Download as pdf or txt
Download as pdf or txt
You are on page 1of 23

Computers them. Engng, Vol. 14, No. 6, pp. 631-653, 1990 009% 1354/90 $3.00+ 0.

00
Printed in Great Britain. All rights reserved Copyright 0 1990 Pergamon Pressplc

SYNTHESIS OF GENERAL DISTILLATION SEQUENCES-


NONSHARP SEPARATIONS
A. AGGARWAL and C.A. FLOUDAS~
Department of Chemical Engineering, Princeton University, Princeton, NJ 08544-5263, U.S.A.

(Received 18 July 1989; final revision received 16 November 1989;


received for publication 30 November 1989)

Abstract-The synthesis problem of distillation sequences that separate a given multicomponent feed
stream into several desired multicomponent product streams, while allowing nonsharp separations in
the columns, is addressed. A superstructure is proposed that contains options for distribution of the light
and heavy key components, all possible sequences and all alternatives for stream splitting, bypassing and
mixing. Light and heavy key component recoveries define the nonsharpness of separation and are treated
explicitly as optimization variables. This superstructure is modeled as a mixed-integer nonlinear program-
ming (MINLP) formulation whose objective is to minimize the total annual cost. Shortcut simulations
and regression analysis are used to develop cost functions to be used in this formulation. The solution
of the MINLP formulation results in an optimal separation sequence that may involve sharp and/or
nonsharp distillation columns. The nature of the proposed mathematical formulation is investigated and
the search for a global optimum solution is performed using a new decomposition approach. The proposed
approach, which is implemented automatically in the procedure NOUS (Not Universally Sharp separ-
ation), is illustrated with six example problems and the results show that using nonsharp separations can
result in savings of lo-30% and even 70% in some cases (see Example 2) as compared to sharp sequences.
It is also illustrated that columns with nonadjacent key components can be included easily by extending
the proposed approach

INTRODUCTION problem. Mahalec and Motard (1977a,b) proposed


the BALTAZAR procedure which uses a stream
The synthesis of optimal separation sequences has
matching algorithm to develop a flowsheet. Alterna-
been one of the most important subjects in process
tive separation sequences are developed to bridge the
synthesis. Separation processes constitute a signifi- discrepancy between the multicomponent streams
cant portion of the total capital investment and encountered at various points in the flowsheet and
operating expenses for a chemical plant and a lot of potential multicomponent product streams required
interest has been generated in the development of by the problem specifications. Despite the insights
systematic approaches that will select optimal separa- provided by these approaches, they do not seem to
tion sequences. yield the desired quality of solutions.
Multicomponent separation processes can be Nath (1977) investigated a subproblem of the gen-
classified into two groups: eral separation problem that involves the separation
Class A-Separation of a single multicomponent of a single multicomponent feed stream into multi-
feed stream into pure substances; and component products and considered systematically
Class B- Separation of several multicomponent the introduction of nonsharp splits. Muraki and
feed streams into several specified multi- Hayakawa (1984) developed a procedure for the
component products. synthesis of sharp split distillation sequences with
bypass to produce two multicomponent products.
Over the last two decades problems of Class A have In the first stage, they used a heuristic procedure to
received a lot of attention. Various synthesis methods identify a column ordering. In the second stage,
have been proposed and Nishida et al. (1981) pro- advantage is taken of the graphical representation
vided a good review of the various approaches. of the separation process through the use of the
Westerberg (1985) and Cheng and Liu (1988) pro- material allocation diagram (MAD) proposed by
vided a review of work performed on distillation- Nath (1977) so as to optimize the separation process
based separation sequences. by the introduction of division and blending. A MAD
There has been much less work for problems be- representation of a nonsharp separation product is
longing to Class B-the general separation problem.
not always unique. Muraki et al. (1986) extended this
In recent years researchers have focused on this work to the synthesis of sequences with one feed and
multiple multicomponent products. They proposed a
method for constructing a modified material allocation
tAuthor to whom all correspondence should be addressed. diagram (MMAD). It should be noted though, that

CACE 14/&-c
631
632 A. ACGARWAL and C. A. FLOUDAS

the above approaches have addressed subproblems of lower and upper bounds for each of these flowsheets.
the general separation problem. The lower bound is established by using a relaxation
Floudas (1987) addressed the general separation of the corresponding nonlinear program and the best
problem. The basic idea was to derive a superstruc- solution of the nonlinear program provides the upper
ture, based upon simple and sharp columns, which bound. Upper and lower bounds very close to each
has embedded in it all possible separation sequences other were considered as implying optimality.
and all possible options of splitting, mixing and It is interesting to note that most of the above-
bypassing of the streams. Simple columns are defined mentioned work assumes sharp splits between com-
as columns with a single feed and one top and one ponents in the separators. When multicomponent
bottom product (no side streams). The formulation products are to be produced, it is frequently not
was a nonlinear nonconvex programming (NLP) necessary to separate the feed into pure components.
problem and the solution of the nonconvex NLP In fact, column bypasses and nonsharp splits may
provided an optimal separation sequence that corre- be essential in reducing the cost of the separation
sponds to a local optimum. Floudas and Anastasiadis sequence. In his review, Westerberg (1985) illustrated
(1988) proposed a mixed-integer linear programming this point by a small example shown in Fig. 1. If a
(MILP) formulation for the general distillation-based mixture of A, B and C is split partially in a single
separation problem. They illustrated that this formu- column and then these products are matched by
lation can be utilized as an initialization procedure splitting into pure component products and mixing
for the determination of a good starting point for the them, it can be shown that the latter solution costs
nonconvex NLP formulation of the general separ- several times the cost of the former both in terms of
ation problem. Wehe and Westerberg (1987) de- capital and operating costs.
scribed an algorithmic procedure for the synthesis of Bamopoulos et al. (1988) studied separation
distillation sequences with bypasses, using sharp sep- sequences with nonsharp separations. They repre-
arators. Their procedure involved examining various sented the problem in terms of a component recovery
structurally different flowsheets and determining matrix (R-matrix) which is an algebraic extension

Al
BI IO

Fig. 1. Advantages of nonsharp separations.


Synthesis of general distillation sequences 633

of the material allocation diagram (MAD) proposed distillation sequence that may involve sharp and/or
by Natb (1977). The R-matrix representation of nonsharp separation tasks. The proposed approach
a nonsharp product is unique, as opposed to the is illustrated with five example problems. It is also
MAD representation. Furthermore, a MAD can be shown (in Example 6) that the proposed approach
typically constructed for the restricted case of a single can be readily extended to include columns with
feed and every component being assigned to at most nonadjacent light and heavy key components where
two products. Manipulation of the component recov- the intermediate components distribute between the
ery matrix generates several plausible flowsheets top and bottom products of the column.
which include the operations of distillation, stream
splitting and stream mixing. They proposed a two- PROBLEM STATEMENT
stage approach. In the first step, a heuristic ordering
of options, coupled with a depth-first technique, is The problem to be addressed can be stated as
proposed. In the second step, a best-first technique f0110ws:
is proposed for searching for the few better schemes.
A single multicomponent feed stream of known
Heuristic rules are employed to optimize the partially
conditions (i.e. flowrate, composition, temperature and
developed sequences. This work provides interesting
pressure) is given which has to be separated into a
insights into the complex domain of nonsharp
number of multicomponen t products of specljied corn -
separations but does not guarantee that all possible
positions. The problem is then to synthesize the optimal
sequences are studied. Cheng and Liu (1988) pro-
distillation sequence, allowing the use of nonsharp
posed a simple heuristic method for the systematic
separators, that separates the singIe multicomponent
synthesis of initial sequences for nonsharp multicom-
feed into several multicomponent products and satisfies
ponent separations. They introduced the component
the criteriu of minimum total unnuuJ cost.
assignment diagram (CAD) for representation of the
problem and the separation speciJication table (SST) The basic assumptions made in this paper are:
for feasibility analysis. Their procedure is useful for
1. The thermodynamic state of the feed streams
generating good initial sequences but it may generate
and distillate and bottom products for all
more columns than required. Muraki and Hayakawa
columns as well as the type of condenser to be
(1987) considered the degrees of separation sharpness
used in each separation are known.
as important variables and proposed a method of
2. The heating and cooling requirements for each
finding optima1 degrees of separation sharpness and
column are provided directly by hot and cold
the ratios of stream divisions for a given separation
utilities (i.e. no heat integration takes place) and
sequence. Muraki and Hayakawa (1988) coupled
all columns operate at fixed pressures designated
this with their earlier two-stage strategy (Muraki and
by the design engineer. These pressures are
Hayakawa, 1984). The first stage involves searching
selected from the range of allowable pressures
for the separation sequence. In the second stage a
which are determined by the available hot and
search is made for optimum values of the degrees of
cold utilities.
separation sharpness and ratios of stream divisions.
3. Each distillation column performs a simple split
These two stages are repeated until an optimum
(i.e. one feed and two products).
process is synthesized.
4. In each column, distribution of only the key
In this paper, a systematic approach based on
components is allowed in the distillate and
mathematical programming techniques is presented
bottom products. The key components are con-
for the synthesis of distillation sequences involving
sidered to be adjacent components (assuming
nonsharp splits of components. The light and heavy
that the components are arranged in order of
key components are allowed to distribute in the top
relative volatility).
and bottom products of each column and recoveries
of the key components are introduced explicitly as Assumption 1 is required for the generation of
optimization variables. A superstructure is developed simulation data in which the feed and product states
that allows for nonsharp splits and involves all are fixed. These conditions are often determined by
possible sequences and options of splitting, mixing the physical properties of the feed and products and
and bypassing. This superstructure can be formulated play an important role in determining the operating
as a mixed-integer nonlinear programming (MINLP) pressures of the columns. In this work, feed streams
problem. The objective function of this formulation to all columns are considered as saturated liquids at
is the total annual cost which is to be minimized. A their bubble point at the pressure of the column. It
set of shortcut simulations is performed to generate is assumed that pumps and exchangers are available
the cost data for each column, and then regression to meet this requirement. The distillate and bottom
analysis is employed to develop cost models which products are saturated liquids at their bubble
adequately represent the cost of each column over a points. All columns have total condensers, Removal
range of flowrates, feed compositions and key com- of Assumption 2 would necessitate the inclusion
ponent recoveries. The solution of this mathematical of heat exchange possibilities between condensers
formulation provides information about the optimal and reboilers of various columns. This would mean
634 A. AGGARWAL and C. A. FLOUDAS

Phase I-nonsharp separation superstructure


In the first phase, a superstructure is derived that
has as columns the required separation tasks accord-
ing to the number of components that are involved.
Since each column performs the separation between
adjacent key components (Assumption 4 in the last
section), the superstructure would consist of a maxi-
COST MODELS
mum of “n - 1” columns for a “n” component feed

1 I
mixture. The basic idea is to avoid having more than
one column with the same set of key components.
This does not exclude schemes with prefractionators
since a component which is the light key for one
t column can be the heavy key for a subsequent column
MIXED-INTEGER NONLINEAR in the sequence. Embedded in the superstructure are
PROWlAMMING (MINLP) all possible sequences for series, parallel or series-
FORMULATION
parallel arrangements. Besides that, all options of
splitting, bypassing and mixing of streams are also
included. In each column the key components are
allowed to distribute between the top and bottom
+
I I products. The nonkey components go entirely to the
top or bottom product depending on whether they
are lighter or heavier than the key components.
A superstructure for a four-component mixture to
Fig. 2. Synthesis strategy. be separated into two multicomponent products is
shown in Fig. 3. Column I has A and B as light
and heavy key components, respectively. Column II
allowing the column pressures to vary within the separates between B and C and column III between
allowable range. This problem could be approached C and D. In Fig. 3, letters within brackets represent
by either considering the columns at discrete pressure components which will be present only if nonsharp
levels (Andrecovich and Westerberg, 1985) or by separation takes place. For example, stream 7 will
considering the pressure of each column as a contin- consist of component B only if the recovery of this
uous variable (Floudas and Paules, 1988). The heat component in column I is incomplete (which means
integration issue for nonsharp separation sequences that column I is a nonsharp separator). The bottom
will be addressed in a separate publication (Aggarwal product of column I (stream 10) can, in principle,
and Floudas, 1990). Assumption 3 excludes complex contain all four components. If the feed to column I,
columns from the scope of this work and comes stream 6, contains all four components, then stream
about due to two primary motivations. First, there 10 would contain all four components for nonsharp
exist a number of industrial examples where only separation and only B, C and D for sharp separation.
simple columns (nonsharp and/or sharp) are used. In this work, sharp separation is taken to mean
One such area is natura1 gas processing (e.g. LPG recoveries of greater than 99%. If stream 6 consists
recovery units). Secondly, we consider the approach of only A and F3then stream 10 would contain A and
presented in this paper as a step in the direction of B or B only depending on whether the separation is
understanding general separation sequences. Future nonsharp or sharp. The operation of columns II and
work would attempt to incorporate complex columns III can be described in a similar fashion. This shows
in this framework. The first part of Assumption 4 is that the superstructure consists of both options for
made in order to facilitate the work and in fact forms each column-sharp or nonsharp separation. The
the basis of determining lower bounds for key com- main steps in the derivation of the superstructure for
ponent recoveries. To relax this assumption, relations a four-component feed stream are as follows:
between recoveries of the key components and distri-
bution of other components would be needed for 1. The feed stream is split into five streams, three
each column. The last section of the paper discusses of which go to the three columns (streams 1, 2
a procedure to incorporate columns with nonconsec- and 3) and the other two arc overall bypasses to
utive key components (to remove the second part of the two products (streams 4 and 5).
Assumption 4). This is illustrated by Example 6. 2. The outlet streams from each column are split
into:
SYNTHESIS STRATEGY
(a) two streams that are directed to the two
multicomponent products (e.g. split stream
The synthesis strategy for distillation sequences 10 into streams 13 and 14);
involving nonsharp splits consists of four phases as (b) two streams to be sent to the other columns.
shown in Fig. 2. There are some exceptions to this step. If the
Synthesis of general distillation sequences 635

A
B
C
D

1 127 20 l(B)CD

Fig. 3. Superstructure for a four-component system.

outlet stream from a column does not con- recoveries of key components. A few representative
tain the key components for some of examples are:
the other columns, then no split stream goes
A sharp sequence with I, II and II in series
to those columns (e.g. stream 16, which is
results from setting streams 2, 3, 12, 17, 26
the top product of column II, can have
and 27 to zero and fixing the recoveries of key
components A, B and C only. Therefore, a
components for all columns at 0.99 (which
split stream (stream 17) goes to column 1
corresponds to sharp separation).
which has A and B as key components, but
Nonsharp sequence with I, II and III in parallel
there is no split stream going to column III
by setting streams 11, 12, 17, 21, 26 and 27 to
since this column has C and D as key
zero and leaving the recoveries as variables.
components whereas stream 16 does not
Sequence with nonsharp coIumns I and II in
contain component D).
parallel and then in series with sharp column III
The creation of such a superstructure can be by setting streams 3, 11, 17, 26 and 27 to zero
generalized for any number of components. The and fixing the key component recoveries for
number of columns is determined from the number column III at the upper bound (0.99).
of components-one column for each separation Sequence with only columns II and III by setting
break point. This superstructure consists of: stream 6 to zero.
(i) an initial splitter for the feed stream; Phase 2--development of cost models
(ii) a splitter at each product stream of each
In this phase the data required for developing the
column;
cost models is generated by performing simulations
(iii) a mixer at the inlet of each column;
for each column in the superstructure. Each column
(iv) a mixer prior to each desired multicomponent
is characterized in terms of the design parameters and
product.
operating variables that will be considered fixed in the
The proposed superstructure is the most general model. The assumptions described earlier fix the
one under the assumptions stated earlier. If for a following characteristics for each column: product
particular problem, practical limitations make certain state, condenser type and column pressure. As
configurations undesirable or infeasible, they can be noted by Andrecovich and Westerberg (1985), the
eliminated by deleting some streams from the super- economically optimal ratio usually lies in the range
structure or imposing restrictions in the mathematical of 1.1-I .25 times the minimum reflux ratio. Since
formulation. From the superstructure shown in Fig. the cost function is normally flat for a short range
3. many configurations with sharp and/or nonsharp around the optimum, the optimum reflux ratio is
separators can be obtained by setting the flowrates often assumed to be 1.2-1.3 times the minimum
of some of the streams to zero and fixing some of the reflux ratio (King, 1971). In this study the optimal
636 A.AGCARWAL and C.A. FLOUDAS

reflux ratio is assumed to be 1.2 times the minimum. ing methods. The heat duties were used to calculate
Then, selection of a set of key component recoveries the investment and operating costs for the heat
allows determination of the rest of the information exchangers. The surface area required was calculated
about a column for a given feed flowrate and by assuming a value of the overall heat transfer
composition. coefficient and a value of the mean temperature
For each distillation task, first a few shortcut difference (AT). Given the cost of utilities, the annual
simulations arc performed to determine lower bounds operating cost for each simulation was calculated.
on key component recoveries which correspond to a In this study, 8000 h of operation per year are
point below which significant distribution of nonkey considered. Thus, for each set of data:
components starts taking place. For this study, the
total capital cost = cost of column
lower bounds were determined such that the recovery
(including trays)
of the nonkey components is 98% or greater. The
+ cost of condenser
work performed for various examples showed that
+ cost of reboiler,
for hydrocarbon systems the lower bounds are nor-
mally around 0.80-0.85. All the simulation data was annual operating cost = cost of cold utilities
obtained using the PROCESS flowsheeting system + cost of hot utilities.
(Simulation Sciences, 1985). A set of 15 shortcut
If a is the payback period and B is the tax
simulations is performed for each column to cover
correction factor, then the total annual cost of each
a range of feed compositions and key component
column (for a fixed feed flowrate, compositions and
recoveries. The recoveries of key components would
fixed set of key component recoveries) can be
range from 0.99 (corresponding to sharp separations)
expressed as follows:
to the lower bounds determined earlier. A good idea
of the range of feed compositions to be covered for
total annual cost = d (total capital cost)
each column can be obtained by analyzing the super-
structure. For example, in Fig. 3, column I can, on
+ j? (annual operating cost).
one hand, receive feed consisting of all four compo-
nents (as from stream 1) and, on the other hand, can To span the possible range of flowrates, 20, 40, 60
receive feed consisting of only components A and B and 80% of the maximum possible flowrates were
(as in the case of a series sequence of III, II and I). studied. As mentioned earlier, simulations need to be
The shortcut simulations provide information about performed for only one flowrate since diameter and
the number of trays, feed tray locations, top and heat duties at different flowrates can be calculated
bottom temperatures and condenser and reboiler heat easily given the data at one flowrate. Since 15 simu-
duties. It should be noted that shortcut simulations lations were performed for each column, 60 sets of
need to be performed for onejowrate onfy since heat data were generated corresponding to every column
duties vary linearly with total flowrate (composition in the superstructure. After the cost of the columns
and recoveries remaining constant) while the number was calculated corresponding to each set of data, the
of trays and temperatures do not depend upon the next step was to identify a good model that represents
flowrate. The additional information needed for cost the simulation data. The cost of a distillation column
analysis is the diameter requirements for the various increases with increasing feed flow following the
cases. One option is the use of a simple formula to six-tenths power law (Peters and Timmerhaus, 1980)
calculate the diameter (Peters and Timmerhaus, 1980; if all other variables remain fixed:
Rathore et al., 1974). The other option is to perform
c = n(F)O.6,
one rigorous simulation corresponding to each short-
cut simulation (15 in all for each column) using the a is a coefficient in the above expression. Andrecovich
number of trays given by the shortcut method. For and Westerberg (1985) showed that for a fixed feed
a given set of key component recoveries and feed composition and recoveries the cost of a column can
composition the diameter is directly proportional to be adequately represented by a fixed-charge model of
the square root of the flowrate (Peters and Timmer- the following form over a range of flowrates:
haus, 1980). So the diameter at different flowrates can
c =a,+a,F.
be calculated easily. In the present work the second
option was selected due to its better accuracy. Also, When the recoveries of the key components as well as
the total simulation work for each column (I 5 short- the feed composition can vary, it seems reasonable to
cut and 15 rigorous simulations) corresponds to only expect that both a, and a, might be represented as
about 30 s of CPU time on an IBM 3081. functions of these variables. Of the various models
The next step involves a representation of the cost to which the cost data was fitted using regression
and sizing requirements of the distillation columns analysis, the following was found to be the best
in the superstructure. The cost module method of representation of the cost of a column:
Guthrie (I 969) can provide column cost based on the
pressure, shell size and material, number, diameter cost = a, + a, + a*+ + u3rhk+ c 6,xi F,
and type of trays. Valve trays were used for all cost- ie, >
Synthesisof generaldistillationsequences 637

where r” is the recovery of the light key component The key components for the columns are given by:
in the top product, rhk is the recovery of the heavy
LK, = {i 1i E Z is the light key for column j } J’ E J,
key component in the bottom product and xi is
the composition of component i in the feed. Z is the Z?K, = (i 1i E Z is the heavy key for cohrmn j ) j E J.
set of components in the feed stream. This model was
found to predict the column costs within 510% of Since the nonkey components are not allowed to
distribute, a set is defined which contains components
the actual costs for the various examples considered
in this work over the range of flowrates, compositions not present in certain streams:
and key component recoveries. It should be noted ND, = {i ) i E Z is not present in stream k} i EZ,
that the above cost expression introduces nonlin-
earities in terms of bilinearities of the flowrates, key k E N.Two variables are associated with each column
component recoveries and compositions. j to define the fractional recovery of the key compo-
nents:
Phase 3-mathematical formulation r),?= fraction of the light key component going out
In the previous sections, the development of the in top product, i E LK,, j E J,
superstructure and the generation of the required
ry = fraction of the heavy key component going out
simulation data has been presented. This section
in bottom product, i E HKj, j E J.
describes the mathematical formulation of the super-
structure as a MINLP problem. This MINLP will Each of the streams k will have associated with it
then be solved to provide the optimal distillation as a variable the flowrate F*. Next, variables are
sequence for the given problem. To derive the needed to represent the composition of the various
mathematical formulation, the following index sets streams in the superstructure. Since at each splitter
and variables will be defined to characterize the the composition of all streams leaving the splitter is
superstructure. the same as the composition of the entering stream,
The components will be denoted by the index set composition variables are needed only for feed and
Z = {i} and the columns by the index set J = {j). The product streams for each column. Refer to the super-
superstructure involves a set of streams k denoted structure for a four-component system shown in
by the index set N = (k}. The splitters are denoted by Fig. 3. The composition of streams l-5 would be
the index set S = {s 1. The splitter so E .S will represent the same as that of the feed stream. Considering the
the initial splitting point of the feed stream. The splitter at the outlet of column I it is clear that the
index set MC = {m’) represents the mixers prior to compositions of streams 11-14 is the same as that
each column and the index set M’ = (m’} represents of stream 10. The same holds for the various other
the final mixers prior to each product. The products splitters in the superstructure.
will be denoted by the index set P = {p}. The follow- Having defined the index sets and variables, the
ing sets will be defined to give the connection of the objective function and the constraints that apply to
sets of splitters and mixers with the various streams it are described next.
in the superstructure:
The objective function
Sp = {/I I E iV is an inlet to splitter s} s o S, An objective function must be derived that can be
used to compare the alternatives provided by the
Sz”’ = {I 1I E N is an outlet from splitter s} s E S, superstructure. The objective function is chosen to
represent the total annual cost. In the section on
Mgr= {i 11E N is an inlet to mixer ml} m’o M’, development of cost models it was shown that by
using regression analysis on the simulation data
MzC = {I 11 E N is an inlet to mixer mC} mc E MC, the following model was developed for each column:

M$ = (I 1I E N is an outlet from mixer mc} mc E MC. cost = a,, + a, + a,r” + ajrhk + 1 bixi F.
( irl >
The inlet and outlet streams of the columns are given A binary variable yj is associated with each column
by: representing the existence of column j. If a particular
column exists in a sequence, the corresponding yi
SU, = (n 1n E N is the inlet to column j > would be set to 1, otherwise it would be zero. This
j o J, ensures that if a particular column is not selected then
no associated cost appears in the objective function.
SU’“p
/ =j~ ‘J” E N is the top product of column j} So for each column j:
1
costi = aojy, + a,j + azjrt + u,,.r]?i’+ c buxik F*
ib, >
SU/bo1= {q )q E N is the bottom product of columnj}
j o J. i E LK,, i’ E HK,, k E SUj,
638 A. AGGARWAL and C. A. FJ_~I.IDA~

where u,,~, a,,,azj, a,i and b, are constant coefficients Mass balances for each component in finai mixers
for each column determined from the regression
analysis. The above cost expression for a column
involves terms linear in the key component recoveries
($ and r>,) and the composition of the feed stream where C,p is the desired quantity of component i in the
to the column (x,~) multiplying the feed flowrate to final product p.
the column. This makes the cost function nonlinear.
The objective function would consist of an
Nonkey specifications
expression as above for each column. Since the aim
is to find the sequence with the minimum total cost, These constraints ensure that the nonkey compo-
the objective function can be expressed as follows: nents do not distribute between the top and bottom
products for the columns:
PC = jElc x,=0 (i,k)ENDik.
i E LK,, i’EHK,. k E SU,,
Summation of mole fractions
where PC is the annualized project cost.
The above objective function is subject to the
following type of constraints:

Mass balance for splitters


Logical constraints
x F!- 1 F,=O SES.
These are constraints based on the binary variables
Id:” IESZ”’
which are associated with the existence or nonexist-
Key component balances for each column j ence of the various columns. The logical constraints
force the flowrate to a column to be zero if the
These constraints take into account the distri-
column is not chosen. These constraints appear in
bution of the key components between the top and
the formulation for each column j :
bottom products. Because of the use of key compo-
nent recoveries as explicit variables, trilinearities are
introduced in this set of constraints:
FPx,P - r!t F,xi,, = 0 i E LK,, n=SU,, pcSU?Q, where 8 is an upper bound.

F gx.,q - r!!kF
I, ” x.111
= 0 i E HKj, n E SU,, q E su,+1. Nonnegativity constraints
To facilitate the application of the global optimum
search technique proposed in this paper, new vari-
ables V;,) are introduced for the component flowrates
for inlet streams to columns by defining: Bounds on recoveries

_&= F,,xin iE1, n e SU,.


(r!$)L C r$ d (rF)” i E LKj, j E J,
Now, the above balances can be written as follows
with only bilinear terms: (rp)’ <rF < (rt“)” i E HK,, j E J.

F,x,,,--$A,,=0 ieLK,, ncSU,, PESUJIOQ, The upper bound on the key component recoveries is
0.99 (corresponding to the case of sharp separations).
F,xi, - r”,“An = 0 i E HKj, n E SU,, qESu~.
This may lead to some discrepancies in the material
balance in cases where some products do not contain
Component balances for each column j
all components or are pure components. To avoid
J;, - FP x,,, - F, xiu = 0 such inconsistencies, the mathematical formulation
would have the upper bound as 1 with the under-
i E I, n E SU,, p E SUYQ, q E SU,+‘.
standing that whenever a solution is found with some
of the recoveries equal to unity, it would actually
Component balances for mixers prior to each column
correspond to 99% recovery. The lower bounds are
In F,x, - ,,I=,A, = 0 i E4 mc E MC. determined as described earlier in the section on
rnc “I= development of cost models (Phase 2).
The full MINLP formulation (P) of the model is
shown next:

min C k E SU,,
jal
Synthesisof generaldistillationsequences

subject to:

Phase 4-search for the global optimum


(a) Remarks on the mathematical formulation. The Floudas ef al. (1989) proposed a new three-step
proposed mathematical formulation is a MINLP approach for searching for the global solution of
formulation since it involves continuous and integer nonconvex NLP and MINLP problems. In the first
variables and features nonlinearities expressed in the step, the variables are partitioned into complicating
form of bilinear terms in the objective function and and noncomplicating variables where complicating
some of the equality constraints. The integer vari- variables are those which, when fixed, render the
ables appear linearly. The solution of the MINLP will original problem easier to solve. Using this partition-
provide the desired optimal sequence. Because of the ing of the variable space, the original problem is
bilinearities present, the problem is nonconvex. This decomposed into two subproblems in the second step.
may imply that a unique global solution cannot be The first subproblem is called the primal problem
guaranteed. and is the original problem solved for a fixed set of
(6) Global optimum search procedure. Previous values of the complicating variables. The second
work on solving nonconvex MINLPs has been based subproblem, called the master problem, is based on
on projection on the integer variables and then the dual representation-of the primal problem and
applying either the Outer Approximation or the consists of Lagrange functions generated from the
Generalized Benders’ Decomposition technique. information provided by the solution of the primal
Duran and Grossmann (1986) described an approach problem. It also consists of those constraints of the
using Outer Approximation for solving MINLPs. original problem involving only the complicating
Kocis and Grossmann (1988) presented an algorithm variables. The third step consists of using the Gener-
using Outer Approximation with Equality Relax- alized Benders’ Decomposition method (Benders,
ation. Paules and Floudas (1989) developed an algor- 1962; Geoffrion, 1972) to iterate between the two
ithmic procedure, APROS, based on the Generalized subproblems until an appropriate stopping criteria
Benders’ Decomposition technique for solving has been met signalling that an optimal solution has
MINLPs. These approaches cannot guarantee opti- been found. At each iteration, the best solution of
mality and very often fail to find the optimal solution the primal subproblems found so far represents an
since at each iteration a nonconvex NLP has to be upper bound on the global minimum. The master
solved. problem provides a lower bound on the global
640 A. AGGARWAL and C. A. FLOUDAS

solution which increases monotonically as more and Table 1. Data for Example 1

more constraints are added to the master problem. Products (kgmol h-‘)
Feed
The stopping criteria is the crossing of the lower and Components (kamol h-‘) PI P2
upper bounds. The main idea behind the approach is Propane 100 30 70
to select the complicating variables in such a manner iso -Butane 100 50 50
that both the primal and master subproblems are n-Butane 100 30 70
convex and hence their respective global solutions Utilities available:
Cooling water @ 305 K (20 K rise), cost = SO. I59/ 106 kJ
can be determined easily. It should be noted, however, Steam @ 343 kPa, cost = %2.45/l@ kJ
that there is no theoretical guarantee that the proposed Payback period = 2.5 yr
global optimum search technique will aIways identtfy Tax correction factor = 0.52

the gtobai optimum solution. Despite this limitation,


the global search technique represents a systematic
approach and is very successful, as demonstrated by In this formulation yr, F, and f, (i E iV, i E I, j E J,
the various example problems. n E SU,) are complicating variables and, therefore,
(c) Selection of the complicating variables. In the fixed. As a result the primal problem (PI) has the
mathematical formulation (P) being considered in following important property:
this paper, nonconvexities exist in the model due to Property-The solution of the primal problem
the presence of bilinearities in the objective function (Pl) is the global optimum of (Pl).
and some of the constraints. For example, if the Proof-With the complicating variables fixed, the
flowrates of streams are fixed, the column balances primal problem (Pl) is a linear optimization problem.
become linear equations. On the other hand, fixing Hence a local solution is the global solution which
the composition of various streams entering and can be found easily using any standard algorithm for
leaving columns also makes these constraints linear. linear programming.
Similar choices are available for the objective func- (e) Formulation of the master problem. With the
tion and other nonlinear constraints. An appropriate selected set of complicating variables, the only
combination has to be chosen so that both the primal constraints the master problem would have,
and master problems become convex. Notice that if besides the Lagrange functions, are the splitter mass
the binary variables, the totalflowrates for ail streams balances and the logical constraints. Bounds can be
and the componentJIowrates for streams which feed the obtained for some of the flowrates in the super-
columns are considered as complicating variables then structure by an analysis of the problem data.
both the primal and master subproblems become linear. Appendix A describes a procedure for determining
(d) Formulation of the primal problem. By select- upper bounds for the overall bypass streams in
ing the binary variables, the total flowrates for all the superstructure. To aid the master problem in
streams and the component flowrates for column feed finding feasible combinations, some additional
streams as complicating variables and fixing their constraints can be derived which are described in
values, the primal problem (PI) is obtained. Appendix B. The complete master problem (Ml) is

c
minj.zJ atj + QjrF + a&F + c b,,x,
,EI
k E SUj,

subject to
F,xi,, - r!yfb = 0 i ELK,, n E SU,, p E su:op, j E J,

Fqxiq-r~~fh=O iEHKi, ncSU,, q E su,b’, j E J,

f,,=F,x,, iczl, nESU,,

fh-FF,x,p-FF,x,q=O iEI, jeJ, ~cSU,, pe.SU~p, qeSU/bOL,

1 Frx, - 1 f, = 0 i E I, mc E MC,
IE.A-+ It M”,’
x,/l= 0 (i, k) E ND,, (Pl)
c Flxi,=CIP ieI, PEP,
,a M’“m
,
,&x,k = 1 k e SF,

Xi&> 0 ieZ, kE(SUjUSUyUSU/b”l),

(r i,“)’ < rz < (r::)” i E LKj, j e J,


(r~~)LGr~‘<(r~~)U iEHK,, jEJ.
Synthesisof general distillation sequences 641

shown below:

subject to
pB 2 L*(y; uk) k = 1,. . . , K’-,
L+(y;IL)<O /c=l,...,Kinfc~,

F,-@yj,(O ~EJ, kESU,,

Fn-Fp-F4=0 neSUj, JESUIT_ qESU,+‘,

c F,- c F,=O mcEMC, (Ml)


lcM$ IEW,’
F,=cC, PEP, iEI,
k%, : I

x_G,=F, IsSU,,
I
c- c J;,- 1 [l-(r~)r]f;“GO nosu,, IESU:OQ,
i eLYK, iEflK,
Fj - C A,, - C [ 1 - (r)S)L]X G 0 n E SU,, 1 E SU,Y,

i E Z, [EN, ~ESU,.

L* and L+ refer to Lagrange functions correspond- extensive communication of data between various
ing to feasible and infeasible primal problems, respect- subproblems. The procedure calls the appropriate
ively. uk are the Lagrange multipliers corresponding solvers, updates relevant parameters and checks the
to the k th feasible primal problem and II* are the stopping criteria. At each iteration, first the reiaxed
Lagrange multipliers for the relaxed primal problems. primal problem is solved by calling the required
pa is a scalar which the master problem tries to mini- solver. The data from the solution of this problem
mize subject to the constraints shown and represents (values of the noncomplicating variables and the
a lower bound on the global solution. As can be seen, Lagrange multipliers of the equations) is stored as
the master problem (Ml) is also linear, permitting an parameters. If the sum of infeasibilities is zero, the
easy solution of the subproblem. If the master prob- primal problem is solved and the values of the
lem gives a set of values of the complicating variables variables and multipliers are stored. This data is used
which does not result in a feasible solution of the to systematically generate and add new Lagrange
primal problem, then a relaxed primal problem is functions to the master problem: The MILP master
solved. In this subproblem, all constraints containing problem is then solved by using the appropriate
the complicating variables are relaxed using slack solver and the new values of the complicating
variables and the objective of this problem is to variables are stored as parameters to be used in
minimize the sum of infeasibilities. The information the primal problem in the next iteration. At each
provided by the solution of this subproblem is used iteration, the stopping criterion is checked. If the
to generate the Lagrange function to be added to the stopping criterion is met, then a final NLP problem
master problem. [which is the full problem (P) with the binary
(f) Automatic implementation of the global optimum variables fixed at values corresponding to the best
search procedure. APROS (Paules and Floudas. 1989) solution found] is solved and the procedure stops.
is a procedure for solving mathematical program- The initial point for this NLP problem is the solution
ming problems that require extensive communication corresponding to the current upper bound. Otherwise
of data between a set of subproblems whose size and it continues on to the next iteration.
structures may vary during the solution procedure. During the implementation of this procedure, it
An extension of this is NOUS (Not Universally was noticed that since the master problem has to
Sharp separation) which represents an algorithmic predict the values of a number of continuous vari-
procedure for the automatic implementation of the ables, it sometimes takes quite a few iterations for the
global optimum search procedure described in this infeasibilities to become zero. To improve on this, a
paper. It uses the procedural modeling language radial search feature is added to the algorithm. If the
GAMS (General Algebraic Modeling System, Brooke sum of infeasibilities becomes smaller than a specified
et al., 1988). value, the full problem (P) is solved as an NLP with
NOUS is a highly systematic procedure which is the values of the continuous complicating variables
interfaced with various nonlinear, linear and mixed- restricted to a small radius around the values given by
integer linear programming solvers and provides the previous master problem. The binary variables
642 A. AGGARWAL and C. A. FLOUDAS

ITASKI LK I HK 1

Fig. 4. Superstructure for Examples 1 and 2.

are kept fixed at values given by the previous master data is generated. The cost data is used to determine
problem. If this full problem is solved successfully, the coefficients for the objective function for the
the primal is solved with the complicating variables mathematical formulation.
fixed at the values provided by the solution of the Then the problem is formulated as an MINLP. The
full NLP problem. This is done to generate the MINLP model consists of 50 variables (including two
appropriate Lagrange multipliers. integer variables) and 40 constraints, out of which 25
The procedure NOUS is general enough to be constraints are nonlinear. The objective function of
applied to examples with any number of components the formulation is also nonlinear. The 50 variables
and any number of products for the general super- are as follows (refer to Fig. 4): 20 flowrates corre-
structure presented in this paper. For different sponding to 20 streams, six component flows (three
examples, only the various sets defining the topology for each column feed stream), 18 composition vari-
of the superstructures have to be modified. The set of ables (three each for feed, top and bottom products
equations remain the same, as does the procedure for for each column), four key component recoveries
implementing the global optimum search technique. (two for each column) and two binary variables (one
for each column). The complete MINLP model for
this example is shown in Appendix C. The global
EXAMPLE 1
optimum search approach proposed in this paper was
applied to this example from three different starting
In the first example, a three-component mixture points. As shown in Table 2, the optimal sequence
has to be separated into two three-component prod- found in all the three cases has a cost of $154,710.
ucts. Data for this problem is provided in Table 1. This sequence involves both columns I and II and is
In the first step, a superstructure is developed which shown in Fig. 5.
is shown in Fig. 4. Then a few shortcut simulations
are performed for each task to determine the lower
bounds on key component recoveries. The results Table 2. NOUS results Cm Example 1
indicated that to avoid the distribution of nonkey Starting cent Optimal cwt
components it was necessary to keep the recoveries of Starting sequence (SJ CS)
the key components greater than 0.85. The range of Sharp sequence. with columns 203.480 156,710
feed compositions for the two tasks is determined by I and II in series
Sharp sequence with columns 210,550 156,710
examining the superstructure as mentioned earlier. II and I in series
Next, a set of shortcut and rigorous simulations are Sharp sequence with columns 216,740 156,710
I and 11 in aarallel
performed for each of the two columns and the cost
Synthesis of general distillation sequences 643

Fig. 5. Optimal sequence using nonsharp separation for Example I.

A detailed enumeration of the various possible the key components for all columns at their upper
sequences from the superstructure is shown in bounds. The results for various sequences are com-
Table 3. For each sequence the cost shown is for the pared in Table 3. The sharp sequence corresponding
corresponding best solution and thus represents a to the best nonsharp sequence has a cost of $210,550.
local optimal solution. The best solution found by This sequence is shown in Fig. 6. In all cases, the
this method is the same as the one resulting from nonsharp sequence results in an annual saving of
the global search procedure. As a comparison, this 18-26%.
problem was solved for the case where the separation Examination of the two sequences depicted in Figs
is sharp. This was done by fixing the recoveries of 5 and 6 shows that there exists a tradeoff between the
amount bypassed and the sharpness of separation.
As the degree of separation reduces the bypass also
Table 3. Enumerationresultsfor Examde I reduces. The cost of a column increases with increase
Annualcost in total flowrate and decreases as the degree of
Nonsharp separation Sharp separation separation decreases. The results indicate that, for
sequence (W (S)
this example, the decrease in cost due to lesser degree
Columns I and II 165,210 203,480
in series (18.81%) of separation outweighs the increase in cost due to
Columns II and I 156.710 210.550 increased total flow to a column.
in series (25.57%) A comparison between the actual costs and costs
Columns I and II 167,670 216,740
in parallel (22.64%) calculated using the coefficients found by regression
The figures in parentheses are percentage savings by using nonsharp
analysis is shown in Table 4 for the various sequences
separationsequences. of Example 1. The results show that for all the three

Fig. 6. Sharp sequence corresponding to the optimal sequence for Example 1.


644 A. AGGARWAL and C. A. FLOUDAS

Table 4. Comparison of casts Table 6. NOUS results for Example 2

Annual cost Starting cost Optimal cost


Startingsequence (S) (S)
Calculated cost Actual cost
Sequence (f) 0) Sharp sequencewith columns 216,480 62,600
I and II in series
Columns I and II 165,210 159,720
Sharp sequence with columns 422,220 62,600
in series
II and I in series
Columns II and I 156,710 159,370 1.67
Sharp sequence with columns 210,140 62,600
in series
I and II in parallel
Columns I and II 167,670 172,190 2.63
in oarallel

Cost cnxfficicnts
Table 7. Enumeration results for Example 2
Coefficient Column I Column II
Annual cost
00 23,947.0 75,835.0
Nonsharp separation Sharp separation
01 - 1399.04 -6615.88
Sequence (S) (8
(12 935.14 338 I .47
a, 773.08 3733.49 Column I only 62,600 No feasible
b, -57.19 163.71 solution
b. 426.56 2889.96 Columns I and II 62,600 (1 only) 216,480
in series (71.08%)
Columns II and I 220.7 IO 422,220
in series (47.73%)
Table 5. Data for Example 2
Columns I and II 62,600 (I only) 210,140
Products(kgmol h-9 in parallel (70.21%)
Feed
The ligures in parentheses are percentage savings by using nonsharp
Comvonents (kgmol h-‘1 Pl P2
separation sequences.
Propane loo 80 20
iso-Butane loo 30 70
n-Butane loo 20 80 composition of desired products is different. From
Utilities available: three different starting points (corresponding to three
COOling water @! 305 K (20 K rise), cost = $0. 159/106 kJ sharp sequences), NOUS leads to a sequence involv-
Steam (a! 343 kPa, cost = S2.45/IOb kJ
Payback period = 2.5 yr ing column I only as the optimal sequence (with a
Tax correction factor = 0.52 cost of $62,600). The results are shown in Table 6.
The results of a detailed enumeration of various
sequences are shown in Table 7. The best sequence
sequences, the calculated costs are within 2-4% of found in this manner also involves column I only with
the actual costs. The cost coefficients for this example a cost of $62,600. This sequence is shown in Fig. 7.
are also given in Table 4. When this problem was solved for sharp separations,
the best sequence found had two columns in it-1 and
EXAMPLE 2
II, with an annual cost of !§210,140. This sequence is
shown in Fig. 8. So by using nonsharp separation, a
In this example, the same feed mixture as in very large reduction (70%) in annual cost results. For
Example 1 has to be separated into a different set of the other two sequences the savings are also large-
products. The data is given in Table 5. The super- 7 1% and 48% as shown by Table 7.
structure, simulation data and mathematical formu- The results show that by allowing for the use of
lation is the same as in Example 1 except that the nonsharp separators and using the methodology

_
LK HK

I(o.851

211.765
Feed

88.235 *

Fig. 7. Optimal sequence using nonsharp separation for Example 2.


Synthesis of general distillation sequences 645

Fig. 8. Optimal sequence using sharp separation for Exampfe 2.

proposed in this paper, a sequence was found which applied to this problem from seven different starting
involves oniy one column as compared to iwo columns points. The results are shown in Table 10. In two
always required for a sharp separation sequence. cases, the algorithm found a sequence costing
$298,520 as the optimal sequence. This sequence
EXAMPLE 3
involves columns I and III in parallel and both of
them in parallel/series with column II and is shown
In this example, a four-component mixture has in Fig. 9. Note that this sequence is slightly better
to be separated into two multicomponent products. than the best one found during a detailed enumer-
Data for this problem is given in Table 8. The super- ation of all possibilities and features uncommon
structure of this problem is the same as the one shown splitting and mixing. Such a sequence would be very
in Fig. 3. Simulations, costing and regression analysis difficult to predict by approaches utilizing heuristics.
is performed as mentioned earlier of each column to In the other five cases the aIgorithm found the
obtain the coefficients of the objective function of the sequence with a cost of $299,120 as the optimal
mathematical formulation. In this case, the (MINLP) sequence. This solution costs only 0.2% more than
formulation has three integer variables and 86 con-
tinuous variables. The total number of constraints is
Table 9. Enumeration results for Example 3
71 out of which 46 are nonlinear.
Annual cost
Table 9 shows the results for the sequences found -
during a detailed enumeration of various sequences. Nonsharp Sharp
separation separation
The best sequence has an annual cost of $299,120. sequence G) @)
This sequence involves columns I and II in series and Columns I, II and III 3 18,580 436,460
in series/parallel with I. The results of Table 9 show in series (27.01%)

that the number of local optimal solutions increases Columns I, III and II 321.160 359,340
in series (10.63%)
rapidly with the number of columns and these sol- Columns III, 1 and II 327,160 355,730
utions differ widely in structure and cost (variations in series (8.03%)
Columns III, II and I 322,720 403,180
of up to 25% in the cost of nonsharp sequences). in series (19.96%)
The proposed global optimum search technique was Columns I, II and III 323,270 363,880
in parallel (I 1.16%)
Columns 1 followed by II and III 339,680 392,040
in parallel (13.36%)
Table 8. Data for Example 3
Columns II followed by I and III 329.600 470,380
Products (kgmol h- ‘) in parallel (29.93 %)
Feed _ Columns III followed by I and II 368,380 401,520
CO”lpO”e”ls (kgmol h- ‘) Pl P2 in parallel (8.25%)
Columns I and II in series, 320,600 350,760
Propane 150 50 loo
and in paratlel with III (8.6%)
iso-Butane 200 100 100
Columns III and II in series, 323,440 353.1 IO
n-Butane loo 40 60
and in parallel with I (8.4%)
iso-Pcntane 150 100 50
Columns I and II in series, 299,120 332,740
Utilities available: and in series/parallel with III (10.1%)
Cooling water (a 305 K (20 K rise), cw.t = %0.15Y/106 kJ Columns III and II in series, 316,110 345,810
Steam ia 446 kPa, cost = %2.807/106 kJ and in series/parallel with I (8.59%)
Payback Period = 2.5 yr
The figures in parentheses are percentage savings by using nonsharp
Tax correction factor = 0.52
separation sequences.
646 A. AGGARWAL and C. A. FLOUDW

Table 10. NOUS results for Example 3 Table 1I. Data for Exam& 4
Starting cost Optimal cost Products (kgmol h-‘)
Starting sequence (0)
Compo”e”ts (k$?h - ‘) PI P2
Sharp sequence with columns 436,460 299,120
I, II and III in series Propane I.50 75 75
Sharp sequence with columns 355.730 298,520 iso-BUta”e 200 LOO 100
III, I and II in series n-Butane 100 40 60
Sharp sequence with columns 363,880 299,I 20 iso-Pentane t50 100 50
I, II and III in parallel
Utilities available:
Sharp sequence with columns 350,760 299,120
Cooling water @ 305 K (20 K rise), cost = $0.159/10*kJ
I and II in series and in
Steam @ 446 kPa, cost = S2.807/I@ kJ
parallel with III
Payback period = 2.5 yr
Sharp sequence with columns 353,110 299,120
Tax corr&tion factor= 0.52
III and II in series and in
parallel with I
Sharp sequence wth columns 332.740 299,120
I and II in series and in Table 12. NOUS results for Examale 4
series/parallel with III
Sharp sequence with columns 345,810 298,520
starting cost Optimal cost
Starting sequence (S) (S)
III and II in series and in
series/parallel with I Sharp sequence with column 377.430 274,410
I followed bv II and 111
in paralkl _
Sharp sequence with columns 366,790 274,410
the best known sequence of $298,520. The various II and III in series
Sharp sequence with columns 284,990 268,770
sequences in Table 9 were also solved for the sharp III and II in series
separation case and the results show that the best Sharp sequmce with columns 295,770 257,900
sequence has an annual cost of $332,740. So the best II and III in parallel
Sharp sequence with columns 277.750 257,900
nonsharp separation sequence is cheaper by 10%. II and rI1 in series/parallel
In all cases the nonsharp sequences result in annual
savings of g-30%.
sequence with a total annual cost of $257,900 as the
EXAMPLE 4
optimal one. This is the same as the best sequence
found by enumerating various sequences as shown
In this example, the same feed mixture as in in Table 13. This sequence consists of columns II and
Example 3 has to be separated into a different set III only and is shown in Fig. 10. The cost of this
of products. The data is given in Table 11. The optimal sequence is 7.15% less than the correspond-
superstructure. simulation data and mathematical ing best sharp sequence which has a cost of $277,750.
formulation is the same as in Example 3. The results So again, allowing nonsharp separations results in
of the global optimum search procedure are listed annual savings of 7-27%. The results in Table 13
in Table 12. From two different starting points, again show a large number of very different local
the global optimum search approach resulted in a solutions.

+ 43.214
.__!_. // .--..--

Fig. 9. Optimal sequence using nonsharp separation for Example 3.


Synthesis of general distillation sequences 647

Table 13. Enumeration results for Example 4 Table 14. Data for Example 5

cost Products (kgmol h- ‘)


(FE,
Nonsharp Sharp
Components h-1) PI P2 P3 P4
separation separation
Sequence (S) (8 n-Butane 25 15 IO -
n -Pentane 25 12.5 12.5 - -
Columns I, II and I11 307,340 42 I.840 n -Hexane 25 5 - 20 -
in series (27.14%)
rt-Heptane 25 - - IO I5
Columns I, III and II 303,200 344,720
in series (12.05%) Utilities available:
Columns III, I and II 296.270 321.750 Propylcnc @ 4.4”C. cost - $2.765/I@ kJ
in series (7.92%) LP steam @ 115°C. wst = $5.414/IO* kJ
Columns I followed by II and III 312,710 377,430 MP steam I& 130.6”C, cost = S7,3%/106 kJ
in parallel (17.15%) HP steam @ 185.6”C, cost = SlO.O36/106 kJ
Columns I and 11 in series, 288,130 343,920 Payback period = 2.5 yr
and III in series/parallel with II (16.22%) Tax correction factor E 0.52
Columns II and III 313,830 366,790
in series (14.44%)
Columns 111 and II 268,770 284,990
in series (5.69%) determined to be 0.8. The problem is formulated as
Columns II and 111 273,760 295,770
in parallel (7.44%)
a MINLP. This formulation has 99 variables (includ-
Coiumns II and III 257,900 277,750 ing three integer variables), 54 nonlinear constraints
in series/parallel (7.15%) and 25 linear constraints.
The figures in parentheses are percentage savings by using nonsharp NOUS was used for this example from five
separation sequences.
different starting points. For two cases, the optimal
solution found has a cost of $256,010. The other three
EXAMPLE 5
cases found a sequence with a cost of S260,430. These
results are shown in Table 15. The best sequence
This example is taken from Cheng and Liu (1988). found by studying the various sequences has an
A four-component feed stream is to be separated into annual cost of $254,570 and is shown in Fig. 12.
four products. None of the products contain all the So the global optimum search approach found a
four components. Product P4 is a pure component solution which is only 0.6% more than the best
product. The data for this example is given in Table known solution from two different starting points.
14. The superstructure for this example is shown in The solution obtained from the other three starting
Fig. 11. One important point which should be noted points is 2.3% more than the best solution available.
is that this superstructure does not contain overall The results of the enumeration are shown in Table 16.
bypass streams from the feed to any of the products. In this case the best sharp sequence is column II
This is because none of the products contain all the followed by columns I and III in parallel with an
four components and hence no such overall bypass annual cost of $261,020. As can be seen from Table
can exist (see Appendix A). Shortcut and rigorous 16, the cost difference between the sharp and non-
simulations are performed to generate the required sharp cases for this sequence is very small. The reason
cost data for developing the cost functions. The lower for this is the particular set of product compositions.
bound on the recoveries of the key components is The recoveries for columns II and III come out to be

feed 600 _ 130.104 157.751

Fig. 10. Optimal sequence using nonsharp separation for Example 4.


648 A. AGGARWAL and C. A. FLOUDAS

Pl
Pz
P3
P4

Pl

P2

P3
Pl P4

ITASKI LK I HK I P2

P3
P4

Fig. 11. Superstructure


for Example 5.

unity for both sharp and nonsharp sequences and different economic assumptions and different costing
hence the only savings can be by having nonsharp procedures.
separation in column I. So depending on the flowrate
to column I the cost differences can be large (17.6%
for the first sequence in Table 16) or very small (0.2% NONADJACENT KEY COMPONENTS
for the last sequence). This example illustrates that
the proposed approach can also handle cases where The first step in the removal of the assumption
the product set consists of pure components, thereby of adjacent key components is the introduction of
requiring combinations of sharp and nonsharp columns with nonadjacent key components in the
separators. superstructure. Such a superstructure, for a three-
For the same problem, the three sequences pro- component mixture to be separated into two multi-
posed by Cheng and Liu (1988) have four separators component products, is shown in Fig. 13. In a
each, which is one more than the number of separ- three-component system, there can be one column
ators for the various sequences studied in this work. with nonadjacent key components: column III.
The general superstructure proposed in this paper A and C would be the key components for this
ensures that for an n-component mixture, a maxi- column. The only feed to this column comes from the
mum of n - 1 sharp and/or nonsharp separators initial mixture. There is no stream from the bottom
would be needed (under the assumption of consecu- of column I coming as feed to column III. This is
tive key components). A comparison of the
cost of various sequences with the sequences found
by other investigators is more difficult because of Table 16. Enumeration results for Example 5

Annual cost
Nonsharp separation Sharp separation
Table 15. NOUS results for Example 5
Sequena (S) 0)
Starting cost Optimal cost
Columns I, II and III 254.570 308.8 10
Starting sequence (0) w in series (17.56%)
Sharp sequence with columns 308,810 260,430 Columns I, III and II 312,960 367,270
I, II and III in series in series (14.79%)
Sharp sequence with columns 367,270 256,010 Columns III, I and II 352.430 385,270
I, III and II in series in series (8.52%)
Sharp sequence with columns 385,270 260.430 Columns 111, II and I 358,820 359,410
III. I and II in series in series (0.16%)
Sharp sequence with columns 359,410 256,010 Columns II followed by 260,430 26 1,020
III, II and I in series I and III in parallel (0.23%)
Sharp sequence with column II 26 1.020 260,430
The Rgures in parentheses are percentage savings by using nonsharp
followed by I and III in parallel
separation sequences.
Synthesisof generaldistillation sequences

32.5

Fig. 12. Optimal sequenceusing nonsharpseparationfor Example 5.

because stream 10 would contain little or no A components by adding columns with nonadjacent
(depending on the degree of sharpness of separation key components.
in column I) and A is a key component for column To complete the mathematical formulation, one
III. A similar reasoning applies to stream 15. The additional item which would be required is an
top product of column III will have components A expression for the distribution of the intermediate
and B but little or no C and hence there is no stream component as the recoveries of the key components
going from stream 22 to column II (which has C are varied. Data from the simulations performed for
as one of the key components). This sort of a generating the cost expression for column III could
superstructure can be generalized for any number of be used to find such an expression.

1 6

A
a I
23 27

Fig. 13. Superstructure


for a three-componentsystem includingnonadjacentkey components.
A. AG~ARWALand C. A. F~ou~as

45.261

110

190

Fig. 14. Best sequence involving column III for Example 6.

EXAMPLE 6 Here, xB. 22 is the composition of component B in


the top product of column III (Fig. 13). This equa-
The above procedure was applied to the three- tion is added as a constraint to the mathematical
component system of Example 1. The superstructure formulation.
would be as described above and is shown in Fig. 13. The best sequence found, which involves column
For column III (with A and C as key components), III, has a cost of $167,540. This sequence consists of
the lower bound on the recoveries of the key columns II and III in a series-parallel arrangement
components is determined by the point below which and is shown in Fig. 14. But the results in Table 17
separation with these key components becomes show that the globally optimal sequence is still the
infeasible. For this example, shortcut simulations one with only columns I and II found in Example 1.
indicated that below recoveries of 0.9, the cost of So the addition of the possibility of columns with
separation starts increasing sharply and soon be- nonadjacent key components did not change the
comes infeasible. Rigorous simulations and costing is globally optimal sequence for the three-component
performed for column III to generate the appropriate system of Example 1.
cost expression. Since the feed composition for this
column is fixed (as explained in the previous section),
the distribution of the intermediate component (com- CONCLUSIONS

ponent B) depends only on the recoveries of the key The problem of synthesizing distillation sequences
components. Using the information available from involving nonsharp separators for separating a multi-
the rigorous simulations, an expression of the follow- component feed stream into several multicomponent
ing form was found to represent the distribution of product streams has been considered. A synthesis
component B within acceptable errors: approach has been proposed that is based upon a
superstructure that contains sharp and nonsharp
separation options for all columns, all possible
sequences and all alternatives for stream splitting,
Table 17. Enumeration results for Example 6
mixing and bypassing. Shortcut simulations for
Annual cost
each column in the superstructure are performed to
Nonsharp separation Sharp separation develop cost models to represent the total annual
(5) G)
column cost in the mathematical formulation. The
Columns I and 11 165,210 203,480
in series (18.81%)
mathematical formulation is an MINLP and its
Columns II and I 156.710 210.550 solution results in an optimal separation sequence for
in s&es (25.57%) the required problem.
Columns I and II 167.670 216,740
in parallel (22.64%)
A new global optimum search approach, based
Columns III and II 167,540 203,030 on a decomposition of the nonconvex MINLP into
in series/parallel (0.16%)
convex subproblems, is used to search for the global
Column 111 followed by 185,490
I and II in mwallel solution of the MINLP formulation and good results
The figures in parentheses are percentage savings by using nonsharp have been obtained for the various examples pre-
separation sequences. sented in this paper. The technique, implemented in
Synthesis of general distillation sequences 651

NOUS, found the global solution for four of the x,~ = Composition of component i in stream 1
five examples from several starting points and for v, = Binary variable denoting the existence or nonexist-
ence of column j
Example 5 found a solution within 0.6% of the
best known solution. The extension of the procedure Greek symbols
to include columns with nonconsecutive key compo- a = Payback period
nents is illustrated through an example problem. @ = Tax correction factor
Future work consists of incorporating heat
integration options and pressure optimization in the
REFERENCES
framework proposed in this paper (Aggarwal and
Floudas, 1990). Incorporation of complex columns in Aggarwal A. and C. A. Floudas, Synthesis of heat inte-
the approach at a later stage is also aimed at so as to grated nonsharp distillation sequences. In preparation
make it more widely applicable. (1990).
Andrecovich M. J. and A. W. Westerberg, An MILP
formulation for heat intenrated distillation seauence
. svn-
Acknowledgement-Financial support from the National
thesis. AIChE Jf 31, 1461 (1985).
Science Foundation under Grants CBT/8612073 and CBT/
Bamopoulos G., R. Nath and R. L. Motard, Heuristic
8820990 is gratefully acknowledged.
synthesis of nonsharp separation sequences. AIChE 3134,
763 (1988).
Benders J. F., Partitioning procedures for solving mixed
NOMENCLATURE variables programming problems. Numer. Mafh. 4, 238
(1962).
ao, = Constant cost coefficient for column j BrbokeA., D. Kendrick and A. Meeraus, GAMS: A Users
a,, = Cost coefficient corresponding to flowrate to Guide. Scientific Press, Palo Alto, Calif. (1988).
column j Cheng S. H. and Y. A. Liu, Studies in chemical process
au = Cost coefficient corresponding to light key recovery design and synthesis: 8. A simple heuristic method for
in column j systematic synthesis of initial sequences for sloppy multi-
ar, = Cost coefficient corresponding to heavy key recov- component separations. Ind. Engng Chem. Res. 27, 2304
ery in column j (1988).
b,, = Cost coefficient corresponding to component i for Duran M. A. and I. E. Grossmann, An outer-approxi-
column j mation algorithm for a class of mixed-integer nonlinear
C,, = Quantity of component i in final product p programs. Math. Program. 36, 307 (1986).
J;, = Flowrate of component i in stream I Floudas C. A., Separation synthesis of multicomponent feed
F, = Total flowrate of stream 1 streams into multicomponent product streams. AIChE _?I
HK, = Index set of heavy key component for column j 33, 540 (1987).
HL&, = Index set of components heavier than the light key Floudas C. A. and S. H. Anastasiadis. Synthesis of distil-
for column j lation sequences with several multicor&onent feed and
I = Index set of components product streams. Chem. Engng Sci. 43, 2407 (1988).
J = Index set of columns Floudas C. A. and G. E. Paules, A mixed-integer nonlinear
LK, = Index set of light key component for column j programming formulation for the synthesis-of heat-inte-
LHK, = Index set of components lighter than the heavy key grated distillation sequences. Computers them. Engng 12,
for column j 531 (1988).
MC = Index set of mixers prior to each column Floudas C. A., A. Aggarwal and A. R. Ciric, Global opti-
&ff = Index set of mixers prior to each final product mum search for nonconvex NLP and MINLP problems.
Mz, = Index set of inlet streams to mixers prior to each Compurers &em. Engng 13, 1117 (1989).
column Geoffrion A. M., Generalized Benders decomposition.
MZ = Index set of outlet streams to mixers prior to each J. Optimization Theory Applic. 10, 237 (1972).
column Guthrie K. M., Capital cost estimating. Chem. Engng Mu
ME, = Index set of inlet streams to mixers prior to each 24, 114 (1969).
final product King C. J., Sepnrofiun Processes. McGraw-Hill, New York
N = Index set of all streams (1971).
ND,, = Index set of combinations for component i not Kocis G. R. and I. E. Grossmann, Global optimization
present in stream k of nonconvex MINLP uroblems in orocess svnthesis.
P = Index set of final products Ind. Engng Chem. Res. i7, 1407 (1988). _
r]r = Fraction of component i going out in top product Mahalec V. and R. L. Motard, Procedures for the initial
of column j, i E LK, design of chemical processing systems. Compulers them.
rbk = Fraction of component i going out in bottom Engng 1, 57 (1977a).
product of column j, i E HK, Mahalec V. and R. L. Motard, Evolutionary search for
(r):)L = Lower bound on fractional recovery of component an optimal limiting process flowsheet. Computers them.
i in top product of column j, i ELK,. Engng 1, 149 (1977b).
(rfik)‘ = Lower bound on fractional recovery of component Muraki M. and T. Hayakawa, Separation processes synthe-
i in bottom product of column j; i E HK, sis for multicomponent products. J. Chem. Engng Japan
(r’,:)U = Upper bound on fractional recovery of component 17, 533 (1984).
i in top product of column j; i ELK, Muraki M. and T. Hayakawa, Multicomponent separation
(ry)l’ = Upper bound on fractional recovery of component process synthesis with separation sharpness. J. Chem.
i in bottom product of column j, i E HK, Engng Japan 20, 195 (1987).
S = Index set of splitters Muraki M. and T. Hayakawa, Synthesis of a multicom-
S:” = Index set of inlet streams to splitters ponent multiproduct separation process with nonsharp
Sy’ = Index set of outlet streams from splitters separators. Chem. Engng Sci. 43, 259 (1988).
SW, = Index set of inlet streams to column j Muraki M., K. Kataoka and T. Hayakawa, Evolutionary
SZJ;Or= Index set of top product from column j synthesis of multicomponent multiproduct separation
SlJy = Index set of bottom product from column j processes. Chem. Engng Sci. 41, 1843 (1986).
652 A. AGGARWAL and C. A. FLOUDAS

Nath R., Studies in the synthesis of separation processes. be added to the master problem. Since the total flowrates of
Ph.D. Dissertation, University of Houston, Texas (1977). all the streams are variables for the master problem, overall
Nishida N., G. Stcphanopoulos and A. W. Westerberg, mass balances can be written for mixers and columns.
A review of process synthesis. AIChE JI 17. 321 (1981).
Paules G. E. IV and C. A. Floudas, APROS: algorithmic Balances for each column j:
development methodology for discrete-continuous opti-
mization problems. Opers Res. 37, 902 (1989). F, - Fp - F, = 0 n E SU,, p E SUFP, q E SU,-.
Peters M. S. and K. D. Timmerhaus, Plant Design and
Economics for Chemical Engineers. McGraw-Hill, New Balances for mixers prior to each column:
York (1980).
Rathore R. N. S., K. A. Van Wormer and G. J. Powers,
Synthesis of distillation systems with energy integration.
AIChE Jf 20, 940 ( 1974). Mass balances for final mixers:
Simulation Sciences Inc., PROCESS Simadafion Program
Input Manual. Fullerton, Calif. (1985).
Treybal R. E., Mass-Transfer Operariorts. McGraw-Hill,
New York (1980).
where C, is the desired quantity of component i in the final
Wehe R. R. and A. W. Westerberg, An algorithmic pro-
product p.
cedure for the synthesis of distillation sequences with
The component flowrates to the columns are also compli-
bypass. Computers &em. Engng 11, 619 (1987).
cating variables. So the following constraints can be added
Westerberg A. W., The synthesis of distillation based separ-
for each column j :
ation sequences. Computers Gem. Engng 9, 421 (1985).

APPENDIX A
Another set of constraints which can be written is de-
In this appendix a procedure will be presented for calcu- signed to prevent the master problem from assigning all the
lating upper bounds for the flowrates of the overall bypass feed going to a column to the top or bottom product only.
streams. We define the following symbols: Two constraints are written for each column. The maximum
F = total feed flowrate, amount of top product can be found by adding:
xi = mole fraction of component i in feed,
(i) for components lighter than the light key-the total
bj = amount bypassed to product j,
amounts in the feed stream;
pu = amount of component i in product j.
(ii) for the light key component--all the entering
The overall bypass streams will have the same composition amount;
as the feed stream. So bixj represents the amount of com- (iii) for the heavy key component-a fraction [I - (r?)“]
ponent i bypassed to productj. This has to be less than the of the entering amount;
total amount of component i in product j, that is:
where (rp)’ is the lower bound on the -very of the heavy
for product j b,x, Gpij for all i.
key component for column j as determined during simula-
So the maximum b, which can still satisfy the above tions. Similarly the upper bound on the bottom flowrate can
condition would be: be determined. Consider column II in the superstructure in
b,=minP”. Fig. 3. B is the light key component and C is the heavy key
I component. Suppose that 0.8 is the lower bound on the light
’ *,
key component recovery as determined during shortcut
This relation can be used to calculate the maximum over-
simulations. Then the top product can have, at most, all of
all bypasses for any superstructure. Let us illustrate this with component A entering the column, all of B and 20% of C.
an example. Consider the system in Example 3. The data So for each column j:
for this problem is given in Table 6. Let b, and b, represent
the maximum amounts which can be bypassed to the two
products. From the given data x, = (0.25,0.333,0.167,0.25):

b Emin%
i xi
50 100 40 100
=mm ( 0.25’0.333’0.167’0.25 >
LHKj = {i 1i E I is lighter than the heavy key for column j )
= 200.
j E J,
Similarly:
IO0 60 50 HLKj = {i 1i E I is heavier than the tight key for column j }
bz= min 100 j E J.
( 0.25 ’ 0.333 ’ 0.167 ’ 0.25 >
= 200.
APPENDIX C
One obvious but important consequence of this property
is that if a product stream does not contain all the compo-
MINLP Formulation for Example 1
nents (i.e. some pii are zero) then the overall bypass to that
product will be zero. This is illustrated in Example 5. (Fig. 4)

mino,,yt+{a,.,+a2.1r~.,+a,.,r~,+b,.,x~.,
APPENDIX B
+b~.,+,,t F,+=o,zY,+ {~+a,&&
To provide more information to the marker problem as an
aid in getting feasible solutions, a number of constraints can + a3.&!12 + bA.2XA.13 + b~,s~.I~I Fz;
Synthesis of general distillation sequences 653

subject to f ~,,a - FM+,, - F I*x B.48- 0,


F, + F2 + F3 + F4 = 300, f c. 13-Fx I. CM - F 18x CL8- 0,
Fe- F,-F8 = 0, 0.333F, + F,+A.,. -fA.5 = 0.

F,--FF,,-Fj,- F,2=O, 0.333F, + F,,-G.,. -.A& = 0,

F,,-FF,5- F,,-FF,,= 0, 0.3331$ + F,,-+.,. -_&.s - 0.

F,,- F,,-Fm= 0, 0.3334 + &,x,,.Y -f~, 1, = 0,


0.3334 + Fmxs., -fB, 13 = 0,
FA,, -r!i.,_fi5
=a
0.333F2 + F,,xc., -fc.,a = 0,
F,.+a,,,-4Lfi,~ = 0,
xc. L6= 0.
~9-%,9- 4:,r,,,
= 0.
X&I. = 09
FM+. IS- +!!zfc. ,3 = 0,
0.333F, + WA,~ + F,,x,,, + F,,.Q,,, + F,,x,,.,s = 30,
fh.5 - 4~~s = 0.
0.333F, + F,xs,, -t F,,xB,, + F,,x,,.,, + F,,x,,.,, = 30,
_&is- F,xn,, = 0. 0.333F3 + F 7x C,B+ F~,RP + F,,-Q.M+ Fwxc.,s= 30.
fc,, - F+c., = 0, x.4,5 + %5 + xc.5 = 1.

f A.13-Fx 13 A.13-0,
- xA.6 + %,6 + xc.6 = 1,

f8.~3
- F,A,.,,= 0, XA.9 + G.9 + xc.9 = 1l

f&.,x - F 13-Q.I3 -- 0, XA.13 + -%,I3 + -%I3 = 1.

fA.5 - F,-G., - 4x,., = 0. XA.14 + Jh.14 + xc.14 = 1,

fe,s - F,x,,, -F&,9 = 0. XA.18 + -%3.lS +xc.m = 1.

fc.s - F~xc,~- Fv+.P = 0, F5 - 3oOy, 4 0,

fA.1, - FMXA.W- F,,x,.,, = 0, F,,- 300~~ G 0.

You might also like