Compurersc&m. Engng, Vol. 14, No. 10, pp. 1165-1184, 1990 009%1354/90 $3.00 + 0.

00
Printed in Great Britain. All rights reserved Copyright 0 I990 Pergamon Press plc

SIMULTANEOUS OPTIMIZATION MODELS FOR

HEAT INTEGRATION-II. HEAT EXCHANGER
NETWORK SYNTHESIS
T. F. YEE and I. E. GnossMaNNt
Department of Chemical Engineering, Carnegie Mellon University, Pittsburgh, PA 15213, U.S.A.

(Received 28 June 1990; received for publication 6 July 1990)

Abstract-In this paper, a mixed integer nonlinear programming (MINLP) model is presented which can
generate networks where utility cost, exchanger areas and selection of matches are optimized simul-
taneously. The proposed model does not rely on the assumption of fixed temperature approaches (HRAT
or EMAT), nor on the prediction of the pinch point for the partitioning into subnetworks. The model
is based on the stage-wise representation introduced in Part I of this series of papers, where within each
stage, potential exchanges between each hot and cold stream can occur. The simplifying assumption on
isothermal mixing to calculate heat transfer area for stream splits allows the feasible space to be defined
by a set of linear constraints. As a result, the model is robust and can be solved with relative ease.
Constraints on the network design that simplify its structure, e.g. no stream splits, forbidden matches,
required and restricted matches as well as the handling of multiple utilities can bc easily included in the
model. In addition, the model can consider matches between pairs of hot streams or pairs of cold streams,
as well as variable inlet and outlet temperatures. Several examples are presented to illustrate the
capabilities of the proposed simultaneous synthesis model. The results show that in many cases, heuristic
rules such as subnetwork partitioning, no placement of exchangers across the pinch, number of units, fail
to hold when the optimization is performed simultaneously.

lNTRODUfflON the solution of an NLP model (Floudas er al., 1986)

to determine the optimal network configuration.
Most of the current synthesis methods for heat
The limitation of a sequential synthesis method is
exchanger networks rely on sequential or step-wise
that different costs associated with the design cannot
procedures (Gundersen and Naess, 1988). In general,
be optimized simultaneously. In other words, trade-
the design problem is decomposed in order to pro-
offs between the different costs, as shown in Fig. 1,
gressively determine targets for synthesizing a net-
cannot be accounted for accurately. In general, early
work. For example, the pinch design method by
decisions on HRAT (the level of energy recovery to
Linnhoff and Hindmarsh (1983) first uses a cost
be achieved by the network) and whether or not to
target to establish a minimum energy consumption,
partition the problem into subnetworks can have
thus fixing the utility requirement for the network
negative effects on the latter decisions of number of
and the pinch location. The problem is then parti-
units and area requirement for the network configur-
tioned into subnetworks disallowing exchangers to be
ation. Sequential design methods, as a result, can
placed across the pinch. Finally, each subnetwork is
often lead to suboptimal networks.
evolved using guidelines and heuristics to synthesize
Floudas and Chic (1989) developed an MINLP
networks with minimum number of units.
model to simultaneously optimize the selection of
Another example is the mathematical program-
process stream matches and the network configur-
ming approach built into the interactive program
ation for a fixed level of energy recovery (HRAT).
MAGNETS (Floudas et al., 1986). The design prob-
The formulation is based on a hyperstructure, which
lem is decomposed into three steps. The first two steps
is similar to the superstructure of Floudas et al.
involve the solution of the LP and MILP transship-
(1986) that embeds all the possible matches. Optimiz-
ment model of Papoulias and Grossmann (1983). For
ation of the model identifies which of the embedded
a particular HRAT value, the LP model determines
matches are needed to minimize the total investment
the minimum utility requirement for the network.
cost of the exchangers.
With the utility consumption fixed at the LP solution,
Dolan et al. (1987, 1989) and Yee and Grossmann
the MILP model is solved to determine the minimum
(1988) proposed methods to account for all types of
number of matches and their corresponding heat
costs simultaneously. Dolan et al. proposed the
loads. Finally, in the third step, heat loads and
matches are fixed and the area cost is minimized by method of simulated annealing as a synthesis tech-
nique, whereas Yee and Grossmann formulated an
extensive MINLP model for retrofit design where the
iAuthor to whom all correspondence should be addressed. piping layout is also considered. In both approaches,

1165
1166 T. F. YEE and I. E. GROSSMANN

0
heated. Specified are also each hot and cold stream’s
heat capacity flow rates and the initial and target
Utility
temperatures stated as either exact values or inequal-
ities. Given also are a set of hot utilities HU and a set
of cold utilities CU and their corresponding tempera-
tures. The objective then is to determine the heat
f t exchanger network which exhibits the least annual

Q+-+@ cost. The solution defines the network by providing

the following:
1. Utilities required.
Fig. 1. Trade-off between costs in design. 2. Stream matches and the number of units.
3. Heat loads and operating temperatures of each
operating cost and capital cost are considered simul- exchanger.
taneously in the search of a least-cost network. 4. Network configuration and flows for all
Furthermore, one does not have to decide whether branches.
subnetworks must be partitioned or not, nor does one 5. Area of each exchanger.
have to specify fixed temperature approaches
(EMAT). Specifically, as shown in Fig. 1, trade-offs As will be shown, constraints on stream matches,
between utility cost, fixed charges for the number of stream splits and number of units can also be
units, and heat transfer area cost are determined specified. In the proposed method, no parameters are
simultaneously. The difficulty, however, as shown by required to be fixed; i.e. level of energy recovery
the results of the two methods, is that it is not trivia1 (HBAT), minimum approach temperature (EMAT),
to establish efficient computational schemes when number of units and matches. Also, there is no need
accounting for all the trade-offs. In the case of the to perform partitioning into subnetworks, and the
simulated annealing method by Dolan et al., a very pinch point location(s) are not pre-determined but
large number of trials is required, while the MINLP rather optimized simultaneously.
mode1 by Yee and Grossmann is very large in size and
has a poor relaxation. REMARKS ON SUPERSTRUCTURE
In this paper, the superstructure representation
proposed in Part I of this series of papers (Yee et al., The proposed strategy involves the development of
1990) is applied to a synthesis model which accounts a stage-wise superstructure and its modeling and
for all the costs simultaneously yet requiring very solution as an MINLP problem to obtain a cost-
reasonable solution times. The costs include fixed optimal network. The reader is referred to Part I of
charges and area costs for the exchangers, and the this series of papers for the detailed discussion of the
cost for the utilities. Based on the simplifying as- superstructure.
sumption for isothermal mixing of streams, it is A superstructure involving two hot and two cold
shown that the problem can be formulated as an streams along with hot and cold utilities is shown in
MINLP which has the desirable feature that all the Fig. 2. Similar assumptions as those used in Part I are
constraints are linear. The solution scheme deter- imposed. The number of stages in the superstructure
mines the network which exhibits least annual cost by can be set for instance to the maximum number of
optimizing simultaneously for utility requirement hot or cold streams, i.e. max{N,, Nc}. Also, as
(HBAT), minimum approach temperature (EMAT), shown in Fig. 2, although in principle, each utility
the number of units, the number of splits and heat stream can be treated as a process stream with
transfer area. Constraints on matches, on the number unknown flow rate, it will be assumed for simplicity
of units and on stream splitting that simplify the that utility streams are placed at the extreme ends of
network structure can be easily incorporated into the the sequence of stages. Finally, the assumption of
model, as well as the specification of inlet or outlet isothermal mixing of streams is imposed. As dis-
temperatures as inequalities. Furthermore, the model cussed in Part I, this simplification eliminates the
can consider the possibilities of matching pairs of requirement for nonlinear heat balances around each
streams of the same type, i.e. hot-to-hot and cold-to- exchanger as well as nonlinear heat mixing equations.
cold as previously proposed by Grimes er al. (1982), Instead, only an overall heat balance around each
Viswanathan and Evans (1987) and Dolan et al. stage is needed. As a result, flow variables are no
(1987). longer required in the model and the model size is
reduced. More importantly, though, the feasible
PROBLEM STATEMENT
space of the problem can now be defined by strictly
linear constraints. The only nonlinearities appear in
The HEN synthesis problem addressed in this the objective function which invovles the cost terms
paper can be stated as follows: for the heat exchanger areas, which are expressed in
Given are a set of hot process streams HP to be terms of stage temperatures. Therefore, the MINLP
cooled and a set of cold process streams CP to be model is robust and can be solved very efticiently.
 Simultaneous optimization models for heat integration-II 1167

Stage 2 I
I

localIon
kxl k-2 k-3

Fig. 2. Heat exchanger network superstructure.

It should be noted that the simplifying assumption flows and temperatures, and solved to determine
for isothermal mixing is rigorous for the case when optimal split flow rates and area distribution for the
the network to be synthesized does not involve stream exchangers. The solution of the suboptimization is
splits. For structures where splits are present, how- then considered as the final cost-optimal network.
ever, the assumption may lead to an overestimation Finally, it should be noted that there are certain
of the area cost since it will restrict trade-offs of area alternatives in the network cotiguration which the
between the exchangers involved with split streams. proposed superstructure neglects. Specifically, the
Also, in some cases this assumption might exclude superstructure does not account for the case of a split
network structures which are only feasible with non- stream going through two or more exchangers in
isothermal mixing. In order to partially overcome this series and the case of stream bypasses. For clarifica-
limitation, the scheme shown in Fig. 3 is proposed. tion, these structures are shown in Fig. 4. In general,
The idea is to use the MINLP model to determine an disregarding stream bypasses is not a significant
optimal structure. If this structure involves split limitation since these are usually not required and
streams, then an NLP suboptimization problem is more importantly, not favorable. In very particular
formulated with the fixed configuration and variable cases, however, the use of bypasses may help to
decrease the number of units, though at the expense
of requiring more area (see Wood et al., 1985).

a) A split stream going through exchangers in series

Solve
NLP optimization
for given
Configuration

Grassroots
Network
( obtained J b) A stream by-pass
Fig. 3. Proposed synthesis strategy. Fig. 4. Limitations of superstructure.
1168 T. F. YEE and I. E. GROSSMANN

The more important configuration which the su- F= heat capacity flow rate,
perstructure neglects is the case where a split stream U= overall heat transfer coefficient,
goes through several exchangers in series. In small CCU = per unit cost for cold utility,
examples where there is not much flexibility in select- CHU = per unit cost for hot utility,
ing structures, this limitation may cause the network CF = fixed charge for exchangers,
to require larger areas. However, for larger problems, C= area cost coefficient,
this restriction is less important since greater flexi- B= exponent for area cost,
bility in matching and selection of configuration can NOK = total number of stages,
usually ensure an equally good network without the R= an upper bound for heat exchange,
particular split structure. This in fact will be demon- F = an upper bound for temperature difference;
strated by an example later in the paper by synthesiz-
(iv) Variables:
ing networks which are as good if not better than
certain solutions in the literature where the reported dt,jk = temperature approach for match (i, j) at
temperature location k,
optimal network involves split streams going through
exchangers in series. dtcu, = temperature approach for the match of hot
stream i and cold utility,
dthu, = temperature approach for the match of
MODEL FORMULATION cold stream j and hot utility,
q,,k = heat exchanged between hot process stream
In this section, the formulation for the MINLP i and cold process stream j in stage k,
synthesis model subject to the simplifying assumption qcu, = heat exchanged between hot stream i and
for isothermal mixing of streams is presented. Binary cold utility,
variables are introduced to designate the existence of qhuj = heat exchanged between hot utility and
each potential heat exchanger in the superstructure, cold stream j,
and to model fixed cost charges for the exchangers. t,,k= temperature of hot stream i at hot end of
Continuous variables are assigned to temperatures stage k,
and heat loads. The general model involves overall t,,k = temperature of cold stream j at hot end of
heat balances for each stream, stream energy balances stage k,
at each stage, assignment of known stage tempera- z,,,, = binary variable to denote existence of
tures, calculation of hot and cold utility loads, logical match (i, j) in stage k,
constraints and calculation of approach tempera- zcui = binary variable to denote that cold utility
tures. The MINLP model is solved to minimize the exchanges heat with hot stream i,
total annual cost comprising of utility cost, fixed zhu, = binary variable to denote that hot utility
charges for each exchanger and heat transfer area exchanges heat with cold stream j,
cost.
For simplicity in the presentation, utility exchang- With these definitions, the formulation can now be
ers are placed at the outlet of the superstructure and presented. For completeness, the relevant equations,
only one type of hot and one type of cold utility are (l-5), which appeared in Part I (Yee et al., 1990) are
assumed. These two assumptions can be easily re- restated without duplicating the discussion. Ad-
laxed to accommodate cases of multiple utilities with ditional constraints for the synthesis model are then
various temperatures. presented along with the modified objective function.
In order to formulate the proposed MINLP model,
Overall heat balance for each stream
similar definitions as in Part I (Yee et al., 1990) are
necessary:
(TIN, - TOUT,)F, = 1 1 q,jk+ qcu,, i E HP,
(i) Indices: !iEST,‘CP
i = hot process or utility stream,
j = cold process or utility stream, (TOW- TIN,&= c 2 qijk+qhuj, jECP. (1)
k = index for stage 1, . . . , NOK and tempera- [Link]’iEHP
ture location 1, , NOK + I; It should be noted that for cases where the inlet or
target temperatures are defined by a range of values,
(ii) Sets:
the corresponding parameters can be substituted by
HP = {i ) i is a hot process stream},
variables in the constraints. The variables then would
CP = (j /j is a cold process stream},
be bounded to reflect the given range.
HU = hot utility,
CU = cold utility, Hear balance at each sfage
ST- {k ( k is a stage in the superstructure,
k = I, . , NOK}; (6.k - t,,k + I 14 = c (~i,kr LEST, iEHP,
JECP
(iii) Parameters:
TIN = inlet temperature of stream, (r,.k - r,.k+ I )4 = 1 4ijk, k 6 ST, i E CP. (2)
TOUT = outlet temperature of stream, lEHP
 Simultaneous optimization models for heat integration--II 1169

Assignment of superstructure inlet temperatures dtcu, d ti,*oK+, - TOUT,,

TIN,= ti.,, i E HP, +r(l -zcui), iEHP,

TINi= ti,NOK+,. jECP. (3) dthuj G TOUTH, - t,.,

Feasibility of temperatures +r(l -zhu,), jeCP. (7)

t,,k 2 t,.k + I) kEST, ~EHP, Note that these constraints can be. expressed as
inequalities because the cost of the exchangers de-
t,.r 2 t++ I? kEST, iECP,
creases with higher values for the temperature ap-
TOUT, < tiNoK+ ,r ieHP, proaches dt. Also, the role of the binary variables in
jECP.
the constraints in (7) is to ensure that nonnegative
TOUT, 3 tj,, , (4)
driving forces exist for an existing match. When a
Hot and cold utility load match (i, j) occurs in stage k, z,,~ equals one and the
constraint becomes active so that the approach tem-
@,,,OK + I - TOUT,)F, = qcu,, i f HP,
perature is properly calculated. However, when the
(TOUT,--t,,)F,=qhu,, jeCP. (5) match does not occur, ziik equals zero, and the
contribution of the upper bound r on the right-hand
Logical constraints
side deems the equation inactive. Similar constraints
Logical constraints and binary variables are needed are used for utility exchangers when the outlet tem-
to determine the existence of process match (i, j) in perature of the utility stream, TOUT, is not strictly
stage k and also any match invoking utility streams. higher (for hot utilities) or lower (for cold utilities)
The t&l binary variables are represented by zijk for than the target temperature of the process stream.
process stream matches, zcui for matches involving Also, in order to avoid infinite areas, small positive
cold utilities and zhuj for matches involving hot bounds are specified for the approach temperature
utilities. An integer value of one for any binary variables dt; that is:
variable designates that the match is present in the
optimal network. The constraints are then as follows: dt,,k 2 5 (8)

qijk - QZ,,k G 0, iEHP, jeCP, keST, where L can be interpreted as the lowest allowable
value of EMAT.
qcu, - Rzcu, < 0, iEHP,

qhu, - Qzhuj < 0, jECP, Objective function

Z ,,t , Zw , zhu, = 0, 1, (6) Finally, the objective function can be defined as the
annual cost for the network. The annual cost involves
where the corresponding upper bound $2 can be set to
the combination of the utility cost, the fixed charges
the smallest heat content of the two streams involved
for exchangers, and the area cost for each exchanger.
in the match.
LMTD terms in the objective function are approxi-
Calculation of approach temperatures mated using the Chen equation (1987):
The area requirement of each match will be incor-
min 1 CCUqcu, + 1 CHUqhu,
porated in the objective function. Calculation of these iEHP jECP
areas requires that approach temperatures be deter-
mined. In the area targeting formulation presented in
Part I, the approach temperatures were calculated
explicitly in the objective function which required the
f c CF,,,jzhuj
use of max operators. Although these terms can be jaCP
handled with the use of smooth approximations, the
synthesis model can avoid their use completely + I .c,, , e;, k z, cij kqijk
through the introduction of approach temperature
variables dt coupled with the use of the binary /(uz,[(df,,t)tdt,,k+I ) (dt, + dt, +I )/21”)‘lBiJ
variables. To ensure feasible driving forces for ex-
changers which are selected in the optimization pro- + c [Link]/(Ui,,, Wcu,WXJT - TINa)
cedure, the binary variables are used to activate or ,eHP
deactivate the following constraints for approach x {dtcu, + (TOUT, - TINcu)}/2]“3)]a~~c”
temperatures:
+ x C,“,jtqhujl(U~“,, [dthu,(TIN,, - TOUT,)
dti,, < tj,k - tj,k + r (1 - zijk),
,cCP
kEST, iEHP, jECP,
x {dthu, + (TfN,,, - TOUT,)}/2]“3)]B”“~~. (9)
dtijk+l~ti.k+l-tj,k+l+r(l-Zijk)t
The proposed MINLP model for the HEN syn-
kcST, iEHP, jECP, thesis problem consists then of minimizing the
1170 T. F. YEE and I. E. GROSSMANN

objective function in (9) subject to the feasible space MINLP model

defined by equations (l-8). The continuous variables
(t, q. dt) are nonnegative and the discrete variables z
are &I. Constraints on stream matches can be easily
1
incorporated into the MINLP formulation by fixing
certain binary values or providing bounds on the heat
load variables. The number of units can be controlled
by adding an integer constraint so that the sum of all
the binary variables is fixed or bounded at a particu- All Yes
Binaries
lar value. Furthermore, stream splitting can be con- Integral?
trolled by limiting the number of matches that can
occur at each stage. To represent the case where no +--- No

splitting is allowed, integer constraints are added to

the formulation so that a maximum of one match can
exist at each stage for each stream:

c z,,~ < 1, i E CP, k E ST.

One way of accounting for the cost of splitting

streams is to introduce assignment variables into the
formulation so that if more than one match exists in
a particular stage, the assignment variable becomes
positive and a cost is added to the objective value. In
this way, there is a corresponding penalty for having v k obtained J
splits in the network configuration.
Fig. 5. Combined Penalty Function and Outer-Approxi-
Finally, the proposed MINLP model can be easily mation Method.
extended to handle inlet and outlet temperatures that
are specified as inequalities (see case 4 of Example 1).
ated in an augmented penalty function. A sequence of
Also, the model can be easily modified to account for
NLP and MILP master problems is then solved in
matches between pairs of hot streams or pairs of cold
which the linear approximations are accumulated in
streams (see Example 5).
the master problem. The cycle of major iterations is
continued until there is no improvement between two
REMARKS successive feasible NLP subproblems. This method
has proved to be effective in solving nonconvex
The attractive feature of the proposed MINLP
MINLP problems and has shown to often lead to the
model is that equations (l-B), which define the feas-
global optimum. In general, it has been observed that
ible space, are all linear. Therefore, there is no need
an important factor leading to a globally optimal
to approximate the feasible region by any lineariza-
solution is that a good solution be obtained at the
tion scheme. This leads in general to a reasonable
level of the relaxed NLP. As a result, even though the
computational time for solving the MINLP problem.
problem is very robust in nature, an initialization
It should be noted, however, that the nonlinearities in
scheme can be used to increase the probability of
the objective function (8) may lead to more than one
obtaining the best relaxed solution. In addition to the
local optimal solution due to their nonconvex nature.
simple procedure presented in Part I (Yee et al., 1990)
In view of the nature of the MINLP model, the
of this series of papers, an alternative procedure
Combined Penalty Function and Outer-Approxi-
which relies on an LP approximation of the MINLP
mation Method by Viswanathan and Grossmann
is outlined in the Appendix. As shown by the
(1990) can be applied to solve the proposed MINLP
examples in the next section, although none of the
model. The solution scheme for the method is shown
solutions obtained has been proven to be globally
in Fig. 5. The initial step involves the solution of the
optimal, they are indeed very satisfactory in terms of
relaxed NLP. If the solution of the relaxed NLP is
minimizing the annual cost.
integer, the algorithm stops. Otherwise, if the relaxed
NLP is noninteger, an MILP master problem based
on the linearization of the relaxed NLP solution is EXAMPLES
then formulated to predict a set of integer values for
the binary variables. This master problem involves Example 1
slack variables that allow the violation of lineariza- Example 1 is from Linnhoff et al. (1982) involving
tions of nonconvex functions and which are incorpor- two hot and two cold streams along with stream and
 Simultaneousoptimization models for heat integration-II 1171

Table 1. Problem data for Example 1

cost
Stream TIN (K) TOUT (K) Fcp (kW Km’) (S kW-’ yr-‘)
HI 443 333 30
HZ 423 303 15 -
Cl 293 408 20
c2 353 413 40 -
Sl 450 450 - 80
WI 293 313 - 20
U = 0.8 (kW m-’ K-l) for all matches except ones involving steam.
U = I.2 (kW me2 K-‘) for matches involving stream.
Annual cost = 1000 x [area (III~)]‘,~for all exchangers except heaters.
Annual cost = 1200 x [area (m’)]” for heaters.

cooling water as utilities. The problem data as well as proposed network, three of the exchangers (2,3,4)
the exchanger cost equations are presented in Table are placed across this pinch and that the minimum
1. Four networks are synthesized to account for cases approach temperature (EMAT at exchanger 3) is just
Of: 2.69 K.
The problem was also solved with MAGNETS
1. No network restrictions.
with a fixed HRAT = 10 K. The solution obtained,
2. No stream splitting allowed.
which is the same as the one reported by Linnhoff et
3. Forbidden, required and restricted matches.
aI. (1982), has an annual cost of $89,832 (see Fig. 7),
4. Target temperature as inequalities.
which is 11% higher than the proposed network in
Results for cases 1 and 3 are presented in the Fig. 6. A drawback of the MAGNETS solution is
MAGNETS User Guide (Grossmann, 1985) and that utility consumption or the level of energy recov-
therefore will be compared with the solutions from ery (HRAT) was fixed throughout the optimization
the simultaneous methodology. procedure. Also, since in MAGNETS, the problem
was decomposed into two subnetworks at the pinch
Case l-no network restrictions
(363-353 K), six units were required as compared to
In constructing the superstructure, the number of five for the simultaneous solution.
stages is fixed at two corresponding to max{ NH, Nc}. It is interesting to note that for HRAT = 10 K, the
Utility exchangers are placed at the two ends of the heuristic estimate of minimum number of units for
superstructure as shown in Fig. 2. The corresponding this problem is seven (e.g. Linnhoff et al., 1982).
MINLP formulation involves 62 constraints and 50 However, for a level of energy recovery correspond-
variables of which nine are binary. Since the cost ing to the threshold case, the heuristic estimate is only
equation does not have an explicity fixed charge, the four. In fact, a four-exchanger network can be ob-
binary variables are only needed to account for the tained using MAGNETS if the HRAT is fixed at the
approach temperatures [see equation (7)]. Since three threshold value of 5.56 K while the EMAT is allowed
of the utility matches do not require binary variables to be less than HRAT. This network is shown in Fig.
as feasibility of approach temperatures is always 8 and has an EMAT at exchanger 2 of only 1.79 K.
guaranteed, nine instead of 12 binary variables are The annual cost is $80,000, which is about S300 less
needed. Also, the value of E in (8) was set to 0.1 as than the network derived by the simultaneous ap-
in the other examples. The model was solved by proach. However, a more complicated structure is
the package DICOPTf + (Viswanathan and required where a bypass at the outlet of exchanger 2
Grossmann, 1990) via GAMS (Brooke et al., 1988) is needed. As mentioned previously, bypasses are
using MINOS5.2 (Murtagh and Saunders, 1985) and undesirable since additional stream splitting is re-
MPSX (IBM, 1979). Three major iterations were quired making the network harder to operate. From
required using a total CPU time of 12.5 s on an IBM a practical standpoint, it is hard to justify the added
3083. The network structure obtained did involve complexity to the network for a nominal savings of
split streams. As a result, a suboptimization was $300 yr-‘. Interestingly, the MAGNETS network of
performed to determine the proper split ratios and Fig. 8 makes use of both of the structures (the bypass
temperatures. The final optimal network is shown in and the split stream going through exchangers 2 and
Fig. 6. This network minimizes the utility to just 4 in series) which are not considered by the proposed
$8000 yr-’ needing only cooling water. It is apparent superstructure. Even so, the network is less than
that the cost data favor the trade-off of requiring 0.5% better in terms of cost. This result indicates that
more area to minimize the utility requirement. The the simplicity of the proposed superstructure in Fig.
total annual cost for the network is $80,274. The level 2 should in general not be. a serious restriction for
of energy recovery corresonds to that of a threshold determining good networks.
problem since only cooling utility is needed. How-
ever, an internal pinch (minimum approach tempera- Case 2-no stream splitting
ture) exists according to the composite curves at As discussed in the preceding paragraph, the com-
358.56-353 K. It is important to note that in the plexity of the network in regards to the number of
1172 T. F. YEE and I. E. GROSSMANN

Cl
I 293

Hl

(27.467)

H2

406.59 313

Annual Utility Cost =$8,000

Annual Capital Cost-$72.274
Total Annual Cost -$80,274

5 400 38.3

Fig. 6. Example I--unrestricted case.

stream splits is also a very important factor for HEN splits are now unnecessary. Once again, only five
design. Although the proposed network of Fig. 6 is units are required and exchangers 2 and 4 are placed
much better in terms of cost than the MAGNETS across the internal pinch (358.56353 K), and EMAT
network of Fig. 7, stream splitting is required. Since for exchanger 4 is just 2.65 K.
a network with stream splitting is more difficult to
operate, it may be desirable to design a network with Case 3-forbidden, required and restricted matches
no stream splits. As mentioned earlier in the paper, It may often be the case that when designing the
a no split network simply corresponds to selecting at HEN, certain restrictions on the network must be
most one unit for each stream in the superstructure. imposed for practical or safety reasons. One such
When the model is constrained so that no stream example is presented in the MAGNETS manual for
splitting is allowed, it may be necessary to incorpor- Example 1. The restriction forbids matching stream
ate more stages in the superstructure in order to allow H2 with cooling water, requires stream HI to ex-
for more flexibility in the rematching of streams. To change a minimum of 300 kW of heat with cooling
do so for Example 1, the number of stages is in- water, and restricts match HI-Cl to a maximum of
creased from two to three. Along with the no split 300 kW. In the proposed formulation, these con-
constraints (IO), the MINLP formulation involved 95 straints can be easily incorporated into the model by
constraints and 66 variables of which 13 are binary. setting bounds and adding constraints to regulate
The optimal solution was obtained in four major heat loads for the matches and fixing the values of the
iterations after 15 CPU s on the IBM 3083. The relevant binary variables. Using a two-stage super-
network is shown in Fig. 9 and has an annual cost of structure, the restricted formulation required 63 con-
$80,909. As compared to the previous design with straints, 50 continuous variables and nine binary
stream splitting, the annual cost is less than 1% variables. The problem was solved in three major
higher, which is insignificant since the two stream iterations using 16.25 CPU s on the IBM 3083. The
 Simultaneous optimization models for heat integration--II 1173

PO8

443 363
HI

Annual Utility Cost = $28,000

Annual Capital Cost = $61,832
Total Annual Cost = $89,832

6 300 7.1

Fig. 7. MAGNETS solution for Example 1. HRAT fixed at 10K.

network, with five units and an annual cost of such that 373 K < TOUT,, < 413 K. To represent
$87,225, is shown in Fig. 10. Again, the solution this in the formulation, the parameter TOUT0 is
compares well with the one obtained from MAG- replaced by a new variable toutc-. Specifically, the
NETS which requires, for HRAT = 10 K, an annual replacement appears in the overall heat balance for
cost of $90,831 with six units. In both of the sol- stream C2 and in the objective function term for
utions, the same utility requirement is needed. It is calculating the area cost for the heater involving C2.
interesting to note that the solution obtained using Furthermore, toutcl is bounded to reflect the allow-
MAGNETS requires 59m2 less area. However, due able range of outlet temperature.
to a different distribution of area in the exchangers With the modification in the formulation, a two-
and economy of scale (a 0.6 exponent on the area cost stage model involving 63 constraints and 41 continu-
equation), the annual cost required by the MAG- ous and nine binary variables was solved in
NETS solution is about 4% more. The results clearly 12.6 CPU s on the IBM 3083 using DICOPT + + . As
show that the trade-off between the number of units shown by the optimal network in Fig. 11, the solution
and area must be considered. Once again, exchangers did indeed take advantage of the range specification
(1,2,4) are placed across the internal pinch and for the target temperature. The solution selected a
EMAT is 3.55 K. minimum heat exchange for stream C2 with a net-
work outlet temperature of 373 K. Note that unlike
Case I-target iemperatures as inequalities the other cases, the utility usage is not minimized
As mentioned previously, in formulating the since 2000 kW is needed from cooling water instead
model, the temperatures for the stream data can be of the 400 kW of cooling water when the outlet
specified as inequalities. To illustrate this point, the temperature of 413 K is specified. The cost of cooling
target temperature for stream C2 in Example 1 is water appears to be sufficiently cheap so that the cost
modified from the fixed value of 413 K to a range of capital is more significant. Hence, the network
1174 T. F. YEE and I. E. GRO~WANN

(15)

c c2
304.21

303
H2 423

Annual Utility Cost 48,000

Annual Capital Cost =$72,000
Total Annual Cost 480,000

Area(m 2 )
56.6

299.8

I 3 I 1800 1 225.0

I 4 I 400 I 11.4

EMAT for exchanger 2 = 1.79K

Fig. 8. MAGNETS solution for Example 1, HRAT fixed at 5.56K.

requires only four units and relatively little area, and binary. The solution procedure using DICOPT+ +
the annual utility cost of $40,000 is more than the required three major iterations and 17.9 CPU s on the
annual capital cost of $36,880. Total annual cost for IBM 3083. The optimal network obtained is shown
the network is $76,880, which corresponds to to a in Fig. 12. The total cost for the network is %715,970,
savings of about $3000 yr-’ as compared to the which is roughly $13,000 less than the previously best
network of case 1 where the target temperature for C2 reported solution of $729,000 by Gundersen and
is fixed at 413 K. Grossmann (1988). In both of the previous papers,
the reported network required six units and 2960 m2
Example 2
of area. The optimal solution from the proposed
Example 2 is from Gundersen and Grossmann method also requires six units, but the area require-
(1988) and was also analyzed by Colberg and Morari ment is 3045.5 mz, or about 85 m* more. However, the
(1990). In both papers, no cost data are presented for total cost actually turns out to be less. One reason for
utility streams and the level of energy recovery is fixed the additional area requirement for the optimal net-
at HRAT = 20°C. As a result, for comparison pur- work is the fact that one of the temperature ap-
poses, the utility usage is fixed for the proposed proaches in exchanger 1 is relatively small, lying
method at HRAT = 2O”C, so that the emphasis is below 2o”C, which leads to a small driving force and
placed on the trade-off between the number of units a large area requirement. The effect of the large area
and area. The problem has the same number of on cost, though, is compensated by the effect of
streams as Example 1 although the cost equation for economy of scale where the incremental area of an
the exchangers involves an explicit fixed charge for exchanger becomes progressively cheaper. Overall, it
each unit. The problem data are shown in Table 2. appears that the optimal network is able to fully take
Using the proposed method, a superstructure with advantage of economy of scale so that, as compared
two stages is constructed and the corresponding to the previously reported solution, even though the
MINLP model is formulated. The formulation in- area requirement is higher, the cost is lower. This
volves 67 constraints and 53 variables with 12 being result clearly illustrates that the minimization of area
 Simultaneous optimization models for heat integration-II 1175

327.02

2S3w1

397.02

Annual Utility Cost = $8000

Annual Capital Cost - $72,909
Total Annual Cost - $80,909

Fig. 9. Example l-no split case.

does not necessarily go hand-in-hand with the mini- for the problem established in Colberg and Morari
mization of cost, even when the same number of (1990) of $716,000.
exchangers is considered.
In addition, it is important to note that in the Example 3
network of Fig. 12, exchanger 5 is placed across the Example 3 is from the MAGNETS User Manual.
pinch at 90-70°C. It is also interesting to note that The main purpose of this example is to analyze the
although the net total heat flow of the composite proposed method in the case where split streams are
streams across the pinch is zero, there is heat trans- required. The problem involves five hot streams and
ferred across the pinch within exchanger 5. It is easy one cold stream along with steam and cooling water.
to verify that across the temperatures 90-7O”C, The problem data are shown in Table 3. Since only
stream Hl transfers 213.5 kW which is absorbed by one large cold stream is present, it is likely that the
stream Cl below the pinch thus cancelling this extra final network will require many split streams, which
heat flow. Also, note that the driving force is rela- is exactly the case for the solution obtained by
tively high in this exchanger. This example clearly MAGNETS shown in Fig. 13. In networks where
shows that not placing exchangers across the pinch is several stream splits may be required, restriction on
a heuristic which may not always hold. the type of split allowed in the model, where the
Another heuristic that may not hold is the one for outlet temperatures at each stage are assumed to be
the minimum number of units. Using the heuristics equal, may have significant impact on the optimal
for the case where the problem is decomposed into network generated.
subnetworks, seven units are predicted. When the The superstructure for the problem was set up with
problem is not decomposed, five units are predicted. five stages. The MINLP formulation contains 222
However, the optimal solution for the problem re- constraints and 104 continuous and 31 binary vari-
quires six units. Finally, it should be noted that the ables. The solution was obtained in three major
capital cost requirement for the optimal network iterations using DICOPT + + , which required
corresponded very closely with the capital cost target 2.78 CPU min on the IBM 3083. The network ob-
1176 T. F. YFE and I. E. GROSSMANN

Cl

(17.225)

(2.239)- b
Wl
c2
Hl 443 .

(27.761)

Forbidden Match: H2-WI

Required Match: Hl-Wi
Restricted Match: Hl -Cl

Annual Utility Cost =$28,000

Annual Capital Cost-$59,225
Total Annual Cost =$87,225
450 408
+

Fig. 10. Example l-restricted case.

tained from the MINLP optimization involves nine ExampIe 4

units and three split streams. However, the number of
units was reduced to seven in the NLP suboptimiza- Example 4 is from Colberg and Morari (1990), a
tion step and the optimal network is shown in Fig. 14. problem involving three hot and four cold streams
The annual cost for the network is $576,640, which along with steam and cooling water. The data are
is slightly higher than the MAGNETS network at shown in Table 4. The interesting aspects of this
$575,595, which was solved with HRAT fixed at 5 K. example are that: (1) the streams have significantly
A comparison between the two networks shows that different heat transfer coefficients; (2) the synthesized
the MAGNETS network requires two additional network for the fixed HRAT from Colberg and
units and four additional split streams. The energy Morari (1990) requires a split stream going through
requirement is lower in the MAGNETS network but exchangers in series as shown in Fig. 4a, a configur-
the investment cost is higher. In fact, its total area is ation the proposed superstructure does not consider.
295.5 m2, which is almost 50% higher than the total In order to synthesize a network for comparison, the
area of 200.9 m2 for the network in Fig. 14. For the level of energy recovery was fixed at HRAT = 20 K
network of Fig. 14, the utility requirement corre- which leads to a pinch at 517497 K. Also, since no
sponds to an HRAT of about 13.1 K with a pinch cost equation was given, the exchanger cost equation
location at 380-366.9K. Similar to the previous from Example 2 of 8600 + 670 x (Area)‘.“’ was used.
examples, the network has exchangers (2, 3 and 4) The superstructure for the problem was con-
which are placed across the pinch. Overall, the results structed with four stages. The formulation involved
of this example are very encouraging since despite the 231 constraints and 151 continuous and 48 binary
simplifying assumption of isothermal mixing, the variables. Solution of the problem to optimality
method indeed obtained a network very close to the required three major iterations and 13.8 CPU min on
optimum. the IBM 3083. Two split streams are required in the
 Simultaneous optimization models for heat integration-II 1177

~2 423 _0 303 )

4 373 5 TOUTc2 < 413

I
Cl
293

Annual Utility Cost = $40,000

Annual Capital Cost = $36.880
Total Annual Cost = $76,880

Fig. 11. Example l-target temperature as inequalities.

solution, and therefore the NLP suboptimization is especially interesting to note that the exchange at unit
performed to determine the optimal split ratios. The 3 has an approach temperature on one side of a mere
final network is shown in Fig. 15. The cost for the 0.88 K. However, the area for the unit is not large
network is $150,998. This compares well with the since the two streams involved have the largest heat
Colberg and Morari network which, using the same transfer coefficients.
cost equation, is at $177,385, roughly 17.5% higher.
Example 5
However, the Colberg and Morari solution does
achieve their objective of minimizing the total area. This example involves the 4SPl problem of Lee
Their network requires 188.9 m2 vs 217.8 m* for the et al. (1970). The data are presented in Table 5 and
network from the proposed approach. The trade-off, involves two hot and two cold proces streams along
though, is that the Colberg and Morari network also with steam and cooling water. The problem is used
requires three additional units and 10 additional split here to illustrate the incorporation of cold-to-cold or
streams. One reason why the number of units is larger hot-to-hot matches in the proposed method. As
is that their problem was partitioned into subnet- discussed previously, several authors have noted that
works. Since the heat transfer coefficients are so it may be desirable in certain cases to have heat
different, certain cross-pinch exchanges may be desir- exchange between two hot or two cold streams.
able. In fact, the optimal solution derived from the Dolan et al. (1987) considered this type of matching
simultaneous approach does indicate this and cross- when they analyzed the 4SPl problem using simu-
pinch exchanges exist in exchangers 2 and 3. It is lated annealing for the case where a match between
hot stream Hl and cold stream Cl is forbidden. For
a minimum approach temperature (EMAT) of 18”F,
Table 2. Problem data for EX~IIIDI~2
they derived a network with a cold-to-cold exchanger
stream TIN (“C) TOUT (“C) Fcp (kW “C-‘)
which required a total annual cost of $13,800. They
HI 150 60 20
HZ 90 60 80
also compared their solutions with the one derived by
Papoulias and Grossmann (1983) for the same restric-
:: 20
25 125
100 25
30 tion and EMAT, and where the objective was to
SI 180 180 -
WI 10 15 - minimize the number of units. In the Papoulias and
U = 0.05 (kW m-* “C-‘) for all matches. Grossmann network, the use of cold-to-cold matches
Cost = 8600 + 670 x [area (II?)]~*’ for all exchangers. was not considered. As a result, more utility was
1178 T. F. YEE and I. E. GROSSMANN

c2 Cl

H2 90 & 70.46 6 66 6'" 60,

1109.46
160

HRAT fixed at 20%

Total Capital Cost = $715,970

Total Area - 3045.4 m2

Fig. 12. Optimal network for Example 2.

required and the network has a higher annual cost of

%21,100.
(f,.k- tl,k + I 14 = c %ik

In the proposed model, hot-to-hot or cold-to-cold

matches can be embedded in the superstructure. As
an example, consider the case of cold-to-cold matches
for which the following modifications are required in
the formulation: 4. Introduce new terms in the objective function to
calculate the cost of the cold-to-cold exchang-
Introduce new heat load variables for cold-to-
ers.
cold matches, qjj,,k, to represent the heat trans-
fer from cold stream j to cold stream jl, where With these modifications, the 4SPl problem was
jl #j. formulated embedding cold-to-cold matches. A
Relax the monotonic decrease of temperatures three-stage representation was used since the problem
along the stages by removing the constraint in involves potentially three “hot” streams. For com-
(4): parison with the results of Dolan et al. (1987) the
‘jk 3 tj.k+ 1, keST, jcCP. (11) minimum approach temperature is set to 18°F. The
MINLP model involves 94 constraints and 70 con-
Introduce the new variables into the overall and
tinuous and 15 binary variables. Solution of the
interval heat balances [equations (1) and (211:
problem using DICOPT+ + required 23.31 CPU s
on the IBM 3083. The optimal solution derived is
(TOUi’J- TIN,)F, = c C q,,x
XEST,GHP shown in Fig. 16. The solution obtained indeed
requires a cold-to-cold match between streams Cl
- $, j,zp (%jl.k - $.,.k) + qhuj, j E CP,
and C2, where C2 is considered the “hot” stream. The
total annual cost for the network is $13,800, which is
 Simultaneous optimization models for heat integration-11

H2

456.63 462.19

1463.69

Pinch location: 380-375K

Annual Utility Cost - $494.900
Annual Capltal Cost - $80,895
Total Cost - $575.595
Total Area I 295.5 III*

Fig. 13. MAGNETS solution for Example 3.

identical to the solution from Dolan et al. (1987). A A second network for the example was obtained
comparison between the two networks shows that the for the case where the specification of minimum
configurations are the same and both networks approach temperature is eliminated. The optimal
achieve minimum energy requirement. The heat load network is shown in Fig. 17. Again, the same network
distribution, however, is slightly different but not configuration is obtained, with one cold-to-cold
enough to significantly affect the capital cost. match involved in the network. The heat load distri-

Table 3. Problem data for Example 3

Stream TIN (K) TOUT (K) Fcp (kW K~-‘) Cost (SkW-‘yr-‘)
Hl 500 320 6 -
HZ 480 380 4 -
H3 460 360 6 -
H4 380 360 20 -
H5 380 320 12
Cl 290 E 18
Sl 700 - 140
WI 300 320 IO
U = I (kW mm2 Km’) for all matches.
Annual cost = 1200 x (area (m2)]06 for all exchangers.
1180 T. F. YEE and I. E. GR~MANN

t 660

TT440.44 H2
H3

(3.698)

“352.22
Wl 380
300 H4
6
32
SJ
320
H5 380

Cl

Pinch Location: 360-366.9K

Annual Utility Cost = $516,660
Annual Capital Cost = $59,780
Total Annual Cost = 576.640
2
Total Area = 200.9 m

Fig. 14. Optimal network for Example 3.

bution, however, is quite different than the previous cost for the network is $11,374, which is about 18%
network, with significant reduction in the utility less as compared to the previous network where
requirement. The annual utility cost is reduced by EMAT is fixed at 18°F.
about $3300. The optimal trade-off, though, requires
an increase of over 50% in heat transfer area. Since
CONCLUSION
area cost is relatively cheap, the annual capital cost
increases by just $859 despite the fact that an EMAT In this paper, a systematic procedure has been
of just 2.15”F exist at exchanger 2. The total annual proposed for the synthesis of heat exchanger net-

Table 4. Problem data for Example 4

Stream TIN (K) ?-OUT(K) Fcp (kW K ‘) h (kWmmzKe’)

Hl 626 586 9.802 I .25
HZ 620 519 2.931 0.05
H3 528 353 6.161 3.20
Cl 497 613 7.179 0.65
c2 389 576 0.641 0.25
c3 326 386 7.627 0.33
c4 313 566 I.690 3.20
SI 650 650 - 3.50
WI 293 308 3.50
I/U -(l/h+,+ l/h,).
Cost = 8600 + 670 x [area (mz)Jos’ for all exchangers.
 Simultaneous optimization models for heat integration-II 1181

65b r,,,.,, “q T3.33 xl3

I .(3.903)1 1 J, > 435.2il

650 1676 /8X

Pinch Location: 517~497U

HRAT fixed at 20 K
Total Cost = $150,998

7 33.6 1.46

6 65.7 0.39

9 296.0 140.06

Fig. 15. Optimal network for Example 4.

works. Unlike previous methods, the proposed ap- minimum approach temperatures (EMAT) and
proach does not rely on a sequential decomposition stream matches are not fixed. The simplifying as-
of the problem since it accounts simultaneously for sumption on isothermal mixing in the superstructure
the trade-offs between energy cost, fixed charges for eliminates flow variables and heat mixing equations
units and cost for exchanger area. The method in- which allows the feasible space for the model to be
volves the optimization of a stage-wise superstructure defined by linear constraints. Thus, the model can be
that is modeled as an MINLP problem. No account solved more efficiently. Given the assumption of
is made for pinch considerations, such as partitioning isothermal mixing of streams, a suboptimization is
into subnetworks or not placing exchangers across performed to determine optimal split ratios when the
the pinch. Energy recovery (HRAT), heat loads, predicted network requires stream splits. A positive

Table 5. Problem data for Example 5

stream TIN (“F) TOUT (“F) Fcp (B.I.u. ‘F--l) @IO00 B.t.u. yr- ‘)
HI 320 200 16.666.8 -
H2 480 280 20,ooo -
Cl 140 320 14.450.1
c2 240 500 11,530
Sl 540 540 12.76
WI 100 I80 5.24

(I = I50 (8.t.u. ftr” “F-‘) for all matches except ones involving steam.
L’ = 200 (B.t.u. ft-* “F-‘) for matches involving steam.
Annual cost = 35 x [area (III’)]~~~ for all exchangers.
 540 540
Sl 320 Sl 320
t t

Hl 00 Hi

180 180

No match allowed for Hl-Cl

EMAT = 18 OF No match allowed for Hi-Cl
Cold-to-cold matches allowed Cold-to-cold matches allowed

Annual Utility Cost = $9988

Annual Utility Cost = $8,898
Annual Capital Cost = $3817
Annual Capital Cost = $4,676
Total Annual Cost
Total Annual Cost = $11,374
Total Area = 832.9 $$13’800
Total Area = 1295.4 ft*

l-kalbad Heatload
ExCh. Area(ft2) Exch. Area(fl2)
OOBiu) OOBiu)

1 6 ( 570.5 1 68.7 ]

Fig. 16. Example S-restricted case with EMAT = 18°F. Fig. 17. Example S-restricted case with no EMAT specification.
 Simultaneous optimization models for heat integration-II 1183

effect from the simplifying assumption is that the Gundersen T. and L. Naess, The synthesis of cost optimal
heat exchanger network synthesis--an industrial review
model will tend to favor structures with no stream
of the state of the art. Computers them. Engng 12,503-530
splits. (1988).
The MINLP model can also easily accommodate IBM Mathematical Programming System ExtendedJ370
constraints on stream matches, heat loads and stream (MPSX/3709) Basic Reference Manual. White Plains
splitting. In addition, the model can consider hot-to- (1979).
Lee K. F., A. H. Masso and D. F. Rudd, Branch and bound
hot or cold-to-cold matches. The limitation of the synthesis of integrated process designs. Ind. Engng Chem.
method lies in the fact that certain configurations are Fundam. 9, 48-58 (1970).
not explicitly included in the superstructure. Linnhoff B. and E. Hindmarsh, The pinch design method
Examples have shown, however, that the limitation is for heat exchanger networks. Chem. Engng Sci. 38,
745-763 (1983).
not severe in view of the combinatorial nature of the
Linnhoff B. er al., A User Guide on Process Integration for
synthesis problem, where several alternative configur- the Eficienr use of Energy. The Institute of Chemical
ations may be very close to the global optimum. The Engineering, U.K. (1982).
MINLP model, which has been applied to problems Murtagh B. A. and M. A. Saunders, MINOS 5.0 User’s
involving up to seven process streams, may or course Guide. Technical Report SOL 83-20, Systems Optimiz-
ation Laboratory, Dept of Operations Research, Stanford
become more expensive to solve in larger problems.
University (1985).
The example problems presented have also shown Papoulias S. A. and I. E. Grossmann, A structural
that pinch considerations may not be relevant for optimization approach in process synthesis-II. Heat
synthesizing the network structure when all the trade- recovery networks. Computers them. Engng 7, 707-721
offs are accounted for simultaneously; this is true (1983).
Viswanathan M. and L. B. Evans, Studies in the heat
even for the case when heat transfer coefficients for integration of chemical process plants. AZChE JI 33,
the streams are the same. Considerations for the 1781-1790 (1987).
economy of scale for area cost and fixed charges for Viswanathan J. and I. E. Grossmann, A combined
the number of units do not necessarily favor the penalty function and outer-approximation method for
MINLP optimization. Comparers them. Engng 14,
minimization of area for which the pinch heuristics 769-782 (1990).
are based. Furthermore, as shown by the examples, Wood R. M., R. J. Wilcox and I. E. Grossmann, A note on
optimal networks often involve exchangers that are the minimum number of units for heat exchanger network
placed across the pinch. synthesis. Chem. Engng Commun. 39, 371-380 (1985).
Yee T. F. and I. E. Grossmann, A screening and optimiz-
Acknowledgmenf-The authors would like to acknowledge ation approach for the retrofit of heat exchanger net-
financial support from the Department of Energy under works. Presented at the Annual AlChE Meeting,
Grant DE-FG-02-85ERl3396. Washington, DC. (1988).
Yee T. F., 1. E. Grossmann and Z. Kravanja, Simultaneous
optimization models for heat integration-I. Area and
REFERENCES
energy targeting and modeling of multi-stream ex-
Brooke A., D. Kendrick and A. Meeraus, GAMS: A User’s changers. Computers Chem. Engng 14, 1151-1164 (1990).
Guide. Scientific Press, Palo Alto, Calif. (1988). Yee T. F., I. E. Grossmann and Z. Kravanja, Simultaneous
Chen J. J. J., Letter to the Editors: Comments on improve- optimization models for heat integration--III. Process
ment on a replacement for the logarithmic mean. Chem. and heat exchanger network optimization Computers
Engng Sci. 42, 2488-2489 (1987). them. Engng (1990). In press.
Colberg R. D. and M. Morari, Area and capital cost targets
for heat exchanger network synthesis with constrained
matches and unequal heat transfer coefficients. Computers APPENDIX
them. Engng 14, l-12 (1990).
Dolan W. B., P. T. Cummings and M. D. LeVan, Heat Initialization Procedure for Sobng the MINLP Models
exchanger network design by simulated annealing. Proc. As shown in Fig. 5, the first step of the Combined Penalty
First Int. Conf. on Foundarions of Computer Aided Process Function/Outer Approximation Method involves the sol-
Operations, Park City, UT (1987). ution of the relaxed NLP problem. Even though this NLP
Dolan W. B., P. T. Cummings and M. D. LeVan, Process formulation is very robust in that it only has linear con-
optimization via simulated annealing: application to net- straints, it is desirable to supply a “good’ initial guess so
work design. AlChE JZ 35, 725-736 (1989). one can increase the likelihood of obtaining the best solution
Floudas C. A. and A. R. Ciric, Strategies for overcoming in cases where multiple local optima may exist. In general,
uncertainties in heat exchanger network synthesis. Com- it has been observed that a good relaxed NLP solution will
parers them. Engng 13, 1133-L 152 (1989). lead to the global optimum for the MINLP model.
Floudas C. A., A. R. Ciric and 1. E. Grossmann, Automatic Following is an initialization procedure that reduces the
synthesis of optimum heat exchanger network configur- MINLP to an LP by assuming fixed temperature driving
ations. AIChE JI 32, 276-290 (1986). forces for each match:
Grimes L. E., M. D. Rychener and A. W. Westerberg, The
synthesis and evolution of networks of heat exchange that 1. Estimate a value of HRAT.
feature the minimum number of units. Chem. Engng 2. Estimate a driving force for each match by:
Commun. 14, 339-360 (1982). (a) determining the LMTD, for each enthalpy interval
Grossmann 1. E., MAGNETS User’s Guide. Carnegie Mel- n using its corresponding temperatures;
lon University, Pittsburgh (1985). (b) using the following weighting equation to calculate
Gundersen T. and I. E. Grossmann, Improved optimization an average driving force for each match (i, j):
strategies for automated heat exchanger network syn-
thesis through physical insights. Presented at the Annual
AfChE Meeting, Washington, D.C. (1988).
1184 T. F. YEE and I. E. GRWSMANN

where q,p is the maximum heat transfer that can occur term of the area by a linear approximation with a
between hot stream i and cold stream j in enthalpy fixed charge. This reduced the MINLP in (I-8) to an
interval n. MILP.
3. For each match i. i in different stases of the suner- 4. Solve the relaxed LP of the MILP in Sten 3.
structure, set the “driving forces yn the objective 5. Use the LP solution along with the estimhted driving
function (8) with the fixed value of the average forces (ALMTD,i) as an initial guess for the relaxed
driving force ALMTD,,, and replace the nonlinear cost NLP problem.