Professional Documents
Culture Documents
Centre for Intelligent Systems Research (CISR), Deakin University, Geelong, VIC 3217, Australia
Department of Mechanical Engineering, Vali-e-Asr University of Rafsanjan, Rafsanjan, Iran
a r t i c l e
i n f o
Article history:
Received 26 June 2014
Accepted 12 September 2014
Available online 17 October 2014
Keywords:
Heat exchanger
Optimization
Genetic algorithm
Firey algorithm
Cuckoo search
a b s t r a c t
This paper comprehensively investigates performance of evolutionary algorithms for design optimization
of shell and tube heat exchangers (STHX). Genetic algorithm (GA), rey algorithm (FA), and cuckoo
search (CS) method are implemented for nding the optimal values for seven key design variables of
the STHX model. -NTU method and Bell-Delaware procedure are used for thermal modeling of STHX
and calculation of shell side heat transfer coefcient and pressure drop. The purpose of STHX optimization is to maximize its thermal efciency. Obtained results for several simulation optimizations indicate
that GA is unable to nd permissible and optimal solutions in the majority of cases. In contrast, design
variables found by FA and CS always lead to maximum STHX efciency. Also computational requirements
of CS method are signicantly less than FA method. As per optimization results, maximum efciency
(83.8%) can be achieved using several design congurations. However, these designs are bearing different
dollar costs. Also it is found that the behavior of the majority of decision variables remains consistent in
different runs of the FA and CS optimization processes.
2014 Elsevier Ltd. All rights reserved.
1. Introduction
Shell and tube heat exchangers (STHX) play a critical role in
operation of many industrial plants including oil reneries, power
stations, and manufacturing sites. By far, they are the most widely
used type of heat exchanger used in different industries. Optimal
design of STHX is a challenging engineering task. Several criteria
such as efciency and capital, operating, and energy costs can be
considered in the design. As mentioned in [1], the design process
has an iterative nature and includes several trials for obtaining a
reasonable conguration that fullls the design specications and
satises the trade-off between pressure drops and thermal
exchange transfers. No doubt, this process is massively time-consuming and expert expensive. Furthermore, there is no guarantee
that the nal design is optimal in terms of considered criteria due
to the limited capability of the design engineers in consideration
and evaluation of all admissible designs. Budget constrains during
the design phase even worsen this. So it is not surprising to see real
world STHX that their designs is far away from being optimal.
Fig. 1 displays the layout and uid ows of a typical STHX. Bafes placed along the tube bundle force the uid to ow through
tubes [2]. Bafes simply intensify the turbulent level and improve
Corresponding author.
http://dx.doi.org/10.1016/j.enconman.2014.09.039
0196-8904/ 2014 Elsevier Ltd. All rights reserved.
282
method by process engineering researchers for design optimization of heat exchangers [814]. Several optimization methods have
been introduced in recent years that outperform genetic algorithm
in term of optimization results. Also some of these methods are
even computationally less demanding. Examples of these methods
are particle swarm optimization [15,16], cuckoo search [17], imperialist competitive algorithm [18], bee colony optimization [19],
and rey algorithm [20]. These methods show different performances in different engineering applications. A conceptual comparison of these methods for several case studies can be found in
[21]. A few of these algorithms have been recently employed for
design and optimization of heat exchangers [2228].
The purpose of this paper is to comprehensively compare performance of the genetic algorithm, rey algorithm, and cuckoo
search method for the design of STHXs. To the best of our knowledge, this is the rst study where rey algorithm and cuckoo
search method are employed for optimal design of STHXs. Seven
design variables are considered as part of the optimization process.
These are tube arrangement, pitch ratio, diameter, length, quantity,
bafe spacing ratio, and bafe cut ratio. Optimization is purely
done for maximizing the efciency. Cost implications of this optimization approach are then analyzed and discussed. Performance
of optimization algorithms is compared on their ability to nd permissible and optimal congurations. The behavior of the seven
design variables are also studied in detail. Simulation experiments
are done for an approximate thermal model of a real world STHX.
The rest of this paper is organized as follows. Section 2 briey
introduces the STHX model used in this study. Optimization algorithms investigated in this study are briey described in Section 3.
Section 4 represents simulations results. Finally, conclusions are
provided in Section 5.
_ p s ; mc
_ p t
C min
minC s ; C t min mc
C
_ p s ; mc
_ p t
C max maxC s ; C t max mc
C min minC h ; C c
where C h and C c are the hot and cold uid heat capacity rates, i.e.,
_ p h and C c mc
_ p c . m
_ is the uid mass ow rate. Specic
C h mc
heats cp are assumed to be constant.
The overall heat transfer coefcient (U o ) in (3) is then computed
as,
Uo
1
do lndo =di
do
do
Ro;f
Ri;f
ho
2kw
di hi di
1
5
where L; N t ; di ; do ; Ri;f ; Ro;f , and kw are the tube length, number, inside
and outside diameter, tube and shell side fouling resistances and
thermal conductivity of tube wall respectively. hi and ho are heat
transfer coefcients for inside and outside ows, respectively.
The total tube outside heat transfer area is calculated as.
At p L do Nt
hi 0:024
kt 0:8 0:4
Re Pr t
di t
for 2500 < Ret < 124; 000. kt and Prt are tube side uid thermal
conductivity and Prandtl number respectively. The tube ow Reynold number (Ret ) is also dened as,
Ret
mt di
lt Ao;t
Ao;t 0:25pdi
1
where mt is the tube mass ow rate and Ao;t is the tube side ow
cross section area per pass,
p !1
q
NTU 1C 2
2 1 e
p
2 1C 1C
2
1 eNTU 1C
U o At
C min
NTU
Nt
np
hs hk J c J l J b J s J r
2
where subscripts s and t stand for shell and tube respectively. The
number of transfer units is dened as,
10
hk ji cp;s
23
ks
m_ s
As
cp;s ls
ls
ls;w
!0:14
11
where ji is the Colburn j-factor for an ideal tube bank. As is also the
cross ow area at the centerline of the shell for one cross ow
between two bafes. lls is the viscosity ratio at bulk to wall tempers;w
ature in the shell side. J c ; Jl ; J b ; Js , and J r in (10) are the correction factors for bafe conguration (cut and spacing), bafe leakage, bundle
and pass partition bypass streams, bigger bafe spacing at the shell
inlet and outlet sections, and the adverse temperature gradient in
laminar ows.
The STHX total cost is made up of capital investment (C inv ) and
operating (C opr ) costs [1],
12
13
C opr
Ny
X
C0
k1 1 i
14
where i and N y are the annual discount rate (%) and the STHX life
time in year. C 0 is the annual operating cost and is calculated as
follows,
C 0 je P hopt
15
where je and hopt are the price of electricity ($/kW h) and annual
operating hours. The pumping power (P) is also calculated in watts
(W),
1 mt
g qt
Dpt
ms
qs
Dps
16
b b0 ec r
3. Optimization algorithms
3.1. Genetic algorithm
Genetic algorithm (GA) is highly likely the most widely used
and researched evolutionary optimization method in the scientic
world. It is a guided stochastic search technique inspired from the
principles of natural ttest selection and population genetics. In
general terms, it is based on the parent and offspring iterations
and their evolutions through generations. GA generates candidate
solutions from the space of all possible solutions and examines
their performance as per the considered objective function. It has
been proven that GA performs strongly well in both constrained
and unconstrained search problems where the number of good
solutions is very limited compared to the size of the search space.
GA converges towards more competitive solutions by applying
elitism, crossover, and mutation mechanisms. GA rst creates a
population (often randomly) of potential solutions (also called
chromosomes) for the optimization problem. This population is
then assessed using the objective function of the interest. Then
GA uses its three operators to create the new population for the
next generation. The best performing chromosome(s) is copied to
the next generation unchanged. This process is called elitism and
makes sure that the best solution(s) is not lost as the optimization
proceeds.
Crossover operator is used for combing good parents and generating offspring. This operator is applied with the hope of retaining
the spirit of good chromosomes. In its simplest form, i.e., single
point, a random point (crossover point) is randomly selected. Then
the operator swaps portions of a pair chromosomes at the crossover point. Alternative crossover methods are multi-points, uniform, and arithmetic. Regardless of the type of applied crossover
operator, its generated offspring only include information held by
the current population. A new operator is required to introduce
and bring new information (solutions) to the population. Mutation
operator creates a new offspring by randomly changing the values
of genes at one or more positions of a selected chromosome. The
pseudo code for GA including three genetic operators is displayed
in Fig. 2.
283
17
r kxi xj k
r
Xd
x xj;k 2
k1 i;k
18
q
xi xj 2 yi yj 2
19
Assuming the jth rey is brighter than ith rey, the movement of xi towards xj is dened as,
c r2i;j
xi xi b0 e
xi xj a i
20
where the second and the third term in right are due to the attraction and randomization. a is a parameter multiplied in the vector of
random numbers i . This vector is generated through drawing numbers from a normal or uniform distribution. As mentioned in [20],
often b0 1 and a 2 0; 1 satisfy most of FA implementations. Note
that (20) is a pure random walk search if b 0. Also other distributions such as Levy ights can be considered for the randomization
terms () in (20).
284
The Lvy ight provides a random walk where its step is drawn
from a Lvy distribution. There are several ways to generate this
random step [20]. The Mantegnas algorithm is one of the most
efcient algorithms for generating symmetric (positive or negative) Lvy distributed steps. In this method, the step length in
(21) is calculated as
22
jv j1=b
where u and
3.3. Cuckoo search
u N0; ru ;
where
21
where a is the step size which depends on the scales of the problem
of interest. Often, a OL=100 satises the search requirements for
most optimization problems. L represents the difference between
the maximum and minimum valid value of the problem of interest.
The product means entry-wise multiplication.
v N0; rv
rv 1 and,
ru
23
C1 b sinpb=2
C1 b=2 b 2b1=2
)1b
24
Cz
t z1 et dt
25
xt1
xti a s xti xbest
r
i
i
26
where xbest
is the current best solution and r is a random number
i
drawn from a normal distribution with zero mean and unit variance. The step length s is also calculated using (22). Further discussion about CS method and its details can be found in [20,17].
4. Simulation results
This section describes the simulation results for optimizing the
design of STHX using GA, FA, and CS method. STHX model used in
simulations is identical to one described and analyzed in [4].
Table 1 summarizes the list of decision variables (STHX parameters) and their range. It is important to note that all these 7 variables are discontinuous due to practical construction constraints.
For instance, tube internal diameter is determined according to relevant standards and suppliers catalogs.
For the three optimization methods, we set the number of iterations (generations) to 30 and 60. The population size is also set to
10, 20, 30, and 50. Accordingly, 8 different sets of experiments are
performed for each optimization method (combination of different
population sizes and iteration numbers). Each experiment (e.g., GA
with 30 iterations and 10 populations) is repeated 50 times and
then statistics of experiments are reported. In total, 400 runs are
simulated and completed for each optimization method. This is
done to make sure conclusions are made based on general and
extensive optimization scenarios rather than a few tailored ones.
Therefore, obtained results and driven conclusions are statistically
meaningful and believable. Simulations are performed using a
Lenovo Thinkpad T420s laptop computer with Intel Core i72640 M CPU @2.8G Hz and 8 GB memory, running Windows 7
Professional.
The purpose of optimization is to maximize the efciency
through nding the best values for seven design parameters listed
in Table 1. For each run, the seven decision variables are randomly
initialized within their range (see Table 1).
285
Tube arrangement
(30, 45,
90)
0.0112
0.0153
1.25
3
100
0.19
0.2
3
8
600
0.32
1.4
0.001
0.001
1
0.001
0.001
Tube inside
diameter (m)
pt/do
Tube length (m)
Tube number
Bafe cut ratio
Bafe spacing ratio
#iter = 30,
#iter = 30,
#iter = 30,
#iter = 30,
#iter = 60,
#iter = 60,
#iter = 60,
#iter = 60,
0
5
10
15
20
25
30
35
40
45
10
15
20
25
30
35
40
45
0
0
10
15
20
25
30
35
40
45
0
0
10
15
20
25
30
35
40
45
50
10
15
20
25
30
35
40
45
50
40
45
50
40
45
50
40
45
50
#iter=30, #pop=50
0
5
10
15
20
25
30
35
#iter=60, #pop=20
100
50
0
0
Efficiency (%)
50
0.97
1.01
1.49
2.53
1.01
1.97
2.97
5.02
50
50
#iter=60, #pop=30
100
100
Efficiency (%)
50
CS
3.77
9.79
21.39
61.39
4.97
19.15
43.73
122.94
50
#iter=60, #pop=10
100
FA
0.90
1.03
1.53
2.53
0.55
1.01
1.56
2.54
50
#iter=30, #pop=30
GA
#iter=30, #pop=20
100
50
50
#pop = 10
#pop = 20
#pop = 30
#pop = 50
#pop = 10
#pop = 20
#pop = 30
#pop = 50
Efficiency (%)
Efficiency (%)
Simulation
GA
FA
CS
50
100
Efficiency (%)
Table 2
The mean of computation time for each optimization run.
#iter=30, #pop=10
100
Efficiency (%)
Efficiency (%)
Efficiency (%)
10
15
20
25
30
35
#iter=60, #pop=50
100
50
0
0
10
15
20
Replicate
Fig. 4. STHX efciency optimization using GA, FA, and CS method for 50 runs.
25
30
Replicate
35
50,000
45,000
45,000
40,000
40,000
286
35,000
30,000
35,000
30,000
25,000
25,000
20,000
20,000
15,000
82
82.5
83
83.5
15,000
82
84
82.5
Efficiency (%)
83
83.5
84
Efficiency (%)
100
Tube Diameter
Tube Arrangement
Fig. 6. The scatter plot of efciency and dollar cost for solutions found by FA (left) and CS (right).
80
60
40
20
0
10
20
30
40
20
10
0
50
10
20
30
40
50
10
20
30
40
50
10
20
30
40
50
Length
Pitch Ratio
2000
1000
2000
0
0
10
20
30
40
50
600
Spacing Ratio
Tube Number
4000
400
200
0
0
10
20
30
40
150
100
50
0
50
Replicate
Replicate
Cut Ratio
1500
1000
500
0
0
10
20
30
40
50
Replicate
Fig. 7. Optimal values of STHX design parameters in 50 runs of FA optimization.
287
100
Tube Diameter
Tube Arrangement
80
60
40
20
0
10
20
30
40
20
10
0
50
10
20
30
40
50
10
20
30
40
50
10
20
30
40
50
1000
2000
0
0
10
20
30
40
50
600
Spacing Ratio
Tube Number
4000
Length
Pitch Ratio
2000
400
200
0
0
10
20
30
40
150
100
50
0
50
Replicate
Replicate
Cut Ratio
1500
1000
500
0
0
10
20
30
40
50
Replicate
Fig. 8. Optimal values of STHX design parameters in 50 runs of CS optimization.
GA is not something to be rectied purely by increasing the number of optimization generations or the population size. Even if the
performance is improved, the computational burden for nding
globally optimal solutions will be massive.1
Fig. 6 displays the scatter plot of efciency and dollar cost for
STHX optimized using FA (left) and CS (right) methods. These
results are from the eighth experiment (#iter = 50, #pop = 60). It
is easy to see that while efciency is almost the same for in the
majority of experiments (83.80%), there is a huge difference in
terms of the dollar cost. The total cost for the majority of solutions
found by FA and CS methods is around $45,000. Also the plot
clearly shows that the total cost increases as the efciency
increases. This is consistent with ndings in [4]. As per results in
this gure, designs identical in terms of efciency can have completely different total costs.
The optimal values for seven design variables obtained using FA
optimization are shown in Fig. 7. These are plotted for fty runs of
FA simulation (#8) to see how their values change from one simulation to another. The followings are observed:
The optimal values for tube arrangement are 30 and 90. The
interesting point is that 45 arrangement is not returned as a
solution for maximizing the efciency.
Tube diameter and pitch ratio often take a value between there
minimum and median in 50 runs. This tendency is in particular
more obvious for the pitch ratio.
Tube lengths between 3 m and 8 m are returned in different
optimization runs. However, there is a tendency towards smaller values.
1
Note that this does not mean that GA is not a suitable tool for STHX design
optimization. GA can generate optimal results if initialization is performed properly
(admissible values are rst picked and assigned to design parameters).
288