Vous êtes sur la page 1sur 20

Computers and Chemical Engineering 25 (2001) 445– 464

www.elsevier.com/locate/compchemeng

A generalized framework for computing bifurcation diagrams


using process simulation programs
Arjun Vadapalli, J.D. Seader *
Department of Chemical and Fuels Engineering, Uni6ersity of Utah, Salt Lake City, UT 84112, USA

Received 21 March 2000; received in revised form 18 December 2000; accepted 19 December 2000

Abstract

Interest in discovering multiple steady-state solutions for reaction and separation processes grew exponentially in the 1990s.
Process simulators like ASPEN PLUS, CHEMCAD, HYSYS, and PRO/II are designed to obtain at best one solution. Simulation
programs can find multiple solutions only by the expenditure of much effort. Here, a bifurcation technique using arclength
continuation is presented that can be incorporated as an add-in subroutine to a simulation program to automatically trace a
solution path, including turning points, to obtain multiple solutions with respect to a user-selected parameter. The technique is
illustrated with applications to the ASPEN PLUS process simulator. The algorithms are based on a predictor– corrector
implementation where the predictors reside in add-in FORTRAN routines and the existing nonlinear equation solvers in ASPEN
PLUS equipment models serve as the correctors. Furthermore, the existing physical property packages in ASPEN PLUS are also
utilized. The method was tested successfully on an adiabatic CSTR example using the ASPEN RCSTR model, and homogeneous
azeotropic, heterogeneous azeotropic, and reactive distillation examples using the ASPEN RADFRAC model. Two of these
examples are presented here. In all four examples, the range of the selected bifurcation parameter covers a region that produces
three multiple solutions. © 2001 Elsevier Science Ltd. All rights reserved.

Keywords: Multiple solutions; Bifurcation; Continuation

1. Introduction form, refers to the case where all input variables are
specified and two or more sets of output variables are
Design of chemical processing systems that satisfy found. For example, in a distillation column, the molar
economical and environmental requirements calls for flow rate, feed composition, feed location, number of
development of new design tools and instrumentation stages, molar reflux rate, molar distillate rate, type of
for identifying the fundamental features of the pro- condenser and column pressure profile are specified,
cesses. One of these features is the phenomenon of and two or more sets of product composition and
multiple steady states, which is the characteristic of column profiles are found. Input multiplicity refers to
numerous physiochemical processes such as chemical the case where one or more output variables are spe-
reactions (catalytic and autocatalytic), mass-transfer cified and multiple solutions are found for the unknown
operations (distillation, extraction, absorption), and re- input variables. Internal state multiplicity refers to the
active mass-transfer processes. Multiple steady states case where multiple sets of internal conditions or profi-
can take place at the levels of a single apparatus, a les are found for the same values of the input and
group of functioning apparatuses, or even an entire output variables. All three types of multiplicities have
chemical process system. been found to occur in steady-state separation opera-
Gani and Jørgensen (1994) define three types of tions that are modeled with nonlinear algebraic and
multiplicities. Output multiplicity, the most common transcendental equations. Multiplicities can also occur
in membrane and crystallization systems, in reaction
* Corresponding author. Tel.: +1-801-5816915; fax: + 1-801-
systems, and in certain cyclic separation operations,
5859291. such as adsorption, when modeled with nonlinear par-
E-mail address: j.seader@m.cc.utah.edu (J.D. Seader). tial differential equations.

0098-1354/01/$ - see front matter © 2001 Elsevier Science Ltd. All rights reserved.
PII: S0098-1354(01)00624-X
446 A. Vadapalli, J.D. Seader / Computers and Chemical Engineering 25 (2001) 445–464

Usually, multiplicity occurs only for certain sets of variables to satisfy the stationarity conditions. Unfortu-
specifications and only over restricted ranges of certain nately, they are of limited utility for the design of
parameters. Often, sets of multiple solutions contain processes that operate near or within regimes character-
both stable and unstable solutions. One or more of the ized by hysteresis and periodic or chaotic behavior.
stable solutions is often clearly superior for practical They have no bifurcation algorithms that can trace
application and one or more of the other stable solu- steady-state solution paths, perform a stability analysis
tions is trivial or very undesirable for practical applica- (evaluation of the eigenvalues of the Jacobian), and
tion. For design purposes, it is important to be aware locate the global optimum.
of the possibility of multiplicity and to discover all One method for finding multiple solutions, described
multiple solutions within the practical domain of oper- by Morgan (1987), but restricted to sets of polynomial
ating variables. Seader and Kumar (1994) discuss in equations, takes the starting system of equations or a
detail the reasons for the existence of multiple solu- partial reduction of it, and applies differential arclength
tions, methods for obtaining them, and their practical continuation using a special homotopy that establishes
implications and significance. parallel continuation paths that lead robustly to all
possible roots, real and complex. This method is widely
available in two software packages, CONSOL (Mor-
2. Motivation gan, 1987) and POLSYS1H in HOMPACK90 (Watson,
Sosonkina, Melville, Morgan, & Walker, 1997). Alter-
Chemical engineers use systems of equations to natively the interval-Newton method in conjunction
model reactors and separation equipment. Frequently, with generalized bisection, as implemented in the public
these equations, whether they be algebraic, transcen- domain software program INTBIS of Kearfott and
dental and/or differential, are nonlinear in nature. Al- Novoa (1990) can locate multiple solutions of general
though linear algebraic equations have only one root sets of nonlinear equations. This method has the advan-
(or none if the coefficient matrix is singular), nonlinear tage that the user restricts the search domain and then
algebraic and/or transcendental equations can have INTBIS finds all of the roots within that domain. A
multiple roots. Except for a small number of special third method involves the use of bifurcation analysis
cases (e.g. quadratic, cubic and quartic equations), a and singularity theory to locate higher-order singular
direct analytical solution is not possible and one must points in addition to limit points, such as real bifurca-
resort to an iterative numerical solution. The most tion points and Hopf bifurcation points, which homo-
commonly applied numerical algorithms for systems of topy algorithms fail to detect. AUTO ( Doedel, Wang,
nonlinear equations, as incorporated into steady-state & Fairgrieve, 1994) is one such package that can per-
process simulators seek, at best, one solution. Addi- form a bifurcation analysis. Doedel, Keller, & Kernevez
tional steady-state solutions can sometimes be found by (1991) present an exhaustive discussion on the numeri-
making case studies, performing a sensitivity study on cal analysis and control of bifurcation problems in
one or more parameters, or varying initial guesses of finite dimensions with several illustrative examples.
the solution. However, this procedure is not likely to Bekiaris, Meski, Radu, and Morari (1993) used bifurca-
produce a smooth continuation or bifurcation curve tion analysis to compute the solution diagram for the
that includes turning points and all multiple solutions. special case of homogeneous azeotropic distillation with
Simulator algorithms either require the user to provide an infinite number of trays and infinite reflux ratio.
the initial guesses for the set of unknowns or provide a This method has come to be known as the infinity/infi-
built-in initialization procedure. In either case, an at- nity analysis. Results obtained from this limiting case
tempt is made to provide an initial guess that is within are then applied to the actual finite column case.
the domain of a practical answer. This works in many For all the above methods, investigators have written
cases and fortunately, the solution obtained is often the specialized programs that include model equations,
only practical solution. However, when multiple solu- thermodynamic routines, and in some cases, corre-
tions exist in the practical domain, i.e. where the mole sponding equation solvers to carry out the necessary
fractions are between zero and one and flow rates and calculations. Also, simplifying assumptions have some-
absolute temperatures and pressures are positive and times been made to successfully solve specific cases.
real, it is desirable to have methods that find all Often, these assumptions have been unsuitable for
solutions. other case studies. The aim here is to utilize existing
In addition to iterative numerical methods, investiga- process models already present in simulators like AS-
tors have considered other mathematical tools that can PEN PLUS, thus saving the user the time of writing
help detect multiple solutions. Some of the steady-state and debugging separate and specialized codes to per-
process simulators (e.g. ASPEN PLUS, CHEMCAD, form a bifurcation analysis. ASPEN PLUS was chosen
HYSYS, and PRO/II) have algorithms that simulta- to illustrate the bifurcation methods developed here
neously adjust the recycle tear variables and the design because it is widely used and its architecture is well
A. Vadapalli, J.D. Seader / Computers and Chemical Engineering 25 (2001) 445–464 447

suited for adding user-supplied subroutines. Before pre- an adiabatic CSTR is presented here to provide a better
senting the details of the bifurcation methods, it is of understanding of the capability and limitation of the
interest to discuss the use of the sensitivity analysis tool sensitivity analysis tool.
of ASPEN PLUS and its shortcomings when attempt-
ing to use it for locating multiple solutions. 3.1. Adiabatic continuous stirred-tank reactor case
study

3. Behavior of the sensitivity analysis tool of ASPEN Consider the liquid-phase hydrolysis of propylene
PLUS in the region of multiplicity oxide to propylene glycol in an adiabatic CSTR of fixed
volume. The influence of the inlet volumetric flow rate
ASPEN PLUS uses sequential modular simulation, of water on propylene oxide conversion and outlet
where output streams for each operation in a process temperature is desired. This problem is solved indepen-
are calculated from given, guessed, or previously com- dently with AUTO, using relevant model equations for
puted input streams and equipment specifications for a CSTR and necessary physical properties, and is com-
the operation. The RCSTR and RADFRAC models of pared to the results obtained from a sensitivity analysis
ASPEN PLUS consist of equations that model a stirred performed with ASPEN PLUS using its built-in RC-
tank reactor and a distillation column, respectively. STR model and physical property estimation models.
Solutions of these equations depend on input stream The reaction, which can be considered irreversible, is:
and equipment specifications. Bifurcation analysis of a
Propylene oxide (A)+ water (B)
particular operation involves varying any one of the
input stream variables or equipment specifications (re- “ Propylene glycol (C) (1)
ferred to as a bifurcation parameter) over a range of
where the rate of the exothermic reaction is second-or-
values and observing its effect on output and/or state
der with respect to propylene oxide, and the rate con-
variables. For example, in the RCSTR program of
stant, k, can be expressed in Arrhenius form:
ASPEN PLUS, material and energy balance equations,
including kinetic expressions, are solved to obtain exit k= A e-E/RT = 9.15
temperature and exit component molar flow rates. For
×1022 exp (−18,700/Tr) m3/ kmol-s
a given reactor volume, a change in the input molar
(2)
flow rate of any one of the reactants affects the values
of all computed variables. where Tr, is the temperature in the reactor in K. The
Since neither bifurcation nor homotopy continuation problem specification is shown in Fig. 1. The feed
tools are present in any of the available commercial consists of a mixture of 1.3202 m3/h propylene oxide,
simulators, the only way to conduct a parametric anal- 1.3202 m3/h methanol (M) and water. Methanol is
ysis of a process is to use a sensitivity analysis tool. A added in order to increase the bubble-point temperature
sensitivity analysis of an exothermic reaction involving of the mixture. Varying the flow rate of water from 5.4
the hydrolysis of propylene oxide to propylene glycol in to 8.2 m3/h varies the total volumetric flow rate, while

Fig. 1. Simulation configuration of the adiabatic CSTR problem.


448 A. Vadapalli, J.D. Seader / Computers and Chemical Engineering 25 (2001) 445–464

keeping the flow rates of propylene oxide and methanol − (0.001379Q − 0.001019)Tr + (0.7317Q − 0.488)
constant. The entering temperature of the feed streams RNET
is 24°C (75°F) and reactor pressure is 1 atmosphere. + =0 (5)
QCA0
The volume of the reactor is 1.1356 m3 (40.1 ft3).
Eq. (3) and Eq. (5) were solved by AUTO with initial
values of Q=8.0 m3/h, CA = 0.081 kmol/m3, CB =
3.2. Comparison of ASPEN PLUS sensiti6ity analysis 34.76 kmol/m3, CC = 2.26 kmol/m3, and Tr =78°C.
with AUTO results Since methanol does not participate in the reaction
mechanism, its mass balance equation was not included
AUTO was used with the following material and with the rest of the mass balance equations. The four
energy balance equations. Material balance equations nonlinear algebraic equations (3 mass balance and 1
for components A, B and C assuming incompressible energy balance) were solved to obtain the solution
flow: vector u, which is made up of elements, CA, CB, CC and
T. The total volumetric flow rate, Q was the selected
QCi0 − QCi −RNET =0 (3) bifurcation parameter.
The bifurcation package of AUTO generates an out-
where i is either component A, B, or C and RNET = put file that contains parameter values (the volumetric
kC 2AVr. Q is the total volumetric flow rate in and out of flow rate of water) and state variables (temperature and
the reactor; Ci0 is the inlet molar concentration and Ci concentrations), along with indicators for turning
is the exit molar concentration. points and algebraic bifurcation points. The sign next
Energy balance equation for the adiabatic reactor: to the step number determines the stability of a particu-
lar solution. Stable points are indicated with the (−)
3 sign and unstable points by the (+ ) sign. In the
− XEB[DHR +DCP(Tr −TR)] =% UiC0 Pi(Tr −Ti0) (4) following example, turning points were found at enter-
1
ing water rates of 7.79 and 6.04 m3/h, with the absence
of any real bifurcation points. The bifurcation diagram
where, XEB is the fractional conversion predicted by the
from AUTO for this problem is presented in Figs. 2
energy balance equation; DHR is the heat of reaction at
and 3 as solid lines. One observes a decrease in the
reference temperature, TR; DCP is the average change in
propylene oxide conversion from 96 to 65% as the
molar heat capacity of the reacting mixture over the
water flow rate increases from a starting point of 5.4
temperature range of TR to Tr; Tr is the exit and reactor
m3/h to the first turning point at 7.79 m3/h. Thereafter,
contents temperature; Ui is the mole fraction of each
an unstable steady state is calculated and the conver-
component in the combined feed; Ti0 is the inlet tem-
sion continues to fall at decreasing flow rates until the
perature of the combined feed; and C0 Pi is the average
second turning point is reached at 6.04 m3/h with a
heat capacity of each component in the inlet stream conversion of 15%. Beyond the second turning point,
over the temperature range of Ti0 to Tr. the flow rate again increases with a marginal drop in
To obtain an energy balance as consistent as possible conversion (15 –4%) corresponding to a second stable
with ASPEN PLUS calculations, the following proce- steady-state region. Calculations are terminated arbi-
dure was employed. The energy balance equation was trarily at a water flow rate of 8.2 m3/h. Thus, two stable
formulated with the help of the RSTOIC model of steady states and one unstable steady state exist for
ASPEN PLUS. RSTOIC computes the exit tempera- total volumetric flows between the two limit points of
ture for a specified conversion of a key component. For 7.79 and 6.04 m3/h, which defines the region of multi-
a fixed set of input variables and equipment specifica- plicity. Unique steady states exist outside the region of
tions, RSTOIC was solved to find exit temperatures for multiplicity.
different fractional conversions. Because conversion A sensitivity analysis run using ASPEN PLUS with
was found to be close to a linear function of the exit the RCSTR model was carried out for each of the two
temperature, the energy balance equation was fitted to options provided in ASPEN PLUS. In the first option,
the equation form XEB =mT + b. Energy balance lines current results are used as initial estimates for the next
of this form were constructed for different total volu- step, which involved a new value of the volumetric
metric flow rates by changing the flow rates of water water flow rate. In the second option, results are reini-
alone while keeping the flow rates of propylene glycol tialized at each step, with ASPEN PLUS supplying a
and methanol constant. The slope and intercept of the new set of initial estimates to begin each step. In Fig. 2,
energy balance plot varies as the total volumetric flow where no reinitialization was done at each step, the
rate changes because of a corresponding change in Ui. initial water rate is 5.4 m3/h and is incremented by 0.1
Thus, slope (m) and intercept (b) are expressed as linear m3/h for each sensitivity step. Open square markers
functions of the total volumetric flow rate (Q). Eq. (4) show the ASPEN PLUS results. Conversion decreases
was reduced to: gradually until the first turning point is approached at
A. Vadapalli, J.D. Seader / Computers and Chemical Engineering 25 (2001) 445–464 449

Fig. 2. Comparison of ASPEN PLUS (with no reinitialization) and AUTO results.

Fig. 3. Comparison of ASPEN PLUS (with reinitialization at each step) and AUTO results.

a water flow rate of 7.8 m3/h. The next sensitivity step, ASPEN PLUS sensitivity analysis finds only one solu-
at a water rate of 7.9 m3/h, witnesses a precipitous tion in the region of multiplicity and is not a reliable
decrease in conversion to 4%. A second series of sensi- tool to accurately predict turning points. In Fig. 3, the
tivity steps starts from a water rate of 8.2 m3/h, and results obtained from the second sensitivity option are
decreases by increments of 0.1 m3/h. The ASPEN compared to the AUTO results. It can be seen that
PLUS results are shown as open triangle markers in ASPEN PLUS calculates the unstable solution sporadi-
Fig. 2. Note that the conversion jumps from 4.0 to 90% cally at a few water flow rates. Therefore, with neither
at a water flow rate of 6.6 m3/h, instead of 6.04 m3/h as sensitivity option is it possible to obtain a smooth
predicted by AUTO. All ASPEN PLUS results are bifurcation curve with all the steady-state solution
close to the AUTO results. Fig. 2 shows that the branches in the region of multiplicity.
450 A. Vadapalli, J.D. Seader / Computers and Chemical Engineering 25 (2001) 445–464

4. Development of arclength continuation algorithms for using the pseudo-arclength continuation technique of
tracking multiple solutions with ASPEN PLUS Keller (1977). Assuming (uj-1, lj-1) and (u%j-1, l%j-1) have
been computed, this well-known technique determines
The traditional way of finding all possible solutions the next solution (uj, lj ) from the equations,
to a set of nonlinear equations has been to use a
homotopy or bifurcation program in which the user f(uj, lj )= 0, (7a)
supplies the model equations. In addition, property (uj -uj-1)Tu%j-1 + (lj -lj-1) l%j-1-Ds = 0 (7b)
estimation equations are necessary. The technique pre-
sented here uses the process model and the property where Ds is the step size along the branch and super-
equations already present in the chemical process simu- script T is the transpose. The advantage of using the
lator, e.g. ASPEN PLUS, to achieve the goal of track- inflated system of Eq. (7) is its capability of computing


ing all possible solutions in a region of multiplicity. past limit points on a branch, because the Jacobian for
Solution branches are computed using an arclength
continuation technique to increment the selected bifur- Eq. (7),
fu f&0xFFFD; n
, is not only nonsingular at a
u’ &0xFFFD;’
cation parameter rather than using sensitivity analysis,
which as shown in Figs. 2 and 3, is an unreliable tool regular solution point of Eq. (6) where fu is nonsingu-
for tracking solutions along a particular branch. Both lar, but also at a simple turning point. Points where the
natural arclength and pseudo – arclength continuation variable l% changes sign are located by a secant itera-
were explored. Both incorporate a predictor-corrector tion scheme and are treated as potential turning points.
method for tracking the observed variables with respect Once a turning point is detected, the solution branch is
to a bifurcation parameter. It was particularly impor- continued in the new direction using pseudo-arclength
tant to provide close approximations to the output continuation. In order to implement this scheme into
variables so that the nonlinear equation solver of AS- ASPEN PLUS, it would be necessary to access the
PEN PLUS could converge to the nearest solution. The derivative information fu and fl, where fu is the Jaco-
user supplies to ASPEN PLUS the predictor method as bian matrix of all functions with respect to state vari-
an add-in, while Newton’s algorithm, included in the ables u, and fl are the function derivatives with respect
ASPEN PLUS simulator, is used for the corrector to parameter l. Alternatively, the derivatives can be
method. The user-supplied predictor algorithm was calculated using a finite differencing scheme by directly
merged with ASPEN PLUS via the inline FORTRAN accessing the functions. Unfortunately, neither was pos-
option available in ASPEN PLUS. sible with ASPEN PLUS and, therefore, an alternative
The convergence behavior of the Newton solver in method was developed. The detection of a turning
the RCSTR model of ASPEN PLUS motivated the point is treated the same for both algorithms as ex-
selection of a natural arclength method over the plained next in some detail. Branch switching is han-
pseudo-arclength method. However, as discussed be- dled differently for the RCSTR and RADFRAC
low, the latter method was used with the RADFRAC models because the latter deals with equilibrium-stage
model. The natural arclength method is used with the operations, whose temperature and composition profi-
RCSTR method because it solves the mass balance les can be suitably exploited to enable the algorithm to
equations first, with the converged results supplied as successfully negotiate around turning points.
initial estimates for solving the energy balance equa- The methods used for detecting turning points and
tion. Separate convergence parameters can be provided for performing branch switching are explained with the
for the mass balance and energy balance equations. The help of Fig. 4, which is a schematic of a typical
NEWTON solver in RADFRAC for the classical bifurcation curve with more than one steady state,
Naphtali –Sandholm option solves the column-describ- where, u i is a component of the vector of state variables
ing equations simultaneously with Newton’s method. u, and l is the bifurcation parameter. The variable u i
Convergence is stabilized using the dogleg strategy of represents a state variable like temperature or compo-
Powell. nent mole fraction, and l represents an input or equip-
In AUTO it is assumed that a solution u0 =u(s0) is ment variable that is being varied over a range of
known for some particular value of l0 =l(s0). With values. The S-shaped curve consists of three steady-
AUTO the first step in bifurcation analysis consists of state solution branches that change directions at the
determining the stationary solution branches. These are two turning points, and hence between the turning
solutions [u(s), l(s)] of the nonlinear system of points there are three sets of state variables that satisfy
equations, the process model. The first is between points S and A,
the second between A and D, and the third between D
f(u, l)=0, u, f  Rn, l  R (6)
and B. Solutions on branches SA and DA cease to exist
where s denotes some parameterization. From the start- to the right of point A. Similarly, solutions on branches
ing point, the branch is traced out in a stepwise manner AD and BD cease to exist to left of point D. For
A. Vadapalli, J.D. Seader / Computers and Chemical Engineering 25 (2001) 445–464 451

parameter values less than at point E, there exists only The criteria for detecting turning points can thus be
one unique solution, as also at points beyond B on summarized as:
branch DB. The branches SA, AD, and DB can be
lim tan − 1(du/du)“90deg (8)
numbered as I, II, and III, respectively (or the reverse Du “ 0
notation can be used and the branches DB, AD, and
4.2. The natural arclength method in the absence of a
SA can be numbered as I, II, and III respectively).
Jacobian matrix
4.1. Detecting a turning point in the absence of a
Jacobian matrix for both algorithms The continuation method makes use of the two most
recent converged solutions. The difference in the two
Turning points are characterized by large changes in state variable vectors, uprevious and ucurrent, is divided by
state variables like temperature or component mole the corresponding difference in the parameter values,
fractions for a small change in the value of the bifurca- lprevious and lcurrent, to obtain the direction vector with
tion parameter. The relative change is measured as the respect to the two solution sets. The following steps
angle tan-1(Du i/Dl). For any two points on a solution summarize the method used to compute any one solu-
branch, the absolute value of the angle formed with the tion branch in the bifurcation diagram,
X-axis will be anywhere between 0 and 90 degrees.
Unless the branch is steep with respect to the X-axis, 1. Step 1: Compute the direction vectors for all compo-
this angle is normally well below 90 degrees. At a limit nents of vector u from
point, since the change in variable u i is large, the angle
is close to 90 degrees. This sudden and large change is du i/dl = (u icurrent-u iprevious)/ (lcurrent-lprevious) (9)
due to the fact that a point like B% in Fig. 4 does not
actually exist and the Newton solver converges to solu- This step is also required for detecting turning points as
tion B, which lies on another solution branch, DB. The discussed in the previous section.
change in variable u i defined by the difference AB is
greater than any possible change when on branch SA 2. Step 2: Compute the guess values u *inext for the next
for the same Dl. In the case of an adiabatic CSTR, this solution along the branch from
change is observed for all exit concentrations and the
reactor temperature. For a distillation column, large
u*inext = u icurrent + du i/dl (lnext-lcurrent) (10)
changes in distillate and bottoms temperature are ob-
served. By thermodynamics, temperature and compo- where the subscripts pre6ious and current correspond to
nent mole fractions in the vapor and liquid are related the two most recent converged solutions, and subscript
to each other, hence causing large changes in distillate next refers to the next converged solution along the
and bottoms product compositions. branch. The guess values u *inext are supplied to the

Fig. 4. Typical bifurcation curve for a chemical process.


452 A. Vadapalli, J.D. Seader / Computers and Chemical Engineering 25 (2001) 445–464

nonlinear-equation solver and convergence is achieved Switching from an unstable branch to a stable one is
in a few steps. It is important to note that since an treated no differently than for the case of switching
adaptive step size strategy is implemented to achieve from a stable to an unstable branch. After performing
the above steps, (lnext-lcurrent) is not equal to (lcurrent- a number of runs, it was found that providing a tem-
lprevious). perature estimate lower in value than at point D en-
The RCSTR program allows the user to supply ables the solver to converge to a point on branch DB.
initial estimates for flow rate and temperature variables, Similarly, providing a temperature estimate higher than
which are calculated in step 2 above. These estimates the temperature at point A causes the solver to con-
are called F-EST and T-EST respectively. In the region verge to a solution on branch SA. The underlying
of multiplicity, where more than one steady-state solu- assumption is that any variable u i will increase or
tion exists, the following convergence characteristics are decrease monotonically as one begins calculations from
observed, branch SA or DB. This means that any solution on
1. Providing estimates of component flow rates and branch SA will always be higher in value than any
reactor temperature does not guarantee convergence solution on branch AD or DB.
to a specific steady-state solution. For certain flow
rates, the steady-state solution obtained is not the 4.3. The pseudo-arclength method in the absence of a
closest solution to a given set of initial estimates. Jacobian matrix
2. The temperature estimate alone is found to be suffi-
cient to ensure convergence to a specific steady-state The following method is explained with the help of a
solution. This behavior might be attributed to the homogeneous azeotropic distillation column studied by
fact that in the RCSTR model the mass balance Güttinger, Dorn, and Morari (1997) to separate the
equations are first solved in an inner loop, while the ternary mixture methanol –methyl butyrate –toluene,
energy balance equations are solved last in the outer which demonstrated the existence of multiple steady
loop. The energy balance equations alone appear to
states caused by nonlinearities in vapor –liquid equi-
control the convergence of the complete system of
librium. Experiments on an industrial pilot-plant-size
equations. Therefore, by only supplying different
column revealed two stable steady states for the same
initial temperature estimates for the reactor, it is
feed flow rate and composition, and the same set of
possible to obtain different steady-state solutions in
operating parameters. Experimental measurements were
the region of multiplicity.
found to be in reasonable agreement with predictions
In the natural arclength algorithm developed for the
obtained from the solution of a model for infinite-stage
RCSTR program, a guess value for the new point C in
columns operating at infinite reflux ratios, referred to as
Fig. 4 is provided by assuming the existence of an
intermediate solution branch lying somewhere between the infinity/infinity analysis. The experiments confirmed
the two solution branches, SA and DB. The intermedi- the existence of output multiplicity, thus supporting
ate solution is found by looking in a direction opposite earlier predictions made with models by various re-
to the direction traversed on branch SA. The main idea searchers. Rigorous equilibrium-stage calculations, per-
is to provide a guess value that lies in the region of formed using the RADFRAC model of ASPEN PLUS
convergence of point C, which is represented as an with the Wilson K-value option, confirm the existence
elliptical region encircling point C in Fig. 4. This is of the two stable steady-states.
accomplished by searching at an angle less than 90 The simulation configuration for the distillation
degrees in the opposite direction. This is implemented problem is shown in Fig. 5. The column, with 40
by providing a guess value of the temperature halfway equilibrium stages plus a total condenser and partial
between points A and B for a bifurcation parameter reboiler, operates at a reflux ratio of R= L/D: 2.0 to
value less than that at point A. The angle formed by the 2.5. Feed is introduced into the column at stage 22,
search direction in this way makes an angle less than 90 numbered from the condenser, at a constant flow rate
degrees with the X-axis. The overall goal is to supply a of 2.0 kg/h. The feed composition given in Fig. 5 is
temperature estimate that lies in the region of conver- similar to that used by Güttinger, Dorn, and Morari
gence of the NEWTON solver of ASPEN PLUS. This (1997). The multiplicity range was found to cover a
heuristic criterion was suggested after having performed distillate flow rate of 1.4 –1.9 kg/h. The sensitivity
a number of tests on the algorithm. A potential prob- analysis tool was used initially to obtain the two stable
lem with this approach is that the Newton solver in steady-state branches. Starting at a distillate flow rate
ASPEN PLUS might possibly converge to a point on of 1.4 kg/h and performing calculations with the AS-
either branch DB or SA. Hence, it might require more PEN PLUS NONIDEAL algorithm (to improve con-
than one attempt to successfully converge to point C. vergence), the distillate flow rate was incremented in
Bifurcation diagrams for other variables like molar flow intervals of 0.1 kg/h until a flow rate of 1.9 kg/h was
rates or concentrations are expected to possess similar reached. Results obtained from this run were stored
characteristics. and the distillate flow rate was then decreased in incre-
A. Vadapalli, J.D. Seader / Computers and Chemical Engineering 25 (2001) 445–464 453

Fig. 5. Simulation configuration of the homogeneous azeotropic distillation column.

ments of 0.1 kg/h from 1.9 to 1.4 kg/h. Significant 1. All three steady-state profiles are distinct. The max-
changes in distillate and bottoms compositions were imum differences in temperature and composition
observed at distillate flow rates of 1.53 and 1.81 kg/h, for profiles III and I is found for the last few stages
which corresponded to turning points like A and D in at the bottom.
Fig. 4. Solving the problem separately with AUTO, 2. The temperature and composition values for the
Güttinger, Dorn, and Morari (1997) confirmed these first few stages from the top are nearly identical for
points as turning points and also calculated the un- all three profiles.
stable distillate and bottoms compositions. They, how- 3. The unstable profile II is characterized by tempera-
ever, did not confirm the unstable branch with ASPEN tures that are close to the stable profile III for stages
PLUS. Here, by using the NEWTON algorithm instead above the feed stage. Stage temperatures for profile
of the NONIDEAL algorithm of RADFRAC, and by III are lower than for the other two profiles for the
providing initial estimates for temperature and for dis- entire column. For stages below the feed stage [22],
tillate and bottom compositions from the bifurcation stage temperatures of profile II increase suddenly
results of Güttinger, Dorn, and Morari (1997), the first and approach those of profile I. Since the tempera-
unstable solution was obtained at a distillate flow rate ture transition seemed to occur close to the feed
of 1.70 kg/h. More points along the unstable branch stage, a study of the effect of feed stage location on
were obtained by using results of the previous run as the unstable state profile was carried out for feed-
the initial estimate for a subsequent run. stage (f.s.) locations of 6, 11, 15, 22 and 28. Results
The temperature and methanol composition profiles obtained from this study are presented in Fig. 8,
of all three steady-states at a distillate flow rate of 1.58 which shows a strong dependency of the unstable
kg/h are shown in Figs. 6 and 7 respectively, where they profile on the feed stage location. This characteristic
are numbered I, II and III. Using the results presented behavior is exploited and used in the pseudo-
by Güttinger, Dorn, and Morari (1997), profiles I and arclength algorithm to enable branch switching. A
III are stable and profile II is unstable. In Fig. 6, the similar behavior was also found for heterogeneous
stage temperatures along profile III of the column vary azeotropic and reactive distillation examples.
by only 1°C along the entire length of the column. The continuation method presented here also makes
Profiles II and I vary from 65°C for the distillate to use of the two most recent converged solutions. The
110°C for the bottoms, with profile II changing difference in the two state variable vectors uprevious and
abruptly at the feed stage [22]. Other observations are: ucurrent and in the corresponding parameter values is
454 A. Vadapalli, J.D. Seader / Computers and Chemical Engineering 25 (2001) 445–464

used to obtain the pseudo-arclength with respect to the 4. Step 2: Compute the direction vectors with respect to
two solution sets. The following steps compute any one s and l,
single solution branch, where ucurrent −uprevious 2 is the
L2-norm of ucurrent − uprevious.
du i/ds = (u icurrent-u iprevious)/Ds (12)
3. Step 1: Compute the pseudo-arclength with vector u
and parameter l,
dl/ds = (lcurrent-lprevious)/Ds (13)

Ds 2 = ucurrent − uprevious 2 +(ucurrent −uprevious)2 (11) du i/dl = (u icurrent-u iprevious)/(lcurrent-lprevious) (14)

Fig. 6. Temperature profiles of all three steady-states of the homogeneous azeotropic distillation column at a distillate flow rate of 1.58 kg/h.

Fig. 7. Methanol composition profiles of all three steady-states of the homogeneous azeotropic distillation column at a distillate flow rate of 1.58
kg/h.
A. Vadapalli, J.D. Seader / Computers and Chemical Engineering 25 (2001) 445–464 455

Fig. 8. The unstable temperature profile for different feed stage locations of the homogeneous azeotropic distillation column.

where, du i/ds and dl/ds are required in step 3 for nearest steady-state solution for a given set of initial
calculating guess values for the next possible solution estimates. The odd convergence behavior of the
and du i/dl is required for detecting turning points. inside-out algorithm with the default option is possi-
bly due to the fact that the model equations are
5. Step 3: Compute the bifurcation parameter and the solved in two nested iteration loops. The NEWTON
guess values ui* for the next solution along the branch, option, on the other hand, solves the column equa-
tions simultaneously, using Newton’s method. The
u*inext =u icurrent +(du i/ds) Dsnext (15) convergence of the NEWTON option is stabilized
using the dogleg strategy of Powell (1970).
lnext =lcurrent +(dl/ds) Dsnext (16) All three branches can be calculated using the NEW-
TON option. When a limit point is exceeded, a new
It is important to note that since an adaptive step-
solution on a different branch is obtained using the
size strategy is implemented to achieve this, Ds is not
equal to Dsnext, which is managed independently in the default option of the inside-out model. The following
algorithm with a maximum value supplied by the user procedure assumes that the user will start most often on
at the start of the run. a stable steady-state branch. This is a fairly good
The RADFRAC model in ASPEN PLUS allows the assumption since the first solution point is obtained
user to supply estimates of temperature and component using the default option of the inside-out algorithm,
mole fractions for each stage as computed in step 3 which rarely converges to an unstable steady-state in
above. The program variables for temperature esti- the region of multiplicity.
mates are called T-EST and those for liquid and vapor There are two possible cases of branch switching. In
mole fraction estimates are X-EST and Y-EST, respec- the first, referring to Fig. 4, a transition is made from a
tively. In the region of multiplicity, where more than stable branch like either SA or DB to the unstable
one steady-state exists, it is possible to converge to a branch AD. The characteristic feature of this case is
specific steady-state solution by specifying both temper- that the old and new solutions lie on different stable
ature and mole fraction estimates for some or all stages. steady-state branches and the unstable branch lies
However, for a given set of initial estimates that are somewhere in between these stable branches. If the
close an actual solution: pseudo-arclength algorithm first calculates branch SA
1. The inside-out method with the default option may in Fig. 4 and encounters a turning point, the default
converge to a solution different from the one origi- option of the inside-out model is used to obtain point
nally sought, usually on the stable steady-state B. Since point C lies on the unstable branch, a very
branch. good initial estimate for point C is obtained by
2. The NEWTON option of the inside-out model is combining column profiles at points A and B. Stages
more predictable and converges usually to the above the feed stage are assigned values from the lower
456 A. Vadapalli, J.D. Seader / Computers and Chemical Engineering 25 (2001) 445–464

steady-state profile and those below the feed stage are to a higher steady-state profile. A brief discussion of the
assigned values from the higher steady-state profile. implementation of the continuation algorithms into the
This procedure was developed from the results obtained UNIX version of ASPEN PLUS, with reference to Fig.
from the homogeneous azeotropic distillation example. 9, is given in the appendix.
With these initial estimates, the NEWTON option con-
verges easily to point C.
When switching from an unstable branch to a stable 4.4. Merits of the continuation algorithm
one, a different approach is applied. For example in
Fig. 4, if the unstable branch is being calculated, start-
ing from point A, then the default option of the inside- 1. The algorithms presented here make complete use of
out model will converge to point E after passing point a simulation program, thus saving the user the need
D. Similarly, if the unstable branch is calculated start- of supplying model equations and thermodynamic
ing from point D, then point B will be calculated after routines.
passing point A. Since branches I and II are already 2. The algorithms calculate the entire bifurcation dia-
calculated, the new branch is searched for in a direction gram systematically, moving from one solution
opposite to that traversed on branch II. For the branch to another. The existence of a new branch
pseudo-arclength algorithm, a shift in the column profi- can be confirmed by looking at the plot after calcu-
les is sufficient to make the NEWTON option converge lating the first few points. If the algorithm retraces
to a point on the new branch. If the default option an old solution curve, then the user can access the
converges to a solution that possesses a lower steady- stored information to restart calculations for the
state profile, then the solution branch being sought is new branch.
the one with the higher steady-state profile. In such a 3. At any one time, the algorithms require only the
situation, the temperatures are shifted up, so that stages previously computed solution and the new solution.
above the feed stage have higher temperatures. A simi- Thus the program can be interrupted at any stage
lar method is followed if the default option converges and restarted from where it was last stopped.

Fig. 9. Implementation of the algorithms with ASPEN PLUS.


A. Vadapalli, J.D. Seader / Computers and Chemical Engineering 25 (2001) 445–464 457

4. The algorithms calculate the bifurcation diagram 5.1. Adiabatic CSTR


using an adaptive step-size algorithm. This allows
the user to start with a small step size and increase The natural arclength algorithm was used to calcu-
the step size with each step until a user-supplied late the bifurcation diagram for the adiabatic CSTR
maximum step size is reached. problem described earlier. A maximum step length for
5. Since ASPEN PLUS cannot compute values outside the bifurcation parameter (feed water flow rate) was
the practical domain, a converged solution is guar- specified as 0.15 m3/h. A value larger than this is likely
anteed to exist in the practical domain. to cause the algorithm to switch solution branches
before completing calculations on the original branch.
The minimum parameter step length was chosen based
4.5. Limitations of the algorithms on the degree of accuracy desired by the user to locate
a particular turning point. A value of 89.9 degrees for
the critical angle was sufficient to locate both turning
1. The algorithms cannot indicate immediately points. All exit component compositions and the exit
whether the solution is stable or unstable. This temperature satisfied this value of critical angle simulta-
shortcoming can be overcome to some extent by neously. Calculations on a branch were started with a
noting that the branches change stability at every value of the step length between the minimum and the
turning point. Therefore, a bifurcation diagram of a maximum, and the step length was increased by 20% at
particular process consisting of three branches can each step until the maximum value was reached. The
have only two possible stability characteristics. The kinetic rate law was available to the main program
three solution branches are either stable – unstable – either by linking a user-supplied FORTRAN subrou-
stable or unstable – stable – unstable. Most investiga- tine or by directly providing these values in the input
tors have reported the former case. data for the RCSTR model.
2. The algorithms cannot detect exact bifurcation The computed bifurcation diagram of reactor tem-
points because the Jacobian generated by ASPEN perature as a function of the volumetric flow rate of
PLUS is not accessible to the user. Therefore, the feed water as the selected bifurcation parameter for this
algorithms calculate only one solution path at any problem is given in Fig. 10, where each point represents
one time. Homotopy continuation algorithms also a converged calculation by ASPEN PLUS. The al-
share this shortcoming. gorithm is implemented in a continuous manner,
whereby the user does not interact with the execution of
the bifurcation program after supplying the required
inputs for the program. AUTO ascertained that
5. Results and discussion branches III and I are stable and II is unstable. As seen
in Fig. 10, turning points occur at 7.84 and 5.71 m3/h.
The algorithms discussed above were used to obtain The temperature falls rapidly from 72.7 to 50.8°C as the
bifurcation diagrams for the two cases discussed above volumetric flow rate of water increases along branch I
(adiabatic CSTR and homogeneous azeotropic distilla- from the starting point of 5.87 m3/h until the first
tion). Vadapalli (1998) gives case studies for heteroge- turning point is encountered at 7.84 m3/h. Beyond that
neous azeotropic distillation and reactive distillation point, the unstable steady state branch II is calculated
examples. The former deals with the dehydration of and the temperature continues to fall, as the flow rate
ethanol using benzene as the entrainer and is performed decreases, until the second turning point is reached at
using the rigorous three-phase distillation algorithm 5.71 m3/h. This turning point is less than the value 6.04
option available in the RADFRAC model of ASPEN m3/h calculated with AUTO. Beyond the second turn-
PLUS. The reactive distillation problem deals with the ing point, the flow rate begins to increase with a
effect of the methanol feed-stage location and the reflux marginal drop in temperature (29.6 to 24°C), and the
ratio on the conversion of methanol and isobutene to second stable steady state, branch III, is calculated.
methyl-tert-butyl-ether (MTBE) and its separation from Corresponding bifurcation diagrams of exit concentra-
inert n-butene (1-butene). The two case studies pre- tions for propylene oxide, water, and propylene glycol
sented here give converged results obtained from AS- against the same bifurcation parameter are shown in
PEN PLUS, and deal with processes that exhibit three Figs. 11–13 respectively. Figs. 10 and 13 look similar
steady states in the region of multiplicity. The branches because the production of propylene glycol depends on
are numbered I, II and III in the order they are the rate of the reaction, which in turn depends on the
calculated. The stability of the branches for some of the temperature of the reactor. Concentration is highest for
cases were established by comparing results obtained volumetric flow rates less than 5.7 m3/h where the
from independent calculations with the bifurcation temperature is higher than 70°C, and is lowest when the
package AUTO. temperature is close to the feed temperature. Corre-
458 A. Vadapalli, J.D. Seader / Computers and Chemical Engineering 25 (2001) 445–464

Fig. 10. Bifurcation diagram of exit reactor temperature for the adiabatic CSTR.

Fig. 11. Bifurcation diagram of the exit propylene oxide concentration for the adiabatic CSTR.

spondingly, the concentrations of both water and pro- less than 5.7 m3/h because at least a 95% conversion is
pylene oxide are the lowest when temperature is the guaranteed in that region. The reactor should not be
highest. All the diagrams presented in this section were operated at water flow rates greater than 7.9 m3/h
constructed from data points obtained from a single because the conversion of propylene oxide is never
run. In all there were 34 ASPEN PLUS simulations higher than 5%. In the region of multiplicity, the oper-
carried out to solve this problem. The conversion plot ating point depends on how the reactor is started up.
similar to Figs. 2 and 3 is shown in Fig. 14. In the The authenticity of the calculated second turning point
region of multiplicity, three operating points are possi- was tested by running independent simulations of AS-
ble. The maximum amount of propylene glycol is pro- PEN PLUS at flow rates less than 5.71 m3/h. ASPEN
duced if the reactor is operated on branch I. It is best PLUS converged to the higher stable steady-state
to operate the reactor with an inlet water flow rate of branch for these tests. A similar procedure was carried
A. Vadapalli, J.D. Seader / Computers and Chemical Engineering 25 (2001) 445–464 459

out for the turning point located at 7.84 m3/h. In this and the maximum arclength, DSMAX, were set to 1.0
case, flow rates greater than 7.84 m3/h were chosen. All and 2.0, respectively. This arbitrary selection of arc-
runs converged to the lower stable steady-state branch. lengths is similar to that used in AUTO. Because the
Hence, for the present problem, the algorithm success- algorithm requires only the two most recently calcu-
fully calculated the complete bifurcation diagram. lated solutions at any one time, it is possible to stop the
algorithm at any time during the computation of the
5.2. Homogeneous azeotropic distillation solution branch, make the necessary changes to the
program parameters, and then restart the calculations
The pseudo-arclength algorithm was used to solve from the last converged solution. Thus, the user does
the homogeneous azeotropic distillation example dis- not have to begin calculations from the first point
cussed earlier. The choice of standard arclength, DS, again. The 10 representative stages were chosen after

Fig. 12. Bifurcation diagram of the exit water concentration for the adiabatic CSTR.

Fig. 13. Bifurcation diagram of the exit propylene glycol concentration for the adiabatic CSTR.
460 A. Vadapalli, J.D. Seader / Computers and Chemical Engineering 25 (2001) 445–464

Fig. 14. Bifurcation diagram of the exit propylene oxide conversion for the adiabatic CSTR.

carefully studying the temperature and composition The bifurcation diagrams of the bottoms and top
profiles presented in Figs. 6 and 7. Only a few impor- compositions compare well with those obtained from
tant stages are selected instead of specifying initial Güttinger, Dorn, and Morari (1997). While they pre-
estimates of temperature and mole fraction for all sented results on a mass basis, results here are on a
stages. The method is particularly useful for large molar basis. Except for the bottoms composition of
columns where specifying estimates for all stages can be methyl-butyrate, all other bifurcation diagrams calcu-
tedious to keep track of for each ASPEN PLUS run. lated using the algorithm look similar to those pre-
The solution of a distillation column model lies in sented by Güttinger, Dorn, and Morari (1997). For this
2MN + 5N +4M+ 5 space, where M is the number of particular bifurcation diagram, shown in Fig. 15 in
components and N the number of stages. Thus, for terms of mole fractions, branches I and II intersect at a
large columns, many variables must be specified. By distillate flow rate of 1.65 kg/h. Hence, the two solution
using a number much smaller than N, the number of branches meet at two points. This however, is not
required initial estimates is reduced significantly. The observed in Fig. 16, which is in terms of mass fractions
scheme is effective because in most columns one stage and where the two branches meet only at the turning
in a section of stages easily represents the change in the point of 1.81 kg/h. Multiplicities are known to exist
entire section. Since the feed-stage location has a direct when some flows are specified on a mass basis (instead
influence on the shape of the unstable profile and in of a molar basis) and are due to the nonlinear mass to
molar flow transformation, a type of multiplicity first
turn enables branch switching, five stages above the
discussed by Jacobsen and Skogestad (1991). In this
feed stage and five stages below the feed stage were
particular problem, flow rates are expressed on a mass
chosen from the 42 stages. This led to the selection of
basis and compositions on a molar basis. However, this
stages 1, 5, 10, 15 and 22 above the feed stage and
intersection is not observed for the other composition
stages 25, 30, 35, 40 and 42 below the feed stage. This
variables. This kind of multiplicity was not reported by
selection represented the column temperature and com-
Güttinger, Dorn, and Morari (1997).
position profiles well because the feed stage was located
Infinity/infinity and AUTO predictions made by Güt-
almost midway between the first and last stages. The tinger, Dorn, and Morari (1997) predict an overlap of
selected bifurcation parameter in this particular prob- branches II and III for the distillate composition of
lem was the distillate flow rate. A critical angle of 89.8 methanol shown in Fig. 17. Although the branches are
degrees was sufficient to locate both turning points. A close to one another, our calculations show them as
more stringent critical angle value could have been distinct. This is also found to be true when plotted on
used, but only at the cost of a smaller specified arc- a mass fraction basis.
length and, hence, additional calculations along a par- The bifurcation diagram for the bottoms temperature
ticular branch. is shown in Fig. 18. This diagram exhibits a very
A. Vadapalli, J.D. Seader / Computers and Chemical Engineering 25 (2001) 445–464 461

interesting behavior along branch I as the turning point another. The large temperature increase can be ex-
of 1.81 kg/h is approached. When starting from a plained by the material balance relationship between
distillate flow rate of 1.51 kg/h. The temperature rises internal vapor, reflux, and distillate flow rates. The
only slowly from 64 to 69°C as the flow rate increases reflux flow rate is kept constant for the entire parameter
to 1.78 kg/h. Thereafter, the temperature rises steeply range and, by material balance, the vapor flow rate is
from 69 to 103.7°C for a change of only 0.03 kg/h in equal to the sum of the distillate and reflux flow rates.
distillate flow rate. Because of this large change in Hence, the increased distillate demand can only be met
temperature for a relatively small parameter change, by an increase in internal vapor flow rate. This is
the algorithm calculates smaller pseudo-arclengths and, possible only by supplying more steam to the reboiler,
hence, the data points in this region are close to one which in turn causes the bottoms temperature to rise.

Fig. 15. Bifurcation diagram of the methyl-butyrate bottoms composition for the homogeneous azeotropic distillation column.

Fig. 16. Bifurcation diagram of the methyl-butyrate bottoms composition for the homogeneous azeotropic distillation column (in terms of mass
fractions).
462 A. Vadapalli, J.D. Seader / Computers and Chemical Engineering 25 (2001) 445–464

Fig. 17. Bifurcation diagram of the methanol distillate composition for the homogeneous azeotropic distillation column.

Fig. 18. Bifurcation diagram of the bottoms temperature for the homogeneous azeotropic distillation column.

The bifurcation diagram for the reboiler duty, shown in of branch I. Because this region is very close to branch
Fig. 19, confirms this along branch I as the turning II, which is unstable, there is a likelihood that the
point is approached. column might oscillate between the two steady states. If
A stability analysis was carried out by Güttinger, so, it is advisable not to operate between distillate flow
Dorn, and Morari (1997), who found two stable steady- rates of 1.77 and 1.81 kg/h.
state branches with an unstable branch between the As can be seen from all the bifurcation diagrams, the
two. Comparing the AUTO calculations performed by rapid change observed in the bottoms temperature is
Güttinger, Dorn, and Morari (1997) with those ob- not observed in corresponding bifurcation diagrams of
tained from ASPEN PLUS, it is ascertained that distillate and bottoms compositions. If the bottoms
branches I and III are stable and branch II is unstable. temperature were to be used as the sole criteria for
Column control might be difficult in the upper region detecting turning points, the algorithm would be likely
A. Vadapalli, J.D. Seader / Computers and Chemical Engineering 25 (2001) 445–464 463

to detect a turning point well before the true turning branches. The natural arclength algorithm developed
point. It is for this reason, as explained above, that here is capable of starting at any position on the
more than one criterion is necessary to accurately de- bifurcation diagram.
tect turning points. Therefore, in addition to the bot- Calculations for the homogeneous azeotropic distilla-
toms temperature, distillate and bottoms mole fractions tion problem were initiated at a stable branch with the
of toluene and methyl-butyrate were used to evaluate pseudo-arclength algorithm. This is reasonable because
possible turning points. As in the case of the adiabatic the user is most likely to use an inside-out algorithm,
CSTR, the algorithm succeeded in calculating all three such as RADFRAC, to find the first solution of the
branches predicted by AUTO and did better than the problem and that algorithm almost always converges to
sensitivity analysis run, where only the two stable a stable steady-state solution. The pseudo-arclength
branches could be calculated. It is also worth noting algorithm was successful in detecting turning points for
that branches I, II, and III do not appear in the same the homogeneous azeotropic distillation example at dis-
order in each of the bifurcation diagrams presented in tillate flow rates of 1.53 and 1.81 kg/h. The turning
this section. This further strengthens the need for such point at 1.81 kg/h was located with an accuracy of
an analysis of a process operating in the region of 0.0005 kg/h, while the turning point at 1.53 kg/h was
multiplicity. found with an accuracy of 0.007 kg/h.

6. Conclusions Appendix A. Implementation of continuation algorithms


in ASPEN PLUS
Two continuation algorithms have been presented
that can be added into ASPEN PLUS to calculate FORTRAN programs based on algorithm pseu-
steady-state branches that exist in regions of multiplic- docodes were linked with the different modules of
ity. For the adiabatic CSTR problem, the natural arc- ASPEN PLUS via in-line FORTRAN, which can be
length algorithm is successful in detecting both turning used to write FORTRAN statements and expressions in
points and all three steady-state branches, as predicted design specifications, FORTRAN blocks, optimization
by the bifurcation package AUTO. The turning points problems, sensitivity blocks and scaling reports. In-line
were located to a degree of accuracy of 0.02 m3/h and FORTRAN statements invoke most FORTRAN fea-
can be improved upon by choosing a smaller value for tures, such as subroutine calls, I/O statements, array
the minimum step length. Unlike the sensitivity analysis variables, and labeled COMMONs. ASPEN PLUS
results presented in Figs. 2 and 3, the continuation checks the FORTRAN code interactively as it is en-
algorithm does not ignore the presence of the interme- tered. Most syntax errors are detected before actually
diate unstable branch lying between the two stable running the simulation. ASPEN PLUS interprets most

Fig. 19. Bifurcation diagram of the reboiler duty for the homogeneous azeotropic distillation column.
464 A. Vadapalli, J.D. Seader / Computers and Chemical Engineering 25 (2001) 445–464

in-line FORTRAN statements. FORTRAN that can- through the flowsheet sequence. The add-in subroutines
not be interpreted is compiled and dynamically linked for the UNIX version of ASPEN PLUS can be down-
to the ASPEN PLUS module. Thus, the overhead for loaded from the web site at: http://www.che.utah.edu/
in-line FORTRAN requiring compilation is small. In  vadapall/C6.html
addition to all the features mentioned above, in-line
FORTRAN also allows access to all stream and block
results. Program diagnostics like convergence results References
that are not available from inline-FORTRAN are ac-
cessed using the Summary File Toolkit routines of Bekiaris, N., Meski, G. A., Radu, C., & Morari, M. (1993)). Multiple
ASPEN PLUS. steady states in homogeneous azeotropic distillation. Ind. Engng
Chem. Res., 32, 2023.
To help run the continuation algorithms, three FOR-
Doedel, E. J., Keller, H. B., & Kernevez, J. P. (1991)). Numerical
TRAN blocks named F-1, F-2, and F-3 were written. analysis and control of bifurcation problems, (I) bifurcation in
Initially, the user does not use these blocks with either finite dimensions. Intl. J. Bifurcation Chaos, 1, 493.
the RCSTR or RADFRAC programs. As shown in Doedel, E. J., Wang, X., & Fairgrieve, T. (1994). Auto94 software for
Fig. 9(a), the problem is specified in the standard continuation and bifuration problems in ordinary differential
equations. In Applied Mathematics Report. Pasadena, California:
ASPEN PLUS format. After running the initial simula- California Institute of Technology.
tion, the bifurcation parameter and a set of observable Gani, R., & Jørgensen, J. B. (1994). Multiplicity in numerical solution
variables for the particular problem are chosen. The of nonlinear models: separation processes. Comput. Chem. Engng,
appropriate interface needed to run the algorithms with 18, S55.
RCSTR and RADFRAC programs is created using Güttinger, T. E., Dorn, C., & Morari, M. (1997)). Experimental study
of multiple steady states in homogeneous azeotropic distillation.
F-3. The interface consists of a set of files that carry the Ind. Engng Chem. Res., 36, 794.
guess values for the current step, the converged values Jacobsen, E. W., & Skogestad, S. (1991). Multiple steady states in
of the previous step and values of the program parame- ideal two-product distillation. AIChE Jl., 37, 499.
ters like ISF, IBR, ISTEP, etc. This step is shown in Kearfott, R. B, Novoa III, M. (1990). Algorithm 681, INTBIS, A
portable interval Newton/bisection package. ACM Trans. Math
Fig. 9(b). After obtaining these files, F-3 is hidden from
Software, 16, 152.
the RCSTR/RADFRAC program using the ‘Hide’ op- Keller, H.B. (1977). Numerical solution of bifurcation and nonlinear
tion for FORTRAN blocks. Next, blocks F-1 and F-2 eigenvalue problems. Applications of Bifurcation Theory, P.
are revealed to run as part of an overall flowsheet with Rabinowitz, Editor, Academic Press, New York, 359.
the ASPEN PLUS programs. As shown in Fig. 9(c) Morgan, A. P. (1987). Sol6ing Polynomial Systems using Continuation
for Engineering and Scientific Problems. Englewood Cliffs, NJ:
block F-1 runs before the program and F-2 after. The Prentice Hall.
results of block F-2 are written into a set of output files. Powell, M.J. (1970). A hybrid method for solving nonlinear equa-
These files are later copied into a set of input data files, tions. Numerical Method for Nonlinear equations, P. Rabinowitz,
using a UNIX shell script. While the main job of F-1 is Editor, Gordon and Breach, London.
Seader, J. D., Kumar, R. (1994). Multiple solutions in process simula-
to enter guess and parameter values for the next itera-
tion. Paper presented at ASPENWORLD 94, 6 – 9 November,
tion, the arclength algorithms reside in block F-2. The 1994, Boston, MA.
UNIX shell script calls ASPEN PLUS a number of Vadapalli, A. (1998). Computing Multiple Solutions in Chemical Pro-
times until the entire bifurcation diagram is calculated. cesses with Aspen Plus. PhD. Thesis in Chemical Engineering,
Each iteration of the shell script represents one pass University of Utah, Salt Lake City, Utah.
Watson, L. T., Sosonkina, M., Melville, R. C., Morgan, A. P., &
through the flowsheet sequence F-1“ (RCSTR or Walker, H. F. (1997). Algorithm 777: HOMPACK90: a suite of
RADFRAC) “F-2. The parameter and initial estimate Fortran 90 codes for globally convergent homotopy algorithms.
values in the input file are replaced at each pass ACM Trans. Math Software, 23, 514.

Vous aimerez peut-être aussi