Professional Documents
Culture Documents
00
q Institution of Chemical Engineers
T
his paper combines mathematical optimization techniques with a better understanding
of the retro® t problem, based on thermodynamic analysis and practical engineering, to
produce a systematic procedure capable of ef® ciently solving industrial size retro® t
problems. The major characteristic of the approach presented is that it offers a systematic and
automatic method for the retro® t design of heat exchanger networks (HENs), combined with a
facility for meaningful user interaction.
The new procedure employs a two-stage approach for retro® t HEN design: The ® rst stage is
the diagnosis stage, during which a minimum number of promising HEN topology modi-
® cations is obtained which enables a desired heat recovery target to be achieved. In the second
stage, the optimization stage, the HEN obtained after implementation of the modi® cations is
optimized using non-linear optimization techniques to minimize the cost of additional surface
area employed.
It has been observed that heat recovery in a HEN is thermodynamically limited by certain
exchanger matches unavoidably tending to a zero degree temperature approach as the heat
recovery increases. These exchanger matches, which are termed as pinching matches, act as a
bottleneck to heat recovery in the HEN. To increase the potential for heat recovery beyond the
limits caused by the pinching matches, the network topology must be altered by repiping of
exchangers, addition of new exchanger matches or creation of stream splits.
Based on the above observation, the diagnosis stage is made up of two steps. In the ® rst step
the HEN bottleneck is identi® ed, and in the second step a mixed integer linear programming
(MILP) formulation is used to select a single modi® cation which will best overcome the
identi® ed bottleneck. These two steps are repeated in a loop to yield the required set of
promising topology modi® cations.
Keywords: retro® t; heat exchanger networks; mathematical programming; network pinch;
pinch technology
design using pinch methods consists of a targeting stage and Ciric and Floudas1 0 proposed a two-stage approach con-
a design stage. During the targeting stage, optimal targets sisting of a match selection stage, and an optimization stage.
for heat recovery and exchanger area requirement are deter- The match selections stage used a mixed integer linear
mined, whilst in the design stage a set of rules and design programming (MILP) formulation to simultaneously select
tools are used to manually design the retro® t HEN to achieve process stream matches and match exchanger assignments
the targets. In contrast to this, mathematical programming for the retro® t HEN, and the exchanger loads obtained in
techniques formulate the design problem as a mathematical this stage remained ® xed during the rest of the design
model using a set of equations or constraints including an procedure. The optimization stage then used a non-linear
objective function. This objective function is generally an programming (NLP) formulation to optimize the matching
estimate of the retro® t HEN cost, and it acts as a measure of order and ¯ ow con® guration of the exchanger matches.
the optimality of the retro® t design. The HEN problem is These two stages were later1 1 combined into a single stage,
then solved by using mathematical optimization techniques using a mixed integer non-linear (MINLP) formulation to
to determine the retro® t HEN design with the lowest total simultaneously optimise all aspects of the network.
cost, subject to the other constraints. Yee and Grossmann1 2 provided an alternative approach
The ® rst pinch retro® t design procedure was proposed by which also had two stages, in this case however a targeting
Tjoe and Linnhoff1 and introduced the concept of area or pre-screening stage and an optimization stage were used.
ef® ciency of heat exchanger networks. The method assumes The pre-screening stage was used to determine the optimal
that the area ef® ciency, a , of the retro® t HEN is equal to that heat recovery level and assess the economic feasibility of
of the existing HEN, and this assumption is used to set the retro® t design. However only the number of new units
targets for design. The design phase itself is conducted required to achieve the `optimum’ investment determined
manually using extensions of the pinch design rules2 . Sil- was carried forward to the optimization stage. During the
angwa3 proposed an extension to Tjoe’ s targeting technique optimization stage, the heat recovery level was allowed to
for cases were the area ef® ciency of the original HEN was vary and an MINLP formulation was used to simultaneously
small, and Polley et al.4 modi® ed Tjoe’ s targeting method optimize the capital-energy trade-off and all the network
by introducing a relationship between pressure drop and parameters.
heat transfer coef® cients to enable the area targets generated
to re¯ ect pressure drop limitations in the process.
The main limitation of the targeting approaches based on Discussion of Existing HEN Retro® t Methods
the area ef® ciency concept is that the area targets they The main advantage of pinch methods lie in the user
produce do not re¯ ect the area distribution within the HEN. interaction provided, whilst their main drawbacks are that
Shokoya and Kotjabasakis5 proposed a technique which they involve potentially time consuming manual proce-
overcame this limitation by incorporating the area distribu- dures, and produce solutions whose quality largely depends
tion of the existing HEN into the targeting mechanism. This on the designer’ s experience and judgement in applying the
techniques also provided additional guidance for the design pinch rules and guidelines. In addition, since pinch methods
task, and enabled the generation of retro® t HEN designs do not explicitly account for the cost of structural changes
which were demonstrably simpler than those obtained using implemented in the retro® t design, they potentially gener-
the method of Tjoe and Linnhoff1 . ate complex designs. The approach of Carlsson et al.6
Carlsson et al.6 proposed a method based on pinch tech- tackles this limitation, but its dependence on accurate piping
nology in which the cost of exchanger area, piping and and other cost data for each potential match could in some
auxiliary equipment, pumping and maintenance associated cases make it impractical; as considerable effort would be
with each potential match, is used to construct a cost matrix. required to generate the necessary data for a moderately
This cost matrix is then used together with a set of rules sized industrial process. The Van Reisen et al.7 decomposi-
(including some from pinch technology) to perform the tion could help to reduce the time required to generate pinch
design. As no targeting is performed in this method, the designs, but the effort required to investigate the alternative
capital-energy trade-off is evaluated by producing several subnetworks could be considerable where large HENs are
designs at varying heat recovery levels. involved.
Most recently, Van Reisen et al.7 introduced a method for In contrast to this, the mathematical methods reviewed
decomposing the original HEN into a number of subnet- have the advantage of being automated procedures, and the
works, so as to reduce the size of the design problem and drawback of providing little scope for user interaction. In
favour the generation of simple retro® t designs. Each of the addition, as they all use cost minimization as their optimi-
subnetworks generated by the decomposition is screened zation objective, they are limited by the two main short-
using pinch targeting techniques to identify the subnetworks comings of the cost estimates used in the models. The ® rst
which would yield the most cost-effective retro® t. The sub- of these shortcomings relates to the area cost estimates.
networks thus identi® ed then become the subject of retro® t Although the area cost estimates used in the models assume
design to which any of the previously discussed design that additional area is provided as new exchanger shells, the
methods can be applied. additional area can in reality be provided in other ways, each
Early applications of mathematical programming to the of which features a different cost. These include the replace-
retro® t problem8 ,9 were based on the grassroots scenario ment of original tube bundles with new tube bundles con-
and thus the cost of modifying the existing topology was taining a greater number of more closely packed tubes, as
not considered in the HEN cost estimates used. However, well as the use of heat transfer enhancement devices in the
dedicated retro® t techniques were later developed which exchanger tubes. The second limitation is that all the models
contained explicit expressions for the costs of topology assume that the cost of moving or repiping an exchanger is
modi® cations. independent of the exchanger involved, however this cost
depends among other things on the distance and complexity identi® ed can only be achieved by a HEN with in® nite
involved in the piping change. Although this assumption exchanger area.
avoids the need to determine modi® cation costs for every The grid representation of a simple HEN and the cor-
possible topology change, the resulting estimates produced responding process composite curves with a heat recovery
cannot reliably differentiate between cheap and costly top- of 40 MW are shown in Figure 2. The composite curves in
ology changes, and this introduces a potential for selecting Figure 3a identify the maximum heat recovery achievable in
non-optimal topology changes. the process as 60 MW, and, as shown in Figure 3b, this level
The mathematical methods reviewed also require expen- of heat recovery also produces a 08 temperature approach in
sive computation to generate solutions, but due to limita- the process exchanger of the network.
tions of non-linear programming techniques, the solutions The HEN in example 2 (Figure 4) is more complicated;
they produce have a signi® cant probability of being non- the composite curves shown in Figure 5a identify the
optimal. This is a consequence of the nature and complexity maximum energy recovery for the process as 250 MW.
of the HEN retro® t problem, and cannot be mitigated by However, as seen in Figure 5b, the network topology can
using more accurate models of the retro® t problem. Conse- only achieve a maximum recovery of 220 MW, irrespective
quently, none of the mathematical approaches described of the area added to the network. If the topology is altered,
offers a viable method for solving even moderately sized as shown in Figure 5c, a heat recovery of 235 MW can be
industrial HEN retro® t problems. achieved, and this exceeds the heat recovery limit (Rm a x ) of
From the above discussion, it is clear that none of the the original topology. This demonstrates that the heat
current methods is able to solve large industrial problems in recovery limit, Rm a x , is associated with the topology of the
an automated procedure, whilst retaining user interaction HEN. The network from example 2 at its Rm a x can be
and control of the solution process. The new approach pre- superimposed on the composite curves, as shown in
sented in this paper, however, achieves this by harnessing Figure 6. Each thick line in this ® gure represents the tem-
the strengths of both pinch and mathematical approaches, perature pro® le of a hot stream in an exchanger. Wherever
and this has been made possible by new insights gained into one of these thick lines touches the cold composite curve,
the behaviour of HENs. a 08 temperature approach exists between hot and cold
streams in that exchanger, and such an exchanger match is
NEW INSIGHTS termed as a pinching match.
A pinching match is de® ned as an exchanger match in
Limits in HEN Topologies which the temperature approach between hot and cold
In most cases, the heat recovery achieved in a HEN can streams unavoidably tends towards a limiting value as heat
be improved by addition of area to some exchangers in the recovery in the HEN is increased. Although the temperature
network. However, when this area is added without altering approach limit used in the examples is 08 , any positive
the HEN topology, it has been observed that the potential for temperature approach limit may be used. Pinching matches
heat recovery improvement often falls short of the full identify the location of the network pinch or bottleneck as
potential existing within the process. This suggests the pre- shown in Figure 7. It must however be emphasized that
sence of a heat recovery limit within the HEN topology, increasing the exchanger area of pinching matches or of
which is independent of the area of the individual exchanger other exchangers in the network will not overcome the
units in the network. This will be illustrated using the network pinch, as the bottleneck in this case is not caused by
concept of composite curves. a capacity limit.
Composite curves as shown in Figure 1 provide a graphi-
cal representation of process stream data. The minimum
vertical separation between the composite curves is the The Network Pinch
global D T m i n or heat recovery approach temperature (HRAT) The network pinch divides the HEN into a heat de® cit
of the process, whilst the overlap of the two curves gives the
heat recovery achieved in an ideal HEN. This heat recovery
increases as the global D T m i n is reduced, until the absolute
minimum D T for feasible heat transfer of 08 is reached. The
maximum energy recovery achievable in the process thus
Figure 4. Original HEN performance for example 2. Figure 5c. Impact of a topology change on the network limits.
For the simple HEN of example 1, the ideal targets and network pinch, and an additional bene® t of the network
Tjoe’ s constant alpha targets are identical, and can be achieved pinch concept is that it helps to identify the topology
by adding area to the single process-process exchanger in changes which can produce this positive impact by over-
the HEN. Since no topology modi® cation is required, this coming the network pinch limits, and this is of immeasur-
target also corresponds to the `zero topology modi® cations’ able assistance in the design task.
target. This is, of course, a trivial example, and as shown in
Figure 9, the resulting retro® t curves are coincident. This
indicates that the ideal and constant alpha target can be Overcoming the Network Pinch
achieved without modi® cation of the original HEN topol- Topology changes in a HEN can take the form of repiping
ogy. In contrast to this, Figure 10 indicates that the constant exchangers in the HEN, inserting new exchangers into the
alpha and ideal area targets in example 2 cannot be achieved HEN, or creating stream splits within the HEN; and examples
without modi® cation of the original HEN topology. of how each of these can be used to overcome the network
If, however, a single change to the HEN topology is pinch are shown in Figures 12 to 14.
allowed, the scope for heat recovery is increased, and the The general rule for identifying topology changes capable
Rm a x of the original topology can be exceeded. This creates of overcoming a network pinch is that they must move heat
a new and larger Rm a x , together with a new set of pinching from below to above the network pinch, and this is ana-
matches and a new network pinch. The retro® t target for logous to the `+/- principle’ of pinch technology2 . Since the
`one topology modi® cation’ de® nes the minimum addi- portion of the HEN below the network pinch is in heat
tional exchanger area required to achieve any feasible heat surplus, any topology change which only creates opportu-
recovery, without making more than one alteration to the nity for heat exchange in the region below the network
HEN topology. As can be seen from the corresponding pinch will not increase the Rm a x of the topology. The same is
retro® t curve, illustrated in Figure 10, the trade-off between true of topology changes which only enable further heat
area added and energy saved also improves when a topology exchange above the network pinch.
change is allowed. This rule is rigorous, and provides a valuable means of
Retro® t targets for any number of topology modi® cations screening out topology changes which cannot improve the
can be de® ned in a similar way, and as shown in Figure 11 heat recovery potential of a network. It must be noted,
all of these could be plotted as retro® t curves. With each however, that in the special case when the original HEN
implementation of topology change, the existing network features signi® cant `criss-crossing’ (or misapplication of
pinch is replaced by a new network pinch which features a driving forces) in a region removed from both the process
new and larger Rm a x , and an improved area-energy trade- and network pinches, unproductive changes in terms of heat
off. This illustrates that the network pinch not only limits the recovery, could produce a reduction in exchanger area
potential for heat potential, but also constrains the trade-off requirement.
between area added in the retro® t design and the resulting In most cases, several topology change options capable of
savings in energy. overcoming the network pinch can be identi® ed, and each
Not all topology changes produce a positive impact on the
Figure 7. The network pinch. Figure 8. Retro® t curves (Tjoe and Linnhoff, 1986)1.
Figure 10. Retro® t curves for example 2. Figure 12. Overcoming the network pinch by exchanger resequencing.
from below to above the network pinch. This however only nature of the retro® t design problem and the speci® c role
identi® es stream splitting as a possible option, and not as the played by topology modi® cation in the design task. In
most promising topology change. particular, the concept of the network pinch has been shown
Pinch technology identi® es the need for stream splitting to provide guidance for selecting topology changes, and this
on the basis of the stream population and stream heat capa- forms the basis of the new retro® t procedure.
city ratios at the process pinch. This is because the tem- The new retro® t procedure consists of two stages; namely
perature driving forces available between hot and cold a diagnosis stage and an optimization stage as shown in
streams are tightest in the region of the process pinch, and Figure 17. The diagnosis stage is used to identify and select
stream splitting often ensures that no D T violation occurs at optimal topology modi® cations to be made to the original
the process pinch. Stream splitting has also been shown to HEN, and it combines the concepts of pinching matches and
be bene® cial when applied to adjacent pinched exchangers the network pinch described in this paper with mathematical
located at the network pinch. programming techniques. The resulting HEN topology is
When a HEN topology at its Rm a x is analysed at the same then passed to the optimization stage, where mathematical
heat recovery using pinch technology, the network pinch programming techniques are used to optimize the topology
and process pinch often do not occur in the same region of and produce the ® nal retro® t design.
the HEN. As such the process pinch and network pinch often A total of four mathematical models1 3 are used in the
point to different streams and exchangers for stream split- procedure, made up of three linear models in the diagnosis
ting. When the two pinches do coincide however, both stage, and one non-linear model in the optimization stage.
pinches indicate the same stream splitting option, and in As illustrated in Figure 18, the linear models are used to
such cases stream splitting has been found to offer the most sequentially identify individual topology changes with the
cost-effective retro® t option. This observation sums up the highest potential for heat recovery, whilst the non-linear
network pinch stream split heuristic, which states that: model is used to optimize the capital-energy trade-off
within a selected design. Although the linear models are
`Whenever adjacent pinching matches occur at coinci-
similar in most respects, they have differences which re¯ ect
dent process and network pinches, stream splitting should
their varying objectives, and are signi® cantly different from
be the topology change implemented.’
those currently published in the literature. A detailed dis-
This heuristic, illustrated in Figure 16, can be used to cussion of these models is beyond the scope of this paper,
identify and select stream splitting options. and only the main features of each stage will be presented.
Figure 14. Overcoming the network pinch by stream splitting. Figure 15. The one topology modi® cation retro® t curve.
Figure 17. The new retro® t procedure. Figure 19. The default search strategy.
is accepted, the stream split heuristic is again applied to Figure 22. Limits of the original HEN in case study 1.
determine whether stream splitting is indicated. After the
resequence options are exhausted, another MILP model (P3)
is used to identify the new exchanger option which would to be submitted to the optimization stage to enable the
produce the largest increase in Rm a x . This search also identi® cation and selection of the most cost-effective design
continues until no further acceptable new exchanger options from among the alternatives.
can be found, and stream splitting conditions are tested for
after each modi® cation is accepted.
Each topology modi® cation selected by the mathematical The Optimization Stage
models yields a new Rm a x which is greater than or equal As stated previously, the primary function of the optimi-
to the previous Rm a x . However, a threshold value for the zation stage is to optimize the HEN topology de® ned in the
increase in Rm a x is speci® ed as an acceptance criteria. Thus diagnosis stage. However, the con® guration and branch
all topology changes which are unable to produce an ¯ owrates of stream splits in the topology are also de® ned
increase in Rm a x greater than or equal to this threshold value during the optimization stage. It will be noted that the
are rejected, subject to the user’ s option of overriding this diagnosis stage only identi® es the exchanger matches involved
decision. Each search for topology modi® cations is con® ned in stream splits, and does not specify any series-parallel
to the set of modi® cations which are identi® ed by the net- arrangements for the stream split. A small superstructure1 4
work pinch rules as being capable of overcoming the derived from the HEN superstructure proposed by Floudas
network pinch. In terms of mathematical programming, et al.1 5 is therefore used to optimize the stream splits and
this means that the superstructure used for the models is de® ne their con® gurations. The optimization stage is not
dynamically generated on the basis of the network pinch however intended to add, delete or move any exchanger
location. This greatly reduces the size of the mathematical units and consequently the basic network topology remains
problem and increases the speed with which solutions are ® xed.
generated. The optimization stage therefore optimizes the heat
The designer may wish to consider topology modi® cation recovery and exchanger area used in the retro® t network,
options other than those selected by the default strategy, and using cost minimization as the optimization objective. Math-
this can be achieved by taking manual control over the ematical programming techniques are particularly suited to
searching procedure. This could lead to the generation of this kind of optimization, as the trade-offs involved are too
several alternative solutions, each of which would have complex for manual evaluation. In the case where several
design alternatives are to be evaluated, each solution is
optimized separately. In this way the cost penalty or bene® ts
Figure 21. Original HEN before debottlenecking. Figure 23. Resequence A, min QH = 98.08, D QREC = 4.4.
Figure 25. Stream split implementation, min QH = 96.30, D QREC = 1.8. Figure 26b. Retro® t Design B.
® rst of these pinching matches lies at the process pinch, the which is much simpler than initiating the design process by
process and network pinches have become coincident, and making such estimates for all new exchanger options.
this ful® ls the conditions of the stream split heuristic. The Design options other than the two discussed above can be
search for resequence options is thus interrupted and the investigated by altering the search strategy. These alter-
stream split implemented. Exchangers 4, 5, and 6 being the native design options can be organised into a search tree
three adjacent pinching matches, are placed in parallel with as shown in Figure 27. In this ® gure, the default strategy
each other to produce a 1.8 MW increase in Rm a x . Once path is shown in bold, and the results of optimization of
again this increase is greater than 0.5 MW and the topology each topology in the search tree at a ® xed furnace duty of
change is therefore accepted. Although the three exchangers 99.6 MW are also shown. The series topology in the search
are initially placed in parallel with each other as shown in tree is the topology that the default strategy would have
Figure 25, the ® nal split con® guration is determined during produced if stream splitting had been forbidden, and this
the optimization stage. topology is shown in Figure 28. This series design can
The search for modi® cations could be stopped at this be compared with the other designs involving stream splits,
point, and the resulting topology submitted to the optimiza- to assess the bene® t of allowing stream splitting in the
tion stage. If, however, the search is continued for another retro® t HEN design. A close look at the exchanger area
step, the next modi® cation selected is the addition of a new requirements of each topology for a furnace duty of 99.6 MW
exchanger. This further increases the Rm a x by 3.9 MW, to reveals that the topology modi® cations selected by the default
give a minimum furnace duty of 92.4 MW. It is not possible strategy consistently reduce the exchanger area requirement
at this stage to determine whether the addition of the new of the HEN. This proves that although the search objective
exchanger can be economically justi® ed, but the optimiza- does not explicitly consider area requirement, it nonethe-
tion stage provides the necessary clari® cation. After optimi- less selects topology changes which reduce exchanger area
zation at a ® xed furnace duty of 99.6 MW, the topology requirement.
without the new exchanger required 1974 m2 of additional The design produced by using the default strategy (Fig-
exchanger area, whilst the one with the new exchanger ure 26b) is selected to compare with the designs produced
required only 1265 m2 . The two retro® t HENs are shown in by Ahmad et al.1 7 and Shokoya and Kotjabasakis5 , which
Figures 26a and 26b and it can be seen that both feature the are shown in Figures 29 and 30 respectively. This com-
same series-parallel split con® guration. parison is shown in Table 1, and reveals that the new method
With this information, the two design options can be requires the least number of modi® cations to the HEN, and
effectively compared to identify the `best’ HEN. The instal- requires exchanger area which is only marginally larger than
lation cost of the new exchanger can be fairly accurately the area matrix method.
estimated at this stage to assist in the decision process,
Figure 28. Design CÐ the series solution. Figure 29. Retro® t design by Ahmad et al.17.
New 3 2 1
ADDRESS
Resequenced 1 2 1 Correspondence concerning this paper should be addressed to Dr. X. X.
Repiped 1 0 0 Zhu, Department of Process Integration, UMIST, PO Box 88, Manchester
Stream Splits 1 (1 ´ 3) 1 (1 ´ 2) 1 (1 ´ 2) M60 1QD, UK.
Existing (+ area) 2 4 5
Additional Area 1990 1257 1265 The manuscript was received 23 November 1995 and accepted for
publication after revision 30 May 1996.