 (1970).
development of an equation-oriented tlowsheet simu- Shacham M., S. Macchietto, L. F. Stutzman and P. Bacock,
lation and optimization package-II. Examples and re- Equation oriented approach to process flowsheeting.
sults. Comput. them. E&ng 10, 3147 (1986b). Comput. them. Engng 6, 79 (1982).
Lin W.. J. D. Seader and T. L. Wayburn, Computing Smith G. J. and W. Morton. Dvnamic simulation usina an
multiole solutions to svstems of in&linked se&ration equation-orientated flowsheeting package. Proc. Chem.
colu&s. AIChE J1 33,-886 (1987). Engng Fundam. ‘87, XVIII Cong.: The Use of Comput. in
Locke M. H., A CAD tool which accommodates an evo- _ -. Eurooean
Chem. Enanp. Federation of Chemical En-
lutionary strategy in engineering design calculations. gineering, Ciardini fiaxos, Italy, April 2630. (1987).
Ph.D. Thesis, Carnegie-Mellon University (198 1). Stadtherr M. A. and C. M. Hilton, Development of a new
Lucia A. and S. Macchietto, New approach to approxi- equation-based process flowsheeting system: numerical
mation of quantities involving physical properties deriva- studies. In Selected Topics on Computer-Aided Process
tives in equation-oriented process design. AIChE JI 29, Design and Analysis (R. S. H. Mah and G. V. Reklaitis,
705 (1983). Eds), AZChE Symp. Ser. 78, 12 (1982).
Lucia A., D. C. Miller and A. Kumar, Thermodynamically Stadtherr M. A. and J. A. Vegeais, Resent progress in
consistent quasi-Newton formulae. Paper 63d, Presented equation-based process flowsheeting. Proc. 1985 Summer
at AIChE Annf Mtg, San Francisco California (1984). Compur. Simulation Co@, Society for Computer Simu-
Lucia A. and K. R. Westman, Low cost solution to lation, Chicago, Illinois (1985).
multistage, multicomponent separation problems by a Stadtherr M. A. and E. S. Wood. Soarse matrix methods for
hvbrid fixed ooint alnorithm. In Proceedinps of rhe Second equation-based chemical process flowsheeting-I. Rear-
I&rnationai Confeyence on Foundations- oi Compuler- dering phase. Compur. them. Engng 8, 9-18 (1984a).
Aided Process Design (A. W. Westerberg and H. H. Stadtherr M. A. and E. S. Wood, Sparse matrix methods for
Chien, Eds). CACHE, &m Arbor (1984).- eouation-based chemical orocess flowsheetina-II. Nu-
Macchietto S., Solution techniques for processes described m’erical phase. Comput. &em. Engng 8, 1!%3? (1984b).
by mixed sets of equations and procedure IChemE Symp. Sun E. T. and M. A. Stadtherr, Issues in nonlinear equation
Ser. 92, 377 (1985). solving in chemical engineering. Comput. them. -Engng
Mah R. S. H. and T. D. Lin. Comoarison of modified 12, 1129-1139 (1988a).
Newton’s methods. Comput. ihem. kngng. 4, 75 (1980). Sun E. T. and M. A. Stadtherr, On sparse finite difference
Marquardt W., P. Holl, D. Butz and E. D. Gilles, DIVA-a schemes applied to chemical process engineering prob-
flow-sheet oriented dynamic process simulator. Chem. lems. Comuut. them. Enem? 12. 849-851 (1988b).
Engng Technol 10, 164 (1987). Van Meulebiouk M. G. G.,A. t?i. Swenke; and j. A. de
Miller D. C. and A. Lucia, The behavior of a hybrid Leeuw den Bouter, Using TISFLO-II for the optimization
tied-point method in a chemical process design environ- of utility generation and supply for a chemical complex.
ment. AZChE JI 31. 329 (19851. 1ChemE Symp. Ser. 74, 7 (1982).
Motard R. L. and H: M. &e. kHESS User’s Guide, 3rd Vickery D. J. and R. Taylor, Path-following approaches to
Edn University of Houston (1971). the solution of multicomponent, multistage separation
Paaani G. and A. d’Arminio Monforte. Some exveri process problems. AIChE JI 32, 547 (1986).
&es with the Montedison equation-or&ted simulator. Wayburn T. L. and J. D. Seader, Solutions of systems of
Proc. Chem. Engng Funabm. ‘87, XVIII Cong.: The Use interlinked distillation columns bv differential homotoor-
of Compur. Chem. Engng, European Federation of Chem- continuation method. In Proceedkgs of the Second Iniei-
ical Engineering, Giardini Naxos, Italy, April 26-30 national Conference on Foundations of Computer-Aided
(1987). Process Design (A. W. Westerberg and H. H. Chien. Eds).
Paloschi J. R., The numerical solution of nonlinear equa- CACHE, A& Arbor (1984a). -
tions representing chemical processes. Ph.D. Thesis, Uni- Wayburn T. L. and J. D. Seader, Degree theory and
versity of London (1982). homotopy-tools for computer-aided process design.
Paloschi J. R., J. D. Perkins and R. W. H. Sargent, Paper No. 27d, Presented at AIChE Ann1 Mtg, San
Steady-state simulation using SPEEDUP. Paper No. 92b, Francisco. California f 1984bj.
Presented at AIChE Spring National Mrg, Houston, Westman i. R., A. L&a anh D. C. Miller, Flash and
Texas (1983). distillation calculations by a Newton-like method. Com-
Pantelides C. C., SPEEDUP-recent advances in process put. them. Engng 8, 219-228 (1984).

You might also like