Carnegie Mellon University

Research Showcase @ CMU

Department of Chemical Engineering Carnegie Institute of Technology

1986

A unified algorithm for flowsheet optimization

Y-D Lang
Carnegie Mellon University

Lorenz T. Biegler

Carnegie Mellon [Link] Design Research Center.

Follow this and additional works at: [Link]

This Technical Report is brought to you for free and open access by the Carnegie Institute of Technology at Research Showcase @ CMU. It has been
accepted for inclusion in Department of Chemical Engineering by an authorized administrator of Research Showcase @ CMU. For more information,
please contact research-showcase@[Link].
NOTICE WARNING CONCERNING COPYRIGHT RESTRICTIONS:
The copyright law of the United States (title 17, U.S. Code) governs the making
of photocopies or other reproductions of copyrighted material. Any copying of this
document without permission of its author may be prohibited by law.
A Unified Algorithm For Flowsheet Optimization

by

Y-D Lang, L. T. Biegler

EDRC-06-18-86 }

September 1986
 A UNIFIED ALGORITHM FOR FLOWSHEET OPTIMIZATION

YD Laity and L T Biefller

Clu:iuical Eiifjiiiuunng Department

CaiiMMjio-Mclloii Univoiiiiy
h. PA 1S213

A flexible ftlgofillmi (or flowsheet optimixaiion is developed on the FLOWTRAN

piocess sinuilaiof. The opiiiniiaiion strategy combines lite feasible and uifeasible

paid appioaches as well as ihe simpler black box approaches. While ihe most
A UNIFII-.D AlXOKllll.M rOK FLOWSHEET OI'TlMlZA'i'lON
efficient optimization strategy is often problem dependent, this paper presents

guidelines thai show which strategy is more efficient for a given problem. Also

V-D Lang and L. T. llegler embedded within the algorithm is a new Broyden strategy for efficiently converging
jrtmont of Chealcal Engineering
Cjmcgic-Mcl Ion University even complex flowsheets, without computing a new Jacobian. This allows for
Pittsburgh. PA 15213
strategies "in-between" the infeasible and feasible path procedures. A ratio test

based on the Kuhn-Tucker convergence test automatically and adaptively adjusts the

optimization strategy.

The implementation on FLOWTRAN is discussed in detail and a number of

examples are run to illustrate the flexibility of the implementation as welt as

demonstrate the effectiveness of the adaptive optimization strategy

SCOPE

flowsheet optimization has bten an important area for chemical process design
Submitted to Computers and Cheaical Engineering
and has its origins in linear programming work in the early 60s (see Griffith and
April. 1986 Stewart (13)1 With the increasing use of flowsheeting tools, process simulation has

become easier tui more widespread. Currently, on commercial simulators, however,

the optimization strategy is either ad hoc (i.e. often approached as a series of case

studies) or involves a detached optimization algorithm that supplies sets of decision

W
a
o
CO
variables as new parametric cases to the process simulator. This "black box"
3 cm
% <* approach was used by Giddy and c o w o i k e i s [1.12J and Friedman and Pinder [11)
°s.
CO O>
CD 2*
CD
CO

1
Ion Uni'
tvani

At
v.-il. . I . M I I s,•.»!(». Hph.m/.i i MUIIMJUIS. as well us Ity Cluliaml | H | an. I I i MMIIMJH In; i»u|M!, io. in iiil«;.i*il»l«; itaili Tins may lui .!^|HM ially iiuu if I-IIIHM IIU: IIOWSIMSMI IS

a.i.i F..1.1. i | l l ) wiiii IM..M- MI|*IUMICJICMI ^ailiem I»J:»IMI ti|iiiiiu/alH»ii :.tt«iitM|i«.N In iliMiculi !«• ciMivi;i<|tt «N 11 MI SOP alijoiiiluii das little ililficuity m haiulltiig the

Hirst: smiins any «|[Link] thai worn foquifttd woio uvaltiaiuil l»y |IUIIIMIUII«J lli«: [Link]! path |MC»IIIUIII. lu&iuad. I M |MO|>O*«I<J a liylMid alijtiniluii UPH) wtiuie th<*

ticci&uiii vjnaiilc* ami iucalcuiaiui»j the entire- flowsluiui. llowsluiui is paiiially convuujod using a fiAeil IMMIII>OI of Wuijsicm iiuiations between

SOP ilciaiimis. InlefusliiMjIy. tins ap|NOacli soiiHtlinios woikutl well evou when
RecoQMKHig not a compleie flowsheet calculation is e*|»cnsivo lot evaluating
compaiciJ tu uifoasible path. However, no cm«wia woie tjivun on how to choose the
•jiailienis. Isaacson ( IS) and Parker end Hughes (20) constructed reduced quatlraiic
number of Weysiom it mat ions for partial convergence or tiow to apply this algorithm
nuuicis Ly inciiviUiiji inuJci perturbation at each base pouu Howvevci. these iet|uiied
on flowsheets where the Wegstein algorithm may be inappropriate for convergence.
converged flowsheets fur each trial point evaluation. Moreover, 'eccni developments

m nonlinear programming algorithms have cast flowsheet optimisation in a new light. j In this paper we develop criteria for intermediate flowsheet convergence and

L'i.'ig Successive [Link]«JUC Programming (SOP) itif% (14). Povwcll (22J). Bema. Locke demonstrate this approach on a number of test problems. More importantly, however,

j-.a tVesteiberg [J] iiuinuMMHied [Link] equation solving simulators thai liowslicei this paper presents a unified strategy for flowsheet optimisation within a fairly

convergence a(*a opt .ni./[Link] can proceed simultaneously. A number of researchers compact and easy to implement structure. Interestingly, this structure incorporates all

na*e «pphea nns concept 10 sequential modular simulators ([Link]) with approaches discussed so far and because of its implementation provides a great deal

tneoufiyog results [Link] deferences exist among these studies in terms of of flexibility in developing optimization strategies tailored to difficult optimisation

calculating giao>ems J'»d implementing the optimnaiion algorithm all of them use *n problems

mfeasibie pain approach to optimisation; i.e. tea/ (or recycle) streams are solved
CONCLUSIONS AND SIGNIFICANCE
Simultaneously w*(h the optimisation problem using SOP to handle both tasks.

A flexible and efficient optimization strategy has been implemented and

More recemi, Begier and Hughes (6.7) advanced the concept of feasible
evaluated using the FLOWTRAN simulator. Due to the structure of the algorithms and
variants. ie converging ihe flowsheet between SOP iterations. On a limned number
in-line FORTRAN capabilities, the optimization implementation allows the following
of test problems tft.s strategy generally required fewer SOP iterations. Reasons for
solution options:
th.s are not c\e»f although it is easy to argue thai converging the flowsheet at each
• "black-box" optimisation
iteration may help to correct problems with SOP resulting from inaccurate gradients
• simultaneous convergence of recycle streams and design constraints using
or »n inefficient line search strategy (such as the one originally proposed by Han and
either Broyden or Newton methods.
Powell). In more recent studies (3.4). simply improving the line search algorithm and
• mfeasibie path optimisation (IP)
allowing (or analytic gradient information where available also improved the
• complete feasible variant optimization (CFV)
pe'formance of the mfeasiLle path approach.
• partially converged flowsheet optimisation with in embedded Broyden
algorithm (EBOPT)
Finally * <sjia ( 16) •[Link] that the feasible variant approach may noi always
 ID tins paiici we outline a swnplu donvaiiOM for an miprovod "l>lacfc-l»ox"

sirjieijy for recycle stream problems. A detailed iloiivatiou is then piesemcd lor an
• flowsheet decision vanahios
cnihcdttetJ Broyden method thai parlially converges flowsheets ai intermediate y guossed lea' stream variables
w - calculated teai stream vanatjius
[Link] iterations In addition, a heunstic strategy is presented that signals when F - oi>|cciive function
h - ie«< equations for converging the flowsheet
[Link] c oiiver gcnce *s desirable 01 not. The resulting EBOPT (Embedded Oruydun
c - additional equality constraints lor optimisation
Optimization) method is lairly general and leads to ihe mfeasible path and feasible g - inequality constraints

vonam algorithms as limiting cases.

Examples of objective and constraint functions can be found in previous studies as

well as in Ihe case studies presented later in the paper. To solve this problem, the
The EBOPT strategy is compared to Ihe infeasible path hybrid algorithm (IPH)

developed by [Link] and some theoretical advantages of EBOPT are demonstrated. In Successive Quadratic Programming algorithm essentialy applies a modified quasi-

particular. EBOPT can generally converge more complex flowsheeting problems more Newton method to converge Ihe optimally or Karush-Kuhn-Tucker (KKT) conditions of

efficiently because of its Broyden capabilities. Also. EBOPT is not as prone to line (MLP). To do this and maintain a consistent active set. the following quadratic

sea»ch failures as IPH is. program (QP) is solved at each iteration:

Fmaiiy. the capabilities of the optimization implementation are demonstrated on (QPl) Mln vT(x l , y l ) T • 1/2 d T Bd

five e*ampic problems. The first problem is characteristic of black-box optimization

e.t. b(xl. yl) • 1
, yl)T d - 0
while trie second illustrates the flowsheet convergence capabilities afforded by the
c(x1,/ Vc(x l , y l ) T d - 0
EBOPT and IP algorithms.
•<* 1 . y l ) • H(*1. y l ) T
The last three examples give a comparison of four algorithms. IP, EBOPT. CFV

and I P H . on reasonably difficult and realistic flowsheet optimization problems. On all

V
problems EBOPT performs better than either IP or CFV. IPH performs best on one

problem but suffers premature line search failures on the other two.

1. Preliminary Theory and Concepts

The flowsheet optimization problem is given by

Here B Is a BFGS (see (10]) update matrix to the Hessian of the Lagrange function
(NIP) Mm F(x.y)
with respect to x »nd y.'A detailed statement of this algorithm may be found in any
SI h(x.y) • y - w(x.y) • 0 of the above references end will not be given here. The version of the algorithm
cL.y) • 0
gU.y) * 0 used in this implementation was developed in Biegler and Cuthrell [ 4 ] and includes
K S" * £ *
the following features:
 1. A I I eHicioMi augmented Lag* [Link]-based line search strategy is used lo cunvOMjunce by slowly converging lecyde alyu* iiitinj mJ ih«iefore io|)iesa<H Hie
guarantee ylobal convergence and allow lull slops in lha region of tha
opinnum The latter ptopariy is not guaranteed by ll»e implementations of most lime consuming pen of Hie opi»»niieuon duOr
Hsn (14) 01 Powell (22].

2. An automatic variable and constraint scaling strategy is included that Since the mfeasible path approach consiiucit information about tha »-y suHace
gives good performance on flowsheet ing problems In addition a condition
number is calculated for lha 8 matrix in QP1 to del ermine when tha and does not require flowsheet convergence until the optimum is found, movement
problem is ill-conditioned or poorly scaled. occurs in both x and y as seen in Fig. 3a. This slap in x and y results from
3. BecA^c of portability, space and availability constraints, tha Ha* we 11 linearizing lha constraints and approximating the surface contours as wall tha
subiouiine VE02AD is used to *o4ve lha OP at each iteration. In Biegier
and Cuirvell ( 4 ) a more reliable and efficient OP coda was used. curvature of tha constraints. As saan above this approximation laads to a
However, there f no noticable differences in function evaluations due to
this substitution and. consequently, tha FLOWTRAN implementation is not straightforward quadratic program. Gradients for (OP1) im found by perturbing lha
affected by this change.
unc on verged flowsheet Also th« expansive vertical staps for flowsheet convergence

This SOP algorithm forms lha core of our unified optimization strategy. are avoided because flowsheet convergence is guaranteed as part of tha solution to

tha optimization problem. In fact, as will be illustrated later, application of tha

To see how the optimization strategies compare from a geometric viawpoint. infaasibla path approach in tha absence of degrees of fraadom is equivalent to
consider the sample flowsheet optimization problem in Figure la. Hera only one Newton's method.
decision variable. *. one tear variable, y. and only one taar aquation, h. are required

for the optimization problem. If one considers this problem from a casa study To prevent an ovaraxtrapolation of tha infaasibla • path approach it may ba

perspective, one can trace a curve for Fix) vs. x (Fig. 1b) where each point on tha advantageous to ensure that tha equality constraints ba converged (or at laast

curve represents a converged flowsheet. Expanding this problem in terms of both x partially converged) at each Iteration. Using tha faasibla variant approach, tha path

and y yeids the contour plot in Fig. 1c. Note that tha optimum lias on the solid for x and y Is given by Figure 3b. Htm vertical staps from flowsheet convergence

ime which represents the tear constraint and a nonlinear projection along this line are introduced and one seas that lha starting point for converging My) • 0 is given

gives the curve m Fig. 1b. by tha OP and is considerably belter than in tha "black-box* approach. Interestingly,

the OP that Is created and solved at each Iteration Is axactly tha same, and requires
[Link] the case-study or "black box" approach tha optimization algorithm is
tha same effort at each Iteration, as with infaasibla path. Nota that with this strategy
merely tied to the outside of the simulator and tha simulator Is responsible for
one assumes that tha flowsheet can ba solved raadily by tha racycla convergence
converging the flowsheet for each evaluation of tha optimization problem. Similarly,
algorithm.
gradient calculations involve perturbation snd convergence of the flowsheet for each

decision variable. Thus, no information about flowsheet convergence is passed In lha next sections we develop a new approach for Improving tha performance

between the optimizer and simulator. In Figure 2 this can be seen in terms of tha of the optimization strategy. This stratagy addresses soma of tha drawbacks with

horizontal steps (in x) made by tha optimizer and tha vertical steps (in y) performed both infeasible path optimization and tha faasibla variant stratagy and also links both

by the simulator. Note that these vertical steps usually represent flowsheet strategies more closely in terms of a unified framework. Before presenting this
strategy, however, u is useful to discuss further the diffaiences betweon foasihle Black Bo-

variants and black bo*" (or case-study) optimization. NFPl • NRP-NX

2. Black BOM Optimization vs. Optimization In • and y Feasible Vaiiant

NFPl • E-NX • NY • NRI

Comparing Figure 2 to Figures 3a and 3b one sees that the main disadvantage

m the optimization path of the first figure Is due lo lack of Interaction between K Infeasible Path

and the dependent variables, y. in the optimization step. In fact, the main difference NFPl • E-NX • NY • 1

between Figures 2 and 3b is simply thai with the feasible variant approach y is where. NFPl number of flowsheet passes pmr iteration

initialized much closer to the converged flowsheet. Generally, this leads lo more NRP number of recycle iterations to converge perturbations in decision
variables
efficient recycle convergence. The improved path however requires flowsheet
NRC number of recycle iterations to converge flowsheet
perturbations m x and y to create a larger QP problem at each iteration. Using SOP
at each new base point (vertical steps in Fig. 2)
with the black box" approach requires the solution of a much smaller QP:
NX number of flowsheet decision variables

f NY number of flowsheet tear variables

Mln
dx 2 x "x x
1 „ E fraction of equivalent flowsheet passes required for
decision variable perturbations (partial flowsheet passes)
NRI number of recycle Iterations to converge flowsheet
at new base point (vertical steps in Fig. 3b)

As seen from the above relationships, flowsheets with few degrees of freedom and
where y - y ( x l ) | h ( x l . y) - 0
many recycle components can be optimized more efficiently with the black box

with the tear equations and tear variables removed. Note that the derivatives in the approach. Note also that NRI is expected to be less than NRC This occurs because

above QP »re reduced gradients and require a converged flowsheet with each the y variables have better initialization with the feasible variant approach and. as

decision variable perturbation. will be seen later, more efficient recycle convergence algorithms can be used with

feasible variants. With the black box approach it is easy, however, to reduce NRC and
Because of the differences in the size of the optimization problem It is easy to
allow the optimization path to be similar to the one followed by the feasible variant
see that the "black box" approach can be superior to the infessible path or feasible
approach. This simply requires keeping track of how the dependent variables, y.
variant methods when the number of tear varibles greatly outnumbers the number of
change with x.
design variables snd the flowsheet is not difficult to converge with conventional

algorithms. The work per iteration for each approach can be approximated by: Consider a perturbation in variable x and a completely converged flowsheet for

that perturbation. For the tear constraints. Mx.y) • 0. we have to a first order

approximation:
 10 it

order cllucit. To avoid these problems. U»o number of Heranonv NHP and NflC. may
dh • 0 - V h\im • V h'dy
• i it
nood io t<e large IO force a amsii m o ' Tiwt it •tp«ci«H y tiu« boctuit flow*)****

convex (jcrxe elgoriihme have only unmmt c unve^u+nce p*op«m«i HHJI lequtrtog M P i
where dx • (0. 0 Ax 0J. Solving this aquation foi dy/A* ( give* column | oi Hie
io be larger ihan it normally e*peci»<J tw •imulattofv
matrix -<V Mx\ y')1) 'V Mx*. y'l*. Therefore, simply by saving the response of ihu y

variables to all perturbations in x. i.a. 3. Development of an Embedded Broyden Strategy

X - tdy%/A*[Link]/&Mt <*Ym%/**mw)

- -(V />UV)VV /KJ»'.KV • To summarize tha previous material and to introduce this section, consider

problem (NLP) again:

we can use the solution of (QP2). d^. and wrila dy as
(NLP) Min FU.y)
d • Y d .

s.t. Mx.y) • y - w(x.y) • 0

c(x.y) • 0
Note from QPi that this equation solves the linaariiation of the tear constraints and gU.y) 1 0

therefore leads to the step in x and y givan In Figure 3b. Also, it is interesting to

note this approach leads to the same slap thai is generated by the Reduced Feasible

Variant (RFV) algorithm described by Biaglar and Hughes ( 7 ) .

In tha "black box" approach tha y variables ware eliminated and tha constraints.

hU.y) • 0. were always sat Is f lad. rnvn for parturbatlons of x. Tha infeasible path (IP)
However, because flowsheet convergence Is the outar loop to saveral levels of

iterative calculations, the convergence error In the tear aquations can ba relatively and tha completa faaslbla variant (CFV) atrataglaa daal with iNLP) explicitly In tha

large. Therefore, in order to calculate the Y matrix correctly, lha perturbation size space of x and y and aolva (QPI) at each Iteration. In addition. CFV converges tha

needs to be chosen accurately. If we include second order corrections and the equations. h(x.y) • 0. by adjusting tha y variables iftir (QPI) Is solved. As mentioned

convergence error. {. in our tear constraints we can write: above, it may ba more efficient to psri/sHy converge tha constraints. htx.y) • 0. at

dh - i - V A ' A J T •(HA**) • each iteration since IP yields a converged flowsheet at the optimum anyway. Also. If

an afficiant and rellabla aquatlon-«olv'€r can ba applied, one can handia both h and c

by convarging tham almuJtaneously.

Rearranging this expression gives an ordar of magnitude estlmata for tha errors in

the Y matrix:
In this section we present en ambaddad Broydan approach for partial
) • (X|AK|)
convergence within flowsheet optimization. This strategy Incorporates tha Infeasible

path and faaslbla variant approaches as limiting casaa and can be vlawad as a
Note that choosing a perturbation alxa too small laada to appreciable error due to
modification of lha hybrid approach proposed by Kisala (18]. In tha previous
convergence noise while a large perturbation alza leads to an error due to second
section we observe that slowly convarging recycle algorithms can laad to
 12 13

inefficiency wiih the black bo* optimization algorithm. These algorithms log. Lot H" h1 V h ' l at iiouii C and consuloi tlie Broyden formula (10J

WiMjbiom o» direct substitution) ere also used in tlie feasible variant and hybrid
k T T
(Hi) ll e ) 6 / A A
apfwoaches. Thus, for flowsheet thai aie difficult lo lolve. one can e«i>oci

conveigcuce problems ai intermediate points.

A •

When tear variables and constraints ere perl of the oplimizeiion problem, we

have enough gradient information from QP1 to allow also for more efficient

convergence routines. In particular. Broyden routines have been used with good , - he,"* VS - h<.V>
results (9. 19. 21] m converging complex flowsheets, even those with additional We note thai this update relation can also be epplied to the nonsqumf metrix without

uesign constraints This section therefore add/asses two points. violating any assumptions (see Dennis & More [ 10]) es to its derivation. Also from
1 HOW can a Broyden algorithm be embedded within the optimization QP1 we see thel the slep from C to D is genereted by H*d • -t»(xc. yc). We cen now
strategy to (partially) converge the flowsheet at intermediate points?
apply the updete formula to get H1 at point 0;
2 What criterion should be used to decide whether (partial) convergence is
necessary at intermediate points?
T
T U-H°o) A * T (q-H°«) e
<B2)
3 1 An embedded Broyden algorithm •To V+ T~~^ J

we now derive a modified Broyden algorithm for partial convergence et

intermediate points. For convenience we will use only the tear variables, y. and leer
Starting from point O we keep x constant and only change y to converge the
equations. nU.y) > 0. in presenting this derivation. Application of this method to flowsheet; thus i a - 0 for the following Iterations. Applying (B1) to H' gives the
additional design constraints, cU.y) • 0. end additional dependent variables is following relations:
straightforward.

To converge or partially converge the equality constraints, consider the step in (B3)
v 6 6
Figure 4. At point C. the gradients end values of the objective and constraint

functions ere evaluated and QP1 is constructed end solved. The search direction

from QP1 and a suitable stepsize leads to point D. from which the equality (I.e. tear and
and any design) constraints may be converged. If one were to apply a Broyden

method to converge the dependent variables et this point one would want to have <B4)

the [Link] (V^h)' at point O to initialize the Broyden method. Since this is not

available and it would be expensive to construct this information, we derive en Note that since J_ is determined by QP1 for the first step t\d Ha does not affect
update strategy based on the gradients evaluated at point C. in the later steps, we write the updete formulae (B2). (83) es:
 IS

with the conventional Broyden approach. Hits method also handiei t)e»ign conairamit
H* • [q-H^\ i'/l'i
easily

3.2 Criteria for using en e«nb«<Sde>d B#oyd«n method

As noted in previous studies, the choice of optimization strategy is often

problem dapandant. If Iha gradients are reasonably accurate and iha flowsheet is
This expression can be simplified mvmn further by noting that
only "mildly" nonlinear, than iha infeasible path approach will converge smoothly.

If. on the other hand, iha flowsheet Is highly nonlinear and difficult lo converge (with

complex units thai are failure prone) then a faasibla path approach with Broydan's
and
method and appropriate safeguards could ba more reliable and efficient. However,

these characteristics mm not always known a priori. Indeed, M shown by Kisala

[18]. partial flowsheet convergence may lead to more efficient performance In

if / u / / Broyden steps are taken. Also for the first step: solving the optimization problem.

However, care must ba taken In dealing with partial flowsheet convergence at

each iteration. In particular, partial convergence can ba detrimental to the Una search
Therefore, the Broyden steps can ba given by: algorithm in determining the stepslie for the next point.
(8b) H l • H * • I** 0 .* 0 ) * < V C

In the SOP algorithm, a given stapsiza along the search direction, d. is accepted

if a "sufficient" decrease is observed with some merit function, p. In the algorithms

of Han and Powell, this function is the exact penalty function; in our algorithm an

augmented La grange function is used. In either casa. it is wall known (sea e.g. Han

114)) that tha search direction from QP1. d. is a descent direction for the merit
Note that when the line search in SOP allows a full step U - l ) . (B5) and (B6)
function. Consequently, finding a nonzero stapsiza Is guaranteed, at least in theory,
differ only in the denominator of the second term. Also, note that In Iha absanca
for tha infeasible path algorithm.
of degrees of freedom (no x variables), thasa aquations reduce to the conventional

Broyden approach. For tha feaslbla variant algorithms, one can also prove that a nonzero stapsiza

will ba found during tha lina search. This can ba shown because all points In tha
Testing of this approach on the problems given below shows that no more than
line search hava converged equality constraints and tha OP solution, from a faasibla
S iterations are required to converge the flowsheet at intermediate points. As shown
point, is also a descant direction for tha lina search function f.
 16 17

However, lot flowsheets thai ate only partially converged, ont csntwl guarantee heuristics and. consequently, will not always guarantee improved fw formeitce

iliat a stepsue with a decreased merit function will l>e found. A simple illustration competed to inf«a»ible path No*ivn K,« ••• u h» ihai ~« f><«»»«-« I«I«« *• *•• r

of this is given in Figure S. From the QP base point at A. one sees the mem encouraging

function can be viewed as a function of 1 along the search direction found by OP 1

Heuristic strategy for partial convergence
Executing a fi*ed number of convergence iterations for a given siepsiie. point B. say.

m*y decrease or increase the objective function. In fact, for a fixed number of
For SOP a common measure of Kuhn Tucker error (KTE) for problem INLP) is
iterations, it is possible that the equality and inequality constraint infeasibilities may given by (22):
also increase. Consequently, it is possible that partial convergence may move the
KTE - |Vf\/| X K*,.l • X I** I • Z l'<,l
mem function value from point 8 to B*. Note that this behavior can occur arbitrarily

close to point A. say. point C. and thus lead to a line search failure - even though
where u. v( and l^ are multipliers calculated for QP1 for g. h and c. respectively.
perfectly reasonable stepsizes exist for infeasible path.
Also. 9(U.y). is defined as max (0. g^[Link]. Reducing KTE to within a zero tolerance

For this reason we apply partial convergence only sfter the line search is a necessary and sufficient condition for satisfying the KJCT conditions for (NLPL

algorithm finds a stepsize. In this way we can avoid line search failures and also Now. from Figure 4. if SOP finds a step from point C to point 0. one needs to

save some work at intermediate points. determine if additional work is required by Broyden's method to move to point

E. Since this method converges only cf and hj( a heuristic measure of how much
As further justification for this safeguard we note that, by itself, the infeasible
improvement can be had is given by the ratio:
path algorithm converges quickly and takes full steps in the neighborhood of the

optimum. For this case, partial convergence is usually not necessary. On the other

hand, at the beginning of the optimization, the search direction may overextrapolate
If this ratio remains small, intermediate convergence is not necessary. If it remains
and lead to a point that is difficult to converge. Here it would be Inefficient to
consistently largo, however, full or partial convergence may help to speed
partially converge the flowsheet during the line search. Instead, allowing the line
convergence. To us« this ratio we propose two triggers for intermediate
search algorithm to find a more reasonable point first will save some effort.
convergence.

Even with this safeguard, one is still not guaranteed that partial convergence
« " ( Z I"AI • Z lv,l)
leads to better performance for the optimization. One way to measure the success

of partially converged points would be to compare merit functions from iteration to

at point C. mov« to point D mnd converge c( and ^ (to point E, say) until the relation
iteration. However, one still needs to know if. a priori, partial convergence is

desirable or even necessary at a given point. In the next section we develop a

strategy for dealing with this task. We should mention that this strategy is based on
is satisfied.
 18 19

Also, ii Is possible lhat th« flowsheet may an ime>meriieie

point Consequently, we Impose • m»«.m«im loi mie» me<1" •' t

toiivcMjcnco In oui case studies successful

bul on moving lo point 0. required more than 5 iterations.

V>i • Z I ' , '

We conclude this section by emphasising that the above strategy is based on

heuristics thet. from our limited experience, work reliably and efficiently. Since very
use Broyden's method lo convarga (to point E. say) cf Bnd hf so thai (BRl) is
little theory governs the concept of partial convergence, we used these guidelines in
satisfied
our implementation. In tha next section we discuss how the constants #% and « }

Otherwise, both poims C and 0 He close to the constraints and Intermediate were chosen and give a statement of tha algorithm.

convergence is not required. Note that by adjusting tha • ' • . one can davalop a full
4. Algorithmic Statement and FLOWTRAN Impte ntatlo
spectrum of methods between tha In feasible path approach ( f ( * 1 . ?>aoo) and tha

feasible variant strategy ( f ^ O . f ( « O L Using tha concepts stated above we now present an algorithmic statement of

tha Embedded Broyden Optimization (EBOPT) strategy and outline the features and
in choosing these parameters. i% should be sat between zero and one to allow
options used In tha FLOWTRAN Implementation. In the algorithmic statement we
for partial convergence. « s should be sat small in order to avoid tha first trigger at
assume the reader It somewhat familiar with tha SOP algorithm and will not dwell
the next Iteration. Our experience Indicates that often vary few Broyden Iterations (1
on its details. Tha reader Is referred to ( 4 ] for tha line search strategy and update
or 2) are required to satisfy (BRl) even If «, Is small, (say 10*1. ^^ on tha other
formulae.
hand, can be sufficiently greater than unity without hampering performance. This

results because point C for the second trigger Is sufficiently close to satisfying the
4.1 Algorithm
constraints: a linearization from that point and a line search usually determine point

D that is reasonably good without partial convergence. In fact, for tha problems we Step 0) Sat U M SOP Iteration counter. 1-0. and initialize the flowsheet with x*

solved, the second trigger for partial convergence was not necessary for good and y*. y* can be found by (partially) converging the flowsheet. Set » as the Kuhn-

performance. Tucker tolerance.

In addition to tha above triggers we have also Included the following conditions Step 1) At U*. y") find tha gradients for F. g. h and c with respect to x and

for Intermediate convergence. First. If tha currant Iteration Is In the nalghborhood of y. This can be dona by direct loop perturbation [ 6 ] or chainruling [ 3 ] .

the optimum, applying Broyden Iterations la usually not necessary since tha
Step 2) Solve (QP1) given above to gat tha search direction d for x and y.
constraints are close to being satisfied anyway. Therefore If KTE( S 10«. say.
Evaluate KTE IHI (BR) at Iteration I using the expressions above. If KTE S , . stop.
where « Is the Kuhn-Tucker tolerance, we do not apply Intermediate convergence.
 21
20

Aim. ih« user • specification ol the opiunuanon piobiem can be nude umply by
Step 3) Porform a Una search with a suitable merit (unction, f. to fiml a
aOUiiMj a few lm«t ol «n line FORTRAN to the input daia lor the iiiiiniiiiun problem
stcpsue. 1. along d. [Link] ^"•«'«kd i . y^y^Xd^.

SCOPT handles up to ilwee tear streams (as does SCVW). which it converges
Step 4) Sol Broyden iioration counter k«0 and evaluate the flowsheet (or
simultaneously, and up to a total of 40 decision and leer variables. Decision
<*'".
variables can be chosen from equipment parameters or feed streams. These

variables are accessed through PUT statements that ere common features in

KJE i 1O« FLOWTRAN.

1
set y" • y\ ao to step 7.
In addition, the user needs to specify a reletive perturbation size (between 10 '

Step 5) k
Set y* • y* and apply the Broyden formulae (B5) and (86) to y and and 10 ' is recommended) and a reletive Kuhn-Tucker tolerance (usually between 10*'

(h.c) until: and 10 *). It should be noted, however, that choosing a Kuhn-Tucker tolerance too

small can result in line search failures and poor steps near the end of the run.
I i
because the gradients mey not be accurate enough to satisfy the tolerance.
then set y"' * yk. If the above relation cannot be satisfied after five iterations

(w>5). set y" • y\ Another option In our implementation deels with the choice of tear variables.

Step 6) Evaluate the gradients at (x l # \ y" 1 ) as in step 1*. Update the In most optimization studies, stream flow rates, pressure and specific enthalpy are

chosen. Temperature Is not chosen because for multiphase streams, enthalpy is not
Hessian matrix for QP1.
realizable from temperature alone. However, for s/ng/t phase streems calculation of

Step 7) Let i»i*1 and go to step 2. enthalpy from temperature is usually direct Bnd a level of iteration and some

convergence noise ere eliminated in the perturbation step. Since perturbing the
4.2 FLOWTRAN Implementation temperature for single phase tear streems can lead to more accurate derivatives, we

The optimization capability in FLOWTRAN was Installed by writing a type 2 have added a T/H option.

(convergence) block. The structure and argument list for this block, called SCOPT.
Finally the t parameters need to be specified for the embedded Broyden
was the same as the existing recycle convergence block, SCVW. Because we did not
algorithm. As mentioned above, fy which defines the second trigger, can be fairly
change any code in FLOWTRAN. we Implemented direct loop perturbation as the most
large. In our experience values of f) > 3 have still led to good performance end
straightforward way for evaluating gradients. Since FLOWTRAN generates end
this trigger was always Inactive. Consequently, we have not used this test et ell.
compiles a FORTRAN main program at run time, it can easily accommodate jn-line

FORTRAN and user written subroutines as part of the input data. This, In turn, allows
On the other hand •f was set. after some testing, to 0.01. As explained above,
the optimizer to evaluate partial flowsheet passes if needed for gradient evaluation.
it takes surprisingly few Broyden iterations to satisfy this test. The most Important
 22 23

The only constraints are bounds on th« decision variables as well as bound* on the
parameter lor determining intermediate convergence Is therefore »,. Sailing f | >f|*0
fraction of feed to ihe bottoms meant.
•fi our impiomom«hon leads to • feasible variant approach. Sailing f | «1.0 yields Ihe

•nfeasible paih approach and ml at medial a values of t , allow partial convergence. In This problem is typical of many simple process optimization problems. Since
our testing, setting , ( to 0.4 yielded good results although this parameter is problem there are no recycles. SOP deals with this model in "black-box" fashion arid solves
dependent. However, on many problems, examination of the output shows that
it completely each lime it requires a function evaluation. Alternately, mn entire
performance of this strategy is not very sensitive to §§. In Table 1. the ranges of
flowsheet could easily have been treated instead of a single unit. The solution of
^% and f j under which the » / n * performance would be achieved are tabulated for the
this problem Is also given in Figure 6. This problem wes solved to a relative Kuhn-
Embedded Broyden Optimiiation (EBOPT) Strategy.
Tucker tolerance of 10 \

5. EM ample Problems
Nole from Table 1 that the performance of the optimization algorithm is

characteristic of the black-box approach. Because the model needs to be solved

The following five example problems were solved by a number of approaches.
several limes It is not surprising that over 16 Simulation Time Equivalents (STE's
A number of similar problems were solved in addition to these. However, for the
measured at the starling point) were required to optimize this three variable problem.
sake of brevity, we chose this set because it represents what can be expected from
Because of the tight tolerance, seven iterations appears to be reasonable for this
the implementation in terms of performance and flexibility. The first problem is
casa.
essentially a black-box implementation on e single unit. The second problem

illustrates how the mfeasible path and Broyden methods can be used to converge
Problem 2 - Cevett Problem Simulation
flowsheets. The last three problems deal with moderately-sized flowsheets, some

with complex models. These allowed comparison of the embedded Broyden strategy To demonstrete the capability of the infeaslble path and EBOPT methods for

(EBOPT) with a number of recent and efficient optimization strategies. Due to the Newton and Broyden convergence, respectively, we selected a modified form of the

flexibility of the algorithm and features of FLOWTRAN. none of these strategies was Cavatt problem, reported by Rosen m%6 Pauls (23). Hare the number of stream

difficult to implement. components waa reduced from 16 to 11. Figure 7 Illustrates the flowsheet where Z1

and 22 ware chosen as tears and Table 2 lists problem data and the converged
Problem 1 - Black Bo* Optimisation
solution. This problem was first solved using the Wegsteln convergence block In

FLOWTRAN with all of the) default options. In this cass 13 Iterations wars required
The first problem deals with a single unit optimization of a 25 tray distillation
to converge tho flowsheet to the default relative tolerance of 0.0005. Using the
column with sidestreams. As illustrated in Figure 6. the distillation column problem,
same tolerance, this flowsheet was also converged using the infeesible path (IP) and
which is solved by a Thiele-Geddes model (FRAKB). seeks .to maximize the degree of
EBOPT methods in 4 and 6 iterations, respectively. However, comparing STE's for
separation of its 5 components among Its overhead, bottoms and sidestreams. The
this problem shows that these methods are not competitive with Wegstein. The
decision variables are the fraction of feed to the two sidestreams and the distillate.
 25
24

EBOPT method retinues 26 flowsheet passes lo construct a Jacobian matrix ai ihe EBOPT. with the heuristic strategy detcnUed in Section 3. only n« c«iu<f to use Hi«

fusi iteration, ihe IP incihod oouils 10 constiucl Ihis Jacobian ai every iteration. Bioyden strategy after iterations 1 and 3 After iiicsc. SOP r**d no irouMe

converging tins problem by itself.

Because of ihe effort required for the initial Jacobian. Broyden's method may

not always be competitive for solving simulation problems. Embedded within an Unfortunately for this problem, the IPH algorithm suflered a line search failure

optimization strategy, however, where the Jacobian Is calculated anyway. Ihe Broyden after 5 iterations. Restarting at this point resulted in a second line search failure

method performs much more efficiently. after 3 additional iterations. In his study. Kisala (18) also reported a line search

failure for Parker's ammonia problem. The reason for this, as explained in Section 3.
For the next three examples we compart the IP and EBOPT strategies with the
may be that, because a fixed number of Wegstein iterations Mt9 applied for each
CFV (Complete Feasible Variant) [7] and IPH Unfeasible Path Hybrid) [18] algorithms.
function evaluation in the line search, a descent direction cannot be guaranteed and
The last two algorithms were implemented by using FLOWTRAN's Wegstein
this method can be prone to failure.
convergence block to (partially) converge the flowsheet between SOP iterations. For

IPH. two wegstem iterations were used between every SOP iteration, as suggested Problem 4 - Methytchlorobenzene Process

by Kisala ( 18], For CFV. the flowsheet was either converged to FLOWTRAN's default
This problem is adapted from an example in the FLOWTRAN manual (24).
tolerance or until 30 Wegstem iterations had been exceeded.
Using the default costs and prices In the costing blocks, the optimization problem

On all problems relative tolerances of 10 4

were used for the Kuhn-Tucker error. illustrated in Figure 9 was formulated. Here six decision variables were chosen for

All problems were recycle flowsheets with complex unit operations and nonideal the optimization. These ere listed along with their initial and optimal values In Table

thermodynamics. 4.

Problem 3 - Ammonia Process A Because this problem contained a rigorous (and often unreliable) absorber model

and the FORTRAN code for FLOWTRAN was not available to us. we were unable to
This problem was adapted from Parker and Hughes (20] »nd has been used in
provide error returns to Ihe optimization algorithm and thus continue in the event of
other studies 19. 18]. The problem statement is given in Figure 8 and in (20).
unit convergence failures. Obviously, error returns are a necessary feature In the
Because of different thermodynamic properties and fewer decision variables, values
Implementation of any flowsheet optimization strategy, and the lack of this capability
of the objective function are slightly lower in this study. The starting point and
reflected how we could solve this problem.
optimal solution for this problem »f given In Table 3. As seen from Table 1. the

double loop flowsheet with »n equilibrium-based reector Is fairly easy to converge From the results In Table 1. one sees that the EBOPT strategy required less

and optimize. Here the EBOPT and CFV approaches are close in performance. effort than either the CFV Of the IP strategies. However, to prevent premature

although EBOPT is slightly superior. Because no Intermediate convergence was termination due to failure In the absorber block. Intermediate recycle convergence

applied for the IP run. more iterations were required than with EBOPT. Interestingly. was suppressed for EBOPT during the first two SQP iterations. The EBOPT algorithm
 26 27

[Link] to apply Broyden's method only afiar iteration 4 in oidar to gal aalitfaciory from Table 1. required only 47% of Hie effort i»»«i ihe infaasihle path algorithm

performance. noeUed Broydon iterations were applied «n•> the J<<1 4ih bin A>n * n d 9t* SOP

iteration However, the accelaianon* after t»>« 3«d 4ih and 5ih i i o i i i o n i were not
Also, duo 10 difficult!** with lha abtorbar. lha IP algorithm could not be
effective (and also did not lead to closer pomisl because they ted to using more
converged from the starling point for EBOPT. From a slightly diffarant starting point
than 5 Broyden iterations without satisfying the ratio lest. This illustrates the
(shown m Table 4). 12 iterations were required to satisfy a Kuhn-Tucker tolerance
difficulty of converging this flowsheet from intermediate points.
slightly above 10 *. Again, because of the unreliable nature of the process units, a

belter and more consistent comparison could not be) made. Similar, but more pronounced results were encountered with the CFV algorithm.

Here the convergence algorithm was unable to converge the flowsheet for the first
The CFV algorithm required over S times the computational effort that EBOPT
three SOP Iterations. For these points the maximum of 30 Wegstein iterations was
required. This represents the difficulty that SCVW has to converge this flowsheet at
exceeded end. consequently. CFV required a lot of computational effort. On the
intermediate points. In fact, for SOP Iterations 1, 2. 5 and 0. CFV required the
other hand, the IPH algorithm did very well for this problem. Because it uses a fixed
maximum of 30 iterations without converging the flowsheet at these base points.
number of recycle iterations at intermediate points, the progress of the optimization

was better than IP. but none of the convergence problems encountered with CFV. or.
Again, as with the previous problem. IPH terminated with a line starch failure
to a lesser extent, with EBOPT. were observed here. Also for this problem there were
after 12 iterations. This could be due to the descent direction line search problem
no apparent difficulties with line search failures.
explained in Section 3.

In summary, partial convergence of the flowsheet at Intermediate optimitation

Problem 5 - Ammonia Process B
iterations led to better results on all of the recycle optimization problems then with
The flowsheet for this problem Is given In Figure 10 along with the problem either the IP or CFV algorithms. However, as shown in section 3. cere must be taken
statement. The decision variables and their initial and optimal values are given in to implement this strategy properly. Therefore, this study illustrates the potential of
Table S. Unlike problem 3. this ammonia process has a single loop design with 3 the EBOPT strategy for flowsheet optimization, although further work may be required
flash units. The unit operation and cost blocks for the reactor are taken from to tune the algorithm for specific problems.
Chapters 9 and 10. respectively, of the FlOWTRAN manual. To make the problem

more interesting, feed rates were chosen as decision variables and s constraint was

imposed on the flow rate of the ammonia product. This type of constraint can be

treated in a straightforward manner by all four of the algorithms compared.

Because of problems with error termination in FLOWTRAN. we suppressed the

Broyden option for the first iteration in the EBOPT run. Even so, this run. as seen
 28 29

( l7)Kai|aluoio. S. "Process Optimization b*y Flowsheet Simulation . Technical

REFERENCES Rosearch Center of Finland. Publication #20. (1984)
(i)Ballmen. S.K end J.L. Caddy. "Optimization of Methanol Process by
Flowsheet Simulation". I&EC Pioc. Des. Dev.. 16, 3. p. 337. (1977) (l8)Kisala. T.P. Sc.O Thesis. Massachusetts Institute of Technology.
Cambridge. MA (1985)
(2]Berna. T.J.. MK Locke and AW. Wester berg, "A New Approach lo
Optima Alton of Chemical Processes". AICh€ J.. 20. 1. p. 37. (1980) ( !9)Meiceife. S.R. and J.O. Perkins. "Information Flow in Modular Flowsheeting
Systems". Trans. I. Cham. E.. 66. p. 210. (1978)
(3)Biagler. L.T.. "Improved Infeasible Path Optimization for Saquantial
Modulars Simulators. Part I: The Interface." Comp. and Cham. Eng.. 9. 3. p. (20)Parker. A.L. and R.R. Hughes. "Approximation Programming of Chemical
245. (1985) Processes". Comp. and Chem. Eng.. 5. 3. p. 123. (1981)

(4]Biaglar, L.T. and J.E. Cuihrall. "Improved Infeasible Path Optimization for (2i)Perkins. J.D.. "Efficient Solution of Design Problems Using a Sequential
Sequential Modular s Simulators. Part II: The Optimization Algorithm." Modular Flowsheet ing Programme". Comp. end Chem. Eng.. 3. p. 375 (1979)
Comp. end Cham. Eng. 9. 3. p. 257. (19851
(22]Powell. M.J.D.. "A Fast Algorithm for Nonlinearly Constrained Optimization
(S]Biegler. L.T. and R.R. Hughes. "Infeasible Path Optimization of Sequential Calculations". 1977 Dundee Conf. on Num. Analysis. (1977)
Modular Simulators". AlChE J.. 28. 6. p. 994. (1982)
(23)Rosen. E.M. and AX. Pauls. "Computer Aided Process Design: The
(6]Biegler. L.T. and R.R. Hughes. "Optimization of Propylene Chlorination FLOWTRAN System". Comp. and Cham. Eng.. 1. 1. p. 11 (1977)
Process: A Case Study Comparison of Four Algorithms". Comp. and Cham
Eng.. 7. 5. p. 645 (1983) (24]Seader. J.D.. W.D. Saidar and AX. Pauls. FLOWTRAN Simulation - An
Introduction. 2nd edition. CACHE Corp. (1977)
(7]Biegler. L.T. and R.R. Hughes. "Feasible Path Optimization for Sequential
Modular Simulators". Comp. and Chem. Eng.. 9. 4. p. 379. (1985)

[8]Challand. T.B.. "Computerized Optimization of Complete Process

Flowsheets". Chem. Eng. Prog.. 79. 6. p. 65. (1983)

[9)Chen. H-S and MA. Stadtherr. "A Simultaneous Modular Approach to

Process Flowsheeting and Optimization". AlChE J.. 31. 11. p. 1843. (1985)

[ I0]0ennis. J.E. and J.J. More. "Quasi-Newton Methods. Motivation and

Theory'. SIAM Review. 19. 1. p. 46 (1977)

( 1 DFnedman. P. and K.L Pmder. "Optimization of a Simulation Model of

Chemical Plants". I&EC Proc. Oes. Oev.. 11. 4. p. 512. (1972)

(i2]Gaines. LO. and J.L. Gaddy. "Process Optimization by Flowsheet

Simulation". I&EC Proc. Oes. Dev.. 15. 1. p. 206 (1976)

[ i3)Griffith. R.E. and R.A. Stewart. "A Nonlinear Programming Technique for
the Optimization of Continuous Processing Systems". Mgt. Scl.. 7. p. 319
(1961)

[M]Han. S-P. "A Globally Convergent Method for Nonlinear Programming".

J. OpL and Appl.. 22. 3. p. 297 (1977)

(15] Isaacson. R.A_ Ph.0 Thesis. University of Wisconsin. Madison (1975)

I I6]jirapongphan. S. Sc.O Thesis. Massachusetts Institute of Technology.

Cambridge. MA (1980)
 IAIIIJ I [Link]>. 1.1 V.i I ton K llpl liaf z . i l i o n

1 I I I |-<1S | l l | c TAHU 7.
N«». Study ( . i i f I'-Uli (IP) rmirr 1 l-ii OV

1 . Disti 1 I J U U I I
7.33361 — --

7 — —

O'U Time (second) 111.09 — — /1 7?

STL'* (6.74 sctondi/SII.) 16.48 — —

358.2 391.2 J3.2 37.2 3i8.2

2. CJVCCC prob l e a 339.4 508.9 263.6 200 339.4

Objective function — _
2995.5 3602.5 646.0 548.3 2995.5
4 6 13
CPU Ti*e (second) 461.27 : 145.19 — 36.86 2395.5 3499.9 1524.2 1199.0 2395.5
STE's (36.83 seconds/STE) 12.52 3.94 1.00
604.1 752.2 636.6 523.0 604.1
). Ammonia A
1129.9 1212.3 1149.3
Objective funcClon
No. of I t e r a t i o n s
-3.67128
11
-3.67115
7
-3.966**
8
-3.67097
6
•s 1082.4 1129.9

1764.7 1812.6 1771.0 1736.6 1764.7

CPU Time (second) 1008.79 652.36 976.06 702.41
STE's (68.27 seconds/STE) 14.78 9.56 2606.7 2636.3 2607.6
14.29 10.29 «7 2589.6 2606.7
Ranges for * ,• 0.036 * » j * 0.55
1844.5 1852.9 1844.0 1839.8 1844.5
0.0018 * \ * 0.0186

831.7 832.4 831.6 831.3 831.7

MCB
Objective function -0.898925* -0.897342 -0.91142* 1 -0.89525 1214.5 1215.0 1214.4 1214.2 1214.5
No. of Iterations 12 7 12 10
CPU Tlae (second) 212.78 120 118.3 71.55 120 120
120.08 241.02 642.21
STE's (19.83 seconds/STE) 10.73 6.06 12.15 32.33 49 49 13 49 49
Ranges for 1 .« » t > 0.268
* 2 > 0.078

5. Amonla B
Objective function -24.9421 -24.9270 -24.9380 -24.9277
No. of Iterations 26 11 7 8
CPU Ti»e (second) 3604.56 1672.57 922.62 4 2480.77
STE's (215.90 seconds/STE) 16.70 7.75 4.27 11.49
Ranges for * ,* i < 0.443

0.0074 s * 2 & 0.0126

« » i r.a C e.J at | n [ . "t due to line si InlU
 TA4HJ ). H . M . I l ul AMMOIII a A Problea T A M O . 4. M.»..|| o( Ht >\ I'roblea

Oft I tow tut lath

ILL.!-1'" '
A. Objective Function -3.67128 -3.23W Objective Function -0.898925 0.W06/ -0 MO)1)
It Design Variables
1. Inlet Temp of Kcactor 180 Design Variables
147.226
2. press of Kcactor 29 SO 3000 1. lottos press, of
Absorber (psls) 32.0401 32.0 32.0
3. Inlet Tea? of High
Pressure Flssh -24.0 -24 2. Top press, of
4. Purge Frsctlon 0.0802432 0.10 Absorber (psls) 31.0401 31.0 31.0
3. Conversion of 3. Spile Fraction
Nitrogen (X) 45.0 41.0 of tecycle 0.377398 0.33 0.33
C. Tear Variables 4. Split Frsctlon
I. Flowrate of Withdrawal 0.622602 0.67 0.67
2193.02 2170. 5. In1st T««p. of
Flash 227.422 270.0 270.0
568.426 612.0
6. Outlet Tee?, of
30.1334 16.3 Hest Exchanger 100.0 120.0 120
85.7868 70.5
104.604 125.0 Tear Variables

2. Temperature 289.883 215.0 Flowratcs:

3. Pressure 2950.0 3000.0 1. HC1 0.0 0.0 0.0

2. Benzene 0.623078 0.0 0.0

3. MCI 82.1304 80.0 90.0

Temperature 100.00 120.0 100.00

Pressure 50.0 50.0 50.0

 TAIili: 5. KCKUU of Awnonla H I'rohtea

SI«|>1« Fluwalicti ( l r i l i l i , n I

1 term Opt t«HJi S t a r t 1 ng | ' i .

la) Mln f{» .y )
|tcilvt function -24.9421 -20.659
' I e . t . h(*.y)-y-w(».y)-U

tx- sign Variable*

I n l e t Tenp.
' •
of Reactor 400.0 400.0

2. I n l e t Teap.
of 1st Flash . 65.0 65.0

3. I n l e t Tc»p.
of 2nd Flash 35.0 35.0

< • •
I n l e t Teap. of
Recycle Compressor 75.3139 107.0

5. Purge Fraction ( I ) 0.814030 1.0

6. I n l e t Press, of
Reactor 1984.69 2000

7. Flowratc of Feed 1 2626.66 2632

8. Flowratc of Feed 2 688.206 1648

C. Tear Variables

1. Flovratc

1303.88 1648.0
lc)
H
2 3921.21 3676.0
contours of F(x.y)
NHj 551.760 424.9

Ar 183.708 143.7

™4 2095.34 1657

2. Temperature 75.3139 60

3. Pressure 1901.49 1930

(*.y) - 0
 -countoura of F(i.y)

Figurt 3a .
Inftaalblt PathOP) Optimisation Strategy

Flgurt 2
Black Box Optimisation Strategy
 Figure *
Partially Converged Strategy
for EBOPT
'l» l .y J ) • Ad) OM/K-
P - 20psls
Sl/F
S2/F- Off

"sufficient
decrease** line
10 S2

Has (Cj In OH - C* In SI
• (C7 • Cg) In S2 • C 9 In »T)

0.1 S Sl/F i 0.7

Figure 5 C 5 - 3.049 0.1 S S2/F * 0.7
C 6 - 1.276 0.1 * OH/F S 0.7
Source of Line Search Failures 0.25 i (SI • S2 • 0H)/F i 0.95
with Partial Convergence C7 - 1.597 20
C 8 - 0.430 SI
C 9 - 3.119
T - 120*F
P - 25PsU

BT
Figure 6

Dlstlllstlon CelusB Opclmlsacloo

Initial

ObJ. Fen - 6.M2 ObJ. Fen - 7.554

Sl/F - 0.2 Sl/f - 0.1
S2/F - 0.2 S2/F - 0.2499
OH/F - 0.3 OM/F - 0.3353
 Klgurr H
I'rocc

Iao
u.

s:
Ste
"2 " 873
H2 - 287
Ar - 7.1
CH4 - 12.9
3
80* F
U7pala
• XL • JS
O •
• • jo; _ _ ,
m Z

4 Mln [Venture Profit over 10 years t 15Z)

s.t. HH3 purity la product 2 95Z

I<3> *
£
MH3 In purge S 1.5 ltmol/hr
teactor teap i 1000*P
Vapor In coapreaaora I 99X
i

•
 Figure 10
nla Proctst ft

5.2Z N,
94Z H:
0.01Z Ar
O.79Z CH,

Max (Before tax profit t 15X over 5 years]

s.t. N H 3 in purge * 4.5 lb aola/hr
No liquid In compressors
100,000 tona NHj/yr
Product
KH 3 Product Purity i 99.9X
1.8 i H 2 / N 2 In S 3.3 combined feed

jsqjoeqy
•

O
d
o c
5 •
<f O

a +
2