Applied Thermal Engineering 20 (2000) 1589±1605

www.elsevier.com/locate/apthermeng

The heat exchanger network design as a paradigm of

technology integration
L. Tantimuratha, A.C. Kokossis*, F.U. MuÈller
Department of Process Integration, UMIST PO Box 88, Manchester, M60 1QD, UK

Abstract

The work presents evidence from experiments that attempt to combine optimisation technologies of
di€erent qualities and advantages. The evidence is illustrated in relevance to the design problem of heat
exchanger networks (HEN). The HEN problem has been addressed separately by Pinch technology,
mathematical programming, and stochastic optimisation. Pinch methods account for conceptual
principles driven by an overall understanding of the problem. Mathematical programming applications
have spanned across the entire spectrum of formulations and include LP, MILP, NLP, and MINLP
problems. Stochastic optimisation is presented mostly in the form of simulated annealing and genetic
algorithms. Considered in isolation, the di€erent optimisation methods are both blessed and ba‚ed in
that they simultaneously enable and disable particular features. The paper presents scienti®c experiments
as a paradigm for the development of an integrated methodology with a purpose to restrain limitations,
compound merits and produce a syncretism of complimentary venues for synthesis and
optimisation. 7 2000 Elsevier Science Ltd. All rights reserved.

Keywords: Heat exchanger network; Stochastic optimisation; Simulated annealing; Technology integration

1. Introduction

A large number of applications have been presented for the design problem of heat recovery
networks. Thermodynamic approaches in the form of Pinch analysis have been ®rst introduced in
the late 1970s with the idea of setting targets prior to design [8]. Pinch analysis reports signi®cant

* Corresponding author. Tel.: +44-161-200-4384; fax: +44-161-236-7439.

E-mail address: [email protected] (A.C. Kokossis).

1359-4311/00/$ - see front matter 7 2000 Elsevier Science Ltd. All rights reserved.
PII: S 1 3 5 9 - 4 3 1 1 ( 0 0 ) 0 0 0 2 5 - 9
1590 L. Tantimuratha et al. / Applied Thermal Engineering 20 (2000) 1589±1605

changes in energy savings and has established a track record of numerous successful applications
in the Chemical Process Industries. Targets are set for the minimum utility consumption, the
minimum heat transfer area of the network, and the minimum number of units. The targets span
di€erent driving forces for heat transfer (minimum approach temperatures DTmin † and include
shortcut calculations as available from a conceptual representation of the network. The individual
targets are further composed into aggregate, ``super'' targets that account for the cumulative cost
of the heat exchanger network. The minimum cost determines the optimal level of driving forces,
DTmin : Although the development of the targets is ecient and straightforward, the development
of appropriate designs is less evident. The idea to develop choices ``close to the targets'' eliminates
rather than to determine choices. The network design entails a great amount of manual
developments that is particularly tedious for large industrial problems.
Mathematical programming strikes a di€erent venue to the design problem. Principles are left
aside and mathematical models are formulated in the form of optimisation problems (LP, MILP,
NLP, or MINLP's). The applications include methods with and without decomposition. The
problem decomposition is motivated by the need to simplify the optimisation. Its most successful
version has been proposed in the form of a MILP±NLP approach: the MILP model determines
targets and matches while the NLP optimises the heat transfer areas and the overall cost [14,7].
Methods without decomposition [20] have resulted in dicult problems where the simultaneous
optimisation (i.e. the simultaneous optimisation of matches and heat transfer area) is traded
against complications in solving a more dicult problem. The network development is fully
determined by the mathematical programming approaches. However, the burden to the
optimisation is signi®cant and the robustness of the applications is open to doubt.
Stochastic optimisation is recently introduced as an alternative approach that improves
robustness at the expense of additional computational times. Simulated annealing (SA) is a general
stochastic technique able to handle large problems without being limited by the non-linearities,
the non-convexities and the discontinuities of the models. The ®rst chemical engineering
applications include pressure relief networks and HEN [6]. The minimum cost layouts are
developed from evolutions of the stream network, the heat loads and heat transfer area and the
consideration of di€erent matches and units. The random search accepts and rejects moves using
standard (and general purpose) criteria. The method apparently yields a ®nal network but is more
demanding in computation than mathematical programming. Although very robust and general,
simulated annealing burrows its way through the ``stochastic'' optimum with the use of seer
computing power rather than insights and engineering knowledge.
This work is motivated by the complimentary nature of the di€erent methods, especially as
this is re¯ected on their several drawbacks and advantages. The paper presents scienti®c
experiments as a paradigm for the development of an integrated methodology with a purpose
to restrain limitations, compound merits and produce a syncretism of complimentary venues
for synthesis and optimisation.

2. Technology integration

The work advocates the combination of di€erent design methodologies. In particular, the
integration experiments consider:
 L. Tantimuratha et al. / Applied Thermal Engineering 20 (2000) 1589±1605 1591

. Stochastic optimisation in the form of a SA algorithm. The algorithm applies the general-
purpose developments by Nielsen et al. [13] that searches with permutations on matches,
streams and heat transfer loads. The search and the annealing algorithm is here adjusted
with the use of the other approaches.
. Mathematical programming based on the MILP formulations proposed by Briones and
Kokossis [3±5]. These constitute models that determine e€ective matches ahead of design and
determine pools of primal solutions one may (or should) consider as a ®rst priority. The
mathematical developments include the ATM model (grassroots design) and the TAME
model (retro®t design).
. Pinch analysis is made available from Supertargets in the form of cost targets. The
Supertargets account for overall costs developed also ahead of design. The costs are often
very close to the optimum and can be used to assess the ``distance from the optimum'' as the
optimisation progresses.
As hinted earlier, the SA by Nielsen et al. [13] is a reference point to signify the general-
purpose development. Subsequent experiments are developed where additional information is
processed into the annealing routines using Pinch Analysis and mathematical programming.

2.1. Integration with mathematical programming

The MILP models furnish information about appropriate and e€ective matches. The
mathematical formulations consider criss-crossing transfer and select matches that minimise
deviations from the targets. In retro®t problems, the models evaluate (assess) existing units,
required investment for new units, and modi®cations. Grassroots and retro®t models are part
of an overall methodology proposed for the analysis of HEN. The methodology allows
engineers to interact and obtain simpler and practical networks [3±5]. The solution includes
information about optimal networks whose exact load distribution requires further
optimisation using an NLP model. The models are used in two di€erent venues that are
explained separately.

2.1.1. Initialisation
SA needs a feasible initial con®guration. This network is further progressed randomly during
the course of the optimisation. Good estimates of the structure are readily available from the
match combinations of the solutions of the MILP models. The MILP models yield preferable
matches and their loads, not the network con®guration. A ``quick and dirty'' network can be
provided using an NLP model. Although not optimal, the initial network is a sensible choice to
start and requires minimal e€ort.

2.1.2. Design pools

Also inexpensive is the set (group) of promising matches. Indeed, the solutions of the MILPs
can be used to produce a pool of promising matches. Instead of selecting matches at random,
the SA can concentrate on selected pairs. The result is a reduced solution space and fewer
combinations. The design pool can alternatively be introduced as an ``option'' with a large
bias. In other words, choices from the pool can be favoured (i.e. by 70%) against matches in
1592 L. Tantimuratha et al. / Applied Thermal Engineering 20 (2000) 1589±1605

complete randomness (say with a probability of 30%). Experiments are presented here for a set
of biases that ranges from 100%, 70%, 50%, and 30%.

2.2. Pinch analysis and simulated annealing

Provided the cost estimates (``targets'') from Pinch Analysis are suciently close to the
optimum, SA can (and should) monitor its current distance from the optimum. The monitoring
concerns the development of the dynamic Markov chain and the annealing schedule.
Parameters associated with these two aspects can be adjusted accordingly. Ideas regarding the
systematic exploitation of this information are presented below.

2.2.1. Dynamic chains

The Markov chain length strongly in¯uences the computational time required in order to
converge each SA run. Smaller values account for signi®cant acceleration but at the possible
expense of a low quality result. Long chains are used to attain equilibrium and are dictated by
the need to ascertain optimal results. However, as the ``destination'' is known in the form of a
target, the occasional distance from the target is an excellent compass for adjusting the chains.
At each annealing temperature, the Markov length is updated as a function of the distance
between the cost target and the best cost found so far:
 
jDCj
†
Markov Length MKL † ˆ f DC ˆ MKLmin exp g 1†
DCmax
The above function suggests that the smaller the distance, the shorter will be the Markov
length. Fig. 1 illustrates the shape of the function with respect to changes in the cost, DC:
MKLmin and MKLmax are the minimum and maximum values of the Markov chain length,
g ˆ ln MKLmax =MKLminw †, DC ˆ C ÿ C target , DCmax ˆ C util ÿ C target , and C, C target , and C util are,
respectively, the total costs from the preceding annealing temperature, the pinch analysis, and
the simplest network. Furthermore, the SA process terminates provided the cost becomes less
than a pre-appointed (default is 90%) fraction of the cost target. With the reduction of the
Markov length, a better speed is expected for the annealing process.

Fig. 1. The function of Markov length and DC:

 L. Tantimuratha et al. / Applied Thermal Engineering 20 (2000) 1589±1605 1593

2.2.2. Tailored annealing schedules

The annealing schedule has an additional e€ect on the convergence speed of the algorithm.
The widely employed logarithmic schedule applies general-purpose analysis that assumes
speci®c estimates for the optimum. These estimates are certainly much less accurate than the
thermodynamic targets available from Pinch Analysis. Therefore, it is natural to upgrade the
primal parameters of the annealing schedule in order to embed the additional information. The
two empirical formula applied in this work are presented in Eqs. (2) and (3):
n†
C
Schedule S1 b i‡1† ˆ a i†
b n† 2†
C

C iÿk † ÿ C n† i †
Schedule S2 b i‡1† ˆ b ; kˆ2 3†
C i † ÿ C n†
The annualised total cost target can be used for de®ning a convergence criterion and for
controlling the annealing temperatures.
(a) The parameter a in S1 is introduced to ascertain the initial annealing temperature, b 0† , is
high enough for a complete randomisation of the network. The ®nal annealing temperature,
b n† , is arbitrarily set to 1; at this level uphill moves terminate and the optimisation stops. As
long as C i † is larger than C iÿ1† , the new temperature is set to the standard exponential
schedule, b i‡1† ˆ 0:95b i † :
(b) The schedule, proposed as S2 in Eq. (3), is completely independent and ¯exible in speed.
The cost value C iÿk† is updated from past values of Markov chains that in turn are
dynamically updated throughout the optimisation process.
The basic assumption in both schedules is that the annealing temperature is slightly lower than
the equilibrium temperature that corresponds to the cost. The experiments have been
conducted separately for Sections 2.1 and 2.2 before they were combined at a single stage. The
latter intertwine
. initial networks from the MILPs
. pools of promising matches from the targeting models (100% bias)
. customised annealing schedules
. dynamic Markov chains.

Remarks. Base cases are the stand-alone applications of SA. Experiments are carried out to
identify gains in
(i) Speed: The time is measured using the internal clock in the computer. The measurement
starts when the SA starts and stops as the process terminates.
(ii) Quality of solutions: The quality of the solutions is assessed at two di€erent fronts:
performance and simplicity of the solution. The performance is assessed with respect to the
attained level of the objective (namely total cost). The number of splits and bypasses values
the network simplicity.
1594 L. Tantimuratha et al. / Applied Thermal Engineering 20 (2000) 1589±1605

Table 1
Process stream and cost data of Example 1a,b,c,d

Stream FCP (kW/8C) Tin (8C) Tout (8C) h (kW/m2.8C)

H1 100 327 40 0.5

H2 160 220 160 0.4
H3 60 220 60 0.14
H4 400 160 45 0.3
C1 100 100 300 0.35
C2 70 35 164 0.7
C3 350 85 138 0.5
C4 60 60 170 0.14
C5 200 140 300 0.6
HU 330 250 0.5
CU 15 30 0.5
a
Optimal DTmin ˆ 25 8C.
b
Cost model of heat exchangers ($/year): a + b
Areac; a = 2000; b = 70, c = 1.0 (Area unit of m2).
c
Cost of utilities: HU (steam) = 60.0 $/kW year; CU (cooling water) = 6.0 $/kW year.
d
Economic data: rate of interest =0%; Project life time = 5 years.

3. Presentation of results
Example 1. The example is a medium size problem with nine process streams and two utilities.
The data is shown in Table 1. The average costs and the computational times for each case are
shown in Fig. 2.
(a) Technology integration improves the speed up to 85% while at the same time the quality
of the solution essentially remains intact (increases run up to +5%).
(b) The use of ``pools'' signi®es the most compelling integration step. Best improvements are
reported for cases where 100% of the choices are developed from the primal pools. The
simple transhipment model works equally well with the more sophisticated area target
models.
(c) The customisation of the Markov length does not disclose signi®cant improvements.
(d) The use of large values for the biases associated with the pools yield simpler networks.
Figs. 3 and 4 illustrate networks that are developed from base cases (i.e. only SA) as against
biased runs. The biased runs typically save 50% with respect to the units involved.

Example 2 The problem consists of 15 process streams and represents a threshold problem.
The threshold temperature is at 68C. The problem data is given in Table 2.
The optimised costs (average values) and the computational time for each case is shown in
Fig. 5. The results indicate that:
(a) The execution time has been reduced to 84% compared to the base case. For the majority of
runs, the costs have not been found to deviate much from the base case (up to +10%).
 L. Tantimuratha et al. / Applied Thermal Engineering 20 (2000) 1589±1605 1595

Fig. 2. Average best cost and time consumed for each case of Example 1.

(b) The most signi®cant improvement in speed is again reported for biased selections from
the primal pools: the higher the bias the better the results. The 100% bias yields reductions
to 68% and 84% as the pools are respectively drawn from the transhipment and the Area
Target Model. The very same scenarios yield the best deviations from the base case (up to
+1.7%).
(c) Networks from biased pools are also simpler. This has been observed earlier in Example
1. Now the case is more evident as one can compare the networks from Figs. 6 and 7. The
network of Fig. 6 involves 20 more heat exchangers, many of them with a small duty and
rather ambiguous merits.
(d) Simplicity is only accomplished with ``pools''. Although acceleration can be observed
using Pinch Analysis, the networks still appear with complicated layouts.

Example 3 A retro®t problem is illustrated next. The problem data are shown in Table 3. The
existing network is given in Fig. 8. Integration experiments have only been made at the
initialisation stage. The matches are obtained by the screening models (TAME model)
proposed by Briones and Kokossis [3±5]. The base case network is shown in Fig. 9.
1596 L. Tantimuratha et al. / Applied Thermal Engineering 20 (2000) 1589±1605

Fig. 3. The network solution of SA stand-alone case of Example 1.

Fig. 4. The network solution of 100% biased transhipment pool of Example 1.

 L. Tantimuratha et al. / Applied Thermal Engineering 20 (2000) 1589±1605 1597

Table 2
Process stream and cost data of Example 2a,b,c,d

Stream FCP (kW/8C) Tin (8C) Tout (8C) h (kW/m2.8C)

H1 5.3 137 14 4
H2 16.8 135 66 4
H3 3327 111 110 3.1
H4 135.4 105 95 1.4
H5 134.6 76 66 2
H6 928 75 74 3
H7 840 66 65 3
H8 12000 65 64 2
H9 324 37 14 4
C1 56.6 4 112 4
C2 108.1 76 100 0.9
C3 205 70 90 4
C4 18.4 38 85 4
C5 280.1 45 70 4
C6 314 4 45 4
HU 150 149 3
CU 0 2 1
a
Optimal DTmin = 68C.
b
Cost model of heat exchangers ($/year): a + b Areac; a = 0; b = 350; c =1.0 (Area unit of m2).
c
Cost of utilities: HU (steam) = 88.0 $/kW year; CU (cooling water) = 0.0 $/kW year.
d
Economic data: rate of interest = 0%; Project life time = 5 years.

(a) Fig. 10 illustrates the acceleration in the speed as mathematical programming is

integrated with SA. The time is reduced by 33% as compared to the base case. The network
costs are also improved.
(b) Complex networks are developed as a standard result. The complexity is accounted to
the fact that many changes are allowed on the existing network. Unlike the grassroots case,
retro®t problems are not expected to a€ord large deviations and the need to use ``primal
pools'' becomes increasingly important.

Example 4 The example illustrates the use of di€erent annealing schedules. The stream data is
shown in Table 4. The design objective is augmented to include the cost of the furnace as a
non-linear expression.
Results are presented in Fig. 11. The schedule S0 is the standard logarithmic schedule. The
solutions for all schedules are good but the computational performances di€er. In all cases the
deviation from the targets is less than 3.5%. The proposed schedules S1 and S2 both resulted
in faster drops of the annealing temperatures. The drop has been particularly signi®cant for
1598 L. Tantimuratha et al. / Applied Thermal Engineering 20 (2000) 1589±1605

Fig. 5. Average best cost and time consumed for each case of Example 2.
L. Tantimuratha et al. / Applied Thermal Engineering 20 (2000) 1589±1605 1599

Fig. 6. The network solution of SA stand-alone case of Example 2.

1600 L. Tantimuratha et al. / Applied Thermal Engineering 20 (2000) 1589±1605

Fig. 7. The network solution of 100% biased transhipment pool of Example 2.

 L. Tantimuratha et al. / Applied Thermal Engineering 20 (2000) 1589±1605 1601

Table 3
Process stream and cost data of Example 3a,b,c

Stream FCP (kW/8C) Tin (8C) Tout (8C) h (kW/m2 8C)

H1 5.3 137 14 4
H2 16.8 135 66 4
H3 3327 111 110 3.1
H4 135.4 105 95 1.4
H5 134.6 76 66 2
H6 928 75 74 3
H7 840 66 65 3
H8 12000 65 64 2
H9 324 37 14 4
C1 56.6 4 112 4
C2 108.1 76 100 0.9
C3 205 70 90 4
C4 18.4 38 85 4
C5 280.1 45 70 4
C6 314 4 45 4
HU 150 149 3
CU 0 2 1
a
Cost model of extra area ($/year): b
Areac; b = 70, c =1.0 (extra) Area unit of m2).
b
Cost of utilities: HU (steam) =88.0 $/kW year; CU (cooling water) = 0.0 $/kW year.
c
Economic data: rate of interest = 0%; Project life time = 5 years.

Table 4
Process stream and cost data of Example 4a,b,c,d

Stream FCP (kW/8C) Tin (8C) Tout (8C) h (kW/m2 8C)

H1 7.913 630.6 338.9 1.6097

H2 5.803 583.3 505.6 0.7265
H3 1.374 555.6 319.4 0.5586
H4 31.65 494.4 447.2 0.7299
H5 6.331 477.8 311.1 0.3541
H6 65.943 422.2 383.3 0.4848
C1 24.8 288.6 650 0.8753
Flue gas 1800 650 ±
CW 300 333 0.6324

a
Optimal DTmin = 58C.
b
Cost model of heat exchangers ($/year): a + b Areac; a = 0; b = 1300; c = 0.6 (Area unit of m2).
c
Cost of utilities: Fuel = 174 $/year; furnace cost = 321.1 Q 0.7 $/year; CU (cooling water) = 4.63 $/kW year.
d
Economic data: Project life time = 5 years.
1602 L. Tantimuratha et al. / Applied Thermal Engineering 20 (2000) 1589±1605

Fig. 8. The existing network of Example 3.

schedule S2. The CPU time is consequently improved by 63% for S1 and by 80% for S2. The
developed networks feature similar qualities and are equally simple.

4. Conclusion

The work presents evidence from attempts to combine optimisation technologies of di€erent
qualities and advantages. Research to integrate merits of complementary technologies has been
presented for utility networks [9±12], water minimisation applications [1,2] and complex
 L. Tantimuratha et al. / Applied Thermal Engineering 20 (2000) 1589±1605 1603

separation systems [15±19]. The evidence is illustrated here in relevance to the design problem
of HEN. The experiments illustrate that SA platforms can signi®cantly be improved with
simple modi®cations that employ knowledge available for the problem (Pinch Analysis) or
knowledge readily obtained from an inexpensive pre-analysis of the problem (screening MILP

Fig. 9. The network solution of SA stand-alone case of Example 3.

1604 L. Tantimuratha et al. / Applied Thermal Engineering 20 (2000) 1589±1605

Fig. 10. Average best cost and time consumed for each case of Example 3.

Fig. 11. Comparison of the annealing schedules S0 (logarithmic schedule), S1, and S2.

models). The improvements are quanti®ed at di€erent levels: (i) as acceleration in the
computational experiments, (ii) as improvement in the optimisation objective, and (iii) as
improvements in the quality of the solution (i.e. simpler solutions). Intelligent implementations
of SA require little e€ort to adjust the moves, the schedule, and the pool of the available
options. The reported improvements explain the merits in integrating technologies and signal
the need for further work along the proposed lines of research.

References

[1] A. Alva-Argaez, A. Kokossis, R. Smith, Wastewater minimisation of industrial systems using an integrated
approach, Comp. Chem. Engng. 22 (1998) S741±744.
 L. Tantimuratha et al. / Applied Thermal Engineering 20 (2000) 1589±1605 1605

[2] A. Alva-Argaez, A. Valianatos, A. Kokossis, A multicontaminant transhipment model for mass exchanger
network synthesis and wastewater minimisation problems, Comp. Chem. Engng. (1999) (in press).
[3] V. Briones, A. Kokossis, Hypertargets: a conceptual programming approach for the optimisation of industrial
heat exchanger networks. Part-I: Grassroots design and network complexity, Chem. Engng. Sci. 54 (1999) 519±
539.
[4] V. Briones, A. Kokossis, Hypertargets: a conceptual programming approach for the optimisation of industrial
heat exchanger networks. Part-II: Retro®t design, Chem. Engng. Sci. 54 (1999) 541±561.
[5] V. Briones, A. Kokossis, Hypertargets: a conceptual programming approach for the optimisation of industrial
heat exchanger networks. Part-III: Industrial applications, Chem. Engng. Sci. 54 (1999) 685±706.
[6] W.B. Dolan, P.T. Cummings, M.D. Le Van, Process optimization via simulated annealing: application to
network design, AIChE 35 (1989) 725±736.
[7] C.A. Floudas, A.R. Ciric, I.E. Grossmann, Automatic synthesis of optimum heat exchanger network
con®gurations, AIChE 32 (1986) 276±290.
[8] B. Linnho€, Pinch analysis Ð a state of the art overview, Trans. IChemE 71 (1993) 503±522.
[9] S. Mavromatis, A. Kokossis, Hardware composites: a new conceptual tool for the analysis and optimisation of
steam turbine networks in chemical process industries. Part I: Principles and construction procedure, Chem.
Engng. Sci. 53 (7) (1998) 405±1434.
[10] S. Mavromatis, A. Kokossis, Hardware composites: a new conceptual tool for the analysis and optimisation of
steam turbine networks in chemical process industries. Part II: Application to operation and design, Chem.
Engng. Sci. 53 (7) (1998) 1435±1461.
[11] S. Mavromatis, A. Kokossis, Conceptual optimisation of utility networks for operational variations. Part I:
Targets and level optimisation, Chem. Engng. Sci. 53 (8) (1998) 1585±1608.
[12] S. Mavromatis, A. Kokossis, Conceptual optimisation of utility networks for operational variations. Part II:
Network development and optimisation, Chem. Engng. Sci. 53 (8) (1998) 1609±1630.
[13] J.S. Nielsen, M.W. Hansen, S.B. Joergensen, Heat exchanger network modelling framework for optimal design
and retro®tting, Comp. Chem. Engng. 20 (1996) S249±S254.
[14] S.A. Popoulias, I.E. Grossmann, A structural optimization approach in process analysis. Part II: Heat recovery
networks, Comp. Chem. Engng. 7 (1983) 707±721.
[15] P. Shah, A. Kokossis, Design targets for separator and reactor±separator systems using conceptual
programming, Comp. Chem. Engng. 21 (1997) S1013±S1018.
[16] P.B. Shah, A. Kokossis, Systematic optimisation technology for high-level screening and scoping of complex
distillation systems, Comp. Chem. Engng. 23 (1999) S113±116.
[17] P.B. Shah, A. Kokossis, Novel designs for ethylene cold-end separation using conceptual programming
technology, Comp. Chem. Engng. 23 (1999) S895±898.
[18] P. Shah, A. Kokossis, A systematic framework for the design of subambient processes, Ind. Eng. Chem. Res.
(1999) (submitted).
[19] P. Shah, A. Kokossis, Systematic synthesis and optimisation of complex distillation processes. AIChE. (1999d)
(submitted).
[20] T.F. Yee, I.E. Grossmann, Z. Kravanja, Simultaneous optimization models for heat integration: Part I, II, III,
Comp. Chem. Engng. 14 (1990) 1151±1200.