Sunteți pe pagina 1din 26

Applied Thermal Engineering 51 (2013) 801e826

Contents lists available at SciVerse ScienceDirect

Applied Thermal Engineering


journal homepage: www.elsevier.com/locate/apthermeng

Optimal heat exchanger network synthesis: A case study comparison


Marcelo Escobar*, Jorge O. Trierweiler
Chemical Engineering Department, Federal University of Rio Grande do Sul (UFRGS)-Brazil Group of Intensification, Modeling, Simulation,
Control and Optimization of Processes (GIMSCOP), R. Eng. Luiz Englert, s/n, 90040-040 Porto Alegre, RS, Brazil

h i g h l i g h t s

< We compare approaches for Heat exchanger network synthesis.


< Systematic initialization strategies are proposed.
< Comparison with other solutions from the literature.
< The suitable solvers and approaches are discussed.
< Simultaneous SYNHEAT model provided the better performance.

a r t i c l e i n f o a b s t r a c t

Article history: Heat exchanger network synthesis (HENS) is one of the most extensively studied synthesis/design
Received 28 May 2012 problem in chemical engineering. This is attributed to the importance of determining energy costs and
Accepted 13 October 2012 improving the energy recovery in chemical processes. Several synthesis procedures for heat exchanger
Available online 29 October 2012
networks have been published in the literature. However, only a modest number can be implemented as
computer tools for the full automatic generation of network configurations. In spite of the computational
Keywords:
methods evolution, some difficulties still need to be overcome. For instance: computational time, and
Heat exchanger network
problem convergence. In this work, the main approaches to solve the problem of heat exchanger network
Optimization methods
Heat integration
synthesis, sequential and simultaneous, were presented and implemented using the General Algebraic
Modeling System (GAMS). Initialization strategies for generating feasible starting points are proposed.
Five case studies with different dimensions, from four to 39 streams, are used as background to compare,
and analyze the different approaches. The results showed the efficiency of the initialization strategies.
Furthermore, the suitable solvers are also analyzed through solving each subproblem. Simultaneous
approaches presented better performance than the sequential one. Finally, the main challenges involved
in using different approaches is discussed to synthesize heat exchangers networks.
 2012 Elsevier Ltd. All rights reserved.

1. Introduction configurations has received considerable attention in the literature


(see Gundersen and Naess [1] and Furman and Sahinidis [2] for
Heat exchanger network synthesis (HENS) has been the subject good reviews).
of a significant amount of research over the last 40 years. It The basic HENS problem was stated by Furman and Sahinidis [2]
represents one of the major components in a chemical processing as follows:
system, because it is strongly related to the energy efficiency of
Given a set of hot process streams HP, which should be cooled from
the process. The heat exchanger network (HEN) has the task of
its supply temperature to its target temperature; a set of cold
integrating the hot and the cold process streams in a process
process streams CP, which should be heated from its supply
reducing the amount of heating and cooling utilities that are
temperature to its target temperature; the heat capacities and flow
required. The problem of synthesizing optimal network
rates of the hot and cold process streams; the utilities available
(e.g., hot utilities HU and cold utilities CU) and their corresponding
* Corresponding author. Tel.: þ55 51 3308 4072; fax: þ55 51 3308 3277.
temperature and costs; develop a heat exchanger network with the
E-mail addresses: escobar@enq.ufrgs.br, escobar029@hotmail.com (M. Escobar), minimum annualized investment and annual operating costs, i.e.
jorge@enq.ufrgs.br (J.O. Trierweiler). minimum Total Annual Cost (TAC).

1359-4311/$ e see front matter  2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.applthermaleng.2012.10.022
802 M. Escobar, J.O. Trierweiler / Applied Thermal Engineering 51 (2013) 801e826

Floudas [3] mentioned that despite the great economic impor-


tance of HENS, it features a number of key difficulties that are
associated with handling: (i) the combinatorial nature of the
problem for identifying the best pairs of hot and cold streams so as
to enhance energy recovery; (ii) forbidden, required, and restricted
matches; (iii) temperature dependence of physical and transport
properties; (iv) different types of streams (e.g. liquid, vapor, and
liquidevapor); and (v) the choice of flow configuration, materials of
construction, and different pressure ratings.
The extensive research effort during the last decades toward
addressing these aforementioned difficulties/issues exhibit varia-
tions in their objectives and solution approaches. The first
approaches during 1960s and early 1970s treated the HEN synthesis
problem as a single task. The main proposed techniques presented
limitations on the theoretical and algorithmic aspects of optimi-
zation that were the bottleneck in expanding the applicability of
the mathematical approaches at that time [3].
The work of Hohmann [4], which received little recognition in
the 1970s, and the work of Linnhoff and Flower [5] introduced the
concept of the pinch point. As a result, the research emphasis
shifted for decomposing the original problem into three separate
tasks: (i) minimum utility cost/consumption; (ii) minimum
number of matches; and (iii) minimum investment cost. Further-
more, HEN synthesis approaches based on decomposition were
developed via thermodynamic [6], and optimization approaches
[7e11]. Fig. 2. Steps of decomposition based HEN synthesis approach.
The main advantage of HEN synthesis approaches based on
decomposition into targets is that they involve simpler subprob-
lems that can be treated in a much easier fashion than the original
single-task problem. However, it may introduce some limitation In Fig. 1 a chronological overview of HEN classical approaches is
since the trade-offs between the utility consumption, the number presented. In 70’s the Pinch Technology was introduced. The
of matches, and the minimum investment cost are not take into sequential approach using in the mathematical programming was
account appropriately and it may result in HEN designs that are developed in the 80’s [9]. Lastly, in the 90’s simultaneous synthesis
suboptimal networks [3]. approaches were proposed [10,11].
Due to the limitations presented by the sequential approach [9], In this work, the main classical approaches available in the
researches in the late 1980s and early 1990s focused on simulta- literature to handle the HENS using mathematical programming
neous approaches [10,11] that attempt to treat the synthesis as were implemented in the modeling system GAMS. Heuristic
a single-task problem considering all trade-offs simultaneously. initialization strategies were developed to converge and compare
These approaches result in more complex MINLP models, however, these approaches. In addition, it was quantitatively evaluated the
advances in theoretical and algorithmic aspects of optimization suitable solvers, comparing the performance for solving each
have allowed large scale MINLP problems to be solved. problem and the quality of the solution obtained. The main idea of

70's 80's 90's 2000's

Pinch Technology Sequential Synthesis Simultaneous Synthesis


Hohmann (1971) Synthesis with Synthesis without
Decomposition Decomposition
Linhoff &Flower (1978) Data Data

Data LP- Transshipment Model ΔTmin


ΔTmin MINLP-Hyperstructure
Minimum Papoulias and Grossmann
ΔTmin Ciric and Floudas (1991)
Utility Cost (1983)

Minimum
Targets:
Minimum MILP- Transshipment Model Total Annual Cost
Number of Papoulias and Grossmann (1983)
1-Minimum Utility Cost
Matches
2-Minimum Number of Matches MINLP- Synheat
3-Minimum Investment Cost Yee and Grossmann
NLP- Superstructure (1990)
Heuristic
Minimum Floudas, Ciric and
Rules Investment Cost Grossmann (1986)

HEN Configuration
HEN Synthesis

HEN Configuration

HEN Evolution

Fig. 1. Historical panorama of classical HEN approaches over the last decades.
M. Escobar, J.O. Trierweiler / Applied Thermal Engineering 51 (2013) 801e826 803

Fig. 3. Graphical representation of the Network Superstructure for hot stream i.

this work was not to beat the best solutions reported in the liter- The overall procedure is sketched in Fig. 2. These three
ature for each case study, but comparing the main difficulties and subproblems are sequentially solved for given process and utility
challenges and also in practical terms, which one is the most data, and a minimum temperature approach (DTmin). The proper
suitable and reliable approach using the current solvers package selection of this parameter is very important because it defines the
available, most of them using deterministic approaches. The trade-off between the capital and energy cost. In the literature,
remainder of this work is organized as follows. Subsection 1.1 typical values for different processes are reported. The optimal
presents the mathematical formulation of the optimization prob- selection, resulting in the HEN with minimum Total Annual Cost
lems. In Section 2 the proposed strategy to solve the different (TAC), may be performed solving the sequential procedure for
optimization problems is presented. Five case studies with different different values.
dimensions are given in Section 3. Results, analysis and discussions Papoulias and Grossmann [7] proposed two linear formulations
are presented in Section 4. Finally, final remarks and conclusions based on transshipment model, problems (P1) and (P2). These
are given in Section 5. models are based on energy balances between hot streams i, cold
streams j, at temperature interval k. The variables Qij denote the
heat load for the match between hot stream i, and cold stream j.
1.1. Approaches for heat exchanger network synthesis CU and Q HU are the cold and hot utility load of each stream at
Qj;k i;k
interval k, while the variable Rk is the residual heat available at each
One of the most prevalent methods for solving the HENS
interval k. In (P2) binary variables yij are defined to denote the
problem is the sequential synthesis method [2]. The basic idea is to
existence of the match between hot stream i, and cold stream j
decompose the problem into three subproblems: (P1) the
resulting in an MILP problem.
minimum utility cost; (P2) the minimum number of matches; and
(P3) the minimum cost network. These problems are then solved
according to the heuristic of finding the minimum cost network
subject to the minimum number of matches, which is subject to the + LP Transshipment Model e Papoulias and Grossmann [7]:
minimum utilities cost [12].

8 PP PP
>
> min CU þ
Ccuj Qj;k HU
Chui Qi;k fMinimum Utility Costg
>
>
>
> k j k i
>
>
>
> subject to
< P P P H P C
Rk  Rk1  CU þ HU ¼ Qi;k  fEnergy Balance of the k  th temperature intervalg
ðP1Þ: Qj;k Qi;k Qj;k
>
>
>
>
j˛CUk i˛HUk i˛HPk j˛CPk
>
>
>
>
CU  0;
Qj;k HU  0;
Qi;k Rk  0 fNonnegativity Constraintsg
>
>
:
R0 ¼ Rk ¼ 0 fAssignmentsg
804 M. Escobar, J.O. Trierweiler / Applied Thermal Engineering 51 (2013) 801e826

+ MILP Transshipment Model e Papoulias and Grossmann [7]:

8 P P
>
> min yij fMinimum number of matchesg
>
>
>
>
i˛HPWHU j˛CPWCU
>
> subject to
>
> P
>
> R  R þ Qijk ¼ Qik H ; i˛HP fEnergy balance for each hot process stream and each interval kg
>
> i;k i;k1 k
>
> j˛CPk WCUk
>
> R R P
>
> i;k1 þ Qijk ¼ Qik H ; i˛HU fEnergy Balance for each hot utility and each interval kg
>
> i;k k
>
> j˛CPk
>
> P
>
> Qijk ¼ Qjk C ; j˛CP fEnergy balance for each cold process stream and each interval kg
>
>
k
< i˛HPk WHUk
P
ðP2Þ: C
Qijk ¼ Qjk ; j˛CUk fEnergy balance for each cold utility and each interval kg
>
>
>
>
j˛HPk
P
>
> Rk  Rik ¼ 0 fDefinition of total residual flows at each intervalg
>
>
>
> j˛HPk WHUk
>
> P
>
> Qij ¼ Qijk ¼ 0 fDefinition of the heat load of each matchg
>
>
>
> k
>
> 0  Qij  Ubij yij fUpper bound constraints for heat loadsg
>
>
>
> Qijk  0; Rik  0 fNonnegativity constraintsg
>
>  
>
>
>
> Integrality conditions
: yij ¼ 0  1
for binary variables

The third subproblem, problem (P3), is a NLP and its solution mass and energy balances, the design equations for each heat
provides the network configuration with minimum investment exchanger (A ¼ Q/U.LMTD), and feasibility constraints for each heat
cost determining the optimal: (i) stream interconnections (e.g. exchanger, i.e. temperature difference at each heat exchanger
series, parallel); (ii) stream heat capacity flowrates (f); (iii) terminal greater than DTmin.
temperatures (T); and (iv) areas of heat exchangers units (A). The
model is based on the superstructure proposed by Floudas et al. [9] + NLP Superstructure Model e Floudas et al. [9]:
depicted in Fig. 3. All possible interconnections must satisfy the

8
>
>
>
>
>
>
>
> !bi;j
>
>
>
> P
nh P
nc Qi;j
>
> min ci;j fMinimum investment cost superstructureg
>
> Ui;j LMTDi;j
>
> i˛HPWHU j˛CPWCU
>
>
>
>
subject to
>
> P 0
>
>
>
>
I
fm;m0 ¼ Fm ; m˛HCT fMass balance for the initial split of each stream mg
>
>
>
>
m
>
> P
00
>
> I
fm;m0 þ
B
fm;m E
0 ;m00  fm;m0 ¼ 0; m˛HCT fMass balance for the mixer before heat exchanger m  m0 g
>
>
>
> m
>
>
>
> P
00
>
> o
fm;m0 þ
B
fm;m E
0 ;m00  fm;m0 ¼ 0; m˛HCT fMass balance for the split after heat exchange m  m0 g
>
>
>
> m
>
> P00
>
< I
fm;m in
0 ; Tm þ
B
fm;m o E I
0 ;m00 Tm;m00  fm;m0 Tm;m0 ¼ 0; m˛HCT fHeat balance for mixers before heat exchange m  m0 g
ðP3Þ: m
 
>
>
>
> Qi;j  fi;jEH Ti;j
IH  T OH ¼ 0; i˛HP; j˛CP fHeat balance for exchanger i  j hot sizeg
>
> i;j
>
>  
>
>
>
> Qi;j  fi;jEH Ti;j
OC  T IC ¼ 0; i˛HP; j˛CP fHeat balance for exchanger i  j cold sizeg
>
> i;j
>
>
>
> 1 ¼ T IH  T OC ; dt 2 ¼ T OH  T IC ; i˛HP; j˛CP
dti;j fDefinitions of temperature approaches in the exchanger i  jg
>
> i;j i;j i;j i;j i;j
>
>
>
> 1  DT
dti;j min ; dti;j  DTmin ; i˛HP; j˛CP
2 fMinimum temperature approaches constraintsg
>
>
>
>
>
>
1
dti;j 2
 dti;j
>
> LMTDi;j ¼ ! ; i˛HP; j˛CP fLMTD definition for heat exchangersg
>
>
>
>
1
dti;j
>
>
>
> ln
>
>
2
dti;j
>
>
>
>
>
> fmI ; fm;m
E O B
0 ; fm;m0 ; fm;m0 ;m00  0; m˛HCT fNonnegativity constraintsg
>
>
>
>
>
:
M. Escobar, J.O. Trierweiler / Applied Thermal Engineering 51 (2013) 801e826 805

STAGE 1 STAGE 2
H1-C1 H1-C1

H1-CU

H1

C1-HU
H1-C2 H1-C2
C1

C2
H2-C1 H2-C1
C2-HU
H2-CU

H2

H2-C2 H2-C2

Fig. 4. Two-stage superstructure for two hot and two cold streams.

Simultaneous approaches are primarily MINLP formulations. i, and cold stream j, at stage k. The number of stages is a param-
The most popular formulations are the HENS without decompo- eter, often chosen as maximum between the number of hot and
sition of Ciric and Floudas [10], problem (P4), and the simulta- cold streams. It is important to mention that in these models, (P3),
neous synthesis formulation of Yee and Grossmann [11] (SYNHEAT (P4), and (P5), the areas are not treated explicitly as a variable and
model), problem (P5). The former is based on the same super- its expressions (design equation) are substituted in the objective
structure presented in Fig. 3, while the later is based on a stage- function. In addition, the logarithmic mean temperature difference
wise superstructure depicted in Fig. 4. The model of Yee and (LMTD) is replaced by the Chen approximation [13], where no
Grossmann [11] assumes isothermal mixing at each stage, which logarithmic terms are involved in order to avoid numerical
significantly simplifies the model formulation since nonlinear heat difficulties.
balances can be eliminated. For each stream, only an overall heat
balance must be performed within each stage. Binary variables zijk + MINLP Hyperstructure Model e Ciric and Floudas [10]:
are used to denote the existence of the match between hot stream

8
>
>
>
>
>
> !bi;j
>
> P P X X
>
>
nh nc Qi;j
>
> min ci;j yi;j þ Ccuj QjCU þ Chui QiHU fMinimum total annual cost hyperstructureg
>
> Ui;j LMTDi;j
>
> i˛HPWHU j˛CPWCU j i
>
>
>
>
>
> subject to
>
>
>
> ð*Þ fall P2 contraintsg
>
>
>
>
>
> ð*Þ fall P3 constraintsg
>
>
< P HU
ðP4Þ: QiHU ¼ Qik ; i˛HU fhot utility loadg
>
> k
>
>
>
>
>
> P
k
>
> QjCU ¼ CU ; j˛CU
Qjk fcold utility loadg
>
>
>
> k
>
>
>
> fi;jEH  Fi yi;j  0; fi;jEC  Fj yi;j  0; i˛HP; j˛C flogical contraintsg
>
>
>
>
>
>
>
> Qi;j Qi;j
>
> fi;jEH  ; fi;jEC  ; i˛HP; j˛CP ffeasibility constraintsg
>
> DTi;jmax DTi;jmax
>
>
>
>
:
806 M. Escobar, J.O. Trierweiler / Applied Thermal Engineering 51 (2013) 801e826

+ MINLP SYNHEAT Model e Yee and Grossmann [11]:

8
>
> PPP P P P P
>
> min cf zijk þ cf zcui þ cf zhuj þ ccu qcui þ chu qhuj þ
>
>
>
> i j k i j i j
>
> !b !b
>
> X  b X fMinimum total annual cost SYNHEAT Modelg
>
> P P P q q qhuj
>
> c
ijk
þ c cui
þ c
>
> Uij LMTDijk Ucui LMTDcui Uhuj LMTDhuj
>
> i j k
>
>
i j
>
> subject
>
> PP  to 
>
>
>
> q ijk þ q cui ¼ F i Tiin  Tiout ; i˛HP
>
>
>
> PP
i k   fOverall heat balance for each streamg
>
> out  T in ; j˛CP
>
> q ijk þ q huj ¼ F j T
>
>
j j
>
>
j k  
>
> P
>
> qijk ¼ Fi ti  ti k kþ1
; i˛HP
>
>
>
> j
  fHeat balance at each stageg
>
> P
>
> qijk ¼ Fj tjk  tjkþ1 ; j˛CP
>
>
>
> i
>
>
>
> tik¼1 ¼ Tiin ; tjk¼NOK ¼ Tjin fAssignment of superstructure inlet temperaturesg
>
>
>
> tik  tikþ1 ; tjk  tjkþ1
>
>
>
> fMonotonic decrease in temperaturesg
>
> tik¼NOKþ1  Tiout ; tjk¼1  Tjout
>
>
>
> 
>
> qcui ¼ Fi tiNOKþ1  Tiout
>
>  
>
> fHot and cold utility loadsg
>
> qhuj ¼ Fj Tjout  tji
<
ðP5Þ: dtijk  DTmin ; dtcui  DTmin ; dthuj  DTmin fMinimum approach temperature constraintsg
>
> qijk  Uzijk ; qcui  Uzcui ; qhuj  Uzhuj
>
>  
>
>
>
> dt  t k  tk þ G 1  z
>
> ijk i j ijk
>
>  
>
>
>
> dtijk  tikþ1  tjkþ1 þ G 1  zijk
>
>
>
> dtcui  tiNOK  tcu out þ Gð1  z Þ
fLogical constraintsg
>
> cui
>
> in þ Gð1  z Þ
>
> dtcui  Ti  tcu out
>
>  cui
>
> out  t 1 þ G 1  z
>
> dthuj  thu huj
>
>
j
 
>
>
>
>
>
in
dthuj  thu  Tj þ out G 1  zhuj
>
>
>
> dtijk  dtijkþ1
>
> LMTDijk ¼ ! fLMTD definition for heat exchangersg
>
>
>
> dtijk
>
> ln
>
> dtijkþ1
>
>
>
>
>
> dt  dt dtijk  dthuj
>
> LMTDcui ¼
ijk

cui
 ; LMTDhuj ¼ ! fLMTD definition for utility exchangersg
>
> dt
>
> ijk dtijk
>
> ln ln
>
> dtcui dthuj
>
>
>
>
>
> T out  t k  T in ; T in  t k  T out fBoundsg
>
> i i i j j j
>
> qijk ; qcui ; qhuj fNonnegativity constraintsg
>
>
>
> zijk ; zcui ; zhuj ¼ 0  1 fIntegrality conditionsg
>
:

2. Models implementation and initialization strategy modeling applications, and it allows building large maintainable
models that can be adapted quickly to new situations. The opti-
In this work it was made an optimization overview aiming to mization problems (P1) to (P5) briefly described in the previous
select the appropriated modeling system and understand how section were implemented in GAMS 22.2 language. The overview
these problems can be solved (see Fig. 5). The selection of of the problems and the general information flow are given in
a flexible modeling system as well as the selection of suitable Fig. 6. The main goal is to compare qualitatively and quantita-
solvers by which a generated model can be solved is often tively the three different approaches considered in this work,
a critical issue in obtaining valuable results in modeling appli- including numerical difficulties involved in its implementation,
cations. It was used the high level General Algebraic Modeling convergence properties, computational efforts, and the quality of
System (GAMS) for mathematical programming and optimiza- the final results. The best solutions found in this work are
tion. It consists of a language compiler and a set of integrated compared to the ones reported in previous works in the
high-performance solvers tailored for complex, large scale literature.
M. Escobar, J.O. Trierweiler / Applied Thermal Engineering 51 (2013) 801e826 807

Fig. 5. Overview of optimization tools.


808 M. Escobar, J.O. Trierweiler / Applied Thermal Engineering 51 (2013) 801e826

2.1. Initialization strategies in this work. The aim of initialization strategy is to prepare
a problem to be successively solved by problems (P3), (P4),
Solving the described optimization problems is not trivial. The and (P5).
linear problems, (P1) and (P2) are not quite complicated, since the The first strategy was developed to converge the NLP problem
optimization packages available can handle them even when no (P3), which is based on the superstructure proposed by Floudas
external initial point is given (GAMS assumes the lower bounds as et al. [9] presented in Fig. 3. In this superstructure, each input
a default initial value if it is not defined). In the NLP case, obtaining stream (hot/cold) has an initial splitter that features a number of
a feasible solution is a difficult task. It is helpful to bound the outlet streams equals to the number of heat exchangers that are
variables as tight as possible in order to decrease the feasible associated with this stream; each heat exchanger has a mixer at its
region. In addition, for deriving the network configuration and to inlet and a splitter at its outlet. The mixer is connected to the
obtain a numerical solution of the NLP formulation, it is clearly initial splitter and to the splitters located after the heat
desirable to start with a “good initial guess”. In the work developed exchangers. Recycle to the same heat exchanger is not allowed.
by Floudas et al. [9] an initialization procedure was proposed. As The splitter at the outlet of each heat exchanger is also connected
pointed out by the authors, that procedure might not yield an initial to a final mixer.
feasible solution. In general, the literature does not give much In this work we have developed a strategy to find a feasible
attention to the initialization strategy and often they do not report starting point for the optimization problem (P3), or at least to
how the problems are solved. Even though, the initialization reduce the number of infeasibilities, since the previous initializa-
strategy is crucial for the convergence of nonlinear problems. In tion proposed in [9] failed to solve the tested medium size prob-
order to compare the approaches in a uniform basis and ease the lems. Initially, the heat capacity flowrates at the initial splitters are
convergence, systematical initialization strategies were proposed replaced by split fractions (xhi,j and xci,j) as follows:

Fig. 6. General information flow in HENS.


M. Escobar, J.O. Trierweiler / Applied Thermal Engineering 51 (2013) 801e826 809

fi;jIH fi;jIC With the mass and energy balance solved sequentially we can
xhi;j ¼ ; xci;j ¼ (1) estimate the temperature difference for each heat exchanger
Fi Fj
terminal, defined by the equations (10) and (11).
1 EH OC
dti;j ¼ Ti;j  Ti;j (10)
where fi;jIH =fi;jIC is the heat capacity flowrate on the hot/cold side from
the initial splitter to the mixer preceding the heat exchanger for the 2
dti;j OH
¼ Ti;j EC
 Ti;j (11)
match between hot stream i and cold stream j (heat exchanger iej),
and Fi/Fj is the total heat capacity flowrate for hot/cold stream i/j. If all temperature difference at the heat exchanger terminals
It is also defined split fractions for the splitter after each heat 1 and dt 2 Þ are strictly positive, then a feasible solution was found
ðdti;j i;j
exchanger: e for which the heat exchanger areas are strictly positive and finite.

BH
fi;j;j BC
fi;j;i
0 0
yhi;j;j0 ¼ ; yci;j;i0 ¼ (2) 2.2. Initialization strategy algorithm
fi;jEH fi;jEC

Based on the ideas described in previous subsection, it was


where fi;jEH =fi;jEC is the heat capacity flowrate entering the heat proposed the overall procedure sketched in Fig. 7. The algorithm for
BH is the heat capacity flowrate on
exchanger iej on hot/cold side, and fi;j;j0 finding a feasible solution is described by the following steps:
the hot side from the splitter after the heat exchanger iej going to the
initial mixer preceding the heat exchanger iej0 . Similarly, fi;j;i
BC is the heat
0 Step 1 (Bounding the variables): the variables temperatures and
capacity flowrate on cold side from the splitter after the heat exchanger heat capacity flowrates are bounded as expressed in Table 1. The
iej going to the initial mixer preceding the heat exchanger i0 ej. boundary step starts by setting the bounds for temperatures and
Using the definitions (1) and (2), the mass balance at the mixer then these bounds are used to set up the bounds for the heat
preceding the heat exchanger for hot and cold side result in capacity flowrates.
equations (3) and (4) respectively. Step 2 (Initial set of split fractions): The split fractions for the hot
X streams are set to maximize the inlet temperatures of the hot
fi;jEH  yhi;j0 ;j fi;jEH0 ¼ xhi;j Fi (3) side of the exchanger, i.e. setting all recycle fractions yhi,j,j to zero
jsj0 (maximizing the inlet temperature avoiding backmixing) and
minimizing the temperature drop. This optimization problem
X
fi;jEC  yci0 ;j;i fiEC
0 ;j ¼ xci;j Fj (4) has analytical solution as expressed in Fig. 7 in the block that
isi0 defines the initial xhi,j values. In a similar fashion, the split
fractions for the cold streams are initiated to minimize the inlet
For any feasible selection the of the split fractions x and y these
temperatures of the cold side, and hence all yci,j,i are set to zero
equations (3) and (4) are solved as a linear system in order to find
and the temperature drop is minimized in order to define the
the heat capacity flowrate entering the heat exchanger at the hot
values of xci,j values.
and cold side of each heat exchanger ðfi;jEH ; fi;jEC Þ. The heat capacity
Step 3 (Solving mass and energy sequentially): With the initial
flowrates from the splitter after each heat exchanger going to the
values for all split fractions, the mass and energy balance are
final mixer fi;jOH can be determined by the mass balance at the
splitter after each heat exchanger as follows:
0 1
X
fi;jOH ¼ fi;jEH @1  yhi;j;j0 A (5)
jsj0

!
X
fi;jOC ¼ fi;jEC 1 yci;j;i0 (6)
isi0

Once we solve the mass balance, we fix all flowrates and then


solve the energy balance in order to find the temperatures. The
energy balance for each heat exchanger is expressed as in (7).
   
Qi;j ¼ fi;jEH Ti;j
EH OH
 Ti;j ¼ fi;jEC Ti;j
OC EC
 Ti;j (7)

OH and T OC from (7) and


Isolating the outlet temperatures Ti;j i;j
substituting into the energy balance at the mixer preceding each
heat exchanger results in the equations (8) and (9), which can be
EH and T EC .
solved as a linear system to find Ti;j i;j

X X
fi;jEH Ti;j
EH
 yhi;j0 ;j fi;jEH0 Ti;j
EH
0 ¼ xhi;j Fi Tiin  yhi;j0 ;j Qi;j0 (8)
jsj0 jsj0

X X
fi;jEC Ti;j
EC
 yci;j0 ;j fiEC EC in
0 ;j Ti0 ;j ¼ xci;j Fj Tj  yci0 ;j;i Qi0 ;j (9)
isi0 isi0 Fig. 7. NLP initialization procedure for a feasible starting point.
810 M. Escobar, J.O. Trierweiler / Applied Thermal Engineering 51 (2013) 801e826

Table 1
Boundary values for the NLP Superstructure variables.

Temperature Heat Capacity Flowrate

Lower Upper Lower Upper

Bound Bound Bound Bound


IH
Ti;j Qi;j Tiin fi;jEH Qi;j Qi;j
Tjin þ þ DTmin IH;Upper OH;Lower IH;Lower OH;Upper
Fj Ti;j  Ti;j Ti;j  Ti;j
OC
Ti;j Qi;j Tiin  DTmin
Tjin þ
Fj
OH
Qi;j Qi;j
Ti;j Tjin þ DTmin Qi;j fi;jEC
Tiin  OC;Upper
Ti;j IC;Lower
 Ti;j OC;Lower
Ti;j IC;Upper
 Ti;j
Fi
IC
Ti;j Tjin Qi;j
Tiin   DTmin
Fi

sequentially solved by solving two linear systems. The first if the interactive procedure fails, all the mass and heat balances are
system defines the unknown flowrates, which are then fixed satisfied, reducing the number of infeasibilities.
and the second linear system is solved for defining the unknown The MINLPs problems involve nonlinearities, and then a good
temperatures. initial point is essential for the convergence. The initialization is
Step 4 (Checking the feasibility of the heat exchangers): For each even more important when the problem size increases. The
process stream and the temperature approaches dti;j 1 and dt 2 are
i;j problem (P4) based on the Hyperstructure [10] is solved using the
checked. If all values, for the existent heat exchangers, are Superstrucutre [9] solution (problem P3) previously described as
positive, then a feasible solution was found and the procedure a feasible initial point. The SYNHEAT Model [11] is easier to solve,
terminates. Otherwise, some split fractions must be iteratively because it is linearly constrained. This problem is initialized using
changed as described in step 5. a sequential procedure that starts with a MILP subproblem,
Step 5 (Definition of a new set of split fractions): Initially, we try generated from the original problem (P5) removing the nonlinear
changing only the initial split fractions xhi,j and xci,j. It starts by terms form the objective function. As a result we have a MILP, for
the selection of the rows of dti;j 2 that present negative values.
which the variables are initialized at their lower bounds. After
Each row is treated independently. Suppose that the row i was solving the MILP subproblem, we fix the binary variables in (P5)
selected. Assuming that the inlet temperature of the cold stream and solve the problem (P5) as an NLP. The solution of this problem
j is already at its minimum value ðTi;j IC ¼ T in Þ. It is possible
j is then used as a feasible initial point. This systematic procedure has
2
increase the dti;j value increasing the heat capacity flowrate been demonstrated in practice quite good and robust as can be seen
entering the heat exchanger on the hot side, which implies in in Section 4.
increasing the split fraction xhi,j. It will decrease the temperature
drop increasing Ti;j OH and hence dt 2 is increased. A different trial 3. Case studies
i;j
of split fractions is selected, and we go back to step 3 and 4, in
which the temperature differences (dt) are recalculated. In order Five examples from literature are presented in this section to
to accept the new value it is necessary to check if the other illustrate the procedures and HENS models. The problem data are
temperature approaches at the same row ðdti;j 2 for jsj0 Þ are not listed in Table 2e6. The case studies were selected with different
0

violated, i.e. are positive. It may happen because increasing the sizes to represent the combinatorial nature of the synthesis problem.
split fraction xhi,j implies in decreasing the split fraction xhi,j0 for Even the case study 05 with 39 process streams is not a really
a feasible solution. If it is not possible to find positive values for large scale problem from an industrial point of view. However, it is
temperatures difference at the same row, we try to increase the one of the largest problem reported in the literature that has been
split fractions yhi,j,j0 from others matches, preferably we should addressed by mathematical programming methods.
recycle streams at higher temperatures. The procedure is
repeated for each row of dt2 with negative values. A similar 4. Results, analysis and discussion
procedure is used to select the split fractions xci,j and yci,j,i
and change the dti;j 1 values. Increasing the split fractions xc
i,j 4.1. Supertargeting: DTmin selection
promotes a smaller temperature drop and the outlet tempera-
ture of the cold stream Ti;j OC decreases increasing dt 1 since the As mentioned before, the parameter DTmin plays an important
i;j
inlet temperature of the hot stream is already in its maximum role in HENS. We performed the selection of this parameter using
IH in
value ðTi;j ¼ Ti Þ.

This procedure described was implemented in GAMS. Split


Table 2
fractions are defined as parameters and subproblems (mass and
Problem data for case study 01 (Gundersen [14]).
energy balance) are sequentially solved, and the temperature
differences are checked for each heat exchanger. Based on the Stream Tin ( C) Tout ( C) F (kW/ C) h (kW/m2  C)

previous algorithm, new fractions are defined and the GAMS file is H1 270 160 18 1
ran again. From our experience it is important to emphasize, that H2 220 60 22 1
C1 50 210 20 1
after the first run only a few negative values of temperature
C2 160 210 50 1
approaches occur. Each stream is treated independently, using trial CU 15 20 1
and error increasing the flow for the heat exchangers with negative HU 250 250 1
temperature approach. Applying this procedure, it was possible to Cost of Heat Exchangers ($ yr1) ¼ 4000 þ 500[Area (m2)]0.83.
find a feasible solution for the majority of the cases tested, not only Cost of Cooling Utility ¼ 20 ($ kW1 yr1).
the five case studies presented here. It should be noticed, that even Cost of Heating Utility ¼ 200 ($ kW1 yr1).
M. Escobar, J.O. Trierweiler / Applied Thermal Engineering 51 (2013) 801e826 811

Table 3 Table 5
Problem data for case study 02 (Papoulias and Grossmann [7]). Problem data for case study 04 (adapted from Sorsak and Kravanja [16]).

Stream Tin (K) Tout (K) F (kW/K) h (kW/m2 K) Stream Tin ( C) Tout ( C) F (kW/K) h (kW/m2 K)

H1 433 366 8.79 1.7 H1 576 437 23.10 0.06


H2 522 411 10.55 1.7 H2 599 399 15.22 0.06
H3 544 422 12.56 1.7 H3 530 382 15.15 0.06
H4 500 339 14.77 1.7 H4 449 237 14.76 0.06
H5 472 339 17.73 1.7 H5 368 177 10.70 0.06
C1 355 450 17.28 1.7 H6 121 114 149.60 1.00
C2 366 478 13.90 1.7 H7 202 185 258.20 1.00
C3 311 494 8.44 1.7 H8 185 113 8.38 1.00
C4 333 433 7.62 1.7 H9 140 120 59.89 1.00
C5 389 495 6.08 1.7 H10 69 66 165.79 1.00
CU 509 509 1.7 H11 120 68 8.74 1.00
HU 311 355 3.41 H12 67 35 7.62 1.00
H13 1034.5 576 21.30 0.06
Cost of Heat Exchangers ($ yr1) ¼ 145.63[Area (m2)]0.6.
C1 123 210 10.61 0.06
Cost of Cooling Utility ¼ 18.12 ($ kW1 yr1).
C2 20 210 6.65 1.20
Cost of Heating Utility ¼ 37.64 ($ kW1 yr1).
C3 156 157 3291.0 2.00
C4 20 182 26.63 1.20
C5 182 318 31.19 1.20
C6 318 320 4011.83 2.00
the Supertargeting procedure proposed by Linnhoff and Hindmarsh C7 322 923.78 17.60 0.06
CU 9 17 1.0
[6], in which the costs are estimated ahead the design for different
HU 927 927 5.0
values of DTmin. In Fig. 8 is presented the estimative of the TAC for
each case study over a range of DTmin. Cost of Heat Exchangers ($ yr1) ¼ 4000 þ 500[Area (m2)]0.83.
Cost of Cooling Utility/Heating Utility ¼ 25 ($ kW1 yr1)/250 ($ kW1 yr1).
It can be noted that the case study 02 is not pinched for the
selected DTmin since only cold utility is needed. This is due the fact
that there is a threshold, recognized by Linnhoff and Hindmarsh [6],
below which the problem is not pinched and either hot or cold Table 6
Problem data for case study 05 (Bjork and Pettersson [15]).
utility (but not both) is consumed. They called them as threshold
problems, and mentioned that it is possible to apply the pinch Stream Tin ( C) Tout ( C) F (kW/K) h (kW/m2 K)
design method to threshold problems with DTmin adjusted to the H1 180 75 30 2
threshold value. Although, in industrial designs, threshold prob- H2 280 120 15 2.5
lems are rare according to Linnhoff and Hindmarsh [6]. The general H3 180 75 30 2
H4 140 45 30 2
results using Pinch Analysis, for the corresponding DTmin selected
H5 220 120 25 1.5
for each case, are presented in Table 7. H6 180 55 10 2
H7 170 45 30 2
H8 180 50 30 2
4.2. Analysis of the optimization problems H9 280 90 15 2
H10 180 60 30 2
H11 120 45 30 2
One of the main drawbacks using mathematical programming
H12 220 120 25 2
to solve the HENS models is the combinatorial nature of the H13 180 55 10 2
problems involved. The number of variables and equations H14 140 45 20 2
substantially increases with the number of process streams. In H15 140 60 70 2
Table 8 is listed the number of variables and equations for the five H16 220 50 15 2.5
H17 220 60 10 2.5
H18 150 70 20 2
H19 140 80 70 2
H20 220 50 35 2
Table 4 H21 180 60 10 2
Problem data for case study 03 (Bjork and Pettersson [15]). H22 150 45 20 2.5
C1 40 230 20 1.5
Stream Tin ( C) Tout ( C) F (kW/K) h (kW/m2 K)
C2 120 260 35 1
H1 180 75 30 2 C3 40 190 35 1.5
H2 280 120 60 1 C4 50 190 30 2
H3 180 75 30 2 C5 50 250 60 2
H4 140 40 30 1 C6 40 150 20 2
H5 220 120 50 1 C7 40 150 20 2
H6 180 55 35 2 C8 120 210 35 2.5
H7 200 60 30 0.4 C9 40 130 35 2.5
H8 120 40 100 0.5 C10 60 120 30 2.5
C1 40 230 20 1 C11 50 150 10 3
C2 100 220 60 1 C12 40 130 20 1
C3 40 190 35 2 C13 120 160 35 1
C4 50 190 30 2 C14 40 90 35 1.75
C5 50 250 60 2 C15 50 90 30 1.5
C6 90 190 50 1 C16 50 150 30 2
C7 160 250 60 3 C17 30 150 50 2
CU 25 40 1 CU 25 40 2
HU 325 325 2 HU 325 325 1

Cost of Heat Exchangers ($ yr1) ¼ 8000 þ 500[Area (m2)]0.75. Cost of Heat Exchangers ($ yr1) ¼ 8000 þ 800[Area (m2)]0.8.
Cost of Cooling Utility/Heating Utility ¼ 10 ($ kW1 yr1)/80 ($ kW1 yr1). Cost of Cooling/Heating Utility ¼ 10 ($ kW1 yr1)/70 ($ kW1 yr1).
812 M. Escobar, J.O. Trierweiler / Applied Thermal Engineering 51 (2013) 801e826

Fig. 8. Supertargeting for the case studies 1 to 5.

case studies considered in this work. For the problems (P2), (P4) only 4 streams, 53 equations, 69 variables, and 12 binary variables
and (P5), the number of binary variables is given in brackets. For are involved in the same MILP optimization problem. Based on the
instance, the fifth case study with 39 process stream involves in its data listed in Table 8 it can be concluded that there are more
MILP transshipment model (P2) 2164 equations, 10832 continuous variables than equations indicating a large number of degrees of
variables, and 413 binary variables. Whilst the first case study with freedom for the optimization.

Table 7
Targets obtained by Pinch Technology for each case study CS.

CS DTmin Pinch temp. min hot utility min cold utility Units Total area Operating cost Capital cost Total annual cost

( C) ( C) (kW) (kW) e (m2) ($ yr1) ($ yr1) ($ yr1)

01 10 165 600 400 7 1226.0 128000.0 282732.1 470732.1


02 10 e 0 1879 10 249.5 34047.0 10034.0 44081.4
03 10 135 8900 6525 16 5280.2 747426.3 376305.4 1524678.3
04 20 130 2170 339 21 4968.0 550978.2 1064750.2 1615728.5
05 5 175 4450 7750 40 6782.9 1010625.0 1581442.9 1591070.1
M. Escobar, J.O. Trierweiler / Applied Thermal Engineering 51 (2013) 801e826 813

Table 8 Table 9
Number of variables and equations for each problem (LP Transshipment, MILP Definitions of indexes used to compare different solvers.
Transshipment, NLP Superstructure, MINLP I-Hyperstructure, and MINLP II-
SYNHEAT Model). Index Efficiency (%) Robustness (%) Quality of Solution (%)

Case LP (P1) MILPa (P2) NLP (P3) MINLPa I (P4) MINLPa II (P5)
Definition 100 XN *
tcs 100
Nmt 100 XN
d*cs þ 3
N cs ¼ 1 tmt;cs N N cs ¼ 1 dmt;cs þ 3
Number of 01 7 69[8] 59 210[8] 69[12]
variables 02 21 844[35] 577 1452[35] 661[135] N is equal to the number of case studies (N ¼ 5). The parameter t consists of the
03 19 1494[71] 315 4057[71] 2143[463] computational time, and d the difference of the objective function for the best
04 33 4084[111] 2013 8035[111] 5346[1203] known solution. The subscripts mt and cs refers to the solver used and the case study
05 25 10832[413] 1494 54442[413] 35049[6397] selected. Nmt refers to the number of case studies solved by the solver mt. The star
Number of 01 5 53 58 189 105 indicates the best values obtained, i.e. the lower values to t, S, and d. The parameter 3
equations 02 19 340 337 971 901 indicates the machine precision dmt;cs ¼ kFobjmt;cs  Fobj*k=3 .
03 17 502 254 1858 2846
04 31 1044 744 2738 7004
solvers. For instance, in some algorithms iteration stands for
05 23 2164 1981 10815 44944
branch and bound nodes, while in other algorithms for simplex
a
Number of binary variables is given in brackets. pivots.
The keys for the methods to solve NLPs are: Successive Linear
Programming (SLP); Successive Quadratic Programming (SQP);
4.3. Solvers evaluation1 Generalized Reduced Gradient (GRG); and Interior Point (IP).
CONOPT [17] is a GRG method and MINOS [17] is a reduced-
Solving the LP transshipment model (P1) is straightforward and gradient method or a projected Lagrangean method both used to
the results, i.e. the minimum utility consumption, are similar to the solve large scale problems. SNOPT [18] is a SQP method. KNITRO
ones obtained using the Pinch Technology listed in Table 7. The [19] is an interior point method, using SQP and trust regions.
main advantages of using the LP model instead of Pinch Technology OQNLP and MSNLP are multi-start heuristic algorithms that
to estimate the minimum utility consumption are the possibility of combine direct search methods with any NLP local solver.
considering operating constraints, and handling forbidden and The Table 12 shows the CPU time and the TAC for each instance,
required or preferential matches. The GAMS license used in this and the best values for each case study are highlighted in bold. It can
work has available 9 different LP solvers, e.g. CPLEX, OSL. They are be seen that once the MILP problem (P2) is solved, and the initiali-
mainly based on the simplex method, which is a well known zation strategy is performed, the problems are quickly solved. The
algorithm for solving linear programs, and the interior point general analysis of the solvers performance expressed in Table 13
methods (also referred to as barrier methods). Currently, large scale points out the GRG solvers MINOS and CONOPT and the SQP solver
LPs can be solved quickly. The performance of the solvers was quite SNOPT as the best NLP solvers for solving the problem (P3). MINOS
similar and all the problems were solved in less than 1 s. had the best weighted average as highlighted in bold in Table 13.
The selection of an appropriated solver for the MILP trans- These solvers had similar computational time and quality of the
shipment model (P2) using GAMS, is trickier. The available options, solution obtained, but in general CONOPT required more iteration
in general based on branch and bound, cutting planes and compared to other two solvers, as expected since it follows a feasible
a combination of both, have different computational performance, path. The solvers KNITRO and OQNLP were computationally expen-
and convergence properties. It is well known form the literature sive. Both did not converge for the fifth case study within a reason-
that this problem may have multiple solutions, however, the global able time. Although, the solver OQNLP presented a satisfactory
optimality can be guaranteed. quality of solution, for those problems that it was able to solve.
In order to compare different solvers capability and have some The methods to solve MINLP problem is generally based on Branch
insight about the suitable selection for solving the problems and Bound (B&B), solving successively NLP relaxations, or some kind
considered, some indexes were proposed. Such indexes are defined of decomposition such as: Generalized Benders decomposition (GBD),
in Table 9. In order to carry out all analysis it was used all solvers for Outer Approximation (OA), and Extended Cutting Planes (ECP).
which it was available a full license, since the demo version restricts BARON, which is a global solver, combines constraint propagation,
the maximum numbers of variables and equations. interval analysis, and duality in its reduce arsenal to provide tight
In Table 10 is listed the computational time and the minimum relaxations in a spatial branch and bound framework [20]. DICOPT
number of units for solving the MILP (Problem P2) for each case (DIscrete and Continuous Optimizer) is based on the outer approxi-
study. The smallest CPU times for each case study is given in bold. mation method with equality relaxation strategies [21]. SBB [22]
In Table 11 is presented the solvers evaluation. The highest combines a standard Branch and Bound method with some NLP
weighted average is given in bold. The results indicated that the solvers in GAMS, e.g. CONOPT, MINOS or SNOPT. BARON (Branch and
solver CPLEX is more robust, and in general faster than the other Reduce Optimization Navigatore [20]) is used for solving non convex
solvers. The weighted average considers a priority order of: the optimization problems to global optimality and it uses MINOS as NLP
computational time (efficiency-35%); ability of solving the solver and CPLEX as MILP solver. Recent research has also focused on
problem (robustness-weighted with 35%); the quality of solution
Table 10
(quality-30%). It is important to mention that the choice of these
Computational time and the minimum number of units for the MILP solvers
weights may result in different conclusions. One could claim about considered for each case study.
the subjective nature of this decision. However, we explicitly re-
Case Study CS01 CS02 CS03 CS04 CS05
ported the quantitative indexes in Table 11, and only the last
column corresponding to the solver evaluation is according to the Solver t(s) t(s) t(s) t(s) t(s)
weights selected. One may not consider the number iterations as BDMLP 0.070 0.85 4.77 a a

potential criteria, because they mean different steps for different CPLEX 0.016 6.97 12.92 2.01 1021.4
a
CoinGlpk 0.032 1.83 11.81 24.16
a
CoinCbc 0.094 37.59 817.28 77.52

1 N. of Units 6 10 18 21 42
All the problems were solved on an Intel Core 2.67 GHz machine with 2.96 GB
a
memory, and default options were used for all the solvers. Convergence failure.
814 M. Escobar, J.O. Trierweiler / Applied Thermal Engineering 51 (2013) 801e826

Table 11 Table 13
MILP solver evaluation for the five case studies. NLP solver evaluation for the five case studies.

Solver Efficiency (%) Robustness (%) Quality of Weighted Solver Efficiency (%) Robustness (%) Quality of Weighted
solution (%) average (%) solution (%) average (%)

BDMLP 44.6 60.0 100.0 66.6 CONOPT 65.5 100.0 99.3 87.7
CPLEX 69.8 100.0 100.0 89.4 MINOS 91.4 100.0 98.9 96.7
CoinGlpk 29.0 80.0 100.0 68.2 SNOPT 82.5 100.0 99.1 93.6
CoinCbc 4.49 80.0 100.0 59.6 KNITRO 17.2 80.0 79.1 57.8
OQNLP 0.8 80.0 79.8 52.2
Weight 0.35 0.35 0.30
Weight 0.35 0.35 0.30

combining derivative free NLP methods such as Tabu Search, Scatter


Search, Simulated Annealing, and Genetic Algorithms, with some and afterward this solution was used as initial point to the NLP
deterministic NLP methods. OQNLP performs Scatter Search to Superstructure, problem (P3). The resulting HEN configurations can
provide starting points for any gradient based local NLP solver. be seen in Fig. A1. An interesting observation is that it was possible to
In the evaluation of the MINLP solvers it was used two MINLP improve the Pinch solution, i.e. to reduce the TAC, using optimization
problems, (P4) and (P5). CPLEX and MINOS were selected as MILP methods. It can be noted that even for this small scale problem, using
and NLP solver, respectively. the sequential procedure directly provided a better solution with
The computational time taken by each MINLP solver and the TAC lower TAC as can be seen in Fig. A2. Due to the “hand-made”
for each instance are given in Table 14, and the best values for each procedure to design HENs using Pinch Technology, it was not used to
case study and model are highlighted in bold. It is also given the the other cases studies. Instead, the proposed initialization strategy
time spent by the initialization strategy ti. The general analysis is was applied to all cases, and the HENs configurations using the
given in Table 15 and the highest weighted average is given in bold. optimization techniques are showed in Appendix A.
The solver DICOPT presented the best performance according to the
evaluation, but the SBB was as robust as DICOPT and had a similar
quality of solution, though it presented in general a slower con- 4.5. Solution comparison
vergence. The solvers BARON and OQNLP did not show a good
performance, they did not converged for the fifth case study, and In order to check the applicability and relevance of the results
they require larger CPU times. obtained by using the proposed strategy to initialize the problems,
in this section we compare the result of this work with solutions
4.4. Solution report given in the literature in the same cases.
The case study CS01 is a classical problem reported by Gundersen
The general results for the best solutions obtained for each case [14]. By using Pinch Technology it was obtained the same solution
study are shown in Table 16. The corresponding HEN configurations previously reported. As mentioned in the previous subsection, using
are depicted in Appendix A. It was possible to obtain a feasible the optimization approaches was possible to obtain even better
solution for all case studies using the initialization strategy pre- solutions, with the heat recovery level predicted by the Pinch
sented in Section 2. It is worth to mention that the simultaneous Technology. The best solution found, using the problem (P4), led to
procedures always presented better solutions when compared with a reduction of 7.7% in the annual costs compared to the Pinch
the sequential one. An interesting observation is that the simulta- solution.
neous solutions using the SYNHEAT model (P5) corresponds to the
best solution found, except for the smallest case studies CS01 and Table 14
CS02. Despite the SYNHEAT model restricts the search space using Computational time and TAC for the MINLP solvers for each case study.
the isothermal mixing assumption, the resulting model features Case study CS01 CS02 CS03 CS04 CS05
a linear feasible region, which make this problem very robust to
Solver t(s) t(s) t(s) t(s) t(s)
solve when compared to the other models with bilinear terms due
to energy balances. (P4) DICOPT 0.42 1.02 3.42 12.5 18.4
SBB 2.70 6.98 17.5 18.2 25.1
For the case study CS01 in particular, two additional designs were a
BARON 101.2 120.8 249.4 400.6
generated. Initially, it was designed a HEN using Pinch Technology OQNLP 7.22 182.3 221.8 a a

(P5) DICOPT 0.63 0.72 3.38 10.3 12.4


Table 12 SBB 0.17 2.35 2.46 50.0 71.8
Computational time and TAC for the NLP solvers for each case study. a
BARON 119.6 178.5 218.5 358.7
a
OQNLP 80.1 15.2 47.6 327.2
Case study CS01 CS02 CS03 CS04 CS05
ti(s) ti(s) ti(s) ti(s) ti(s)
Solver t(s) t(s) t(s) t(s) t(s)
(P4) 0.11 7.2 13.1 21.7 1022.4
CONOPT 0.10 0.25 0.30 0.54 1.05
(P5) 0.78 2.7 9.3 54.6 978.5
MINOS 0.09 0.22 0.18 0.16 1.03
SNOPT 0.08 0.15 0.22 0.20 2.03 Solver TAC TAC TAC TAC TAC
KNITRO 0.17 0.39 90.5 207.6 3000.0a
($ yr1) ($ yr1) ($ yr1) ($ yr1) ($ yr1)
OQNLP 2.53 22.95 90.5 738.9 3000.0a
(P4) DICOPT 368649.0 43705 1574948 1518187 2197653
Solver TAC TAC TAC TAC TAC
SBB 361982.8 43705 1769655 1758489 2197653
($ yr1) ($ yr1) ($ yr1) ($ yr1) ($ yr1) a
BARON 362429.7 43242 1574948 1589810
a a
CONOPT 376387 43841 1613300 1538058 3157356 OQNLP 362459.9 43708 1594011
MINOS 376387 43708 1624325 1581014 3086047 (P5) DICOPT 399509.7 43570 1507654 1461006 2055421
SNOPT 376387 43821 1613300 1562215 3126166 SBB 366006.7 43685 1506667 1467675 2055421
a a
KNITRO 376387 43885 1613300 1579267 BARON 377541.3 43689 1506667 1461276
a a
OQNLP 376387 43708 1627248 1518187 OQNLP 399509.8 43689 1518149 1467503
a a
Maximum time reached with no convergence. No convergence.
M. Escobar, J.O. Trierweiler / Applied Thermal Engineering 51 (2013) 801e826 815

Table 15 Table 17
MINLP solver evaluation for the five case studies. Solution comparison for case studies CS02 and CS05.

Solver Efficiency (%) Robustness (%) Quality of Weighted Case study CS02 Case study CS05
solution (%) average (%)
Method Annual cost Method Annual cost
DICOPT 89.9 100.0 98.8 96.1 ($ yr1) ($ yr1)
SBB 45.8 100.0 97.0 80.1
Papoulias and Grossmann [7] 43934 Pettersson [26] 1997054
BARON 1.0 80.0 79.2 52.1
Lin and Miller [23] 43329 Luo et al. [27] 1965000
OQNLP 2.1 70.0 68.7 45.8
Pariyani et al. [24] 43439 Ernst et al. [28] 1943536
Weight 0.35 0.35 0.30 Yerramsetty and Murty [25] 43538
This work 43242 This work 2055421
This worka 1997849
The case study CS02 is the popular 10 SP1 problem, studied by a
Improved solution.
many research workers [7,23e25]. The best solution found, using
SYNHEAT, is presented in Table 17 along with other published capacity flowrates. This procedure allows the quick evaluation of
values. The cost of the best network found in this work is 43242 the objective function. The problem is solved in two levels, an upper
$ yr1, which is slightly better than the other reported values. The level to identify subnetworks and in a lower level each subnetworks
network structure is depicted in Fig. A4. is optimized using a hybrid genetic algorithm. The authors obtained
For the case study CS03, the best solution reported in Bjork and an initial solution of 2062 M$ yr1, which is to be noted worse than
Petterson [15] has an annual cost of 1513854 $ yr1, while the best our initial solution with a cost of 2055 M$ yr1. After the optimi-
solution found in this work with a cost of 1506667.4 $ yr1 is only zation of 11 subnetworks, the final solution has the cost of
marginally better. 2012 M$ yr1. After some further work they mention in the text
The case study CS04 was adapted from the work of Sorsak and a cost reduction to a level of 1965 M$ yr1 with 44 units. However,
Kravanja [16]. The best solution found in this work is worse, about 15% in the network depicted in their paper they show an annual cost of
higher than the reported value of 1259.9 k$ yr1. A point worth noting 1943 M$ yr1 with 44 units. This is the same solution obtained in
here is that in their work they comprised different heat exchanger the work of Ernst et al. [28], in which Luo is also an author. They do
types with different costs, and therefore a direct comparison is not not mention the DTmin used. This solution is only 3.2% higher than
possible. However, comparing the heating recovery level of the the solution obtained in this work with 44 units for a DTmin of 5  C.
solution with the targets predicted by the Pinch Technology it is The minimum predicted by the Transhipment model of Papoulias
possible to consider the solution obtained as a fair solution. and Grossmann [7] is 40 units. However, in our configuration both
Bjork and Pettersson [15] and Pettersson [26] developed utility consumptions are higher than the minimum allowed by
a genetic algorithm which seems effective for large scale HENs. thermodynamic. It clearly shows that our initial configuration can
They solve linear subproblems for identifying potential matches be further optimized. We identified 11 subnetworks in this initial
and subnetworks, which are then optimally designed separately. solution (shown in Fig. A11). Optimizing the subnetworks it was
For the case study CS05 Bjork and Pettersson [15] found a network possible to reduce the costs to 1979.0 M$ yr1, which compares very
with a TAC of 2073 M$ yr1 with 48 units. The solution corresponds well with the reported values in the literature. It is only 1.15% higher
to a DTmin of 20.9  C. Pettersson [26] improved the previous solu- than the best solution reported in [28].
tion to the one with cost of 1997.1 M$ yr1. It is important to mention that the best solutions listed in
In the work of Luo et al. [27] is used a monogenetic algorithm for Table 16 refer to the solutions obtained in this work. These solu-
solving the case study CS05. Their formulation is based on an tions are preliminary solutions. Some additional efforts can be
explicit solution for the superstructure of Yee and Grossmann [11]. made for improving this initial solution. The potential of
In that case the isothermal mixing assumption is rescinded. In improvement can be analyzed in advance by comparing the solu-
addition, binary variables are avoided by using non-differential tion with the targets predicted by the Pinch Technology.
formulation with conditional expressions and the “min” operator.
The decision variables are reduced to the areas and the heat 4.6. Methods evaluation

Table 16 It was performed a qualitative evaluation over the optimization


General reporting of best solutions for the case studies.
methods to solve the HENS problems. All methods are suitable to
Case Hot Cold Number Area Operating Capital Total annual treat the design, but they have different features. The sequential
study utility utility of units (m2) cost cost ($ yr1) cost ($ yr1)
procedure is easier to implement and solve. It has a lower
(kW) (kW) ($ yr1)
computational effort, and it is very robust. But even for small scale
Sequential procedure (P1)e(P3) case studies, they converge to suboptimal solutions that can be
01 600 400 6 1151 128000.0 248387.2 376387.2
improved using simultaneous approaches. Comparing the different
02 0 1879 10 268.7 34046.8 9661.8 43708.6
03 8900 6525 18 6418.8 777250.0 836050.0 1613300.0 simultaneous approaches, the SYNHEAT model presented in
04 2170 339 21 5430.9 550978.2 967208.9 1518187.2 general the best performance in terms of computational time and
05 4450 7750 42 9627.2 389000.0 2697247.9 3086247.9 robustness. Even though the simultaneous methods are computa-
Simultaneous-hyperstructure (P4)
tional demanding and they may fail to converge, they allow
01 600 400 6 1112.2 128000.0 233982.8 361982.8
02 0 1879 10 251.6 34046.8 9195.7 43242.5
obtaining better solutions when compared to the sequential
03 9304.5 6929.5 18 5618.1 813651.0 761297.1 1574948.2 procedure. After this study, it is strongly recommended the model
04 2170 339 21 5430.9 550978.2 967208.9 1518187.2 SYNHEAT for synthesizing HENs. This model and its simplifying
05 8561 11861 43 3615.5 717855.3 1479797.9 2197653.3 assumptions were discussed, and this study is available online at
Simultaneous-SYNHEAT (P5)
the website minlp.org.2 This collaborative website has a major goal
01 600 400 6 1061.8 128000 238006.7 366006.7
02 0 1879 10 259.4 34046.8 9523.6 43570.3 to create a library of optimization problems that can be generally
03 9924 7549 17 4513.4 869374.5 637292.9 1506667.4
04 1938 107 21 4895.8 487172.7 973833.5 1461006.2
05 6226 9526 44 3991.8 531072.7 1524348.4 2055421.1 2
http://minlp.org/library/problem/index.php?i¼93&lib¼MINLP.
816 M. Escobar, J.O. Trierweiler / Applied Thermal Engineering 51 (2013) 801e826

formulated as MINLP, providing high level description of the obtained. Extensive tests were performed, taking into account all
problem formulations with corresponding GAMS (input) files. possible combinations (case study-solver), for which a full solver
Some examples are used to illustrate the models and the results are license was available. The tests resulted in 25 LPs, 25 MILPs, 25 NLPs
reported. and 40 MINLPs in a total of 115 instances to be solved. The main
conclusions were that the solver CPLEX was the best solver to linear
4.7. Final remarks problems, and for NLP problems the solvers MINOS and CONOPT had
the best performance. For the MINLP problems, the solvers DICOPT
During the systematical implementation of the optimization and SBB were the more suitable with the best performances.
problems it was faced some difficulties and some remarks are given The methods were compared and it was possible to conclude that
below: the sequential problem is easier to solve and implement, quite
robust, with a low computational effort compared with the simul-
 The LP and the MILP models need heat balances in temperature taneous methods, but with a poor quality of solution since it was
intervals. To increase the search space and hence the flexibility always possible to improve the solution using the simultaneous
to find the heat loads and matches solving the MILP, it is framework. The simultaneous methods are harder to treat, but in
possible use a lower DTmin even a zero value; general they provide better solutions. Comparing the two different
 On solving the MILP model it is interesting limit the maximum simultaneous approaches, the SYNHEAT model presented the best
number of heat exchangers per process stream, avoiding performance, despite the isothermal mixing assumption. It is due the
complicated configurations and significantly decreasing the fact that this problem is easier to solve since it is linear constrained.
number of variables in the NLP step. In addition, multiple Finally, despite the difficult involved in obtaining a solution
solutions may be obtained using integer cuts; using optimization methods for HENS, the proposed systematic
 Using GAMS it is interesting to use the model property called framework strategy was shown to find feasible and applicable
“holdfixed” to fix the variables already known and pass to the solutions to all case studies for each optimization method.
solver as parameters reducing the dimension problem easing
the problem solvability; Acknowledgements
 Solving the NLP it is possible to provide a DTmin relaxation, i.e.
allow small violations, in order to give more flexibility to the The authors are very grateful for the financial support from
solver allowing heat cross trough the pinch if the total annual CNPQ and PETROBRAS.
cost suggests it;
 Solving the MINLP SYNHEAT, a priori does not take into account Nomenclature
pinch information, but it may be desired to find a solution with
a utility consumption near to the predicted by pinch analysis
(or LP transshipment model). Therefore, it is interesting to add Abbreviations
extra constraints limiting the maximum utility consumption; HEN heat exchanger network
 The problem can be solved with different parameters setting: HENS heat exchanger network synthesis
(i) minimum allowable temperature difference (DTmin); (ii) LMTD log mean temperature difference
number of stages (for SYNHEAT); TAC total annual cost
 To avoid complex networks one may introduce constraints to
the model to: (i) limit the maximum number of matches; and Sets
(ii) avoid stream splits. It is important to compare the costs CP set of cold process stream j
with the ones without these constraints to check if it is an CU set of cold utility
economically reasonable solution; HP set of hot process stream i
 Especially for large scale problems it can be identified HU set of hot utility
subnetworks, which can be optimally designed separately in
order to optimize the whole configuration. Variables [units]
dti;j1 [ C] temperature approach for the inlet for i and outlet for j in

5. Conclusions the exchanger for i and j


dti;j2 [ C] temperature approach for the inlet for j and outlet for i in

Although HENS has been one of the most studied problems in the exchanger for i and j
process synthesis, even to small scale problems finding feasible dti,j,k [ C] temperature approach between hot stream i, cold stream
solution using optimization methods has been troublesome. In this j, at location k
work, the main optimization methods for HENS were presented in dtcui [ C] temperature approach between hot stream i, and cold
a uniform notation and these problems were implemented in the utility
modeling system GAMS. dthuj [ C] temperature approach between cold stream j, and hot
A new strategy for the initialization of the variables for the utility
synthesis of HENs has been developed. This approach has been fi;jEH [kW/K] hot flow capacity through the exchanger in the match i
applied successfully in order to find an initial feasible solution and j
providing to the solver a good starting point to start allowing the fi;jEC [kW/K] cold flow capacity through the exchanger in the match i
algorithm to evolute toward the optimum. Five case studies with and j
I
different dimensions were used to illustrate the convergence fm;m 0 [kW/K] flow capacity from initial splitter to the mixer

strategy and compare the performance of each approach. In addi- preceding exchanger for m and m0
tion, it was analyzed the solver selection to each particular problem E
fm;m0 [kW/K] flow capacity through the exchanger in the match
through a quantitative comparison among the different optimiza- m and m0
tion methods for HENS. For comparison purposes, some indexes E
fm;m0 ;m00 [kW/K] flow capacity from splitter after the exchanger for
were used to distinguish the performances. They were based on m and m0 to the mixer preceding exchanger in the
numerical aspects, solvable capacity and quality of the solution match m and m00
M. Escobar, J.O. Trierweiler / Applied Thermal Engineering 51 (2013) 801e826 817

o
fm;m 0 [kW/K] flow capacity from splitter after the exchanger for m Binary variables
and m0 to the final mixer yi,j existence of the match between hot stream i and cold
LMTDi,j [ C] log mean temperature difference between hot stream i stream j
and cold stream j zi,j,k existence of the match between hot stream i, cold stream
LMTDi,j,k [ C] log mean temperature difference between hot stream j, at stage k
i, cold stream j, at stage k zcui existence of the match between hot stream i, and cold
LMTDcui [ C] log mean temperature difference between hot stream utility
i, and cold utility zhuj existence of the match between cold stream j, and hot utility
LMTDhuj [ C] log mean temperature difference between cold
stream j, and hot utility Parameters [units]
Qi,j [kW] heat load between hot stream i and cold stream j bi,j exponent for the area cost in exchanger iej
Qi,j,k [kW] heat load between hot stream i and cold stream j in ci,j cost for heat math between hot stream i and cold stream j
temperature interval k Ccuj [$/kW yr] utility cost coefficient for cooling utility j
QCU [kW] minimum cold utility requirement Chui [$/kW yr] utility cost coefficient for heating utility i
QHU [kW]minimum hot utility requirement Fm [kW/K] flow capacity of hot/cold stream m
qi,j,k [kW]heat load between hot stream i and cold stream j at stage k Fi, Fj [kW/K] flow capacity of hot stream i, j
qcui [kW] heat load between hot stream i and cold utility hi, hj [kW/m2 K] heat transfer coefficient for hot stream i, j
qhuj [kW] heat load between cold stream j and hot utility Tiin ; Tjin [ C] inlet temperature of hot stream i/cold stream j
Rk [kW] heat residual load out of temperature interval k Tiout ; Tjout [ C] outlet temperature of hot stream i/cold stream j
Ri,k [kW] heat residual of hot process stream/utility i out of DTi;jmax [ C] maximum possible temperature drop through
temperature interval k exchanger iej
IH [ C] inlet temperature of hot stream i to the exchanger for i and j
Ti;j U [kW] upper bound for heat exchangers
OH [ C] outlet temperature of hot stream i from the exchanger for
Ti;j G [ C] upper bound for temperature difference
i and j Lbi,j [kW] lower bound on the heat load between stream i and
IC [ C] inlet temperature of cold stream j to the exchanger for i
Ti;j stream j
and j Ubi,j [kW] upper bound on the heat load between stream i and
OC [ C] outlet temperature of cold stream j from the exchanger
Ti;j stream j
for i and j Ui,j [kW/m2 K] overall heat transfer coefficient between hot stream
I
Tm;m  0 i and cold stream j
0 [ C]inlet temperature to the exchanger for m and m
o
Tm;m [  C]outlet temperature to the exchanger for m and m0
0

tik [ C] temperature of hot stream i at hot end of stage k


tjk [ C] temperature of cold stream j at hot end of stage k Appendix A. Network configurations

a 800kW 1000kW 180kW


270 160 AH1,C1=318.14 m2
H1
AH1,C2=56.55 m2
1100kW 2200kW
220kW
220 60 AH2,C1=610.16 m2
H2
AH2,C2=210.60 m2
210 50
C1 AH1,CU=3.79 m2
600kW
210 160 AH2,CU=12.86 m2
C2
AC2,HU=35.99 m2

TAC= 3.9037252 x 105 $/year

b 1000kW

0.56
800kW 180kW AH1,C1=144.98 m2
270 160
H1 AH1,C2=118.17 m2
1100kW 2200kW
220kW
220 60 AH2,C1=610.16 m2
H2
AH2,C2=251.21 m2
210 50
C1 AH1,CU=3.79 m2
600kW
210 AH2,CU=12.86 m2
160
C2 AC2,HU=35.99 m2
0.31

TAC= 3.8045216 x 105 $/year

Fig. A1. HEN configuration for case study 01 using Pinch Technology (a), using NLP Superstructure using Pinch solution as initial point (b).
818 M. Escobar, J.O. Trierweiler / Applied Thermal Engineering 51 (2013) 801e826

a 1400kW 580kW
270 160
H1 AH1,C1=232.00 m2
1100kW 2020kW 400kW
220 60 AH1,C2=201.28 m2
H2
600kW AH2,C1=456.96 m2
210 50
C1 AH2,C2=210.60 m2
0.89
0.08 AH2,CU=23.39 m2

210 160 AC1,HU=26.86 m2


C2

TAC= 3.7638724 x 105 $/year

180kW

b 0.02
1800kW
270 0.1 160
AH1,C1=34.74 m2
H1
700kW 2420kW 400kW AH1,C2=291.22 m2
220 60
H2 AH2,C1=603.58 m2
600kW
210 50 AH2,C2=132.35 m2
C1
0.07 AH2,CU=23.39 m2

210 160 AC1,HU=26.86 m2


C2
0.69
5
TAC= 3.6198278 x 10 $/year

1000kW

c 178.2kW
160
270
H1
801.8kW AH1,C1,1=145.54 m2
AH1,C2,1=104.16 m2
50
C1 AH2,C2,1=287.75 m2
600kW AH1,C1,2=17.70 m2
210 160
C2 AH2,C1,2=459.95 m2
AH2,CU=23.39 m2
400kW
220 60 AC2,HU=23.37 m2
H2
2021.8kW
1098.2kW
Stage 1 Stage 2

TAC= 3.66006681 x 105 $/year

Fig. A2. HEN configuration for case study 01 using Superstructure (a), Hyperstructure (b) and Synheat Model (c).
M. Escobar, J.O. Trierweiler / Applied Thermal Engineering 51 (2013) 801e826 819

588.93 kW
433 366
H1
173.1 kW
0.24 353.5 kW
522 644.5 kW 411
H2
341.3 kW 610.4kW
AH1,C4=25.10 m2
0.256
544 1191.0 kW 422
H3 AH2,C3=4.33 m2
1215.5 kW 1170.8 .kW
339 AH2,C4=3.26 m2
500
H4
AH2,C5=19.57 m2
1641.6.kW 718.2.kW
472 339
H5 AH3,C2=6.70 m2

450 355 AH3,C3=24.7 m2


C1
0.81 AH4,C2=38.23 m2
478 366
C2 AH5,C1=82.94 m2
494 311
C3 AH4,CU=32.68 m2

433 333 AH5,CU=33.16 m2


C4
495 389
C5

TAC= 4.3708618 x 104 $/year

Fig. A3. HEN configuration for the case study 02 using Superstructure.

588.93 kW
433 366
H1
644.5 kW 173.1 kW 354.4 kW
522 411
H2
340.9 kW 610.4kW
AH1,C4=25.10 m2
0.265
544 1190.1 kW 422
H3 AH2,C3=4.33 m2
1215.9 kW 1170.8 .kW
339 AH2,C4=2.44 m2
500
H4
AH2,C5=16.50 m2
1641.6.kW 718.2.kW
472 339
H5 AH3,C2=5.34 m2

450 355 AH3,C3=19.90 m2


C1
0.834
AH4,C2=32.21 m2
478 366
C2 AH5,C1=82.94 m2
0.28
494 311
C3 AH4,CU=32.68 m2

433 0.14 333 AH5,CU=33.16 m2


C4
495 389
C5

TAC= 4.3242542 x 104 $/year

Fig. A4. HEN configuration for the case study 02 using Hyperstructure.
820 M. Escobar, J.O. Trierweiler / Applied Thermal Engineering 51 (2013) 801e826

Fig. A5. HEN configuration for the case study 02 using Synheat.

Fig. A6. HEN configuration for the case study 03 using Superstructure.
M. Escobar, J.O. Trierweiler / Applied Thermal Engineering 51 (2013) 801e826 821

3150kW
180 75
H1
4232.7 kW

0.49 1221 kW 0.29 4146.2 kW


280 120
H2
2067.27 kW 1082.7 kW 75
180
H3
900kW 2100kW
40
140
H4
1017.3 kW 3982.7
120
220
H5
1796 kW

0.43 2578.9
180 55
H6
4200kW
60
200
H7
0.15 2316.7 447.3 kW
40
120
H8
3904.6
230 40
C1

220 0.40
100
C2
190
40
C3
190 50
C4

0.25
250
50
C5

190 853.8 90
C6
5400kW
250 60
C7

TAC= 1.57494821 x 106 $/year

Fig. A7. HEN configuration for the case study 03 using Hyperstructure.
3150 kW
180 75
H1
4865.67 kW 4754.33 kW
280 120 AH2,C5,1=242.284 m2
H2
3150 kW 75 AH5,C1,1=25.94 m2
180
H3
1802 kW 900kW 1198 kW AH6,C3,1=291.67 m2
40
140
H4 AH7,C4,1=1260 m2
795.065 kW 4204.9 kW 120
220 AH2,C2,2=475.43 m2
H5
4375 kW
180 55 AH1,C5,4=292.72 m2
H6
4200 kW 60 AH3,C5,4=292.72 m2
200
H7 AH5,C6,4=596.5 m2
795.065 kW 854.329 kW 6350.6 kW
40
120
H8 AH4,C1,5=168.125 m2
1202.9 kW
230 40
C1 AH8,C6,7=134.167 m2
2445.67 kW
220 AH8,C5,8=42.2 m2
100
C2
190 875kW AH4,CU=70.65 m2
40
C3
AH8,CU=474.78 m2
190 50
C4
AHU,C1=13.94 m2

250 AHU,C2=31.08 m2
50
C5
AHU,C3=6.54 m2
190 90
C6 AHU,C7=94.66 m2
5400kW
250 60
C7

TAC= 1.50666740 x 106 $/year

Fig. A8. HEN configuration for the case study 03 using Synheat.

Fig. A9. HEN configuration for the case study 04 using Superstructure and Hyperstructure (a), and Synheat (b).
M. Escobar, J.O. Trierweiler / Applied Thermal Engineering 51 (2013) 801e826 823

950kW
0.54 0.26
180 2200kW
H1 75
150 2250kW
280
H2 120
1750
0.19 1400
180 0.71
H3 75
1900kW 950kW
140
H4 45
220 2500kW
H5 120
180 250 1000
H6 55
170 1750kW 2000kW
H7 45
1750
2150kW
180 0.47 50
H8
280 2250kW 600kW
H9 90
2350kW 1250kW
180 60
H10
2250
120
H11 45
1800 700
220
H12 120
1250
180 55
H13
H14 140 45
950 1050

1800
140 1800 1900kW
H15 60
1300
0.51
1250
220 50
H16
220 1600kW
H17 60
150
0.09
1050
150 70
H18
0.29 3150
80
H19 140
3150 2800kW
220 50
H20
1050 50
180 60
H21
150 2100kW
H22 45

0.36
230 C1
40
0.71 0.35
2150kW 120
260 C2
0.51

0.16 0.16
40
190 C3
0.51
0.38 1450 50
190 C4
2300kW 0.31
50
250 C5
0.46 0.49
0.54 40
150 C6
0.5
150 C7
40
210 C8
120
130 C9
40
C10
120 60
150 C11
50
130 C12
40
160 C13
120
90 C14
40
90 C15
50
0.68
50 C16
150
30
150 C17

TAC= 3.08624799 x 106 $/year

Fig. A10. HEN configuration for the case study 05 using Superstructure.
824 M. Escobar, J.O. Trierweiler / Applied Thermal Engineering 51 (2013) 801e826

2050 kW 1100 kW
180
H1 75
280 2400 kW
H2 120
1400 kW 1750 kW
180
H3 75
1900 kW 950 kW
H4 140 45
1900 kW
220 120
H5 250 kW
1000kW
180 55
H6
2000kW 1750kW
170 45
H7
1750kW 2150kW
0.19
180 50
H8
280 2850kW
H9 90
2227.44 1372.6 kW
180
H10 60
2250kW
120 45
H11 423.2
220 2076.8 kW
H12 120
1250 kW
180 55
H13
1900 kW
H14 140 45
0.1
1800
0.44 661.9
1338.1 kW
140 0.04 1800
H15 60
1600
0.43
0.15 950 kW
220 50
H16 1600 kW
220
H17 60
1061.9
0.73
0.21 538
150 70
H18 1050
0.28 3150
H19 140 80
3150
220 50
H20
2800 kW
1150 kW 50 kW
180 60
H21
150 45
H22
1100 kW
0.62
0.24
230 C1
40
2823.19 kW
120
260 C2
0.28

0.44 40
190 C3
988.13 kW 0.4
50
190 C4

2549.38 kW 0.59
50
250 C5
0.28 0.02
0.29
0.55 40
150 C6
950 kW
150 C7
40
210 120 C8
130 C9
40
120 C10
60
150 C11
50
130 C12
40
160 C13
120
90 C14
0.5 40
50
90 C15
1250 kW 50
150 C16
150 30
C17
0.28

TAC= 2.1976329 x 106 $/year

Fig. A11. HEN configuration for the case study 05 using Hyperstructure.
M. Escobar, J.O. Trierweiler / Applied Thermal Engineering 51 (2013) 801e826 825

150kW

3000 kW
180
H1 75
2155.52 kW 244.48 kW
280 120
H2
3150 kW
180
H3 75
1900kW 2850 kW
140
H4 45
876.71 kW
233.2 kW
1400 kW
220
H5 120
1250 kW
180 55
H6
3750kW
170 45
H7
3900kW
180 50
H8
2139.58 kW 710.42 kW
280 90
H9
2600 kW 1000 kW
180 60
H10
2250 kW
120
H11 45
2273.19 kW
220
H12 120
460.58 kW 226.81 kW
789.42
180 55
H13
150 kW
1750 kW
H14 140 45
1866.76
1933.24kW
140 1800 kW
H15 60
2207.55 342.45 kW
220 50
H16 1600
220
H17 60

1600 kW
150 70
H18
3000

1200
80
H19 140
5204.7 kW 745.32 kW
220 50
H20
200
1000
180 60
H21
150 45
H22
1200 kW
230 C1
40
604.9 kW
120
260 C2

560.58 kW
40
190 C3

439.42 kW
50
190 C4

2987.8 kW
50
250 C5

100 kW
40
150 C6
333.24 kW
150 C7
40

210 120 C8

130 C9
40
C10
120 60
150 C11
50

130 C12
40
160 C13
120
90 C14
40
90 C15
50
50 C16
150

30
150 C17

TAC= 2.05542113 x 106 $/year

Fig. A12. HEN configuration for the case study 05 using Synheat.
826 M. Escobar, J.O. Trierweiler / Applied Thermal Engineering 51 (2013) 801e826

References [15] K.M. Bjork, F. Pettersson, Optimization of large-scale heat exchanger network
synthesis problems, in: IAESTED International Conference, Modeling and
Simulation, 313, 2003.
[1] T. Gundersen, L. Naess, The synthesis of cost optimal heat exchange networks e
[16] A. Sorsak, Z. Kravanja, MINLP retrofit of heat exchanger networks comprising
an industrial review of the stage of the art, Computers & Chemical Engineering
different exchanger types, Computers & Chemical Engineering 28 (2004) 235e251.
12 (1988) 503.
[17] B.A. Murtagh, M.A. Saunders, MINOS 5.5 User’s Guide. Technical Report SOL
[2] K.C. Furman, N.V. Sahinidis, A critical review and annotated bibliography for
83-20R, Systems Optimization Laboratory, Department of Operations
heat exchanger network synthesis in the 20th century, Industrial & Engi-
Research, Stanford University, Stanford, California, 94305-4022, 1998.
neering Chemistry Research 41 (2002) 2335.
[18] P.E. Gill, W. Murray, M.A. Saunders, SNOPT: An SQP Algorithm for Large-scale
[3] C.A. Floudas, Nonlinear and Mixed-integer Optimization: Fundamentals and
Constrained Optimization. Numerical Analysis Report 97-2, Department of
Applications, Oxford University Press, New York, 1995.
Mathematics, University of California, San Diego, 1997.
[4] E.C. Hohmann, Optimum Networks for Heat Exchange. Ph.D. Thesis, University
[19] R.A. Waltz, J.L. Morales, J. Nocedal, D. Orban, An interior algorithm for
of Southern California, Los Angeles, CA, 1971.
nonlinear optimization that combines line search and trust region steps,
[5] B. Linnhoff, J.R. Flower, Synthesis of heat exchanger networks I. Systematic
Technical Report 2003-6.
generation of energy optimal networks, AIChE Journal 24 (1978) 633.
[20] N.V. Sahinidis, BARON: a general purpose global optimization software
[6] B. Linnhoff, E. Hindmarsh, The pinch design method for heat exchanger
package, Journal of Global Optimization 8 (2) (1996) 201e205.
networks, Chemical Engineering Science 38 (5) (1983) 745.
[21] J. Viswanathan, I.E. Grossmann, Combined penalty function and outer-
[7] S. Papoulias, I.E. Grossmann, A structural optimization approach in process
approximation method for MINLP optimization, Computers & Chemical
synthesis II. Heat recovery networks, Computers & Chemical Engineering 7 (6)
Engineering 14 (1990) 769.
(1983) 1707.
[22] M.R. Bussiek, A.S. Drud, SBB: A New Solver for Mixed Integer Nonlinear
[8] J. Cerda, A.W. Westerberg, Synthesizing heat exchanger networks having
Programming (2001). Slides: gams.com/presentations/or01/sbb.pdf.
restricted stream/stream matches using transportation problem formulations,
[23] B. Lin, D.C. Miller, Solving heat exchanger network synthesis problems with
Chemical Engineering Science 38 (10) (1983) 1723e1740.
Tabu search, Computers & Chemical Engineering 28 (2004) 1451e1464.
[9] C.A. Floudas, A.R. Ciric, I.E. Grossmann, Automatic synthesis of optimum
heat exchanger network configurations, AIChE Journal 32 (2) (1986) [24] A. Pariyani, A. Gupta, P. Ghosh, Design of heat exchanger networks using
randomized algorithm, Computers & Chemical Engineering 30 (2006) 1046e1053.
276.
[25] K.M. Yerramsetty, C.V.S. Murty, Synthesis of cost-optimal heat exchanger
[10] C.R. Ciric, C.A. Floudas, Heat exchanger network synthesis without decom-
networks using differential evolution, Computers & Chemical Engineering 32
position, Computers & Chemical Engineering 15 (6) (1991) 385.
(2008) 1861e1876.
[11] T.F. Yee, I.E. Grossmann, Simultaneous optimization models for heat inte-
[26] F. Pettersson, Synthesis of large-scale heat exchanger networks using
grations II. Heat exchanger network synthesis, Computers & Chemical Engi-
a sequential match reduction approach, Computers & Chemical Engineering
neering 14 (10) (1990) 1165.
29 (2005) 993e1007.
[12] L.T. Biegler, I.E. Grossmann, A.W. Westerberg, Systematic Methods of
[27] X. Luo, G. Fieg, K. Cai, X. Guan, Pettersson, Synthesis of large-scale heat
Chemical Process Design, Prentice Hall, Upper Saddle River, New Jersey,
exchanger networks by a monogenetic algorithm, Computer Aided Chemical
1997.
Engineering 27 (2005) 729e734.
[13] J.J.J. Chen, Comments on improvements on a replacement for the logarithmic
[28] P. Ernst, G. Fieg, X. Luo, Synthesis of large-scale heat exchanger networks
mean, Chemical Engineering Science 42 (1987) 2488.
using a sequential match reduction approach, Heat and Mass Transfer 46
[14] T. Gundersen, A Process Integration, PRIMER, SINTEF Energy Research, Inter-
(2010) 1087e1096.
national Energy Agency, 2000.

S-ar putea să vă placă și