SIMULATION OF COMPACT HEAT EXCHANGERS USING GLOBAL

REGRESSION AND SOFT COMPUTING
A Dissertation
Submitted to the Graduate School
of the University of Notre Dame
in Partial Fulfillment of the Requirements
for the Degree of
Doctor of Philosophy
by
Arturo J. Pacheco-Vega, B.S., M.S.
Mihir Sen, Director
Department of Aerospace and Mechanical Engineering
Notre Dame, Indiana
April 2002
SIMULATION OF COMPACT HEAT EXCHANGERS USING GLOBAL
REGRESSION AND SOFT COMPUTING
Abstract
by
Arturo J. Pacheco-Vega
This dissertation investigates enhancement in accuracy of heat rate predictions
in compact fin-tube heat exchangers. The sources of error from a conventional
approach based on correlating heat transfer coefficients, sometimes of 25–30%, are
studied first. These include the idealized assumptions in the procedure by which
correlations are found, the data compression that occurs through the correlation
process, and the multiplicity of solutions for a proposed correlating function obtained
using local regression.
To remove the non-uniqueness of regression results, a methodology based on
global optimization techniques that include genetic algorithms, simulated annealing
and interval analysis is introduced. Applications of globally-determined correlations
to single-phase and condensing coils demonstrate improved accuracy with errors of
only 3%. The issue of degradation due to the idealized assumptions in the procedure
is addressed by correlating the error generated from these with flow rates and inlet
temperatures. Heat rate estimations are improved by a factor of three. The en-
hancement in predictions by direct correlations of heat rates instead of heat transfer
coefficients is also investigated. Using global regression, direct-heat-rate and heat-
transfer-coefficient-based correlations are first determined and later applied to heat
exchanger data. The direct heat-rate estimations are more accurate.
Arturo J. Pacheco-Vega
The effect of the functional form of the correlation is explored through a set of
possible correlating functions. Two globally-determined correlations, a power-law
and an inverse linear, are selected from this set to demonstrate that their accuracy
in estimations are comparable, with errors of 2.7% and 3.4% respectively.
The compression of experimental information is addressed using artificial neural
networks. The network is trained using information of operational and geometrical
parameters. Comparisons against conventional correlations demonstrate enhance-
ments in accuracy of sometimes 400%. Cluster analysis has also been applied to
condensing heat exchangers with validation of its excellent characteristics for data
classification. Finally, the problem of accuracy in estimations from neural networks
using limited data is analyzed. A cross-validation methodology to estimate expected
errors is developed. This approach produces an upper bound on the estimated error
in heat rates as demonstrated from its application to evaporating coils.
To my daughters
Elia, Nadia and Andrea
for you are my little treasure
ii
CONTENTS
TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi
FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
ACKNOWLEDGMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x
NOMENCLATURE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii
CHAPTER 1: GENERAL INTRODUCTION . . . . . . . . . . . . . . . . . . 1
1.1 Heat transfer characteristics and steady state simulation of heat ex-
changers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Soft Computing technologies and thermal engineering . . . . . . . . . 6
1.3 Interval analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Overview of dissertation . . . . . . . . . . . . . . . . . . . . . . . . . 9
CHAPTER 2: PROBLEM DESCRIPTION . . . . . . . . . . . . . . . . . . . 11
2.1 Complexity of heat exchanger analysis . . . . . . . . . . . . . . . . . 12
2.2 Coefficient of heat transfer and correlations in heat exchanger analysis 13
2.2.1 Correlations of the heat transfer coefficient . . . . . . . . . . . 14
2.2.2 The overall heat transfer rate equation . . . . . . . . . . . . . 15
2.2.3 Functional form of the overall heat transfer coefficient . . . . . 16
2.2.4 Experimental separation of the overall thermal resistance . . . 17
2.2.5 Techniques to find correlations from experiments . . . . . . . . 18
2.3 Objectives and scope of the present dissertation . . . . . . . . . . . . 22
CHAPTER 3: HEAT EXCHANGER EXPERIMENTAL DATA . . . . . . . . 23
3.1 Experimental data for heat exchanger 1 . . . . . . . . . . . . . . . . . 23
3.2 Published experimental data and correlations for heat exchanger 2 . . 26
3.3 Manufacturer’s experimental data for heat exchanger 3 . . . . . . . . 29
iii
CHAPTER 4: NONLINEAR REGRESSION ANALYSIS AND GLOBAL OP-
TIMIZATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.1 Multiplicity of solutions in nonlinear least-squares analysis . . . . . . 33
4.1.1 Multiplicity of “best” correlations in heat exchanger 1 . . . . . 33
4.1.2 Multiplicity of “best” correlations in heat exchanger 2 . . . . . 40
4.2 Global optimization techniques . . . . . . . . . . . . . . . . . . . . . 41
4.2.1 Genetic algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.2.2 Simulated annealing . . . . . . . . . . . . . . . . . . . . . . . 43
4.2.3 Interval analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.3 Global-based-regression for correlations . . . . . . . . . . . . . . . . . 47
4.3.1 Application to heat exchanger 1 . . . . . . . . . . . . . . . . . 48
4.3.2 Application to heat exchanger 2 . . . . . . . . . . . . . . . . . 51
4.4 Coupling between hydrodynamics and temperature . . . . . . . . . . 56
4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
CHAPTER 5: DIRECT CORRELATION OF HEAT RATE . . . . . . . . . . 64
5.1 Direct correlation of heat rate in heat exchanger 2 . . . . . . . . . . . 65
5.2 Proposed functional forms for direct correlation in heat exchanger 1 . 68
5.3 Three- and four-parameter search by genetic algorithms . . . . . . . . 69
5.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
CHAPTER 6: HEAT RATE PREDICTIONS IN HUMID AIR-WATER HEAT
EXCHANGERS USING NEURAL NETWORKS . . . . . . . . . . . . . . 77
6.1 Artificial neural networks . . . . . . . . . . . . . . . . . . . . . . . . . 78
6.1.1 Separation of data for training and testing . . . . . . . . . . . 80
6.1.2 Heat exchanger analysis . . . . . . . . . . . . . . . . . . . . . 83
6.2 Artificial neural network results . . . . . . . . . . . . . . . . . . . . . 85
6.3 Clustering of experimental data . . . . . . . . . . . . . . . . . . . . . 91
6.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
CHAPTER 7: NEURAL NETWORK HEAT RATE ESTIMATIONS WITH
LIMITED EXPERIMENTAL DATA . . . . . . . . . . . . . . . . . . . . . 97
7.1 Artificial neural network analysis . . . . . . . . . . . . . . . . . . . . 98
7.2 Predictions of the heat rate with ANNs . . . . . . . . . . . . . . . . . 99
7.3 Error sources in ANN estimation . . . . . . . . . . . . . . . . . . . . 102
7.4 ANN-based error estimate methodology . . . . . . . . . . . . . . . . . 105
7.5 Validation of error estimate procedure . . . . . . . . . . . . . . . . . 108
7.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
CHAPTER 8: GENERAL CONCLUSIONS . . . . . . . . . . . . . . . . . . . 111
8.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
8.2 Recommendations for future work . . . . . . . . . . . . . . . . . . . . 114
iv
APPENDIX A: GENETIC ALGORITHMS . . . . . . . . . . . . . . . . . . . 117
A.1 Optimization of a function of one variable . . . . . . . . . . . . . . . 118
A.2 Optimization of a function of two variables . . . . . . . . . . . . . . . 120
A.3 Properties of crossover . . . . . . . . . . . . . . . . . . . . . . . . . . 120
APPENDIX B: SIMULATED ANNEALING . . . . . . . . . . . . . . . . . . . 132
B.1 Optimization of a function of one variable . . . . . . . . . . . . . . . 134
B.2 Optimization of a function of two variables . . . . . . . . . . . . . . . 136
APPENDIX C: INTERVAL ANALYSIS . . . . . . . . . . . . . . . . . . . . . 137
C.1 Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
C.2 Optimization of a function of one variable . . . . . . . . . . . . . . . 142
C.3 Optimization of a function of two variables . . . . . . . . . . . . . . . 144
APPENDIX D: EXTRAPOLATION IN NEURAL NETWORKS . . . . . . . 147
D.1 Neural network analysis of a two-dimensional function . . . . . . . . . 147
D.2 Neural network analysis of the Sieder and Tate (1936) data . . . . . . 149
BIBLIOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
v
TABLES
4.1 COMPARISON OF CORRELATIONS FOR SINGLE PHASE HEAT
EXCHANGER. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.2 COMPARISON OF PERCENTAGE ERRORS IN (U)
−1
AND
˙
Q
PREDICTIONS BETWEEN VARIOUS CORRELATIONS FOR SIN-
GLE PHASE HEAT EXCHANGER. . . . . . . . . . . . . . . . . . . 50
4.3 COMPARISON OF PERCENTAGE ERRORS IN j
s
, j
t
AND
˙
Q
PREDICTIONS BETWEEN VARIOUS CORRELATIONS AND THE
ANN. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.4 VALUES OF THE D-CONSTANTS IN EQ. (4.18). . . . . . . . . . . 58
4.5 COMPARISON OF PERCENTAGE ERRORS IN
˙
Q WITH = 0
AND -CORRECTION INCLUDED. . . . . . . . . . . . . . . . . . . 60
5.1 CORRELATION FUNCTIONS . . . . . . . . . . . . . . . . . . . . . 70
5.2 COMPARISON OF DIFFERENT CORRELATIONS. . . . . . . . . . 72
6.1 PERCENTAGE OF DATA SETS OF DIFFERENT SURFACE CON-
DITIONS THAT FALL IN DIFFERENT CLUSTERS. . . . . . . . . 92
A.1 REPRESENTATION OF ALL POSSIBLE MAPPINGS OF (x, y)
TO (x

, y

) OBTAINED BY CROSSOVER USING n
b
= 2 BITS PER
VARIABLE. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
A.2 NUMBER OF AVENUES N
ave
AS A FUNCTION OF THE NUM-
BER OF BITS n
b
AND CROSSOVER VALUES cp. . . . . . . . . . . 126
vi
FIGURES
1.1 Single-row fin-tube heat exchanger. . . . . . . . . . . . . . . . . . . . 2
3.1 Heat exchanger experimental facility. . . . . . . . . . . . . . . . . . . 25
3.2 Schematic of a compact fin-tube heat exchanger. . . . . . . . . . . . . 27
3.3 Schematic of a compact heat exchanger with air and R22 as fluids. . . 30
4.1 Cross-flow heat exchanger. . . . . . . . . . . . . . . . . . . . . . . . . 34
4.2 Section of S
U
surface. . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.3 Experimental vs. predicted heat flow rates for power-law correlations;
+ A, ◦ B . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.4 Ratios of predicted air- and water-side Nusselt numbers. . . . . . . . 39
4.5 Section of the surface S
js
(n
1
, n
2
, n
3
, n
4
); A is the global minimum; B
is a local minimum. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.6 Comparison in heat rate predictions from global-based correlations;
◦ genetic algorithms; simulated annealing; + interval method. . . . 52
4.7 Experimental vs. predicted heat transfer rate by the IM-based corre-
lation with = 0. Straight line is the perfect prediction. . . . . . . . . 53
4.8 Experimental vs. predicted j
s
for heat exchanger with dry surface; +
Eq. (4.12); McQuiston (1978b); ◦ Gray & Webb (1986). Straight
line is the perfect prediction. . . . . . . . . . . . . . . . . . . . . . . . 55
4.9 Experimental vs. predicted perturbation parameter . Straight line
is the perfect prediction. . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.10 Comparison in heat rate predictions from IM-based correlation with
-correction included. . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.11 Experimental vs. predicted heat transfer rate by the IM-based corre-
lation with -correction included. Straight line is the perfect prediction. 62
5.1 Experimental vs. predicted
˙
Q for a heat exchanger for dry surface.
Straight line is the perfect prediction. . . . . . . . . . . . . . . . . . . 67
vii
5.2 Change in S
˙
Q
∗ with generation; the numbers correspond to N
C
. . . . 72
5.3 Experimental vs. predicted heat flow rates for a power law correlation. 73
5.4 Experimental vs. predicted heat flow rates for an inverse linear cor-
relation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.5 Experimental vs. predicted heat flow rates for a quadratic correlation. 75
6.1 A 5-5-3-3 neural network used. . . . . . . . . . . . . . . . . . . . . . . 81
6.2 ANN prediction errors vs. percentage of data used for training; ––
error E
a
using training data; –◦– error E
b
using data not used for
training; –– error E using complete data. Error bars indicate stan-
dard deviations. The E
a
and E curves have been shifted horizontally
by −1% and 1%, respectively, for clarity. . . . . . . . . . . . . . . . . 84
6.3 Experimental vs. predicted j
s
for heat exchanger with dry surface; +
ANN; McQuiston (1978b); ◦ Gray & Webb (1986). Straight line is
the perfect prediction. . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.4 Experiments vs. predictions for heat exchanger with dropwise con-
densation; + ANN; McQuiston (1978b). Straight line is the perfect
prediction. (a) Sensible heat j
s
, (b) total heat j
t
. . . . . . . . . . . . 88
6.5 Experiments vs. predictions for heat exchanger with film condensa-
tion; + ANN; McQuiston (1978b). Straight line is the perfect pre-
diction. (a) Sensible heat j
s
, (b) total heat j
t
. . . . . . . . . . . . . . 90
6.6 Experimental vs. predicted
˙
Q
t
for a heat exchanger under dry surface
conditions. Straight line is the perfect prediction. . . . . . . . . . . . 93
6.7 Experimental vs. predicted
˙
Q
t
for a heat exchanger under dropwise
condensation. Straight line is the perfect prediction. . . . . . . . . . . 94
6.8 Experimental vs. predicted
˙
Q
t
for a heat exchanger under film con-
densation. Straight line is the perfect prediction. . . . . . . . . . . . . 95
7.1 Configuration of a 11-11-7-1 neural network for fin-tube heat exchanger.100
7.2 Experimental vs. predicted
˙
Q
t
. Straight line is the perfect prediction. 101
7.3 Variation in
˙
Q
t
predictions from a 11-11-7-1 ANN with ˙ m
a
and T
in
a,db
. 103
7.4 Bar diagram of error estimation. . . . . . . . . . . . . . . . . . . . . . 106
7.5 Comparison of bar diagrams of error estimation from the ANN-based
methodology using different number of training cycles; 350,000
cycles; 400,000 cycles; ◦ 450,000 cycles. . . . . . . . . . . . . . . . 108
viii
A.1 Flow chart for the genetic algorithm. . . . . . . . . . . . . . . . . . . 119
A.2 Genetic algorithm for function f(x) =
sin
2
(10x)
(x+1)
. . . . . . . . . . . . . . 121
A.3 Plot of function “Peaks.” . . . . . . . . . . . . . . . . . . . . . . . . . 122
A.4 Crossover operation. . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
A.5 Mapping in a two-dimensional space using a representation of string-
length of 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
A.6 Mapping in a two-dimensional space using a representation of string-
length of 3 for cp = 1. . . . . . . . . . . . . . . . . . . . . . . . . . . 128
A.7 Mapping in a two-dimensional space using a representation of string-
length of 3 for cp = 2. . . . . . . . . . . . . . . . . . . . . . . . . . . 129
A.8 Mapping in a two-dimensional space using a representation of string-
length of 4 for cp = 1. . . . . . . . . . . . . . . . . . . . . . . . . . . 129
A.9 Mapping in a two-dimensional space using a representation of string-
length of 4 for cp = 2. . . . . . . . . . . . . . . . . . . . . . . . . . . 130
A.10 Mapping in a two-dimensional space using a representation of string-
length of 4 for cp = 3. . . . . . . . . . . . . . . . . . . . . . . . . . . 130
A.11 Mean distance l
t
as a function of the number of bits n
b
. . . . . . . . . 131
B.1 Flow chart for the simulated annealing. . . . . . . . . . . . . . . . . . 133
B.2 Simulated Annealing for function f(x) =
sin
2
(10x)
(x+1)
. Jumps are marked
as (a), (b), (c), (d), (e). . . . . . . . . . . . . . . . . . . . . . . . . . . 135
C.1 Flow chart for the interval method. . . . . . . . . . . . . . . . . . . . 141
C.2 Plot of function f(x) =
sin
2
(10x)
(x+1)
, and its first and second derivatives. . 143
C.3 Interval Newton method for function f(x) =
sin
2
(10x)
(x+1)
. . . . . . . . . . 145
C.4 Plot of function “Three-hump camel”. . . . . . . . . . . . . . . . . . 146
D.1 Neural network and polynomial approximations of function f(x).
− f(x); ∗ ANN; – – polynomial . . . . . . . . . . . . . . . . . . . . . 148
D.2 Region in parameter space of the Sieder and Tate (1936) experimental
data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
D.3 Comparison in Nu-estimations between neural networks and the Sieder
and Tate (1936) correlation. ◦ Estimations inside the cloud of exper-
imental data; + Estimations outside the cloud of experimental data. . 151
ix
ACKNOWLEDGMENTS
I would like to express my gratitude to Dr. Mihir Sen for helping me to success-
fully complete this dissertation. He is more than an advisor; he is a guide and a
friend. His broad knowledge, interest, tenacity, enthusiasm, criticism, and constant
encouragement have been essential in my formation as a researcher. Moreover, his
friendship towards me and my family is something for which I am specially grateful.
I also would like to thank my co-advisor Dr. K.T. Yang whose energy, enthusiasm,
insight and vast experience have been a source of inspiration. My appreciation goes
to Dr. Stephen Batill and Dr. Bill Goodwine for spending part of their time in
reading and making valuable suggestions to improve the quality of my work.
Thanks to Rod McClain for his help during the experimental part of this inves-
tigation, and to Donald Peterson and Jeff Squyres for creating a thesis-format-class
file to be used in conjunction with L
A
T
E
X which has been very useful in writing this
dissertation.
I want to thank all my friends at Notre Dame for the moments we have shared
during these years. I risk neglecting to mention some of them, so I will trust that
they will know I appreciate their friendship. Also, many thanks to Elia’s parents,
Nico and Chela, for their support.
My parents, Arturo and Obdulia, are examples of dedication, sacrifice, perse-
verance and, above all, of love for their children. They have provided me with
solid foundations that include persistence and honor. For all of this, and for their
constant encouragement to pursue higher goals my endless gratitude.
x
Juan, Rafael, Raul and Sigfrido, for whom I have great respect and admiration,
are my brothers and best friends. Their support and brotherhood has been very
important to successfully complete this stage in my life. The dedication of this
document barely touches my appreciation for them.
I have been a very fortunate man for sharing my life with Elia, my wife. She is an
example of courage, commitment, sacrifice, and love. Together we have had to face
many adversities to succeed in life and Elia has been essential in achieving them.
Without her love, relentless support, understanding and continuous encouragement,
it had not been possible to accomplish this goal. I would like to express my deepest
gratitude, admiration and love for her.
My daughters Elia, Nadia and Andrea have given me the most precious moments
in my life. They are an endless source of inspiration, and the joy of my everyday
life. I am proud of them as I hope they will, in the future, be proud of me. This
dissertation is specially dedicated to them.
My endless gratitude goes to my late grandfather Juan. His support and love
for me and my family was instrumental for reaching this goal.
Grief also struck me several times while at Notre Dame. One of the most painful
moments in my life came when my dearest aunt Lenny passed away. She was always
an example of goodness, love, support, and commitment. My brothers, my wife,
my daughters and I were her family, and she was a fundamental part of mine. Her
presence is greatly missed. This work is also dedicated to her memory.
I thank the people of M´exico and of the United States for a Fulbright-CONACyT
(Consejo Nacional de Ciencia y Tecnolog´ıa) scholarship that allowed me to pursue
my doctoral degree at the University of Notre Dame. I also gratefully acknowledge
other financial support by BRGD-TNDR, and Dr. Robert Nelson and the AME
xi
department at UND. Thanks to Tyler Refrigeration Corporation for sharing their
experimental data.
Last but not least, this work is dedicated to God and the Virgin Mary who in
the most difficult times, when progress was slow and my spirit was low, they gave
me the strength to stand and continue. Looking back in time, I can just say that
this has been an extraordinary experience.
xii
NOMENCLATURE
A total heat transfer area for either side [m
2
] in Chapter 2
A total air-side heat transfer area [m
2
] in Chapter 3
A
a
total air-side heat transfer area [m
2
]
A
i
, A
o
inner and outer heat transfer areas [m
2
] in Chapter 2
A
tb
surface area of tubes without fins [m
2
]
A
w
water-side heat transfer area [m
2
]
C vector of unknown constants
c
p
specific heat [J/kg K]
D tube inside diameter for heat exchanger 1 [m]
D tube outer diameter for heat exchanger 2 [m]
D tube diameter for heat exchanger 3 [m]
T mass diffusivity of fluid [m
2
/s]
D
h
hydraulic diameter [m]
E error corresponding to procedure (c)
E
a
error corresponding to procedure (a)
E
b
error corresponding to procedure (b)
G
c
air mass velocity based in free-flow area [kg/m
2
s]
F interval extension of f in IM
f, f
s
, f
t
, f
1
, f
2
functions
f
C
, g
C
functional forms for correlations
G interval extension of g in IM
xiii
G

interval extension of Hessian of f in IM
g gradient of f
H height of heat exchanger [m]
h coefficient of heat transfer [W/m
2
K]
h
fg
enthalpy of transformation, liquid to vapor [J/kg]
h
if
enthalpy of transformation, solid to liquid [J/kg]
h
t
total (enthalpy) heat transfer coefficient [kg/m
2
]
I = (θ
w
, θ
a
,
˙
M
w
,
˙
M
a
) in Chapter 5
j Colburn j-factor
k thermal conductivity of fluid [W/m K]
L length of heat exchanger [m]
L length of duct [m] in Chapter 4
m number of independent variables
˙ m mass flow rate [kg/s]
M population size in GA
M number of experimental data sets in Chapter 7
M total number of experimental data sets in Chapter 6
˙
M normalized mass flow rate
M
a
number of data sets used for training
M
b
number of data sets used for testing but not for training
M
1
number of data sets with dry surface
M
2
number of data sets with dropwise condensation
M
3
number of data sets with film condensation
N number of independent inputs in Chapter 7
N interval-Newton image, an interval vector
N
c
correlation number
xiv
N
cir
number of circuits
N
col
number of columns
N
row
number of rows
N
s
number of operations before step vector reduction in SA
N
t
number of operations before adjustment in temperature in SA
Nu Nusselt number
N
1
number of experiments, for data set 1
N
2
number of experiments, for data set 2
n
b
number of bits in GA
n
1
, n
2
, n
3
, n
4
correlation constants
O
i
output variable for run i
O
i,k
= ¦j
s
, j
t
,
˙
Q¦ for run i
P perimeter of duct [m]
Pr Prandtl number
P
s
percentage of splitting [%]
p probability in SA
p
c
probability of crossover in GA for data set 1
p
m
probability of mutation in GA
˙
Q heat transfer rate between fluids [W]
˙
Q

normalized heat transfer rate between fluids
R Euclidean distance to centroid in Chapter 7
R thermal resistance [K/W] in Chapter 2
R = Nu
¸
¸
A
/Nu
¸
¸
B
in Chapter 4
1 set of real numbers
Re Reynolds number
Re

normalized Reynolds number
xv
Re
D
Reynolds number of air based in the diameter
Re
δ
Reynolds number based in the fin spacing
r
t
reduction factor in SA
Sc Schmidt number
S
cv
absolute value of percentage error in cross-validation
S
e
root mean square of percentage error in Chapter 7
S
e
, S
e
˙
Q
root mean square error
S

variance of perturbation parameter error
S
j
variance of the Colburn j-factor error
S
U
variance of thermal resistance error [m
4
K
2
/W
2
]
S
U
mean value of S
U
S
˙
Q
variance of total heat transfer rate error [W
2
]
S
˙
Q
∗ variance of normalized heat rate error
S
˙
Q
∗ mean value of S
˙
Q

T fluid temperature [

C]
T

annealing temperature-like parameter in SA
T

normalized temperature
t fin thickness [m]
U overall coefficient of heat transfer [W/m
2
K]
V average fluid velocity [m/s]
W width of heat exchanger [m]
W synaptic weight between nodes in Chapter 6
w humidity ratio [kg/kg]
X an interval vector
x
a
tube spacing in the longitudinal direction [m]
x
b
tube spacing in the transverse direction [m]
xvi
z section coordinate in S
U
surface in Section 4.1.1
z section coordinate in S
js
surface in Section 4.1.2
Greek symbols
α thermal diffusivity [m
2
/s]
∆T
m
characteristic temperature difference [

C]
δ fin spacing [m]
perturbation parameter
η fin effectiveness
µ dynamic viscosity of fluid [kg/m s]
ν kinematic viscosity [m
2
/s]
ρ density of fluid [kg/m
3
]
θ bias in Chapter 6
θ normalized temperature difference in Chapter 5
σ standard deviation of S
˙
Q
∗ in Chapter 5
σ standard deviation in procedure (c) in Chapter 6
σ
a
standard deviation in procedure (a)
σ
b
standard deviation in procedure (b)
σ
f
ratio of free-flow cross-sectional area to frontal area
σ
S
U
standard deviation of S
U
σ
S
js
standard deviation of S
js
Subscripts and superscripts
A, B side A, B
a air side
db dry bulb
xvii
e experimental value
i inner
in inlet
l latent
o outer
out outlet
p predicted value
r refrigerant side
s sensible
t total
w water side
wb wet bulb
max maximum value
min minimum value
xviii
CHAPTER 1
GENERAL INTRODUCTION
A heat exchanger is a complex device that provides the transfer of thermal en-
ergy between two or more fluids at different temperatures. Heat exchangers are
used, either individually or as components of larger thermal systems, in a wide
variety of industrial, commercial and household applications, e.g. refrigeration, ven-
tilating and air-conditioning systems, power generation, process, manufacturing and
space industries, as well as in environmental engineering. Although a wide variety
of heat exchangers exists in the market, the present dissertation will concentrate in
the commonly used compact heat exchanger. A typical single-row fin-tube compact
heat exchanger is shown schematically in Figure 1.1. The complexity of these sys-
tems stems from their geometrical configuration, the physical phenomena present
in the transfer of heat, and the large number of variables involved in their opera-
tion. As a consequence of such complexity, no analytical solutions entirely based on
a first-principles approach are available; most calculations are based on empirical
information from the manufacturers of these equipments and, presently, the vast
majority of the analyses for predicting their behavior include assumptions and con-
ditions that are not consistent with the phenomena occurring in them under actual
states of operation. The shortcomings in the current approach thus lead, in many
cases, to unsatisfactory predictions of the heat transfer with errors that can be some-
times of the order of 25–30% or higher. Since heat exchanger performance is very
1
often a key factor in the overall thermal system design, constant improvements in
their modeling and simulation are definitely needed to increase the accuracy of their
predictions and, consequently, to improve the reliability and efficiency of thermal
systems for the specific application. This research is specifically concerned with the
enhancement in the accuracy of predictions of the steady state performance of plate
fin-tube compact heat exchangers.
δ
t
L
W
a
t
e
r

i
n
W
a
t
e
r

o
u
t
A
i
r
W
H
D
Figure 1.1. Single-row fin-tube heat exchanger.
2
The use of soft computing technologies has received much attention in recent
years as an alternative approach to cope with real-world problems. Soft computing,
being a collection of techniques that includes fuzzy logic, artificial neural networks
(ANNs), genetic algorithms (GAs), simulated annealing (SA) and probabilistic rea-
soning, which exploit in different ways the tolerance for imprecision and uncertainty,
has provided an option to solve problems that are computationally complex or math-
ematically intractable. It has been used in a wide variety of applications in both
the natural and engineering sciences, among which are decision making, pattern
recognition, system control, information processing, natural languages, optimiza-
tion, speech recognition, vision and robotics. The excellent inherent characteristics
provided by these methodologies allow them to recognize and model the nonlinear
phenomena present in complex processes. Applications to thermal systems include
some work in the design of thermal equipment with both GAs and SA, and their
simulation and control applying ANNs. A main advantage provided by soft com-
puting stems from the fact that its constituent methodologies are for the most part
complementary and synergistic rather than competitive, and can be combined to
increase even more the quality of their results.
Interval methods have been studied intensively for a few decades to provide re-
liable estimates in several application areas. The main virtue of these techniques,
which are based on the interval analysis, is the mathematical and computational
guarantee in the estimations to complex nonlinear problems. Because of such qual-
ities, interval methods have begun to be used in a variety of disciplines including
signal processing, digital control, motion planning, parameter estimation, process
design and simulation, and thermal engineering. Within the context of heat ex-
changers, however, no application of the method has been reported in the open
literature. Thus, it appears promising to use these technologies in the description,
3
identification, analysis, modeling and simulation of the heat transfer processes in
heat exchangers.
The remaining sections of this chapter provide an overview of the literature
on the current state in the analysis of heat exchangers, the use of soft computing
methodologies and the application of interval analysis which are relevant for the
present examination. First, the literature review concentrates on investigations re-
lated to the study of the heat transfer characteristics and the steady state simulation
of single-phase, condensing and evaporating heat exchangers using standard tech-
niques. These include analytical, numerical and experimental approaches applied to
both heating and cooling coils as well as current methodologies used to obtain corre-
lations from experimental data. Next, the application of soft computing techniques,
specifically artificial neural networks, genetic algorithms, simulated annealing and
clustering, to several areas of science and in particular to thermal systems focusing
on compact heat exchangers is reported. Finally, information related to the devel-
opment of interval analysis and its application to various fields, among them heat
transfer, are also reviewed.
1.1 Heat transfer characteristics and steady state simulation of heat exchangers
Several different approaches have been devised to study heat exchangers. Early
attempts include analytical solutions of the Nusselt number for a large collection of
duct shapes under laminar flow (Shah and London, 1978), with either constant wall
temperature or constant wall heat flux boundary conditions, using different tech-
niques such as conformal mapping (Sastry, 1964; 1965) or Galerkin integral methods
(Haji-Sheikh et al., 1983). Many simplified models for heat exchangers have also
been proposed. An analytical approach for predicting the air-side performance of a
single-phase heat exchanger with louvered fins was developed by Sahnoun and Webb
4
(1992). Their model predicted heat transfer coefficients with errors of as much as
25%. For calculating the air-side heat transfer in heat exchangers under condensing
conditions an analytic method has been described by Ramadhyani (1998). Recently,
Srinivasan and Shah (1997) examined condensation phenomena occurring in com-
pact heat exchangers.
Other attempts to analyze transport phenomena in the air-side, within the fin-
tube passages (Kushida et al., 1986; Bastani et al., 1992; Torikoshi et al., 1994), and
in the water-side, inside the tube bends (Goering et al., 1997), have been carried
out with CFD techniques assuming isothermal fins. Ranganayakulu and Seetharamu
(1999) performed a steady state simulation of a single-phase heat exchanger using
finite elements. Their analysis included the effect of one-dimensional heat conduc-
tion at the wall, nonuniformity in the inlet fluid flow, and a few different models of
temperature distributions. For condensing heat exchangers very few numerical stud-
ies of the thermal-hydraulic characteristics have been carried out. Wagner (1998)
developed a moving grid algorithm to analyze the change of phase in a counterflow
heat exchanger, whereas a three-dimensional numerical study of heat transfer un-
der dry and wet conditions was reported by Jang et al. (1998). Unfortunately, the
common thread in the aforementioned studies is the inclusion of assumptions and
situations that are far from being compatible with the phenomena that exist in heat
exchangers.
Due to the fact that no analytical or accurate numerical solutions are available,
the information has been usually derived experimentally. A large amount of ex-
perimental information about transport phenomena in single-phase, cooling, and
evaporator heat exchangers are reported in the open literature (Webb, 1980; Kaka¸ c
et al., 1981; Kays and London, 1984; Shah et al., 1990). For instance, Beecher
and Fagan (1987) determined performance data for single-phase finned-tube heat
5
exchangers; Jacobi and Goldschmidt (1990) characterized, experimentally, heat and
mass transfer performance of a condensing heat exchanger. Similar studies were
also examined by McQuiston (1978a), Mirth and Ramadhyani (1995), and Yan and
Sheen (2000). Thermal performance data for evaporators have been developed by
Panchal and Rabas (1993). These findings are all based on the experimentally de-
termined overall, air-side and water-side heat transfer coefficients.
A few studies have been carried out to find correlations for the performance of
compact heat exchangers. The most representative examples are those for single-
phase operating conditions (Gray and Webb, 1986), for heat exchangers operating
under wet conditions (McQuiston, 1978b), and for evaporators (Kandlikar, 1991).
Some applications of correlations to the modeling of condensers and evaporators
are those reported by Kandlikar (1990), Srinivasan and Shah (1997), Corberan and
Melon (1998), and Yan and Lin (1999), among others.
1.2 Soft Computing technologies and thermal engineering
From the collection of techniques in soft computing, the attention will be focused
in neural networks, genetic algorithms, simulated annealing and cluster analysis.
ANNs have been developed in recent years and used successfully in many appli-
cation fields, among them, pattern recognition, optimization, business, engineering
mechanics, system control, and thermal engineering (Sumpter et al., 1994; Refenes,
1995; Kulkarni et al., 1998; Zeng, 1998; Zhang et al., 1998; Haykin, 1999; Kalogirou,
1999; Sen and Yang, 1999; Smith, 1999; Jain et al., 2000). Heat transfer analyses
have been carried out introducing an ANN-based methodology to estimate the con-
vective heat transfer coefficients using crystal thermography (Jambunathan, et al.,
1996), and the mass transfer coefficients in bubble columns with Newtonian and
non-Newtonian fluids (Alvarez et al., 2000). To investigate the critical heat flux in
6
water-subcooled flow boiling, Mazzola (1997) used a combined approach of ANNs
and standard correlations. Both steady and unsteady heat conduction studies were
performed using ANNs by Gobovic and Zaghlonl (1993), and Kuroe and Kimura
(1995). Other specific applications include manufacturing and materials processing
(Mahajan and Wang, 1993; Marwah et al., 1996), and heat transfer data analysis
in natural convection problems (Thibault and Grandjean, 1991). ANNs have been
employed in the modeling of thermal devices including solar receivers (Fowler et
al., 1997), thermosyphon water heaters (Kalogirou et al., 1999) and HVAC systems
(Mistry and Nair, 1993). Previous work on the prediction of heat rates in heat ex-
changers without condensation have been reported by Zhao (1995), and D´ıaz (2000).
Control problems in heat exchangers and HVAC systems have also been studied with
the ANN technique (Jeannette et al., 1998; D´ıaz, 2000).
Since the first conception of GAs by Holland (1975), its use in statistics, electro-
magnetics, optimization, engineering and machine learning (Goldberg, 1989; Chat-
terjee and Laudatto, 1997; Johnson and RahmatSamii, 1997), among several other
areas, has increased at a fast pace. In the thermal sciences, one of the most impor-
tant applications of evolutionary algorithms has been in the area of optimization of
various kinds (Sen and Yang, 1999), e.g. to find the optimum shape in cooling chan-
nels (von Wolfersdorf et al. 1997) and fin profiles (Fabbri, 1997) or in the cooling
of electronic components (Queipo et al., 1994). The analysis of high heat flux flow
boiling systems (Castrogiovanni and Sforza, 1997), the study of inverse heat transfer
problems (Raudensky et al., 1995; Okamoto et al., 1996; Li and Yang, 1997), and
the numerical solution of the conduction equation (Davalos and Rubinsky, 1996) are
problems that have also been approached with this technique. Some work has been
done in the design of thermal equipment (Schmit et al., 1996; Tayal et al., 1999),
and their control (Abdel-Magid and Dawoud, 1995) using this methodology.
7
After its introduction by Kirkpatrick et al. (1983) and
˘
Cerny (1985), indepen-
dently, the method of simulated annealing has been successfully applied to an in-
creasingly number of disciplines, such as economics, manufacturing, telecommunica-
tions, and civil engineering (Ingber, 1996; Mavridou and Pardalos, 1997; Khandel-
wal and Kesavadas, 1999; Svergun, 1999; Ceranic et al., 2001; Randall et al., 2002),
among others, where the process of optimization, in one form or another, has been
necessary. Several applications to thermal systems exist in the literature. These
include the layout configuration of heat generating electronic components (Camp-
bell et al., 1997), the optimal design of heat exchangers (Chaudhuri, et al., 1997),
the regulation of temperature fields in rapid prototyping technologies (Fourligkas
and Doumanidis, 1999), and the optimization of energy supply systems (Gonzalez-
Monroy and Cordoba, 2000).
The popularity of cluster analysis has increased substantially over the last few
years. Its methodologies have been used in many disciplines, among them, pattern
classification, image processing, information retrieval, marketing, biology, chem-
istry, and engineering (Anderberg, 1973; Kaufman and Rousseeuw, 1990; Jain et
al., 1999). However, the literature contains few applications to thermal systems.
Exceptions include studies of temperature time series data associated with climatic
variability (Fovell, 1997), and critical heat flux classification (Moon and Chang,
1994), among very few others.
1.3 Interval analysis
The use of interval computations in various areas of science such as manufactur-
ing, quantum mechanics, economics, control, and engineering, (Ratschek and Rokne,
1993; Femia et al., 1995; Kearfott and Kreinovich, 1996; Nakagiri and Suzuki, 1997;
Floudas, 1999; Vehi et al., 2001), just to mention a few, has grown rapidly in the
8
last decade. In thermal engineering the methodology has been applied mainly to
the area of chemical processes to predict the phase stability using cubic equations
of state (Hua et al., 1998), and to model vapor-liquid equilibrium for the design of
separation systems (Gau et al., 2000). Both the prediction of azeotropes to model
reactive systems, and the simulation of adsorption phenomena in porous materials
with density-functional theory were carried out by Maier et al. (2000), and Maier
and Stadtherr (2001) using this technique.
1.4 Overview of dissertation
The present dissertation is divided into 8 chapters. The sequence of the presen-
tation of the material is as follows: Chapter 2 describes the complexities involved
in the transfer of heat in heat exchangers as well as the advantages and limitations
of current methodologies used to analyze their nonlinear thermal behavior. Some
generalities on the experimental data, and the procedures, that were obtained from
three different compact heat exchangers that will be employed in this research work
are presented in Chapter 3. In Chapter 4 the analysis of two different heat exchang-
ers is carried out first to demonstrate that, for most forms of correlating functions
that are used, a least squares analysis of the error gives results that are not unique.
Later, a methodology based on global regression analyses to find the “best” possi-
ble correlation from a given functional form is introduced. Chapter 5 investigates
the possibility of enhancement in the predictions by direct correlation of the heat
rates instead of correlating the usual heat transfer coefficients. A genetic algorithm
approach is used to find the optimum solution from a set of proposed correlations
to find out which functional form may be the most suitable to represent the heat
transfer characteristics in these devices. Chapter 6 illustrates the application of arti-
ficial neural networks to predict accurately the steady-state behavior of condensing
9
heat exchangers. In Chapter 7 the applicability of ANNs is extended to simulate
the phenomena in heat exchangers with limited data. Chapter 8 summarizes the
results obtained in this study. Suggestions for further work are also provided.
10
CHAPTER 2
PROBLEM DESCRIPTION
The previous chapter provided an overview of the different research works that
have contributed to the understanding of the steady state performance of fin-tube
heat exchangers and the current state in the procedures used nowadays to predict
their thermal behavior. Information on the application of several emerging tech-
nologies that belong to soft computing as well as on interval analysis to various
disciplines, in particular to thermal engineering, that may be useful to simulate
accurately the thermal characteristics of heat exchangers was also included.
One of the main objectives of the Hydronics group at the University of Notre
Dame is the study of the steady state behavior of heat exchangers and the devel-
opment of new methodologies that enable constructing more accurate models to
simulate their nonlinear response, so that these can later be used in analysis, design
and control of thermal systems. The present work includes analytical, experimental
and numerical investigations to increase our understanding of the physical phenom-
ena involved in the transfer of heat in compact fin-tube heat exchangers, to improve
the accuracy of current prediction models, as well as to develop new models for their
steady state simulation.
The current chapter describes the complexities involved in the transfer of heat
in heat exchangers as well as the limitations of the conventional approach used
nowadays to simulate their behavior. First, an explanation for the lack of reliable
11
analytical solutions to the heat exchanger analysis on the basis of its complexity
is stated. Later, a thorough analysis about the advantages and limitations of the
current methodology based on the use of the heat transfer coefficient and the com-
pression of its information be means of correlations is provided. Finally, the chapter
also considers the possibility of enhancement in the accuracy of the prediction of
heat transfer rates in these devices by means of new procedures which are struc-
tured on the basis of global optimization techniques as well as by soft computing
methodologies.
2.1 Complexity of heat exchanger analysis
For the design of a thermal system it is often necessary, for selection purposes, to
predict the heat transfer rates of heat exchangers under specific operating conditions.
The heat rate, for any heat exchanger, is a function that depends in a nonlinear
fashion on the flow rates, inlet temperatures and the geometry.
For a given device exchanging thermal energy between two fluids, the interac-
tions among the geometrical components and the flow, in each side, generate a
complicated hydrodynamics that has strong influence upon the transfer of heat. For
instance, in the external side of a compact heat exchanger, tube-tube and fin-tube
interactions create complex vortex structures in the proximity of their unions, wakes
behind the tubes, flow development along both fins and tube-walls, and even tur-
bulence in the flow; in its internal side, on the other hand, the tube bends induce
secondary flows and flow readjustment downstream of them, while hydrodynamic
entrance regions and sometimes turbulent flows are also present.
The result of this is the appearance of non-uniform local heat transfer rates and
complicated temperature fields. Moreover, other phenomena like condensation on
the fins and tubes as well as partial evaporation inside the tubes appear under some
12
states of operation in cooling or evaporating coils, e.g. in air-conditioning and re-
frigeration applications. These phenomena increase the computational difficulty of
the problem since it is necessary to account for latent heat released during the con-
densation process, the hydrodynamic effect of the film and droplets on modifying
the flow pattern, as well as the inhomogeneous distribution of the flow due to the
coexistence of liquid-vapor states of the refrigerant. Although the governing equa-
tions can be formulated, because of the aforementioned complexities it is virtually
impossible to solve them with reasonable computational effort under actual states
of operation.
As a consequence, no analytical solutions entirely based on first-principles are
available and merely estimations of simplified models using CFD techniques have
been carried out, e.g. Bastani et al. (1992), Jang et al. (1998), and Ranganayakulu
and Seetharamu (1999), among others. Currently most calculations are based on
experimental information of the heat rate, which is the quantity that completely
characterizes the heat exchanger performance, as a function of the variables of the
system, e.g. two flow rates, two inlet temperatures and the geometry. For each heat
exchanger, such information is determined by the manufacturers for specific fluids,
and presented to the user, i.e. the thermal system engineer who needs it for design
and prediction purposes under different conditions.
The next section illustrates the conventional analysis of these devices based on
experimental information of the heat transfer coefficients as a way to characterize
the performance of heat exchangers, and describes the flaws in their use.
2.2 Coefficient of heat transfer and correlations in heat exchanger analysis
The use of two heat transfer coefficients represented by nondimensional correla-
tions, one for the inside and another for the outside, has been the universal vehicle to
13
characterize the performance of heat exchangers. From these, the heat rate can be
obtained assuming a characteristic temperature difference for each side. Since this is
the common procedure, it is clear that the accurate determination of the heat trans-
fer coefficient is of utmost importance to the heat exchanger analysis. This section
examines the inconsistencies of the current approach based on the use of correlated
heat transfer coefficients and shows that the correct estimation of the heat transfer
rate may be compromised by the assumptions made during the different stages of
the procedure.
2.2.1 Correlations of the heat transfer coefficient
The information about the heat transfer coefficients is normally compressed and
transported from the manufacturer to the thermal systems engineer by means of
correlations. In the most common procedure, correlations of experimental data are
developed in terms of dimensionless quantities like Nusselt Nu, Prandtl Pr, and
Reynolds Re, numbers in the interest of compactness to achieve greater generality.
The functional form of the correlation equation, e.g. Nu = f(Re, Pr) and sometimes
geometrical parameters, is selected on the basis of simplicity and common usage,
power-laws being an example, and cannot be totally justified from first principles.
The main advantage offered by nondimensional correlations is that their appli-
cation to different fluids other than those used to construct them, and to a wide
range of flow rates, is straightforward. On the other hand, the lack of precision in
the estimations of the heat rate from correlations is the main drawback. There are
several reasons for this. The limited number of parameters used in the functional
form of the correlation, e.g. power laws, which is used to represent the heat transfer
phenomena cannot account for all the complexities involved. Furthermore, the un-
certainties due to property variations, where these have to be computed twice, once
14
to build the correlation and later to apply it for other operating conditions, increase
the error bands in the predictions.
2.2.2 The overall heat transfer rate equation
In the conventional procedure, the heat transfer rate is determined by means
of two energy balance equations, one for each fluid, and a conveniently established
overall heat rate equation. The first two describe the change in enthalpy of each
stream whereas the third provides the thermal interaction between the two fluids
as a function of their temperature difference. The heat rate equation is formulated
in terms of the inner and outer heat transfer coefficients combined into a single
overall heat transfer coefficient and a characteristic temperature difference between
the working fluids as
˙
Q = UA∆T
m
. (2.1)
In Eq. (2.1), U is the average overall heat transfer coefficient, which depends on
the mass flow rates of both fluids, A is a suitable area (either the internal- or
external-side area), and ∆T
m
is an appropriate mean value of the characteristic
temperature difference. Thus, both U and ∆T
m
are independent of location within
the heat exchanger. It is to be noticed that Eq. (2.1) is written as a product of two
independent functions, i.e. hydrodynamic- and temperature-dependent parts. On
the other hand, also note that although ∆T
m
is given in terms of the inlet and outlet
bulk temperatures of both fluids, which can all be measured from experiments, a
specific form for it must be determined.
The form of ∆T
m
is dependent on the geometry of the heat exchanger, and gen-
erally it can be obtained analytically (Bowman et al., 1940; Jakob, 1957) by means
of energy balances under suitable assumptions, e.g. steady state conditions, con-
stant thermophysical properties, spatially homogeneous heat transfer coefficients,
15
constant mass flow rates of both fluids, and similarity in temperature profiles. Hence,
Eq. (2.1) is exact only if the aforementioned assumptions are true. In reality, how-
ever, it is well-known not to be the case due to the fact that sudden changes in
operating conditions, development of hydrodynamic and thermal boundary layers,
and thermal dependence of properties induce distortions of the velocity and tem-
perature profiles and hence inhomogeneous local heat transfer coefficients.
2.2.3 Functional form of the overall heat transfer coefficient
As mentioned at the begining of the current section (Section 2.2), the inner and
outer heat transfer coefficients are usually combined into a single overall transfer
coefficient to calculate the heat rate using Eq. (2.1). Although there are a number
of ways in which this can be done, the common approach is to use a particular
functional form in which the reciprocal of the overall transfer coefficient is an additive
function of two independent expressions, one dependent on the inner and another
on the outer mass flow rates, ˙ m
i
and ˙ m
o
respectively, e.g. 1/U = f
1
( ˙ m
i
) + f
2
( ˙ m
o
).
The explicit form, known as the thermal resistance equation in which each quantity
can be thought to be a thermal resistance on either side of the heat exchanger, may
be written as
1
UA
=
1
h
i
A
i
+
1
ηh
o
A
o
(2.2)
where the terms of the right hand side are the internal and external resitances
respectively, η is the fin effectiveness, h
i
and h
o
are the inner and outer heat transfer
coefficients, and A
i
, and A
o
are the total inner- and outer-side areas respectively.
The main advantage of this functional form, from an experimental standpoint,
is that if one of the resistances is zero or known a priori, then the value of the other
is immediately determined. On the other hand, the summing function implies that
the heat transfer coefficient on one side does not depend on the mass flow rate of
16
the other side and hence assumes the existence of a wall temperature separating the
two fluids. For unidirectional parallel- and counter-flow arrangements this local wall
temperature, although sometimes unknown, is physically well-defined. In most heat
exchanger configurations, such as the fin-tube type, however, it is well-known not
to be the case. The inclusion of a tube-arrangement and a set of fins makes it very
difficult not only to measure the temperature at the wall but moreover to specify
where it should be measured. Furthermore, property variations with temperature
breaks the assumption of independence of the transfer coefficients. Nevertheless, it
is possible to define a hypothetical local wall temperature that allows us to still use
the aforementioned approach to compute the heat rate. In such a case, although
the coefficients of heat transfer remain coupled, it is possible to find them in a least
squares sense. The hypothetical wall temperature will then be defined such that it
fits the experimental data the best. The accurate determination of both coefficients
is a challenging problem.
2.2.4 Experimental separation of the overall thermal resistance
There are several ways to try to separate the internal and external thermal re-
sistances to determine both transfer coefficients. Experimentally, the manufacturer
may use a condensing fluid or a high Reynolds number turbulent flow in one side
of the heat exchanger, so that the corresponding thermal resistance would be either
very close to zero or known from previously determined forced-convection correla-
tions. Once this is known, the heat transfer coefficient on the other side may be
calculated. Even though experiments and analysis are easy to do, a main drawback
is that it is not possible to compute the heat transfer rate under other operating
conditions than those used to generate the information. If the user needs predic-
tions at different temperatures where there is no change of phase, there would be a
17
significant difference in the estimations of the heat rate from a correlation that was
built from experimental data corresponding to conditions of a condensing fluid.
A better procedure is to perform enough experiments that cover a wide range of
operating conditions in both sides of the heat exchanger. This is the most general
experimental methodology but also the most difficult to handle from the thermal-
resistance-separation perspective. The reason for this is that, since both thermal
resistances are non-zero and unknown, and hence coupled, it would be necessary to
provide a mechanism to separate them. One way to overcome this problem is to
propose first a functional form for each transfer coefficient and later to perform a
regression analysis. However, as will be demonstrated in detail in Chapter 4, for
most correlation forms that are used, a least squares analysis of the error provides
non-unique solutions.
2.2.5 Techniques to find correlations from experiments
Several methods have been proposed to find correlations for each heat transfer
coefficient from experimental data. These are either based on the Wilson plot tech-
nique (Wilson, 1915) or on one of its variations (Briggs and Young, 1969; Khartabil
and Christensen, 1992), that can be used when experiments lead to thermal resis-
tances that are about the same order of magnitude; or from multiple linear regression
analyses that can be applied when conditions of very high Reynolds number turbu-
lent flow or a condensing fluid provide a thermal resistance of one side of the heat
exchanger considerably smaller than that of the other side (Kays and London, 1984;
Shah, 1985).
For discussion purposes to illustrate the limitations of these approaches, let us
consider a two-fluid heat exchanger with fluid media A for one side of the wall and
B for the other side of it, where either fluid can be a gas or a liquid. Also the fluid
18
on either side, A or B, may be regarded as the over-tube- or in-tube-side fluid. The
corresponding thermal resistances for each side are R
A
and R
B
respectively, being
these related to the overall resistance by R = R
A
+ R
B
. Their explicit forms are
similar to those given in Eq. (2.2). By assuming a functional form for each heat
transfer coefficient, e.g. power laws are commonly used, as
Nu
A
= n
1
Re
n
2
A
f
1
(Pr
A
, Geometry),
Nu
B
= n
3
Re
n
4
B
f
2
(Pr
B
, Geometry) (2.3)
where f
1
and f
2
are functions, the thermal resistance equation becomes a nonlinear
regression model. The task is to find the four correlation constants, n
1
to n
4
, such
that predictions from the model and experiments are close. Since correlations are
developed from experimental information, the number of constants that can be
obtained for them are linked to the type of data used. As pointed out in Section
2.2.4, there are several ways experiments can be carried out to separate the thermal
resistances. These may lead to one of the following cases: either each resistance is
zero, constant or variable. In the latter cases the thermal resistance may be known
or unknown. From knowledge of these, assumptions have to be made to carry out
regression analyses for finding the corresponding correlation constants. Thus, the
different cases can be grouped as:
Side A
_
¸
¸
_
¸
¸
_
A
1
: R
A
is constant and unknown
A
2
: R
A
is variable and known
A
3
: R
A
is variable and unknown
Side B
_
¸
¸
_
¸
¸
_
B
1
: R
B
is constant and unknown
B
2
: R
B
is variable and known
B
3
: R
B
is variable and unknown
19
where cases of R
A
= 0 and R
B
= 0 can be included as particular instances in A
2
and B
2
respectively.
For combinations (A
1
, B
1
), (A
1
, B
2
), (A
2
, B
1
), and (A
2
, B
2
), there is no infor-
mation available in the open literature. On the other hand, conditions (A
1
, B
3
)
and (A
3
, B
1
), and (A
2
, B
3
) and (A
3
, B
2
), are not distinguished here. Thus, in what
follows, when referring to a particular combination, e.g. (A
3
, B
1
), it may also be
referring to its equivalent case, e.g. (A
1
, B
3
).
From experimental data corresponding to case (A
3
, B
1
), Wilson (1915) devised
a graphic technique to find two parameters, n
1
and R
B
, using a straight-line fit and
assuming n
2
to be known. From the same type of data, Khartabil and Christensen
(1992) proposed a nonlinear regression scheme in which the problem was cast into
two linear regressions and a nonlinear equation to find three parameters, n
1
, n
2
and R
B
. Several investigators (Rich, 1973; Wang et al., 1996; Chang et al., 1997;
Benelmir, et al., 1997), among others, have applied either one of these techniques
to find correlations for single-phase and condensing heat exchangers.
The most widely applied technique in heat exchanger analyses, (Kays and Lon-
don, 1984; Shah, 1985), uses experiments that generate conditions for (A
3
, B
2
).
In this procedure, R
B
is usually determined from turbulent-flow correlations while
multiple linear regressions are carried out to compute two or more constants of only
one correlation, e.g. n
1
and n
2
and the constants for the geometrical parameters
(Gray and Webb, 1986; Kayansayan, 1994; Kim et al., 1998; Abu Madi et al., 1998)
for heat exchangers operating under single-phase situations and (Idem et al., 1990;
Jacobi and Goldschmidt, 1990) for those under condensing conditions. A particular
case is that of McQuiston (1978b), who determined R
B
from the Dittus-Boelter
correlation but the conditions used were those to achieve a constant value of the
thermal resistance.
20
Briggs and Young (1969) used data that belong to the case (A
3
, B
3
) to develop a
technique based on modifications to the Wilson plot as two successive linear regres-
sion analyses to find three constants, n
1
, n
3
and n
4
, in an iterative manner taking
two of them at a time and assuming the fourth one, n
2
, to be known. Further ap-
plications of this method in heat exchangers include those by Kedzierski and Kim
(1996) and Kumar et al. (2001).
From the above, it is clear that correlations developed from techniques that use
either measurements corresponding to cases (A
3
, B
1
), in which R
B
is unknown but
constant, or (A
3
, B
2
), will have severe limitations regarding the variety of heat ex-
changer operating conditions to which they can be applied as was noted before in
Section 2.2.4. On the other hand, although in principle the technique introduced by
Briggs and Young (1969) may be used to find correlations for a wider range of appli-
cations, in reality this is not the case due to the assumptions made in the regression
process (Shah, 1990; W´ ojs and Tietze, 1997). Moreover, all of the aforementioned
techniques oversimplify the regression problem by fixing a priori the values of some
constants which assumes the heat exchanger to behave in a certain way. This is
a severe flaw of these methods since laminar and turbulent flows as well as other
thermal phenomena appear under different operating conditions.
Hence the use of these methodologies to build correlations are among the reasons
for their wide error bands in the predictions of the heat transfer rate under different
heat exchanger operation states. Unfortunately, a major difficulty to achieve better
levels of accuracy in heat rate estimations for heat exchangers is compounded by
the fact that no robust procedures to develop general correlations, or alternative
approaches are available as yet.
21
2.3 Objectives and scope of the present dissertation
The overall objective of the present dissertation is to provide methodologies that
allow accurate estimations of the heat transfer rate in heat exchangers for the anal-
ysis, design and control of thermal systems. To achieve such a goal, this study
will make use of techniques that belong to the areas of optimization and artificial
intelligence, i.e. global optimization methods and soft computing technologies; in
particular, interval analysis, genetic algorithms, simulated annealing, artificial neu-
ral networks and cluster analysis. The procedures developed on the basis of the
aforementioned methods will be applied to heating and cooling coils with differ-
ent geometrical complexities where there exists either one of the following cases:
no-change-of-state of both fluids, phase-change of the fluid inside the tubes, or con-
densation on the exterior surfaces.
The importance of this investigation is that it will provide both the manufacturer
of these equipments and the thermal system engineer with appropriate tools to
accurately simulate the heat transfer characteristics of heat exchangers for their
design, which is the interest of the former, as well as for their analysis and selection,
which is the preference of the latter.
22
CHAPTER 3
HEAT EXCHANGER EXPERIMENTAL DATA
In order to fulfill the objectives of the current dissertation, several sets of ex-
perimental data that were obtained from three fin-tube heat exchangers, referred
here as heat exchanger 1, heat exchanger 2, and heat exchanger 3, whose levels
of complexity differ among them will be used. Heat exchanger 1 is a single-row
single-column type commonly used for heating purposes while the other two, heat
exchanger 2 and heat exchanger 3, have multi-row multi-column configurations and
are usually applied in air conditioning and refrigeration applications. The following
sections illustrate the different experimental procedures necessary to generate the
corresponding data sets.
3.1 Experimental data for heat exchanger 1
Extensive experimental data were obtained in a variable-speed open wind tun-
nel facility, a schematic of which is illustrated in Figure 3.1, built specifically for
compact heat exchanger studies (Zhao, 1995). The heat exchanger considered is a
single-row fin-tube type, shown schematically in Figure 1.1; it is a type T water
coil heat exchanger with a single circuit of 12 tubes connected by bends, nominally
457 mm 610 mm in size, manufactured by the Trane Company. The fluids used
were atmospheric air at room-temperature flowing through the fin passages and hot
water as the fluid inside the tubes. Air is forced into the wind tunnel by a blower
23
that is controlled by a variable speed drive. In order to straighten the air flow, one
honeycomb is located upstream and another downstream of the heat exchanger. A
Pitot tube located upstream of the heat exchanger is connected to a differential
pressure transducer to give the air flow rate. The water flow rate is measured by a
turbine flow meter. Water temperature is raised by a PID-controlled electrical resis-
tance heater. Type T single thermocouples are used for sensing the air temperature.
Because the inlet air temperature is very close to the room temperature, only one
thermocouple is used there, while the average reading of five thermocouples is used
for the outlet air temperature. A computerized data acquisition system collects all
the experimental data.
The measured variables obtained from the transducer outputs, are the inlet
and outlet water temperatures, T
in
w
and T
out
w
respectively, the inlet and outlet air
temperatures, T
in
a
and T
out
a
respectively, the mass flow rate of water, ˙ m
w
, and the
mass flow rate of air, ˙ m
a
. The heat transfer rate,
˙
Q, can be determined by
˙
Q = ˙ m
w
c
p,w
_
T
in
w
−T
out
w
_
(3.1)
= ˙ m
a
c
p,a
_
T
out
a
−T
in
a
_
(3.2)
where c
p,w
and c
p,a
are the specific heats of water and air respectively. An average
value for
˙
Q is used since the air and water sides give slightly different values, but
well within 10%.
Two different data sets were determined for this heat exchanger following the
aforementioned procedure. The first one (called data set 1) was acquired by Zhao
(1995), while the second set (data set 2) was obtained as part of the overall inves-
tigation reported here. A total of N
1
= 259 measurements for the data set 1, and
N
2
= 214 experiments for data set 2 were recorded for different air and water flow
rates and water inlet temperatures. The regimes covered by these for the water
24
T
a
out
T
w
in
T
w
out
T
a
in
m
w
m
a
(b)
Fan
Motor
Heat exchanger
(a)
Air
Figure 3.1. Heat exchanger experimental facility.
25
flow, were laminar, transitional and turbulent, although slightly different upper and
lower bounds were obtained for each set of measurements. Data set 1 (Zhao, 1995),
will be used in Chapter 4 to demonstrate the need for global regressions due to the
non-uniqueness in the correlation constants that can be found from a least-squares
process, while data set 2 will be used in Chapter 5 to indicate that other correlation
forms can be used as alternatives to the commonly applied power laws, to enable
good fitting of the experimental data.
3.2 Published experimental data and correlations for heat exchanger 2
Detailed heat and mass transfer measurements from five different multiple-row
multiple-column plate-fin type heat exchangers with staggered tubes were obtained
and published by McQuiston (1976, 1978a). Because of their importance to appli-
cations such as air-conditioning and refrigeration these data and the corresponding
correlations have become standards in the field. To carry out these tests, atmo-
spheric air flowing through the fin passages and water inside the tubes were used
as working fluids. The experiments were made for a wide range of operating states.
The conditions were such that, for certain cases, condensation would occur on the
fins. The nature of the air-side surface, i.e. whether dry, covered with condensed
water droplets, or covered with a water film, was determined by direct observations
and recorded. All five heat exchangers had a nominal size of 127 mm 305 mm,
were geometrically similar but with different fin spacings. A schematic with the
geometrical parameters and dimensions is shown in Figure 3.2. Details about the
experimental facility and instrumentation are given in McQuiston (1976, 1978a).
It must be noted that, even when there is no condensation, analysis of the data
shows that the specific humidities at the inlet and outlet vary by as much as 150%.
This discrepancy may be due to heat losses between the locations at which the
26
temperatures were measured and the heat exchanger. The data for dropwise and
film condensation could also have been similarly affected.
δ
t
x
a
D
x
b
W
a
t
e
r

i
n
W
a
t
e
r

o
u
t
A
i
r
Figure 3.2. Schematic of a compact fin-tube heat exchanger.
Since McQuiston (1978a) was interested in predicting the air-side heat transfer,
he used high Reynolds-number turbulent flow on the water side so that its heat
transfer coefficient, which could be estimated from the Dittus-Boelter correlation,
was much higher than that on the air side and thus introduced little error. He defined
27
the air-side heat transfer coefficients using the log-mean temperature difference for
the dry surface. A similar procedure was followed for determining the air-side heat
transfer coefficients under condensing conditions using the enthalpy difference as
the driving potential.
The experimentally determined transfer coefficients were summarized in terms
of the nondimensionalized Colburn j-factors, being these defined as
j
s
=
h
a
G
c
c
p,a
Pr
2/3
a
, j
t
=
h
t,a
G
c
Sc
2/3
a
(3.3)
where j
s
is the Colburn j-factor for the sensible heat and j
t
is that for the total
heat. Of course, in the dry case the two are the same. In the above equations, h
a
is
the heat transfer coefficient and h
t,a
is the total (enthalpy) heat transfer coefficient
respectively, G
c
is the air-flow mass velocity, c
p,a
is the specific heat of air, Pr
a
is the
Prandtl number of air defined by Pr
a
= µ
a
c
p,a
/k
a
, and Sc
a
is the Schmidt number of
air given by Sc
a
= µ
a

a
T
a
, µ
a
is the dynamic viscosity of air, ρ
a
is the air density,
k
a
is the thermal conductivity of air, and T
a
is the mass diffusivity of water vapor
in air. Other reported variables are the inlet and outlet air dry-bulb temperatures
T
in
a,db
and T
out
a,db
, inlet and outlet air wet-bulb temperatures T
in
a,wb
and T
out
a,wb
, inlet and
outlet water temperatures T
in
w
and T
out
w
, and the total heat rate
˙
Q. Geometrical
factors like fin spacing δ, and tube outside diameter D, as well as the Reynolds
numbers Re
D
and Re
δ
defined as
Re
D
=
G
c
D
µ
a
, Re
δ
=
G
c
δ
µ
a
(3.4)
were also included.
A total of M = 327 experimental runs were reported by McQuiston (1978a).
From these M
1
= 91 corresponded to dry-surface conditions, M
2
= 117 corresponded
to dropwise and M
3
= 119 to film condensation.
28
From the previously mentioned experimental data, the correlations that were
obtained by McQuiston (1978b) are
j
s
= 0.0014 + 0.2618 Re
−0.4
D
_
A
A
tb
_
−0.15
f
s
(δ) (3.5)
j
t
= 0.0014 + 0.2618 Re
−0.4
D
_
A
A
tb
_
−0.15
f
t
(δ) (3.6)
where
f
s
=
_
¸
¸
_
¸
¸
_
1.0 for dry surface
(0.90 + 4.3 10
−5
Re
1.25
δ
)
_
δ
δ−t
_
−1
for dropwise condensation
0.84 + 4.0 10
−5
Re
1.25
δ
for filmwise condensation
(3.7)
f
t
=
_
¸
¸
_
¸
¸
_
1.0 for dry surface
(0.80 + 4.0 10
−5
Re
1.25
δ
)
_
δ
δ−t
_
4
for dropwise condensation
(0.95 + 4.0 10
−5
Re
1.25
δ
)
_
δ
δ−t
_
2
for filmwise condensation
(3.8)
A
A
tb
=
4
π
x
a
D
h
x
b
D
σ
f
(3.9)
where A is the total air-side area, A
tb
is the surface area of tubes without fins, x
a
and x
b
are the two tube-center distances shown in Figure 3.2, t is the fin thickness,
and σ
f
is the ratio of the free-flow area of air to the frontal area. The range of
validity of the correlation was also given.
Gray and Webb (1986) added data from other sources to find a different corre-
lation
j
s
= 0.14 Re
−0.328
D
_
x
b
x
a
_
−0.502
_
δ −t
D
_
0.0312
for dry surface (3.10)
with its corresponding range of applicability.
3.3 Manufacturer’s experimental data for heat exchanger 3
Limited experimental data from a series of tests of several multi-row multi-
column fin-plate type heat exchangers with staggered tubes, schematically shown
29
in Figure 3.3, were provided by the manufacturer Tyler Refrigeration Corporation.
The type of heat exchangers tested were all for refrigeration applications with atmo-
spheric air flowing outside the tubes and between the fin passages, and refrigerant
Freon 22 as the fluid inside the tubes. In the air-side, under certain conditions,
not only condensation but also frost formation over the fins and tubes would occur.
Inside the tubes, in the refrigerant-side, regions of liquid and vapor also coexist.
The heat exchangers had nominal sizes in the range of 818 mm. length by 100 mm.
height, to 3, 328 mm. length by 230 mm. height, and were similar in shape. All the
geometrical parameters are shown in Figure 3.3.
.
.
.

t
δ
x
a
x
b
D
1 2 3
...
1
2
3
N
row
W
L
Air
H
R22 out
R22 in
N
col
Figure 3.3. Schematic of a compact heat exchanger with air and R22 as fluids.
30
The experiments were conducted for a limited number of operating conditions
with large variations of the geometrical parameters. Different geometries were con-
sidered by varying the lengths, L, fin spacings, δ, number of rows, N
row
, number
of columns, N
col
, number of circuits, N
cir
, and tube-center distances, x
a
and x
b
.
However, the tube diameter, D, was kept unaltered. The measured variables were
the inlet and outlet air dry-bulb temperatures, T
in
a,db
and T
out
a,db
respectively, the inlet
refrigerant temperature, T
in
r
, the inlet humidity ratio of air, w
in
a
, the mass flow rate
of air, ˙ m
a
, and the rate of defrosted water, ˙ m
w
. The sensible and latent heat transfer
rates,
˙
Q
s
and
˙
Q
l
were determined for each geometry under conditions of thermal
equilibrium between the air-side and the refrigerant-side by means of
˙
Q
s
= ˙ m
a
c
p,a
_
T
in
a,db
−T
out
a,db
_
(3.11)
˙
Q
l
= ˙ m
w
(h
if
+h
fg
) (3.12)
where c
p,a
is the specific heat of air, h
if
is the latent heat of transformation of the
water from solid to liquid and h
fg
is the latent heat from liquid to vapor. The total
heat rate is given by
˙
Q
t
=
˙
Q
s
+
˙
Q
l
(3.13)
The total number of experiments carried out were only M = 38 by varying the
N = 11 independent variables L, δ, N
row
, N
col
, N
cir
, x
a
, x
b
, ˙ m
a
, T
in
a,db
, w
in
a
, T
in
r
.
31
CHAPTER 4
NONLINEAR REGRESSION ANALYSIS AND GLOBAL OPTIMIZATION
Chapters 2 and 3 have introduced, respectively, the motivation for the current
research work and the experimental data that will be used to achieve the goals set.
In Chapter 2, particularly, it was indicated that among the reasons for the lack of
accuracy in estimations of the performance of heat exchangers from the conventional
approach were the assumptions considered along the process. Another important
reason was that presently there are no sufficiently robust techniques that allow
finding the correlation constants corresponding to both coefficients of heat transfer.
The current chapter addresses one of the objectives of this dissertation, that
is the enhancement in accuracy of the predictions of heat exchanger performance
from correlations. To this end, experimental data sets of a single-phase compact
heat exchanger and of a condensing coil will be used. First the non-uniqueness of
the result from a local least-squares correlating procedure is demonstrated. Later,
the chapter introduces a methodology based on global optimization and provides a
description of three techniques that, given a correlation form, will enable the closest
fit to the experimental measurements. Genetic algorithms, simulated annealing,
and interval analysis will then be applied to illustrate the advantages of using this
global approach to find the values of the correlation constants for both transfer
coefficients. Next, an extension of the analysis of Section 2.2 is provided to address
the issue of temperature dependence of properties and inhomogeneity of the heat
32
transfer coefficients that break the validity of the summation function and couple the
hydrodynamics. Finally it will be shown that such interaction can be correlated and,
when taken into account, improves even more the predictions of the heat transfer
rate.
4.1 Multiplicity of solutions in nonlinear least-squares analysis
In the following sections the least-squares methodology will be applied to two
different heat exchangers, a single-row fin-tube type (called heat exchanger 1) de-
scribed in Section 3.1, whose data according to the classification given in Section
2.2.5 belong to case (A
3
, B
3
), and a multi-row multi-column one (heat exchanger 2)
given in Section 3.2, whose data correspond to the case (A
3
, B
2
).
4.1.1 Multiplicity of “best” correlations in heat exchanger 1
Let us consider the experimental data set 1 for heat exchanger 1. For prediction
purposes the heat rate depends on the inlet temperatures and mass flow rates as
˙
Q =
˙
Q(T
in
a
, T
in
w
, ˙ m
a
, ˙ m
w
). A common way to correlate experimental data in heat
exchangers is to use power laws to represent both the inner, water, and the outer,
air, heat transfer coefficients, h
w
and h
a
respectively. The overall heat transfer
coefficient, U, is related to these by
1
UA
a
=
1
h
w
A
w
+
1
ηh
a
A
a
(4.1)
where A
w
and A
a
are the inner, water-side, and outer, air-side, heat transfer areas,
and η is the fin efficiency. Assuming the heat transfer coefficients to be constant
along the length of the heat exchanger tube, shown schematically in Figure 4.1, the
energy balance between the fluids, as applied to the water-side, is given by
− ˙ m
w
c
p,w
dT
w
= UP(T
w
−T
a
)dx (4.2)
33
where T
w
and T
a
are water and air bulk temperatures at the location of the differen-
tial element. If the air temperature is approximated by T
a
= T
in
a
, integration of Eq.
(4.2) between the corresponding limits for x and T
w
lead to the following expression
for the overall heat transfer coefficient
U =
˙ m
w
c
p,w
A
a
ln
_
T
in
w
−T
in
a
T
out
w
−T
in
a
_
. (4.3)
After combining Eq. (4.3) with the two energy balances, one for the air-stream and
another for the water-stream, given in Eqs. (3.1) and (3.2), the total heat transfer
rate is found to be
˙
Q = ˙ m
w
c
p,w
_
T
in
w
−T
in
a
_
_
1 −exp
_
−UPL
˙ m
w
c
p,w
__
(4.4)
where P is the perimeter of a section of the duct, and L is its length.
x =0 x = L
T
a
out
T
w
out
dx
m C T
w p,w w
m C (T + dT )
w p,w w w
T
a
in
m ,V
a a
T
w
in
m ,V
w w
dQ=UP(T - T )dx
w a
Figure 4.1. Cross-flow heat exchanger.
By taking the functions f
1
= Pr
1/3
a
and f
2
= Pr
0.3
w
in Eq. (2.3), where subscripts
A=a and B=w refer to the air-side and water-side respectively, it can be re-written
as
34
ηNu
a
= n
1
Re
n
2
a
Pr
1/3
a
,
Nu
w
= n
3
Re
n
4
w
Pr
0.3
w
(4.5)
to provide the nondimensional forms of the transfer coefficients h
a
and h
w
. In Eq.
(4.5), the dimensionless Nusselt, Reynolds, and Prandtl numbers are defined as
follows:
Re
a
=
V
a
δ
ν
a
, Nu
a
=
h
a
δ
k
a
, Pr
a
=
ν
a
α
a
,
Re
w
=
V
w
D
ν
w
, Nu
w
=
h
w
D
k
w
, Pr
w
=
ν
w
α
w
, (4.6)
where V is the average flow velocity, δ is the spacing between fins, D is the tube
inner diameter, and α, ν and k are the thermal diffusivity, kinematic viscosity and
thermal conductivity of the fluids, respectively. Experimental data set 1, in Section
3.1, were confined to 200 < Re
a
< 700 and 800 < Re
w
< 4.5 10
4
so that all results
here are restricted to those ranges.
After substituting Eqs. (4.5) and (4.6) into the overall resistance equation, Eq.
(4.1), the overall heat transfer coefficient is given by
1
U
=
δ
k
a
1
n
1
Re
n
2
a
Pr
1/3
a
+
A
a
A
w
D
k
w
1
n
3
Re
n
4
w
Pr
0.3
w
(4.7)
where Eq. (4.7) is nonlinear with respect to the set of unknown constants.
In order to find the constants n
1
, n
2
, n
3
and n
4
a least-squares method must be
used to minimize the difference between the measured and predicted values of the
total thermal resistance. This is done through the average squared error defined as
S
U
=
1
N
1
N
1

j=1
_
1
U
e
j

1
U
p
j
_
2
(4.8)
where U
e
j
and U
p
j
, for j = 1, . . . , N
1
, are the experimental measurements and the
corresponding values predicted by Eq. (4.7), respectively. In the multidimensional
35
parameter space for the correlation constants, the function S
U
(C) can be seen as a
smooth manifold where C = (n
1
, n
2
, n
3
, n
4
) is the vector of unknown parameters;
thus, a search must be pursued to find C such that S
U
(C) is a minimum.
This procedure was carried out for the N
1
= 259 data sets, described in Section
3.1 as data set 1, by applying a gradient-based approach (Vanderplaats, 1984).
Results from this exercise indicated a multiplicity of solutions for the set of constants
each one of these depicting a case of a local minimum least-squares error in S
U
. For
example, Figure 4.2 shows a section along a line joining two of these minima. The
coordinate z is such that C = C
A
(1 − z) + C
B
z, where C
A
and C
B
correspond to
the two minima.
−0.2 0 0.2 0.4 0.6 0.8 1 1.2
0
0.5
1
1.5
2
2.5
3
x 10
−5
z
A B
S


(
m


K

/
W

)
U
4
2
2
Figure 4.2. Section of S
U
surface.
36
The S
U
values between minima A and B, shown in the figure, are within 17.9%
of each other, and one is to be preferred over the other only if the difference is
considered to be significant and larger than experimental error. In this particular
case the two sets of correlations corresponding to each one of the minima are
Correlation A: ηNu
a
= 0.1018 Re
0.591
a
Pr
1/3
a
Nu
w
= 0.0299 Re
0.787
w
Pr
0.3
w
,
Correlation B: ηNu
a
= 0.0910 Re
0.626
a
Pr
1/3
a
Nu
w
= 0.0916 Re
0.631
w
Pr
0.3
w
.
Figure 4.3 shows the predicted heat rates for the two correlations,
˙
Q
p
, plotted
against the experimental values,
˙
Q
e
. Both correlations appear to work fairly well.
Even though
˙
Q is well predicted by either of the two correlations, the inner or outer
heat transfer coefficients calculated are very different. Figure 4.4 shows the ratios
of the air-side and water-side Nusselt numbers
R
a
=
Nu
a
¸
¸
A
Nu
a
¸
¸
B
, R
w
=
Nu
w
¸
¸
A
Nu
w
¸
¸
B
(4.9)
calculated from correlations A and B as functions of the Reynolds numbers Re
a
and
Re
w
respectively. The predictions of the two correlations differ by about 12.5% on
the air side and 10–40% in the water side.
There are several reasons for this non-uniqueness in the solutions for S
U
. As
noted in Section 2.2.3, experimentally, the cause for the multiplicity in the heat
transfer coefficients is due to the impossibility of an accurate determination of the
temperature at the wall separating the two fluids. Mathematically, however, such
multiplicity is due to the nonlinearity of the function to be minimized in which
the functional form of the assumed correlation plays a fundamental role. Although
37
0 2000 4000 6000 8000 10000
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
10000
−10%
+10%
−10%
+10%
Q
e
(W)
Q
p

(
W
)
Figure 4.3. Experimental vs. predicted heat flow rates for power-law correlations;
+ A, ◦ B
38
many nonlinear models can be transformed into a linear form, the function S
U
in
Eq. (4.8) is not linearizable due to its mathematical formulation.
0 1 2 3 4 5 x 10
4
0.5
0.6
0.7
0.8
0.9
1
1.1
1.2
1.3
1.4
1.5
Re
w
R
a
R
w
Re
a
0 2 4 6 8 10 x 10
2
Figure 4.4. Ratios of predicted air- and water-side Nusselt numbers.
This raises the question as to which of the local minima is the “correct” one. A
possible conclusion is that we should use the one that provides the smallest value
of the function S
U
, i.e. the global minimum, and hence the best possible heat rate
prediction. This leads to the search of the global minimum which can be done by
39
means of recently developed global optimization techniques e.g. genetic algorithms
(GAs), simulated annealing (SA) and interval methods (IMs).
4.1.2 Multiplicity of “best” correlations in heat exchanger 2
Let us now look at a regression analysis to find the best correlation for the
dry surface conditions of heat exchanger 2, the procedure for the other cases being
similar. The general form of this correlation, as proposed by McQuiston (1978b), is
j
s
= n
1
+n
2
Re
−n
3
D
_
A
A
tb
_
−n
4
. (4.10)
In a least-squares method the difference between the measured and the predicted
values of j
s
is minimized to determine the constants n
1
, n
2
, n
3
and n
4
in the above
equation. This is done by calculating the variance of the error defined as
S
js
=
1
M
1
M
1

i=1
[(j
s
)
e
i
−(j
s
)
p
i
]
2
(4.11)
where (j
s
)
e
i
, for i = 1, . . . , M
1
, are the experimental measurements and (j
s
)
p
i
, for
i = 1, . . . , M
1
, are the values predicted by Eq. (4.10). S
js
(C) is a smooth manifold
in a five-dimensional space, where C = (n
1
, n
2
, n
3
, n
4
) is the vector of unknown
constants, and a search must find C such that S
js
(C) is a minimum.
This procedure was carried out for the M
1
data sets. It was found that S
js
had
multiple local minima, the following two being examples.
Correlation A: j
s
= −0.0218 + 0.0606 Re
−0.0778
D
_
A
A
tb
_
−0.0187
(4.12)
Correlation B: j
s
= −0.0057 + 0.0562 Re
−0.1507
D
_
A
A
tb
_
−0.0436
. (4.13)
Superficially, the two correlations appear to be quantitatively different, and are both
different from Eq. (3.5) of Section 3.2.
Figure 4.5 shows a section of the S
js
surface that passes through the two minima,
A and B. For clarity the location coordinate z is such that C = C
A
(1 −z) + C
B
z,
40
−0.2 0 0.2 0.4 0.6 0.8 1 1.2
0
0.5
1
1.5
2
2.5
3
3.5
4
× 10
−8
z
A B
S
j
s
Figure 4.5. Section of the surface S
js
(n
1
, n
2
, n
3
, n
4
); A is the global minimum; B is
a local minimum.
where C
A
= (n
1
, n
2
, n
3
, n
4
)
A
and C
B
= (n
1
, n
2
, n
3
, n
4
)
B
correspond to the two
minima. These two minimum values of S
js
are within 17.1% of each other and the
corresponding j
s
predictions are within 3.7%. Thus the error in heat rate prediction
would increase somewhat if the higher minimum is chosen instead of the lower one.
4.2 Global optimization techniques
The current section introduces three global optimization techniques that have
proved to be suitable for applications involving complex phenomena. Genetic algo-
rithms (GAs), simulated annealing (SA) and interval methods (IM) are among the
41
global search techniques that have received increased attention in application areas
where the process of optimization in one form or another is necessary.
GAs are stochastic adaptive algorithms that work simultaneously with a popu-
lation of solutions to find a near global optimum. Applications of GAs to thermal
engineering include the prediction of heat rates in heat exchangers (Pacheco-Vega
et al., 1998), flow boiling (Castrogiovanni and Sforza, 1997), and electronic cooling
(Queipo et al., 1994).
Simulated annealing (SA) is a probabilistic method derived from statistical me-
chanics which distinguishes among different local optima to find the global optimum
solution to complex optimization problems. Some examples of its applications are:
the design of heat exchangers (Chaudhuri et al., 1997), and the study of intermolec-
ular interactions (Nilar, 1991).
Interval methods are deterministic techniques that are capable of solving global
optimization problems. The prediction of phase stability (Hua et al., 1998), and the
modeling of vapor-liquid equilibrium in chemical processes (Gau et al., 2000), are
among those few studies that have applied this method.
4.2.1 Genetic algorithm
Genetic algorithms (GAs) are based on the Darwinian evolutionary principle of
natural selection and natural genetics wherein the fittest members of a species sur-
vive and are favored to produce offspring. Developed by Holland (1975), originally
to study the adaptive processes of natural systems, GAs have been applied to solve
optimization problems in which multiple extrema are involved. GAs are discussed in
detail by Goldberg (1989), Davis (1991), Michalewicz (1992), and Mitchell (1996).
In the context of GAs, the members of a species may be regarded as candidate
solutions to a particular problem. The vocabulary used in GAs is borrowed from
42
natural genetics. Thus, a solution is encoded as a string, n
b
, of “0”s and “1”s, called
chromosome. The objective function, f, is named fitness function and a collection
of strings, M, is called population. A generation, G, is an iteration of the genetic
algorithm. The idea behind GAs is that after creating an initial population of
possible solutions, the members of the population are ranked according to how well
they satisfy a certain criterion, i.e. an objective function f. The fittest members
are more favored to combine amongst themselves to form the next generation or
offspring which tends to represent better solutions to the problem. A generation is
completed when the offspring replaces the parents from the preceding generation.
The evolution is achieved by means of genetic operations. These are crossover and
mutation. The crossover operator uses the concept of mating in genetics to alter
the composition of the offspring by exchanging part of strings from the parents. It
works based on a preselected probability p
c
. A detailed analysis on the properties
of crossover can be seen in Section A.3 of Appendix A. The mutation can be seen
as the operator which provides variation to the population and works by changing
a gene, based on a preselected probability p
m
, from 0 to 1 or vice versa.
An outline of the procedure with its corresponding flow chart, shown in Figure
A.1, can be found in Appendix A. Two examples are also given in the appendix to
illustrate how the technique works.
4.2.2 Simulated annealing
Simulated annealing (SA) is a generalized Monte Carlo technique that stochas-
tically simulates the slow cooling of a physical system to avoid getting trapped in
local optima. Detailed discussions of the simulated annealing are provided in van
Laarhoven and Aarts (1987), Aarts and Korst (1989), and Nirwan and Hou (1997).
43
This technique, introduced in Kirkpatrick et al. (1983) and
˘
Cerny, (1985) inde-
pendently, is analogous to the way liquids freeze and crystallize, or metals cool and
anneal. At high temperatures the energetic particles of a liquid are free to move
and rearrange. If the liquid is cooled slowly, the particles lose mobility as a result of
decreasing energy. Finally the particles settle down to an equilibrium state result-
ing in the formation of a crystal which is supposed to have the minimum energy.
This state would correspond to the optimal solution in a mathematical optimiza-
tion problem. However, if a liquid metal is cooled quickly i.e. quenched, it will not
reach the minimum energy state but rather an amorphous one with a higher level
of energy corresponding in the mathematical sense to a local extremum.
In order to use the analogy between thermodynamical systems and mathematical
optimization problems, a description of possible system configurations, a thermo-
static operation that schedules the decrease of the temperature, a random relaxation
process to look for the equilibrium solution at each temperature and an objective
function, e.g. f, which is the analog to the energy in thermodynamics, must be
provided (van Laarhoven and Aarts, 1987; Aarts and Korst, 1989; Nirwan and Hou,
1997).
Although at a given temperature SA behaves like a local search, such exploration
is prevented from being trapped in local minima by allowing uphill-climb moves with
certain probability p. This probability is related to the Boltzmann distribution such
that p = exp (−∆f/T

) (Aarts and Korst, 1989), where ∆f is the change in the
objective function from a current configuration to a newly produced solution, and
T

is a control parameter, which by analogy with the original application is known
as the temperature of the system and is independent from the objective function
involved. The criterion used to accept or reject a new solution is the Metropolis
algorithm (Metropolis et al., 1953).
44
It is to be noted that T

greatly influences the algorithm by allowing or rejecting
certain moves. The key to a good simulated annealing algorithm lies therefore in
the way T

is initially selected and later reduced. A good initial value of T

is the
one that provides an acceptance rate of the function being evaluated of about 90 %.
Several authors, among them Kirkpatrick et al. (1983), recommend a reduction
factor of r
t
= 0.9 to assure that a SA will get the global optimum.
An outline of the procedure is given in Appendix B. The corresponding flow
chart is shown in Figure B.1 of Appendix B. Two illustrative examples are also
provided.
4.2.3 Interval analysis
Interval methods are techniques based on interval analysis that provide enclo-
sures of solutions, within a desired tolerance, to nonlinear problems with mathe-
matical and computational certainty. Developed upon the idea of bounding round-
off errors during calculations when using real numbers, interval analysis involves
computations using intervals, where these are defined in terms of their endpoints.
Thus, a real interval X is defined as a bounded closed set of real numbers, i.e.
X = [x
L
, x
U
] = ¦x ∈ 1 [ x
L
≤ x ≤ x
U
¦, where x
L
≤ x
U
, and x
L
and x
U
are its lower
and an upper bounds respectively. It is to be noted that in the present section upper-
case symbols denote intervals while lower-case quantities are real numbers. Further-
more, a real interval vector of m components is defined as X = (X
1
, X
2
, . . . , X
m
)
T
.
Detailed discussions on interval analysis, which include definitions of interval arith-
metic operations, are provided in Moore (1966), Neumaier (1990), Hansen (1992),
and Kearfott (1996). It is to be noted that a key feature of interval methods is the
ability to compute inclusion monotone interval extensions, F(X), of real functions,
45
f(x) (Schnepper and Stadtherr, 1996). Thus, the interval extension F(X) encloses
all possible values of f(x) for x ∈ X.
A particularly useful technique that has been used to solve optimization problems
and nonlinear systems of equations (Moore, 1966) is the interval-Newton method
(INM). Given a nonlinear system of equations with a finite number of roots within
some initial interval, the method is capable of narrowly enclose all possible solutions
to the system (Moore, 1966; Neumaier, 1990; Hansen, 1992; Kearfott, 1996).
Within the context of a multidimensional optimization problem, e.g. minimiza-
tion of a function f(x) where x is a vector of m independent variables, the INM
is used to find the roots of the gradient of the objective function; that is, of
g(x) ≡ ∇f(x) = 0, since one of these will be the global minimum. In order to
increase the efficiency of the search, the method is usually combined with an inter-
val branch-and-bound approach (Schnepper and Stadtherr, 1996) so that roots of
g(x) = 0 that may not be the global minimum do not need to be found.
For a global search, the algorithm has to be applied to a sequence of intervals,
beginning with some initial interval X
0
. Hence, for an interval X
k
in the sequence,
the first step is to apply the so-called function-range test, where an interval exten-
sion, G(X
k
), of the gradient g(x
k
) is calculated. The interval X
k
will be discarded
if G(X
k
) does not contain a zero, and will be kept if zero is element of it. The
objective-range test is applied next. Here the interval extension of f(x
k
), F(X
k
),
is computed. In this test, if the lower bound of the F(X
k
) is greater than a known
upper bound of the “global” minimum of f(x
k
), then the interval is discarded since
it would not contain the global minimum. Otherwise it is kept for further testing.
The last step is the interval-Newton (root inclusion) test, in which the linear interval
equation system
G

(X
k
)(N
k
−x
k
) = −g(x
k
) (4.14)
46
has to be solved for the new interval N
k
using, e.g., an interval Gauss elimination or
a Gauss-Seidel-like technique (Hansen, 1992; Kearfott, 1996). The latter procedure,
developed by Hansen and Sengupta (1981), has been used in this investigation.
In Eq. (4.14), G

(X
k
) is an interval extension of the Jacobian of the gradient g,
i.e. a Hessian matrix of f over the current interval, and x
k
is an interior point of
X
k
which is usually its midpoint. Since any root x

∈ X
k
is also contained in N
k
(Neumaier, 1990; Hansen, 1992; Kearfott, 1996), the next step is an iteration of the
form X
k+1
= X
k
∩ N
k
. The above iteration step along with Eq. (4.14) are perhaps
the keys of the method. From them, the outcome may be either the existence of a
unique solution to g(x) = 0, a condition of no solution due to X
k
∩ N
k
= ∅, or a
case where N
k
lies partially within X
k
and further testing is required for the new
interval X
k+1
. Mathematical proofs of the existence and uniqueness of a solution are
provided in the references Neumaier (1990), Hansen (1992), and Kearfott (1996).
A more detailed discussion of interval analysis is presented in Appendix C. A
flow chart of the algorithm presented in Hansen (1979), which uses the INM to
find the global minimum of a function of one variable, is shown in Figure C.1 of
Appendix C. Two examples are also given to indicate how the method works. The
present analysis uses a multidimensional extension of it (Hansen, 1980).
4.3 Global-based-regression for correlations
Let us illustrate the application of the global optimization approach with the
same two examples used to show the nonuniqueness of the regression process. For
the data collected from heat exchanger 1 the process is carried out with all three
methods, while for heat exchanger 2 the global regression is done by applying the
GA only.
47
4.3.1 Application to heat exchanger 1
Genetic algorithms (GAs) are used first to find the four correlation constants of
Eq. (4.5). To this end, the population size is chosen as M = 30 and the length of
each chromosome as n
b
= 30 bits. The latter selection encodes ¦n
1
, n
2
, n
3
, n
4
¦ into
a string of 120 bits (Goldberg, 1989) and, since the search domain was considered
to be (0,1.5), provides a resolution of 1.4 10
−9
for each correlating parameter.
It is to be noted that the search domain was determined from a preliminary trial
in order to enclose any physical solution. The probability of crossover was set to
p
c
= 1, while that of mutation was p
m
= 0.03. These allow the genetic algorithm
to search in the parameter space efficiently. The maximum number of generations
was set to G = 3000. The results obtained from the GA-analysis are summarized
in Table 4.1. Shown are the values of the four constants and the variance S
U
for
the best solution found by the algorithm. An average value of the variance over
10 runs of the algorithm S
U
as well as its standard deviation σ
S
U
, necessary since
slightly different results are obtained each time the code is run due to the stochastic
nature of the method, are also shown in the table. The best solution was found at
G = 2959 generations.
The correlation constants are computed next with simulated annealing (SA).
After some initial trials, the initial temperature-like parameter was set to T

=
800, the temperature reduction factor used in the reduction schedule was r
t
= 0.9,
the tolerance to 1 10
−9
, the maximum number of iterations was G = 500, the
number of operations before temperature adjustment N
t
= 30, and the number of
cycles for a given temperature N
s
= 20. With these, it is possible to achieve a
good schedule in the temperature reduction and, therefore, a good chance to find
the global minimum. As in the case of the GA analysis, the search domain of
¦n
1
, n
2
, n
3
, n
4
¦ is (0,1.5) for each one of its elements. The corresponding values of
48
n
1
to n
4
, S
U
, S
U
over 10 different runs and σ
S
U
are presented in Table 4.1. The
“best” number of iterations needed to achieve the global optimum was G = 373.
Since G here has a different meaning from that corresponding to the GA, CPU times
necessary to achieve convergence are given below.
Finally, the interval method is used to determine the corresponding constants.
In order to apply the interval method, an initial interval for the vector of cor-
relation constants must be provided. The current analysis considers the same
search domain that was used for the other two methods, which is (0,1.5) for each
element of ¦n
1
, n
2
, n
3
, n
4
¦. The tolerance for enclosing the solutions was set to
1 10
−9
. A narrow enclosure on S
U
was determined to be S
U
(n
1
, n
2
, n
3
, n
4
) =
[5.4901 10
−6
, 5.4901 10
−6
] and the point value of S
U
, shown in Table 4.1, was
S
U
= 5.4901 10
−6
. As in the case of the GA and SA, the corresponding results
given by the interval Newton method are also shown in Table 4.1.
Table 4.1. COMPARISON OF CORRELATIONS FOR SINGLE PHASE HEAT
EXCHANGER.
Method n
1
n
2
n
3
n
4
S
U
S
U
Zhao (1995) 0.0183 0.7520 0.1368 0.5850 - -
GA 0.0249 0.8135 0.1206 0.5574 5.7822 10
−6
6.1934 10
−6
SA 0.0303 0.7850 0.1015 0.5919 5.4901 10
−6
5.4920 10
−6
IM 0.0304 0.7848 0.1015 0.5920 5.4901 10
−6
5.4901 10
−6
From the results illustrated in Table 4.1, it can be seen that all three methods
provide values of S
U
within 5% of each other, being both IM and SA equally ac-
curate. On the other hand, the differences in S
U
between both GA- and SA-based
correlations with respect to the IM-based correlation, which gives the least value,
are 11% and 1% respectively. The reason of such differences is the larger variability
of S
U
obtained from the GA, with a value of σ
S
U
= 5.3 10
−7
, as compared to that
49
of the SA with σ
S
U
= 3.1 10
−9
. Since the IM is a deterministic method σ
S
U
is
naturally zero.
In terms of CPU-times necessary to find a “best” solution, the GA needed only
2hr, followed by SA with 5hr, and IM being the most expensive with 120hr. Since
all the codes have been written in MATLAB, the CPU-times reported here are all
based on MATLAB computations using a SUN Microsystems Ultra 30. It should be
further noticed that the CPU-times can be reduced substantially if a faster machine
is used and/or the code is written in a language like FORTRAN or C which can
be compiled. As reported above, computationally speaking the interval method is
found to be much more costly than GAs and SA. The reason for this is that at each
iteration not only the objective function but also other quantities like gradients and
Hessian matrices must be calculated. Furthermore, the interval Newton method has
to be used to determine the stationary points within the interval thus increasing the
computational cost.
Table 4.2. COMPARISON OF PERCENTAGE ERRORS IN (U)
−1
AND
˙
Q
PREDICTIONS BETWEEN VARIOUS CORRELATIONS FOR SINGLE PHASE
HEAT EXCHANGER.
Prediction method (U)
−1
˙
Q
Zhao (1995) 21.41 9.16
Genetic algorithm 3.79 2.96
Simulated annealing 3.64 2.87
Interval analysis 3.63 2.87
Table 4.2 illustrate the rms percentage errors in both the reciprocal of the over-
all heat transfer coefficient as well as the heat rate predicted by the correlations
obtained from all three methods. Also shown for comparison purposes are the re-
sults provided by Zhao’s (1995) correlation which was determined from the same
50
experimental data. It is to be noted that the predictions given by global optimiza-
tion methods are much better than those provided by Zhao’s (1995) correlation with
percentage errors confined to less than 3.8% and 3% in (U)
−1
and
˙
Q respectively.
Qualitative comparisons in the accuracy of the predictions from global-based
correlations are presented in Figure 4.6. Shown are the percentage errors in heat
rate estimations against the experimental values. From the figure, it can be seen
that all three techniques give approximately the same error in the predictions for
small values of the heat rates while the GA estimations seem to deviate a little bit
more than those obtained from the other two methods at larger values. Figure 4.7
illustrates actual comparisons of
˙
Q between the correlation obtained from IM and
experiments. The straight line in it indicates the equality between the predictions
and the experimental measurements. The symbol , in the caption of the figure, is
a perturbation parameter that will be introduced in Section 4.4.
4.3.2 Application to heat exchanger 2
In the current analysis, multiplicity of minima of the error surface comes from
the mathematical form of the correlating function assumed, Eq. (4.10), rather than
the lack of accurate wall-temperature measurements. This correlating form induces
the nonlinear function S
js
, being minimized, to be not linearizable. As before,
faced with this, it is natural to search for the global minimum among the various
local minima of S
js
. A genetic algorithm (Pacheco-Vega et al., 1998) can be used
to search for the global minimum of the function S
js
since Section 4.3.1 showed
that both IM- and SA-based correlations were almost equally accurate with the
GA-based correlation being also very close, and the GA resulted to be the least
computationally expensive. In fact, correlation A given in Section 4.1.2 is found to
be the global minimum.
51
0 2000 4000 6000 8000 10000
−10
−8
−6
−4
−2
0
2
4
6
8
10
E
r
r
o
r

(
%
)
Q
e
(W)
Figure 4.6. Comparison in heat rate predictions from global-based correlations; ◦
genetic algorithms; simulated annealing; + interval method.
52
0 2000 4000 6000 8000 10000
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
10000
−10%
+10%
Q
e
(W)
Q
p

(
W
)
Figure 4.7. Experimental vs. predicted heat transfer rate by the IM-based correlation
with = 0. Straight line is the perfect prediction.
53
The same global search procedure has been followed to find the best correlations
for dropwise and film condensation by first fixing the constants found for the dry
surface and then searching for the additional constants that Eqs. (3.7) and (3.8)
have. The results of the correlations for the dry surface are shown in Table 4.3 as
“Eq. (4.12),” and for the wet surfaces as “Minimizing S
j
.”
Table 4.3. COMPARISON OF PERCENTAGE ERRORS IN j
s
, j
t
AND
˙
Q PRE-
DICTIONS BETWEEN VARIOUS CORRELATIONS AND THE ANN.
Surface Prediction method j
s
j
t
˙
Q
McQuiston (1978b) 14.57 14.57 6.07
Gray & Webb (1986) 11.62 11.62 4.95
Dry Eq. (4.12) 8.1 8.1 3.84
Eq. (5.4) 9.3 9.3 3.84
ANN 1.002 1.002 0.928
McQuiston (1978b) 8.50 7.55 8.10
Dropwise Minimizing S
j
9.12 9.43 9.91
Minimizing S
˙
Q
− 17.76 4.27
ANN 3.32 3.87 1.446
McQuiston (1978b) 9.01 14.98 10.25
Filmwise Minimizing S
j
8.32 10.89 10.44
Minimizing S
˙
Q
− 18.97 4.34
ANN 2.58 3.15 1.960
Combined ANN 4.58 5.05 2.69
Table 4.3 shows the ability of the correlations to predict the same data from
which they were derived. The errors indicated are the root-mean-square (rms)
values of the percentage differences between the predicted and experimental data.
The results of the correlations of McQuiston (1978b) for all surfaces and Gray and
Webb (1986) for dry surfaces only are also shown; as claimed, the latter is seen to
perform better than the former. The correlations found here using the global search
technique are much better than the previous correlations for the dry surface, slightly
worse for dropwise and slightly better for filmwise condensations. Some of the
54
0 0.005 0.01 0.015 0.02 0.025
0
0.005
0.01
0.015
0.02
0.025
j
s
e
j
s
p
−10%
+10%
−10%
+10%
Figure 4.8. Experimental vs. predicted j
s
for heat exchanger with dry surface; + Eq.
(4.12); McQuiston (1978b); ◦ Gray & Webb (1986). Straight line is the perfect
prediction.
differences could be due to possible elimination of outliers by previous investigators;
none of the published data has been excluded here. Qualitative comparisons of j
s
under dry-surface conditions between the previously published correlations and the
GA-based-correlation results are shown in Figure 4.8. As indicated quantitatively
in Table 4.3, the scatter provided by the global-based correlation is smaller than
that given by McQuiston’s (1978b) and Gray and Webb’s (1986) correlations. The
corresponding comparisons for the wet surfaces will be shown later in Chapter 6.
55
4.4 Coupling between hydrodynamics and temperature
In Section 4.2 it was demonstrated that through global optimization it is possible
to obtain the “best” solution for the correlation parameters of any functional form.
To achieve such a goal, the common procedure of assuming the overall heat transfer
coefficient to depend on the mass flow rates of both fluids only, and also, that such
dependence could be formulated as a summation of thermal resistances was consid-
ered. This approach is exact under specific conditions such as constant properties,
spatially homogeneous transfer coefficients, and similarity in temperature profiles.
However, as an example, the properties of many fluids, e.g. the viscosity of wa-
ter, have large variations depending on the temperature values thus increasing the
complexity in the description of the heat transfer phenomena due to the coupling
between the hydrodynamics inside and outside of the tubes in a heat exchanger.
Nevertheless, if such interaction is not very strong, then it may be possible use the
global-regression methodology, introduced in the present chapter, combined with a
small correction factor that will account for such coupling. To find this factor the
current section will consider the experimental data set 1, as was done in Sections
4.1.1 and 4.3.1, along with the global-based correlation developed from it for heat
exchanger 1.
Let us now deviate from the ideal case mentioned above and introduce a small
perturbation parameter, ( ˙ m
w
, ˙ m
a
, T
w
, T
a
), in the equation that describes the ther-
mal interaction between the fluids. To determine the heat rate it is possible to
write
˙
Q = UA∆T
m
[1 +( ˙ m
w
, ˙ m
a
, T
w
, T
a
)] (4.15)
where the product UA∆T
m
is the heat rate, given in Eq. (2.1), which is based on
ideal assumptions. U = U( ˙ m
w
, ˙ m
a
) is the overall heat transfer coefficient dependent
56
only on the mass flow rates, A is either the water-side or air-side heat transfer area,
∆T
m
is a characteristic temperature difference, which from an energy balance also
based on the aforementioned assumptions, of heat exchanger 1 (see Figure 1.1) is
given by
∆T
m
=
T
in
w
−T
out
w
ln
_
T
in
w
−T
in
a
T
out
w
−T
in
a
_. (4.16)
In Eq. (4.15) = ( ˙ m
w
, ˙ m
a
, T
w
, T
a
) is a perturbation parameter which depends on
the mass flow rates and temperatures of both fluids. On the other hand, U may be
estimated by any global-based correlation. The correlation employed in this section
is the one given by the interval analysis technique which was chosen due to the
reliability of the method (Gau et al., 2000). It is to be noted that the perturbation
parameter include the effects due to variation of properties with temperature in the
local heat transfer coefficients, the inhomogeneity of these, and the approximation
errors of the assumed correlation form.
In order to find the functional form and the corresponding constants for , a
regression analysis can be carried out. Here, the variance of the error is defined as
S

=
1
N
1
N
1

i=1
(
e
i

p
i
)
2
(4.17)
where
e
i
, for i = 1, . . . , N
1
, are the values determined from Eq. (4.15), since both the
heat rates from experiments and correlations are known, and
p
i
, for i = 1, . . . , N
1
,
are the estimations by the yet-to-be-determined equation that fits the data the best.
The -values are confined to (-0.08,0.08).
After a number of trials, the final mathematical form of the perturbation param-
eter , obtained on the basis of an artificial neural network approach (Pacheco-Vega
et al., 2001b), is given explicitly as
57

p
= C
1
+C
2
_
1 + exp
_
C
3
+
5

i=1
D
1
i
1 + exp(D
2
i
+D
3
i
Re

w
+ D
4
i
Re

a
+ D
5
i
T
in,∗
w
+ D
6
i
T
in,∗
a
)
__
−1
(4.18)
where C
1
= −0.099, C
2
= 0.19, C
3
= 3.6 and Ds are given in Table 4.4.
Table 4.4. VALUES OF THE D-CONSTANTS IN EQ. (4.18).
Constant i = 1 i = 2 i = 3 i = 4 i = 5
D
1
i
7.1 −2.9 −2.1 −5.3 8.3
D
2
i
8.8 1.0 5.3 −0.011 1.7
D
3
i
14.6 −4.8 8.1 1.8 1.4
D
4
i
−5.8 1.2 −0.24 0.55 −0.31
D
5
i
−1.6 2.8 −13.2 2.9 8.7
D
6
i
−1.3 −12.7 15.0 −0.58 −3.0
In Eq. (4.18), the quantities Re

w
, Re

a
, T
in,∗
w
, and T
in,∗
a
, are all normalized vari-
ables in the range (0,1). These are defined as
Re

w
=
Re
w
−Re
w
¸
¸
min
Re
w
¸
¸
max
−Re
w
¸
¸
min
, Re

a
=
Re
a
−Re
a
¸
¸
min
Re
a
¸
¸
max
−Re
a
¸
¸
min
,
T
in,∗
w
=
T
in
w
−T
in
w
¸
¸
min
T
in
w
¸
¸
max
−T
in
w
¸
¸
min
, T
in,∗
a
=
T
in
a
−T
in
a
¸
¸
min
T
in
a
¸
¸
max
−T
in
a
¸
¸
min
(4.19)
where the subindices max and min stand for the maximum and minimum values of
the Reynolds numbers, Re
w
and Re
a
, and the inlet temperatures, T
in
w
and T
in
a
, of
water and air respectively.
A comparison between the experimental values of the perturbation parameter
and the predictions given by Eq. (4.18) are shown in Figure 4.9. From the figure it
is observed that the data could be correlated up to a certain degree indicating the
existence of certain patterns in besides the expected random noise. A consequence
58
of the above is that errors in the predictions from correlations cannot be attributed
to uncertainties in the experiments only but to the deficiencies of the conventional
analyses.
−0.1 −0.05 0 0.05 0.1
−0.1
−0.08
−0.06
−0.04
−0.02
0
0.02
0.04
0.06
0.08
0.1
ε
e
ε
p
+10%
−10%
+10%
−10%
Figure 4.9. Experimental vs. predicted perturbation parameter . Straight line is
the perfect prediction.
The rms error of the percentage differences between predictions and measure-
ments of the heat rates are presented in Table 4.5. From the table it can be seen
an improvement in accuracy of 68.6% when is taken into account. On the other
59
hand, the absolute dimensional rms error in the heat rate predictions is shrunk from
a value of 102.66 W if the parameter is zero, to 30.62 W if the correction is included.
This progress can also be noticed by the substantial reduction in the scatter of heat
rate estimations from the IM-based correlation when comparing Figures 4.6 and
4.10. Figures 4.7 and 4.11 also illustrate such improvement. In these, the results on
the heat transfer rate predictions are shown respectively for = 0 and ,= 0.
Table 4.5. COMPARISON OF PERCENTAGE ERRORS IN
˙
Q WITH = 0 AND
-CORRECTION INCLUDED.
Correction
˙
Q
= 0 2.87
correlated- 0.90
4.5 Summary
Non-uniqueness of the result for the correlation constants obtained from a least-
squares procedure, due to the lack of accurate tube-wall temperatures and the im-
possibility of linearizing the assumed functional form, is found to be another reason
for the increase of the error bands in the heat rate predictions. In this case, the
search for the optimum set of values should be global rather than local. It has been
shown that there can be small but significant differences between the two.
This chapter has introduced a methodology based on global optimization to find
correlations for both transfer coefficients from experimental data. All three tech-
niques considered are suitable to search and find, through regression analyses, the
optimum values of the complete set of correlation constants for both heat transfer
coefficients that enable accurate predictions of the heat transfer rates in a heat ex-
changer. The differences between their heat rate estimations are within 3.5%. Each
60
0 2000 4000 6000 8000 10000
−10
−8
−6
−4
−2
0
2
4
6
8
10
E
r
r
o
r

(
%
)
e
Q (W)
Figure 4.10. Comparison in heat rate predictions from IM-based correlation with
-correction included.
method has shown to have advantages over the others. For instance both genetic
algorithms and simulated annealing do not require any gradient information to carry
out the global search and hence are faster than the interval method. Moreover they
do not even need analytic expressions for the function itself and can also search in
a discrete parameter space. GAs have advantage over SA in being more efficient in
the search as they produce a set of possible solutions each iteration. However, GAs
are not guaranteed to find the global optimum but only the region in which the
61
0 2000 4000 6000 8000 10000
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
10000
−10%
+10%
Q
e
(W)
Q
p

(
W
)
Figure 4.11. Experimental vs. predicted heat transfer rate by the IM-based correla-
tion with -correction included. Straight line is the perfect prediction.
62
optimum it is located with a high degree of probability. SA on the other hand, is
probabilistically guaranteed to find an optimal solution to a problem (Ingber, 1993;
Yang, 2000) and hence is more reliable than the GA. The major virtue of IMs is that,
unlike GAs and SA, they guarantee that the global solution has been found despite
of errors from rounding, approximation and uncertainty of data. Other advantages
of the method are that in the case of a multimodal function in which several global
minima exist, IM is capable of finding all those minima whereas both GAs and SA
cannot. The need for providing analytic expressions of the functions, gradients and
Hessians are among the limitations of the method. A main drawback is that in
higher dimensions the algorithm is computationally very expensive.
The current work has demonstrated that through global optimization it is pos-
sible to predict accurately the performance of heat exchangers with errors in the
heat rate estimations confined to less than 3%. It must pointed out, however, that
the conventional approach to compute the heat rate from correlations, i.e. two heat
transfer coefficients and a characteristic temperature difference, include assump-
tions and conditions, like constant properties, similarity in temperature profiles,
and spatial homogeneity of the transfer coefficients, that are not consistent with
the phenomena occurring in them under actual states of operation. Temperature
dependence of properties increment the complexity of the heat transfer phenomena
by coupling the hydrodynamics occurring inside and outside the tubes. A thorough
analysis has been carried out to demonstrate that such interaction is indeed one
reason for the wide error bands in estimations from correlations. It has been shown
that this interaction, defined in terms of a perturbation parameter, can be correlated
to improve even more the predictions of the performance in any heat exchanger. The
error is reduced to about a third when this effect is taken into account.
63
CHAPTER 5
DIRECT CORRELATION OF HEAT RATE
In Chapter 4, it was indicated that correlations found from heat exchanger ex-
perimental data using a regression approach are often non-unique and that, in such
cases, a global error minimization algorithm, rather than a local procedure, should
be used. The chapter also reported the application of a global regression method-
ology that enables improved accuracy in the estimations of the behavior of heat
exchangers. The current chapter attempts to increase the exactness of heat transfer
rate predictions by correlating the heat transfer rates directly.
In the previous chapter, the proposed global approach was used to find the
correlation constants of heat transfer coefficients by correlating these. From their
knowledge the heat rate could be calculated. As mentioned in Section 2.2, the heat
transfer coefficients are defined based on the assumption of a characteristic temper-
ature difference, and their usefulness depends on the similarity of the temperature
profiles. In other words, if the temperature distributions are always identical except
for a multiplicative constant, this constant is the only quantity needed to determine
the heat flux at the wall. With property variations, condensation, laminar and tur-
bulent flows, and many other phenomena present, this is often not the case. The
situation is worse for mass transfer during condensation, since its rate is clearly not
proportional to the enthalpy difference. Thus it may be better to develop correla-
tions to predict the heat rates directly, rather than intermediate quantities like the
64
transfer coefficients. In addition, the user of the information is usually interested
only in the overall heat rate and not in intermediate variables like the heat transfer
coefficients.
The goal of the present section is to correlate the heat rate directly using a
genetic algorithm based global procedure. To this end, the same two heat exchangers
that were considered in Chapter 4 will be used here as examples. First, data from
heat exchanger 2 will be applied to explore the advantages of predicting the heat
rate directly instead of using j-factors to determine the transfer coefficients, from
which the heat rates have to be calculated, as is usually done. Later, by taking
the experimental data from heat exchanger 1, the applicability of the methodology
introduced in the previous chapter is extended to find correlations of the heat rate as
direct functions of the normalized versions of the measured variables, instead of the
common nondimensional groups. This procedure has the advantage of being more
accurate for the specific fluids in the parameter range in which the experimental
data were taken, since fluid property values do not have to be determined, although
it may be less general in terms of fluids used. As an example of such property
variations, consider for instance that the viscosity of the water at 40

C is 65% of
what it is at 20

C. Furthermore, due to the fact that the correlation that is obtained
can only be as good as its assumed functional form, predictions of various three- and
four-constant correlation functions will also be investigated to find the correlation
that provides the closest fit to the experiments.
5.1 Direct correlation of heat rate in heat exchanger 2
As mentioned above, using the method of the previous chapter to determine the
heat rate, which is what the user of the information is usually interested in, one
must find the j-factor first, then the heat transfer coefficient, and finally the heat
65
rate. The procedure gives some generality to the results since different fluids and
temperatures can be used. If accuracy is the goal, however, it may be better to
correlate the heat rate directly, as will be investigated in this section. As was done
in Chapter 4 for the same heat exchanger, the dry-surface case is discussed in some
detail, and for the wet surfaces the procedure is similar.
Using the water-side heat transfer coefficient defined on the basis of the log-mean
temperature difference (McQuiston, 1978a), the heat rate is
˙
Q = (T
in
w
−T
in
a
)
1 −exp ¦UA[( ˙ m
w
c
p,w
)
−1
−( ˙ m
a
c
p,a
)
−1

( ˙ m
a
c
p,a
)
−1
−( ˙ m
w
c
p,w
)
−1
exp ¦UA[( ˙ m
w
c
p,w
)
−1
−( ˙ m
a
c
p,a
)
−1

(5.1)
where
U =
_
A
A
w
h
w
+
Pr
2/3
a
η G
c
c
p,a
j
s
_
−1
(5.2)
and j
s
is given in Eq. (4.10). To find the constants in the correlation, the mean
square error of the heat transfer rates
S
˙
Q
=
1
M
1
M
1

k=1
_
˙
Q
e
k

˙
Q
p
k
_
2
(5.3)
must be minimized. In Eq. (5.3) k = 1, ..., M
1
, where
˙
Q
p
are the predictions, and
˙
Q
e
are the experimental values of the heat rates. A genetic-algorithm based global
search results in a correlation, which when converted to a j-factor is
j
s
= −0.0479 + 0.0971 Re
−0.0631
D
_
A
A
tb
_
−0.0137
(5.4)
which is different from Eqs. (3.5), (4.12) and (4.13), all of which have been obtained
from the same data.
The results of using direct correlation of the heat rate are also shown in Table 4.3
of the previous chapter as “Eq. (5.4).” The percentage error is 3.84%, which is the
same as that given by Eq. (4.12). However, when looking at the absolute dimensional
rms error, direct correlation of heat rate shows an error of only 158.6 W, compared
66
0 500 1000 1500 2000 2500 3000 3500 4000
0
500
1000
1500
2000
2500
3000
3500
4000
4500
Q
p

(
W
)
−10%
+10%
Q
e
(W)
Figure 5.1. Experimental vs. predicted
˙
Q for a heat exchanger for dry surface.
Straight line is the perfect prediction.
to an error of 272 W from the j-factor correlation; this is an improvement of 41.7%.
This is in agreement with the qualitative results shown in Figure 5.1 which reflects a
better accuracy given by the direct correlation of heat rate. The global minimization
of S
j
does not necessarily imply the global minimization of S
˙
Q
. The difference of
the predictions in j
s
is 14.8% between the two methods.
A similar procedure was followed to find the correlations for j
t
under wet-surface
conditions, and the errors are given in Table 4.3 of Section 4.3.2 under “Minimizing
S
˙
Q
”. The direct correlation of dimensional heat rate gives predictions that have
67
56.9% less error compared to the j-factor correlation for dropwise condensation
and 58.6% for film condensation. Thus, predicting the heat rate rather than j-
factors gives more accurate predictions. This is somewhat expected since the j-factor
assumes the existence of the heat transfer coefficient, while
˙
Q does not.
5.2 Proposed functional forms for direct correlation in heat exchanger 1
The present section uses the approach introduced in Pacheco-Vega et al. (1998),
that is, the traditional non-dimensional groups are avoided in order to work with
normalized variables defined in the following manner:
θ
w
=
T
in
w
−T
min
T
max
−T
min
, θ
a
=
T
in
a
−T
min
T
max
−T
min
,
˙
M
w
=
˙ m
w
˙ m
w
¸
¸
max
,
˙
M
a
=
˙ m
a
˙ m
a
¸
¸
max
,
˙
Q

=
˙
Q
˙
Q
¸
¸
max
(5.5)
where T
min
and T
max
are the minimum and maximum temperatures measured, and
˙ m
w
¸
¸
max
, ˙ m
a
¸
¸
max
and
˙
Q
¸
¸
max
are the maximum values measured of the corresponding
quantities. All normalized variables are thus in the range [0, 1]. In general, a
correlation is a function f
C
such that
˙
Q

= f
C
(I), where I = (θ
w
, θ
a
,
˙
M
w
,
˙
M
a
) is the
input vector and C is the vector of constants that define the function.
The power law is one of many different forms that are commonly used in the
heat transfer literature, but there are numerous other possibilities, some of which
will be explored while restricting ourselves to three and four constants. To reduce
the range of possibilities it will be assumed that the combined effect of the internal
and external thermal resistances is additive; thus the heat rate
˙
Q

is proportional
to the temperature difference θ
w
−θ
a
if all other parameters are held constant, e.g.
in the following form
θ
w
−θ
a
˙
Q

= g
C
(
˙
M
a
,
˙
M
w
). (5.6)
68
Table 5.1 shows the functional forms, g
C
, of the correlations considered. In the
table, N
c
is the correlation number. The linear, quadratic and logarithmic functional
forms have three free constants while the others are all four-constant. For each
functional form, the corresponding parameters are determined by minimizing the
variance of the error which is defined as
S
˙
Q
∗ =
1
N
2
N
2

j=1
_
˙
Q
∗,e
j

˙
Q
∗,p
j
_
2
(5.7)
where
˙
Q
∗,e
j
, for j = 1, . . . , N
2
, are the experimental values and
˙
Q
∗,p
j
, for j = 1, . . . , N
2
,
are the estimations determined from the set of correlations, g
C
, given in Table 5.1.
It is to be noted that the data set 2, N
2
= 214 measurements for heat exchanger 1,
provided in Section 3.1 is used in the current analysis.
5.3 Three- and four-parameter search by genetic algorithms
The genetic algorithm is used to find the values of the constants, represented
as a r-dimensional vector C ∈ ( where ( ⊂ 1
r
, for which S
˙
Q
∗(C) is a global
minimum. An outline of the procedure (Goldberg, 1989; Koza, 1992; Michalewicz,
1992; Mitchell, 1996), also shown schematically in Figure A.1 of Appendix A, is:
• An initial population of Cs in a specific domain is randomly generated, each
member of which is a set of n
b
binary numbers that forms a n
b
-bit vector called
a chromosome.
• S
˙
Q
∗, calculated for each member of the population, determines its fitness or
effectiveness. The probability distribution for the next generation, on the basis
of which parents are selected, is calculated from the fitness values. Selection is
done “with replacement,” meaning that the same chromosome can be selected
more than once to become a parent.
69
Table 5.1. CORRELATION FUNCTIONS
N
C
. Correlation Functional form (g
C
)
1. Linear
θ
w
−θ
a
˙
Q

= n
1
−n
2
˙
M
w
−n
3
˙
M
a
2. Quadratic
θ
w
−θ
a
˙
Q

= n
1
−n
2
˙
M
2
w
−n
3
˙
M
2
a
3. Exponential
θ
w
−θ
a
˙
Q

= n
1
e
−n
2
˙
Mw
+n
3
e
−n
4
˙
Ma
4. Logarithmic
θ
w
−θ
a
˙
Q

= n
1
−n
2
ln
˙
M
w
−n
3
ln
˙
M
a
5. Power-law
θ
w
−θ
a
˙
Q

= n
1
˙
M
−n
2
w
+n
3
˙
M
−n
4
a
6. Inverse linear
θ
w
−θ
a
˙
Q

=
1
n
1
+ n
2
˙
M
w
+
1
n
3
+ n
4
˙
M
a
7. Inverse quadratic
θ
w
−θ
a
˙
Q

=
1
n
1
+ n
2
˙
M
2
w
+
1
n
3
+ n
1
˙
M
2
a
8. Inverse exponential
θ
w
−θ
a
˙
Q

=
1
n
1
+ e
n
2
˙
Mw
+
1
n
3
+ e
n
1
˙
Ma
9. Inverse logarithmic
θ
w
−θ
a
˙
Q

=
1
n
1
+ n
2
ln
˙
M
w
+
1
n
3
+n
4
ln
˙
M
a
70
• Crossover is applied to the parents based on a preselected probability p
c
. The
crossover point in both is the same, and bits corresponding to the rest of the
strings at that point are interchanged.
• Mutation is performed on a bit by bit basis with a preselected probability p
m
,
so that a mutated bit is changed from 0 to 1 or vice versa.
Once crossover and mutation have been applied to the complete population a new
population is created. Each iteration of this process is a generation. In Figure A.1
of Appendix A, the index j refers to a member in the population of size M, where
M is even, and the variable G indicates the number of the current generation. The
present results are with M = 20, kept the same for each generation. Each member
of the population is a value of C. For r = 4, C is encoded into a string of 120
bits (Goldberg, 1989), so that there are four segments of 30 bits, each of which
corresponds to an element of this vector. This gives a resolution of 2.0 10
−8
. The
probabilities of crossover and mutation were set at p
c
= 1.0 and p
m
= 0.03, and
the program was run to 1000 generations. The search domain of C depends on the
correlation, and was determined from a preliminary trial. As examples, in the power
law the elements of C are in [0, 1], while for the inverse quadratic correlation they
lie in [0, 22]. Figure 5.2 shows the decrease in the variance with each generation
for each of the nine correlations. The variances are seen to level off after a certain
number of generations.
The final results are summarized in Table 5.2 in ascending order of the variance
S
˙
Q
∗. Shown are the values of the constants, the variance S
˙
Q
∗, and the generation at
which they were found. Because of the random numbers used in the procedure, the
computer program gives slightly different results each time it is run. A number of
runs were made for each correlation; the average value of the variance, S
˙
Q
∗, as well
as the standard deviation σ are also in Table 5.2.
71
0 200 400 600 800 1000
0
0.005
0.01
0.015
2
1
4 9 7 3
6 5
8
S
Q
*
G
g
Figure 5.2. Change in S
˙
Q
∗ with generation; the numbers correspond to N
C
.
Table 5.2. COMPARISON OF DIFFERENT CORRELATIONS.
N
C
n
1
n
2
n
3
n
4
S
˙
Q

G
g
S
˙
Q

σ
5 0.1875 0.9997 0.5722 0.5847 4.407 10
−5
975 1.408 10
−4
7.916 10
−5
6 −0.0171 5.3946 0.4414 1.3666 1.601 10
−4
984 1.941 10
−4
4.902 10
−5
8 −0.9276 3.8522 −0.4476 0.6097 6.721 10
−4
912 1.278 10
−3
7.029 10
−4
3 3.4367 6.8201 1.7347 0.8398 1.208 10
−3
966 1.255 10
−3
1.471 10
−4
7 0.2891 20.3781 0.7159 0.7578 1.508 10
−3
990 1.509 10
−3
1.056 10
−6
9 0.4050 0.0625 −0.5603 0.2048 1.680 10
−3
830 2.111 10
−3
1.131 10
−3
4 0.6875 0.4714 0.4902 − 2.411 10
−3
917 2.412 10
−3
8.798 10
−7
1 2.3087 0.8533 0.8218 − 7.551 10
−3
976 7.553 10
−3
2.330 10
−6
2 1.8229 0.6156 0.5937 − 1.196 10
−2
896 1.196 10
−2
1.508 10
−6
72
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
1.2
1.4
−10%
+10%
Q
e
*
Q
p
*
Figure 5.3. Experimental vs. predicted heat flow rates for a power law correlation.
There is little difference between the first- and second-place, power-law and
inverse linear, correlations. Figures 5.3 and 5.4 show the predictions of these cor-
relations versus the actual experimental values, and there appears to be little to
choose between them; the deviation from the exact values are 2.7% and 3.4% on the
average, respectively, so that either could do just as well in predictions. Figure 5.5,
on the other hand, shows the worst case prediction, with an error of 51% relative to
the exact.
73
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
1.2
1.4
−10%
+10%
Q
e
*
Q
p
*
Figure 5.4. Experimental vs. predicted heat flow rates for an inverse linear correla-
tion.
74
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
1.2
1.4
−10%
+10%
Q
e
*
Q
p
*
Figure 5.5. Experimental vs. predicted heat flow rates for a quadratic correlation.
75
5.4 Summary
As has been mentioned in this document, the prediction of the heat rates usually
depend on the heat transfer coefficient. However, this is defined upon the assump-
tion of a characteristic temperature, and its existence relies on a similarity in the
temperature profiles. For a heat exchanger, this similarity has to hold for both fluids
at the same time under all conditions of operation. This assumption is extremely
difficult to achieve due to the nonlinearity of the phenomena involved. For this
reason it may be better to correlate the heat transfer rate directly.
This chapter has shown that correlating the heat rate in a direct fashion im-
proves the accuracy of the estimations from globally-determined correlations. In
some instances, this enhancement has been as large as 59%. This is an important
achievement since, for practical purposes, the heat rate is the quantity that com-
pletely characterizes the heat exchanger and thus is the relevant information needed
by the design engineer.
On the other hand, power laws are commonly used in the heat transfer literature,
though there are numerous other possibilities. Direct correlations of the heat rate
have been carried out to determine the constants for a set of proposed correlation
forms, using a global search based on the genetic algorithm. In addition, it was
chosen to correlate the relevant variables in normalized terms, instead of using the
usual nondimensional variables. The purpose of this is to incorporate the variation
of fluid properties within the correlation, so that there is no need to use property
tables in addition to the correlations. Two correlations, one being a power law and
the other an inverse linear, are found to predict the experimental data very well
with errors of the order of 2.7% and 3.4%.
76
CHAPTER 6
HEAT RATE PREDICTIONS IN HUMID AIR-WATER HEAT EXCHANGERS
USING NEURAL NETWORKS
Chapter 5 has demonstrated that by predicting the heat rate in a direct fash-
ion, instead of through intermediate variables like the heat transfer coefficients, the
accuracy of the estimations from globally-determined correlations is enhanced. On
the other hand, Chapter 4 established the weakness of the correlation procedure in
regards to the multiplicity of solutions obtained from them and the inconvenience of
using heat transfer coefficients for estimating the heat rates. A further drawback of
the conventional procedure is the compression of information from correlations. As
mentioned in Chapter 2, the common procedure is to compress information about
the heat transfer coefficients by means of correlations so that variations with respect
to the operating parameters can be taken into account through standard nondimen-
sional groups. Usually the form of the correlation cannot be completely justified
from first principles; it is selected on the basis of simplicity and common usage.
Moreover, the limited number of correlating constants cannot entirely represent the
different scales of the heat transfer phenomena in heat exchangers, and generate
errors in the estimations of the heat rate that are larger than the experimental
error.
This problem may be alleviated by an alternative approach: artificial neural
networks (ANNs). ANNs have been developed in recent years and used success-
fully in many application areas, among them thermal engineering (Sen and Yang,
77
1999). Some examples are heat transfer data analysis (Thibault and Grandjean,
1991), manufacturing and materials processing (Mahajan and Wang, 1993; Marwah
et al., 1996), solar receivers (Fowler et al., 1997), convective heat transfer coefficients
(Jambunathan, et al., 1996), and HVAC control (Jeannette et al., 1998). Previous
work on the prediction of heat rates in heat exchangers without condensation has
been reported by Zhao et al. (1995) and D´ıaz et al. (1999). The most attractive
advantage of the method is that it allows the modeling of complex systems without
requiring detailed knowledge of the physical processes. Because of this, and its in-
herent characteristics for handling incomplete information, generalizing the acquired
knowledge, and tolerance for imprecision, ANNs may provide attractive options to
capture the intrinsic patterns that govern the behavior of heat exchangers. As an
example of its application, the problem of accuracy in condensing heat exchanger
(heat exchanger 2) predictions using ANNs will be addressed.
Thus, the current chapter concentrates on using ANNs for the prediction of the
performance of heat exchangers with condensation. As has been the case throughout
this document, experimental data related to this problem, which was given in Section
3.2, will be used here. First, some generalities about the ANN will be given. Later
the focus will be on the issue of separation of data for training and testing the
neural network. Finally, the heat rate will be computed using this soft computing
approach.
6.1 Artificial neural networks
ANNs offer an attractive alternative to the correlation method discussed in the
previous chapters to predict the performance of heat exchangers. Although there are
many different types of neural networks, the feedforward configuration has become
the most widely used in engineering applications (Zeng, 1998). An ANN consists
78
of a series of layers, each with a number of nodes, the first and last layers being
the input and output layers while the remaining are hidden layers. In a multilayer
feedforward network, the nodes of each layer are connected only to those of the
layer before and the one after. The connections are associated with weights and the
nodes with biases. The nodes, or neurons, provide nonlinear mappings between their
inputs and outputs by means of activation functions. Although several appropriate
activations functions exist (Haykin, 1999), the sigmoidal logistic function which gives
a balance between linear and nonlinear behavior is the most commonly used. Such
nonlinear mapping capability, and the fact that the neurons are massively connected
enable the ANN to estimate any function without need of an explicit mathematical
model of the physical phenomenon. The training process for the ANN is carried
out in two steps: feedforward and backpropagation. During the feedforward stage,
a set of input data is supplied to the input nodes and the information is transferred
forward through the network to the nodes in the output layer where the produced
output is compared to the given data. During the backpropagation stage, the error
between output and target values is propagated backwards. The weights and biases
are repeatedly adjusted in order to minimize this error (D´ıaz et al., 1999) for which
the scheme used in this study is the backpropagation algorithm (Rumelhart et al.,
1986). Feedforward followed by backpropagation of all the data comprises a training
cycle. Further details are in Sen and Yang (1999). The mathematical background,
the procedures for training and testing the ANN, and an account of its history can
be found in the text by Haykin (1999).
It is to be noted that the present study uses the sigmoid activation function for
the nodes in the hidden and output layers while a linear function, i.e. f(x) = x, is
applied for those in the input layer. Also, before training is performed, all variables
are normalized to be within the [0.15, 0.85] range so that the nodes operate in the
79
relatively linear portion of the sigmoid function to avoid computational difficulties
in its asymptotic limits (Marwah et al., 1996; D´ıaz, 2000). During training, a critical
task is to find the ANN structure, i.e. the number of hidden layers and nodes in
these, such that the network would be large enough to learn the application but
small enough to generalize well. The current work follows the approach of D´ıaz
(2000), in which a particular configuration is chosen from a preselected set after
the performance of each ANN structure has been analyzed. Other methods, such
as pruning algorithms (Castellano et al., 1997), incremental approaches (Marwah
et al., 1996), or evolutionary programming (Yao and Liu, 1997), can be used to
optimize the network configuration.
6.1.1 Separation of data for training and testing
After a number of trials, the ANN structure chosen for the present analysis
consists of four layers: the input layer at the left, the output layer at the right and
two hidden layers. This 5-5-3-1 configuration, where the numbers stand for the
number of nodes in each layer, is similar to the schematic shown in Figure 6.1 which
is used in the next section for the heat exchanger analysis, the only difference being
that the fourth layer has only one output, the heat rate. The inputs to the network
correspond to the air-flow Reynolds number Re
D
, inlet air dry-bulb temperature
T
in
a,db
, inlet air wet-bulb temperature T
in
a,wb
, inlet water temperature T
in
w
, and fin
spacing δ. The output is the total heat rate
˙
Q. For testing the trained ANN, the
variables Re
D
, T
in
a,db
, T
in
a,wb
, T
in
w
, and δ are input and the corresponding values of
˙
Q
are predicted.
A total of M = 327 experimental runs were reported by McQuiston (1978a) for
three different surface conditions. The data can be separated in different ways into
training and testing data. D´ıaz et al. (1999), for example, used 75% of the total data
80
input layer
a,wb
T
a,db
in
T
w
in
T
in
Re
D
δ
hidden layers output layer
j
s
j
t
Q
.
W
θ
Figure 6.1. A 5-5-3-3 neural network used.
sets available for training and the rest for testing. This has two disadvantages: first,
the data set available for training is smaller than the total amount of information
available so that the neural network model is not the best possible, and second, if
the training data do not include the extreme values the estimations may fall in the
extrapolated range and are hence less reliable. The reader is referred to Appendix
D which provides an analysis on the extrapolation capabilities of the neural network
approach. The issue of separating the complete data into training and testing sets
is further analyzed in the following way.
The M available sets of experimental data are first randomly sorted to avoid
introducing any bias in the selection process and their order is then fixed. Only the
first M
a
of these are chosen for training and the rest M
b
= M − M
a
kept aside for
81
the moment. The fraction used for training is thus P
s
= M
a
/M. The rms values of
the relative output errors
S
e
˙
Q
=
_
_
1
M
a
Ma

i=1
_
˙
Q
p
i

˙
Q
e
i
˙
Q
e
i
_
2
_
_
1
2
(6.1)
are calculated at each cycle of the training process in order to evaluate the perfor-
mance of the network and to update the weights. Here i = 1, ..., M
a
, where
˙
Q
p
are
the predictions, and
˙
Q
e
are the experimental values of the heat rates. The stopping
criterion for the training process, however, is based on the absolute value of the
percentage error between
˙
Q
e
and
˙
Q
p
. When this error levels off, i.e. the change of
rate per cycle is less than 1%, the number of training cycles is recorded. The above
is achieved with 250,000 cycles, which is kept the same for the rest of the procedure.
After the training is finished, three different data sets are tested: (a) the same M
a
data that were used for training are tested, (b) the M
b
data left out of the training
process are tested, and (c) the complete M data sets are tested. In each case
the percentage error between the predicted and experimental values are calculated,
being E
a
, E
b
and E, respectively. Without reordering the M data sets the procedure
described above is repeated for different values of the percentage of splitting, i.e.
P
s
= 10%, 20%, ..., 90%, 95% and 99%. The exact shape of the error vs. P
s
curve
depends on the initial order of the data sets, but some general features can be
identified.
To determine the overall characteristics of the error, the curves were calculated
ten times and the results averaged to remove the influence of the initial random
ordering of the data sets. Figure 6.2 shows the average error in prediction calculated
in the three different ways as a function of the training fraction P
s
. The error bars
indicate the standard deviations, σ
a
, σ
b
and σ corresponding to E
a
, E
b
and E,
respectively, that resulted from the ten different curves. From this figure, it can be
82
seen that as P
s
increases the prediction errors for all three cases asymptote, on the
average, to approximately the same values. For any P
s
, E
a
is always small since
the training and testing data are identical. At small values of P
s
, E
b
and E are
both very large indicating that an insufficient fraction of the data has been used
for training. As M
a
increases, better predictions are obtained. Beyond P
s
= 60%
approximately the differences in the prediction errors for all data sets are small. The
same trend is observed for the σs which become smaller as P
s
→ 100%, indicating
that the prediction is somewhat insensitive to the initial random ordering of the
data. Near the end, however, at P
s
= 99% for example, σ
b
becomes large again
because now M
b
is very small and the error depends greatly on the data set that is
picked for testing.
From this exercise one can deduce that, in this case at least, if more than 60%
of the available data is used for training, the results will be essentially the same.
The rest of the chapter uses all of the available data for training, and the same for
testing. This gives us the best prediction over the widest parameter range. For the
correlations also, as is commonly done, the entire data set was used for finding the
correlation and testing its predictions.
6.1.2 Heat exchanger analysis
From the total of M = 327 experimental runs that were reported by McQuiston
(1978a), M
1
= 91 corresponded to dry-surface conditions, M
2
= 117 corresponded
to dropwise and M
3
= 119 to film condensation. Figure 6.1 shows a schematic of the
feedforward neural network configuration used for the present analysis. As before,
there are four layers in this configuration. The inputs to the network correspond
to the same physical variables described in the previous section. The outputs cor-
respond to j
s
, j
t
and the total heat rate
˙
Q. Thus, j
s
, j
t
, and
˙
Q are functions of
83
0 20 40 60 80 100
0
5
10
15
20
25
30
35
E
r
r
o
r

(
%
)
P (%)
s
Figure 6.2. ANN prediction errors vs. percentage of data used for training; ––
error E
a
using training data; –◦– error E
b
using data not used for training; ––
error E using complete data. Error bars indicate standard deviations. The E
a
and
E curves have been shifted horizontally by −1% and 1%, respectively, for clarity.
84
Re
D
, T
in
a,db
, T
in
a,wb
, T
in
w
, and δ. For testing the trained ANN, the physical variables
are input and the corresponding j
p
s
and j
p
t
and
˙
Q
p
are predicted. The j-factors are
not necessary for the heat rate predictions, but are calculated here merely for the
purpose of comparison with the previously discussed correlations.
As mentioned before, the performance of the network is evaluated by computing
the rms values of the output errors
S
e
=
_
_
1
M
k
M
k

i=1
_
O
p
i,k
−O
e
i,k
O
e
i,k
_
2
_
_
1
2
(6.2)
at each stage of the training, and the absolute values of the percentage differences
between O
p
i,k
and O
e
i,k
. Now i = 1, ..., M
k
, k = 1, 2, 3, where O
p
i,k
= ¦j
s
, j
t
,
˙

p
are
the set of predictions, and O
e
i,k
= ¦j
s
, j
t
,
˙

e
are the experimental output values.
Several ANN configurations were tested, as in D´ıaz et al. (1999), and the best results
were given by the fully-connected 5-5-3-3 configuration which was chosen. With a
number of 800,000 cycles the absolute value of the percentage error levels off. This
number of cycles provides a maximum error of 4%. It is to be noted that although
three different ANN models could be used to represent the functional relationships
between the independent variables and each output, i.e. j
s
, j
t
and
˙
Q, the three-node
output network, shown in Figure 6.1, has been used here for the sake of consistency
in the analysis.
6.2 Artificial neural network results
The ANN results are shown in Table 4.3 of Chapter 4 along with those of corre-
lations previously discussed. For all three surfaces, the ANN predictions are much
better than any of the correlations with maximum errors confined to less than 2%
in the estimations of the heat rate, and 3.9% in predictions of the j-factors. It is
of interest to note that the ANN gives better prediction for dry surfaces than for
85
wet. This is expected since the physical phenomena associated with condensation
are more complex.
j
s
p
j
s
e
0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016
0
0.002
0.004
0.006
0.008
0.01
0.012
0.014
0.016
−10%
+10%
Figure 6.3. Experimental vs. predicted j
s
for heat exchanger with dry surface; +
ANN; McQuiston (1978b); ◦ Gray & Webb (1986). Straight line is the perfect
prediction.
To determine whether the network based on training data separated by some
physical condition would perform better than another trained with the combined
data set, the entire set of runs M was combined to train a single ANN. The error,
shown in Table 4.3 under “Combined,” is larger than the ANN predictions for indi-
vidual cases, indicating that this ANN has more difficulty in differentiating between
the different physics involved. However, even then the predictions of the total heat
rate have errors of the order of only 2.7%.
86
j
s
e
0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018
0
0.002
0.004
0.006
0.008
0.01
0.012
0.014
0.016
0.018
−10%
+10%
j
s
p
(a)
87
0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018
0
0.002
0.004
0.006
0.008
0.01
0.012
0.014
0.016
0.018
−10%
+10%
j
t
p
j
t
e
(b)
Figure 6.4. Experiments vs. predictions for heat exchanger with dropwise conden-
sation; + ANN; McQuiston (1978b). Straight line is the perfect prediction. (a)
Sensible heat j
s
, (b) total heat j
t
.
88
Comparisons of j
s
and j
t
under dry- and wet-surface conditions between the
previously published correlations and the ANN results are shown in Figures 6.3–
6.5. For the dry surface there is only one j-factor, but under condensing conditions
both j
s
and j
t
are shown. The corresponding estimations of the heat transfer rates
are given in Figures 6.6–6.8. Note that in these the reported quantity is the total
heat rate,
˙
Q
t
, which in the case of the dry surface corresponds to the sensible heat
transfer rate,
˙
Q, reported in Chapters 4 and 5. In all the figures, the straight lines
indicate equality between prediction and experiment. The accuracy and precision
given by the ANN is notable. There are some data points in Figures 6.3–6.5 that
are outliers and can clearly be eliminated to improve the predictions, if desired.
0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018
0
0.002
0.004
0.006
0.008
0.01
0.012
0.014
0.016
0.018
−10%
+10%
j
s
p
j
s
e
(a)
89
0 0.005 0.01 0.015 0.02
0
0.002
0.004
0.006
0.008
0.01
0.012
0.014
0.016
0.018
0.02
−10%
+10%
j
t
e
j
t
p
(b)
Figure 6.5. Experiments vs. predictions for heat exchanger with film condensation;
+ ANN; McQuiston (1978b). Straight line is the perfect prediction. (a) Sensible
heat j
s
, (b) total heat j
t
.
90
6.3 Clustering of experimental data
The experimental measurements reported by McQuiston (1978a) were classified
according to a visual procedure. For dry surface conditions this method is relatively
reliable, but under wet conditions it is very difficult to distinguish between different
forms of condensation by simple observation. One can ask if this visual information
is really necessary, or if the data can be separated using some pattern recognition
procedure. One of the simplest is numerical clustering using a k-means algorithm
(Theodoridis and Koutroumbas, 1999). Hence, this method is applied to classify
the data into three groups, with all variables normalized in the [0, 1] range.
Table 6.1 shows the percentage of data sets of all three surface conditions that
belong to a particular cluster I, II, or III. These results show that cluster I clearly
corresponds to dry surface conditions, while the situation is less crisp for clusters II
and III. The difference between the dry and wet wall data would have been sharper
if only two clusters (dry and wet wall) had been used. Each of the three clusters is
analyzed separately to train three ANNs using their respective data sets, M
I
= 81
for cluster I, M
II
= 66 for cluster II and M
III
= 180 for cluster III. In each case a 5-
5-3-1 ANN configuraton was chosen. The training and testing were performed in the
same way as in the previous sections. The total heat rate errors given by the ANN
procedure are 0.866% for cluster I, 1.626% for cluster II and 1.242% for cluster
III which are of the same order for data separated by visual observation. Visual
information is thus not really necessary and the clustering technique is a viable
alternative that produces equally good predictions. Note that, as mentioned before,
when the data corresponding to the three surface conditions are combined to train
a single ANN, such network generates errors in the predictions larger than those
from the invididual cases indicating that the ANN has more difficulty to distinguish
91
between the different conditions. Thus, cluster analysis may be used along with the
neural network approach in order to further improve its results.
Table 6.1. PERCENTAGE OF DATA SETS OF DIFFERENT SURFACE CON-
DITIONS THAT FALL IN DIFFERENT CLUSTERS.
Surface I II III
Dry 89.01 10.99 0.00
Dropwise 0.00 29.06 70.94
Filmwise 0.00 18.49 81.51
6.4 Summary
Data compression during information transfer between the manufacturer of heat
exchangers and the designer, usually done through correlations, increases the error
in the final predictions. Artificial neural networks do not have the drawbacks of
correlations, and are an attractive alternative that can be used to accurately model
the thermal behavior of heat exchangers without need to assume a functional form
for the correlation nor an accurate mathematical model of the details of the process.
It has been shown here that they can predict the behavior of heat exchangers under
dry and wet conditions and their predictions are remarkably more accurate than
those from correlations. The error in the heat rate estimations from the neural
network methodology is, in some instances, 6.5 times smaller than that generated
by McQuiston’s (1978b) correlations and 4 times smaller than the one provided
by global-based-search correlations, which were obtained from the same data. The
network is able to catch the complex physics in a heat exchanger very well. In the
end, the error in the predictions of the neural network, confined to less than 2% in all
cases studied, is probably of the same order as the uncertainty in the measurements,
which is the best that can be expected.
92
0 500 1000 1500 2000 2500 3000 3500 4000
0
500
1000
1500
2000
2500
3000
3500
4000
4500
−10%
+10%
e
Q
t
(W)
°
p
Q
t

(
W
)
°
Figure 6.6. Experimental vs. predicted
˙
Q
t
for a heat exchanger under dry surface
conditions. Straight line is the perfect prediction.
93
0 1000 2000 3000 4000 5000
0
1000
2000
3000
4000
5000
6000
−10%
+10%
Q (W)
t
e
°
Q

(
W
)
t
p
°
Figure 6.7. Experimental vs. predicted
˙
Q
t
for a heat exchanger under dropwise
condensation. Straight line is the perfect prediction.
94
0 1000 2000 3000 4000 5000
0
1000
2000
3000
4000
5000
6000
−10%
+10%
Q
e
t
(W)
°
Q
p t

(
W
)
°
Figure 6.8. Experimental vs. predicted
˙
Q
t
for a heat exchanger under film conden-
sation. Straight line is the perfect prediction.
95
ANNs also have some limitations. They do not provide any physical insight
into the phenomena in which they are used, but then neither do correlations. The
training of the network may be computationally expensive though its subsequent
use for prediction purposes is not. For users with PCs, which is how most design is
currently done, it is quick and easy to use. One issue that needs to be looked at in
greater detail is the range of validity of the predictions as a function of the range
and density of the training data set. Though research in this and related areas is
ongoing, the neural network technique is currently a viable and accurate alternative
to conventional correlations.
The nature of the experimental data used here were classified by direct obser-
vations according the nature of the air-side surface, i.e. whether dry, covered with
condensed water droplets, or covered with a water film. This chapter has used
cluster analysis to classify the same data in an algorithmic way. The results show
that the visual information is not really necessary since the data can be separated
using these kind of procedures. The use of these procedures may help to increase
the quality of the modeling of heat exchangers by combining them with the neural
network technique.
96
CHAPTER 7
NEURAL NETWORK HEAT RATE ESTIMATIONS WITH LIMITED
EXPERIMENTAL DATA
Chapter 6 has demonstrated the usefulness of the ANN approach to accurately
model the thermal characteristics of condensing heat exchangers. It is seen that,
because of its inherent attributes, ANNs can correlate given experimental data with
errors of the same order as the uncertainty of the measurements. It is to be noted,
however, that networks trained from few tests may give large errors.
Empirical information necessary for design and selection of heat exchangers is
generated by means of experiments. Ideally, the experimental data should uniformly
cover the entire region of parameter space where predictions are to be made. How-
ever, the set of geometrical and operational parameters is extremely large and, even
if each quantity is allowed to assume a few different values, the resulting number of
required measurements is prohibitive. A common alternative to reduce the number
of experiments to a practical and economical level is the use of systematic methodolo-
gies such as those in the design of experiments (Montgomery, 1991). The drawback
of these methods, however, is that they fundamentally sacrifice information about
the interaction of the variables of the system on behalf of reducing the number of
tests and thus generate data which are limited and cannot always represent complex
interactions in a high-dimensional independent-parameter space.
Limited data offer only partial information about the true phenomena in heat ex-
changers and therefore empirical models, either in the form of correlations or ANNs
97
constructed from them, cannot accurately reproduce their behavior under different
conditions. Several techniques have been proposed in the context of neural networks
to determine confidence intervals of estimations using the convex hull of the data
(Courrieu, 1994), linearization theories (Chryssolouris et al., 1996), wavelets (Shao
et al., 1997) or self-organizing maps (Chinnam and Ding, 1998), among others.
These, however, are not adequate for undersized data, i.e., the number of data sets
is of the same order of magnitude as that of the independent parameters. Recently,
Niyogi and Girosi (1999) discussed mathematically the problem of approximating
functions from scattered data using linear superpositions of non-linearly parameter-
ized functions.
The current chapter focuses on expanding the applicability of the ANN method
to the prediction of the performance of heat exchangers used for refrigeration appli-
cations using very limited amount of data. To this end, typical experimental data
provided by a heat exchanger manufacturer which are denoted as data for heat ex-
changer 3 in Section 3.3 will be used. Initially, the ANN approach will be applied to
the data to show its capability in the representation of heat rates. Later, a method-
ology based on the cross-validation procedure is developed to estimate the expected
error of the ANN approach constructed from undersized data when determining
heat exchanger behavior under different conditions. Finally, the proposed technique
will be applied to the available experimental measurements to confirm the validity
of the method.
7.1 Artificial neural network analysis
As in the previous chapter, the current investigation also considers the fully-
connected feedforward neural network. The architecture of this ANN, schematically
illustrated in Figure 7.1, has one input layer, two hidden layers and one output layer.
98
Also, the sigmoid activation function for the nodes in the hidden and output layers
is used and the weights are adjusted by the backpropagation algorithm. Moreover,
as in Chapter 6, all variables are normalized in the [0.15, 0.85] range. Further details
are in Pacheco-Vega et al. (2001a).
The present section follows the training procedure described in Chapter 6. Thus,
the configuration of the ANN and the number of training cycles are determined by
a trial and error process (D´ıaz et al., 1999) in which either may be changed if the
performance of the network during training is not good enough. The performance
is evaluated by calculating the rms values of the output errors
S
e
=
_
1
M
M

i=1
_
O
p
i
−O
e
i
O
e
i
_
2
_
1
2
(7.1)
at each stage of the training and the percentage errors between the output and
target values. Here i = 1, ..., M, where O
p
i
= ¦
˙
Q
t
¦
p
i
are the predictions, O
e
i
= ¦
˙
Q
t
¦
e
i
are the experimental output values, and M is the total number of training data sets.
After some trials, the best results were obtained by the 11-11-7-1 configuration with
a number of 400, 000 training cycles (Pacheco-Vega et al., 1999), necessary for the
absolute value of the percentage error to change very little, i.e. 1%, with each cycle.
7.2 Predictions of the heat rate with ANNs
The capability of the ANN to model complex phenomena with very few data sets
can be illustrated by taking the total of M = 38 experimental runs and training
a fully-connected 11-11-7-1 ANN, shown in Figure 7.1. The N = 11 input nodes
correspond to the variables: ˙ m
a
, T
in
a,db
, w
in
a
, T
in
r
, L, δ, x
a
and x
b
, N
row
, N
col
, and
N
cir
, with the geometrical quantities scaled by tube diameter D; the output node
corresponds to the total heat rate
˙
Q
t
. The resulting function
˙
Q
t
=
˙
Q
t
( ˙ m
a
, T
in
a,db
, w
in
a
, T
in
r
, L/D, δ/D, N
row
, N
col
, N
cir
, x
a
/D, x
b
/D)
99
input layer hidden layers output layer
Q
t
.
N
row
T
a,db
in
T
r
in
a
w
in
m
a
L
D
δ
D
N
col
N
cir
x
D
a
x
D
b
Figure 7.1. Configuration of a 11-11-7-1 neural network for fin-tube heat exchanger.
100
is a manifold in a twelve-dimensional space. The prediction of
˙
Q
t
obtained from the
trained ANN plotted against the available experimental data is shown in Figure 7.2.
The straight line indicates the equality between the predicted and experimental
values of the heat transfer rates. It can be noticed that both the accuracy and
precision of the results are remarkable. The root-mean-square error of the percentage
difference between the predictions and measurements is less than 1.5%, approaching
the uncertainties in the experiments.
0 1000 2000 3000 4000 5000 6000 7000 8000
0
1000
2000
3000
4000
5000
6000
7000
8000
−10%
+10%
Q (W)
e
t
Q


(
W
)
p t
Figure 7.2. Experimental vs. predicted
˙
Q
t
. Straight line is the perfect prediction.
101
To demonstrate that the ANN has indeed captured the intrinsic patterns of the
heat transfer phenomena, one of the M = 38 data sets was randomly selected as
an example. Using this data set, two different cases were analyzed with the already
trained neural network and their results depicted in Figures 7.3(a) and 7.3(b). In
the first, the mass flow rate of air, ˙ m
a
, was the only input parameter to the ANN
that was allowed to vary; while in the second, with the exception of the air dry-bulb
inlet temperature, T
in
a,db
, all the input parameters to the network were fixed. For the
first case analyzed, ˙ m
a
was increased up to 10% whereas for the second, T
in
a,db
was
increased up to 2.5

C, both from their corresponding original values. Figure 7.3(a)
illustrates the variation of
˙
Q
t
with ˙ m
a
. From the figure, it can be seen that as ˙ m
a
increases so does the heat rate. This is expected to happen since increasing the
mass flow rate of air will cause the heat transfer coefficient to become larger. The
variation of the heat rate with the dry-bulb inlet temperature of air, shown in Figure
7.3(b), is also consistent since the characteristic temperature difference between the
air and refrigerant becomes larger as T
in
a,db
increases with a corresponding increment
in
˙
Q
t
. On the other hand, it to be noted that there is a larger change in
˙
Q
t
with
˙ m
a
than that with T
in
a,db
. For a 10% increment in these, the heat rate increases up
to 3% with respect to the mass flow rate and 2.1% with the dry-bulb temperature.
7.3 Error sources in ANN estimation
In order to use ANNs as a reliable tool for thermal analysis and design of heat
exchangers, one has to take into account the factors that influence its predictions.
As noted by several authors (Kramer and Leonard, 1990; Chryssolouris et al., 1996;
Shao et al., 1997; Chinnam and Ding, 1998; Niyogi and Girosi, 1999), the perfor-
mance of neural networks is influenced by noise corruption, spatial distribution and
size of the data used to construct the ANN model, and the characteristics of the
102
m (kg/s)
a
Q


(
W
)
t
0.39 0.4 0.41 0.42 0.43 0.44
5200
5220
5240
5260
5280
5300
5320
5340
5360
5380
5400
(a) Predicted
˙
Q
t
from an ANN vs. ˙ m
a
.
Q


(
W
)
t
T (°C)
a,db
in
4.2 4.3 4.4 4.5 4.6 4.7 4.8
5200
5220
5240
5260
5280
5300
5320
5340
5360
5380
5400
(b) Predicted
˙
Q
t
from an ANN vs.
˙
T
in
a,db
.
Figure 7.3. Variation in
˙
Q
t
predictions from a 11-11-7-1 ANN with ˙ m
a
and T
in
a,db
.
103
ANN, i.e. number of layers, number of hidden nodes, its architecture, etc. Noisy
data associated with uncertainties in measurements are generated in the experi-
ments. The noise can be maintained at a very small value if the experiments are
carried out with care and using accurate instruments. The fact that the ANN is
comprised of a finite number of hidden layers and nodes per layer to approximate
an unknown function also introduces an error. The magnitude of this error depends
on the representational capability of the ANN which may increase due to overfitting
as the size of the ANN becomes large. Another source of error stems from the fact
that only finite data are available for training. As the number of training data sets
increases, the error decreases. Niyogi and Girosi (1999) demonstrated that it is not
possible to reduce the upper bounds on errors due to ANN size and limited training
data simultaneously. Thus if we want to rely on the ability of the ANN to general-
ize the relationship between the input and output quantities that govern the heat
transfer phenomena in a heat exchanger we will have to be very careful in providing
an adequate training set. On the other hand, as a consequence of being applied to
regions beyond the range of available training data, ANNs are very likely to have
a poor performance on their predictions (see Appendix D for an analysis on the
extrapolation capability of ANNs). In fact, for a fixed neural network configuration,
we may have two limits depending on the availability of the training data and the
number of inputs to the ANN. The first arises when the number of measurements
is very large. In such a case, if the network is used outside the convex hull of the
training data (Courrieu, 1994), then the error in the predictions of the ANN will be
large because there are no data in the region to support the predictions. Inside the
domain given by the convex hull, the empty spaces where measurements are absent
are small in size and there is little degradation of the ANN predictions. The second
limit appears when there are very few training experiments. The voids inside the
104
convex hull of the data are large enough so that they contribute to the inaccuracy
of predictions made by the ANN in these regions.
7.4 ANN-based error estimate methodology
Since the size, density and distribution of the training data are intrinsically linked
to the reliability of the ANN predictions, it is crucial to determine the importance or
influence of each data set. The present section uses a variant of the statistical tool
known as cross-validation (Stone, 1974) to establish a methodology to determine the
validity of the ANN predictions by relying on the importance that each training data
set has in the trained ANN. This is a method that estimates the error of statistical
prediction models (Gong, 1986) and has been widely applied to different types of
neural networks in a number of ways; Prechelt (1998) presents a guide to select the
criteria for automatic early stopping in order to detect overfitting when training
multilayer perceptrons using cross-validation, while Leonard et al. (1992) use the
technique to find the optimal architecture in radial basis functions networks.
From the M available sets of experimental data, (M − 1) are used to train the
ANN. After the training is finished, the data set left out is predicted and the result
compared to the experimental value. The percentage error, S
cv
, given in the present
case in terms of heat rates as
S
cv
=
¸
¸
¸
¸
¸
˙
Q
p

˙
Q
e
˙
Q
e
¸
¸
¸
¸
¸
100 , (7.2)
is a measure of the importance of that particular set of data with respect to all
the measurements. If, for instance, a large value of this error is obtained, the point
excluded during the training process is important and its absence will produce an
ANN with poor estimation and generalization capabilities. On the other hand, if
the associated error S
cv
is small, it means that the data set has enough support from
its neighbors that its presence is not very important. This procedure is repeated M
105
times, once for each training data set. As a last step, a neural network model of the
error surface associated with S
cv
which can estimate the error associated for a new
set of design variables is generated using the above mentioned procedure.
R
S
c
v

(
%
)
0 0.2 0.4 0.6 0.8 1 1.2 1.4
0
10
20
30
40
50
60
70
Figure 7.4. Bar diagram of error estimation.
The methodology proposed in this section is now tested by taking the M = 38
data sets provided by the manufacturer, and following the procedure outlined above.
Note that 38 ANNs will then be produced, trained and tested. For each neural
network the configuration is a 11-11-7-1. The number of training cycles was set to
106
400,000 as in Section 7.2. The error surface generated from this exercise is shown
in Figure 7.4 as a bar-plot of S
cv
as a function of the data set. The latter is ordered
according to its Euclidean distance R in 11-dimensional space from the centroid of
all M sets. It is to be noted that the horizontal axis of Figure 7.4 is in the range
[0, 1.4], the reason being that, since we are working in a Banach space, R has been
calculated with all the input variables being normalized in the [0, 1] range. From
the figure it can be observed that the error values are in the range of zero to 60%
indicating the relative importance that each data set has in building the estimation
capability of the ANN model. From the manufacturer perspective, however, Figure
7.4 provides an indication of where new experiments may be needed in order to
decrease the estimation errors from empirical models. Note also that the error S
cv
is independent of R with void spaces appearing in different regions.
The error estimate procedure can be used to indicate if an ANN is overfitting
under a selected number of training cycles, or if such number was adequate. The
underlying idea is that if the topology of the S
cv
-surface varies significantly when
the number of training cycles is changed from the original, it will be an indication
that the network is overfitting. If such variation is small it will not be the case.
The error estimate methodology was used with the same M = 38 data sets
considered in the previous analysis, where the original number of training cycles
was 400,000. Thus, two different tests, namely test 1 and test 2, were carried out
using 350,000 and 450,000 cycles respectively. For both the network architecture
is the same 11-11-7-1 used previously. The corresponding results are presented in
Figure 7.5. Shown are the topologies of the S
cv
-surface that were generated from
tests 1 and 2, and the corresponding S
cv
-surface, illustrated also in Figure 7.4, that
was constructed with the original number of training cycles. As can be observed
from the figure, all three S
cv
-surfaces agree among each other quite well indicating
107
very little variability with the number of training cycles. Taking the original error-
surface as the basis, the rms percentage errors of the other two are confined to 5.6%
for 87% of the M = 38 data sets.
R
S
c
v

(
%
)
0 0.2 0.4 0.6 0.8 1 1.2 1.4
0
10
20
30
40
50
60
70
Figure 7.5. Comparison of bar diagrams of error estimation from the ANN-based
methodology using different number of training cycles; 350,000 cycles; 400,000
cycles; ◦ 450,000 cycles.
7.5 Validation of error estimate procedure
In order to confirm the estimate of the procedure presented in the previous
section, we take one experiment out of the manufacturer’s data and look at the
108
remaining measurements as if they were the total available. Thus, we have M = 37.
The experiment taken out is put aside and considered as a “test experiment.” The
error estimation technique is applied to the M data sets and the actual error of the
test experiment compared with the estimation.
Three typical test experiments were selected in the small, medium and large
error range, respectively. For the first test, the estimated error was determined to
be 0.761% while the actual error was calculated as 0.112%. The second test gave
23.3% and 6.19% respectively, while the 48.78% and 14.21% were the values for the
third one. All the networks considered had a 11-11-7-1 configuration and 400,000
training cycles were used to ensure a reasonable level of error. For all three tests
the error estimates are larger than the actual errors, indicating that the method
provides an upper bound estimate for the error in ANN predictions.
7.6 Summary
In the present study we have applied the ANN approach to accurately model
the thermal characteristics of refrigerating heat exchangers. Because of the inherent
attributes of the ANN technique, which traditional analyses including standard
correlations do not have, ANNs can correlate given experimental data with errors
of the same order as the uncertainty of the measurements. Even when discrete
variables such as those in heat exchanger geometry are involved, as is the case here,
its ability to recognize patterns allows the neural network to capture all the complex
physics without need to assume mathematical models of the process. These features,
in principle, make the ANN approach suitable for use in the estimation of heat rates
under different conditions.
Correlations and neural networks, being empirical models of complex systems are
constructed based on experimental information. Their effectiveness in estimating the
109
heat rates under operating conditions different from those used in their development,
depend on the number and distribution of these measurements. Models constructed
from large, dense and well distributed measurements will tend to have smaller errors,
while those built from undersized data will perform poorly. Limited data arise from
the fact that in industrial applications, such as heat exchanger manufacture, it is
not economically possible to perform a large number of experiments.
We have presented a methodology for the estimation of errors from an ANN
trained with a limited number of data sets. A low value of the estimated error
is an indication that there are sufficient experimental data to support the ANN
prediction; while a large error will indicate absence of enough points to aid in the
ANN predictions and more experiments would be needed in order to improve these.
The procedure proposed here can help the manufacturer to plan new measurements
by showing where these are needed.
110
CHAPTER 8
GENERAL CONCLUSIONS
8.1 Conclusions
The present dissertation has dealt with the enhancement in the accuracy of pre-
dictions of the steady state performance of plate fin-tube compact heat exchangers.
To this end, two types of techniques have been used. The first encompasses global
optimization methods that are used to find the parameters of a given functional
form. The second is a adaptive function itself which belongs to the soft computing
methodology: the artificial neural network.
Heat exchangers are important complex systems for which the accurate deter-
mination of their performance is essential. Due to the fact that no first-principles
analytical solutions are available, a common practice is to use nondimensional cor-
relations for the inner and outer heat transfer coefficients, together with a charac-
teristic temperature difference, to compute the heat rates. Correlations developed
from experimental information, however, often produce unsatisfactory predictions
with errors that can be much larger than the uncertainties in the measurements.
Such errors, sometimes as large as 25-30%, are due to the assumptions made during
the different stages of the procedure. These include the use of average coefficients
and constant thermophysical properties, as well as the data compression that occurs
through the correlation process. A complete analysis of the conventional methodol-
ogy has shown that non-ideal effects, which are present in the transfer of heat, are
111
important and can be correlated to improve the heat rate predictions. It has been
demonstrated that when these are taken into account, the error from correlations
are reduced by a third. The current work has also found that another important
reason for the lack of accuracy in the heat rate estimations from correlations is the
multiplicity of the result in the correlation constants obtained from a local regression
process.
In this project, a global-search-based methodology that enables a better accu-
racy in the estimations of the heat rates by finding the optimum values of the
correlation parameters for the transfer coefficients was introduced. The heat rate
estimations from correlations determined from this global-search-based procedure
are confined to less than 3%. Results from the application to a single-phase and
a condensing heat exchanger have shown that all three techniques used, genetic
algorithms, simulated annealing, and interval methods, are suitable in this regard.
The differences among their heat rate predictions are within 3.5%. An extension
of the aforementioned methodology is also presented to develop correlations able
to predict the heat rates directly rather than through intermediate quantities like
the heat transfer coefficients. Results indicate that correlating the heat rate in a
direct way improves the accuracy of the estimations from globally-determined cor-
relations. In some instances, the enhancement in accuracy has been as large as
59%. Furthermore, determination of the correlation constants for a proposed set of
functional forms show that it is possible to have more than one correlation capable
of estimating the heat rates accurately. Two correlations, one a power law and the
other an inverse linear, are found to predict the experimental data very well with
errors of the order of 2.7% and 3.4%.
In the current dissertation the inaccuracy from correlations due to compression
of information has been addressed using the neural network technique. ANNs do
112
not have the deficiencies of correlations; moreover they do not need explicit math-
ematical representations nor detailed information of the process. Results from the
application of ANNs to condensing heat exchanger data have demonstrated that,
because of its inherent attributes that allow them to recognize and simulate nonlin-
ear phenomena in complex processes, ANNs can correlate given experimental data
with errors of the same order as the uncertainty of the measurements. In all the
cases investigated here, the estimation errors from the ANN approach are restricted
to less than 2%. This represents an enhancement in the accuracy of 400% as com-
pared to the results provided by global-based-search correlations. Even with limited
experimental measurements involving discrete variables, such as those included in
the heat exchanger geometry, the ANN approach is suitable to simulate the behav-
ior of heat exchangers as illustrated from its application to refrigerating coils. The
error in the estimation of the heat transfer rates is of only 1.5%.
To extend the applicability of the ANN approach for industrial purposes, this
work has introduced a methodology based on the cross-validation technique for the
estimation errors of ANNs trained with a limited amount of measurements. This
procedure finds regions where not enough data are available to construct a reliable
neural network. The results indicate that the proposed approach gives an upper
bound of the estimated error in the heat rates.
The focus of this research work has been to develop methodologies that allow
accurate steady-state estimations of the heat transfer rate in heat exchangers. Al-
though these methodologies have been directed for the study of heat exchangers,
concentrating particularly in the most widely used compact fin-tube type, they can
well be applied to other types thermal devices and extended to analyze thermal
networks.
113
8.2 Recommendations for future work
There are several topics that have been left out of the scope in the present
dissertation. These are related to the analysis of heat exchangers, and the analysis
and complementary applications of the techniques used in the current work in order
to improve their capabilities. Thus, some recommendations for further work in these
areas are provided next.
Although the current dissertation has used optimization techniques to find reli-
able models, i.e. correlations for the simulation of heat exchangers, the step ahead
is the geometrical optimization of the heat exchanger. This is a problem that can
be addressed by means of genetic algorithms and simulated annealing to search for
the number of tube rows and columns, fin spacing, diameter, etc., to maximize the
heat transfer rate from the model, e.g. an ANN, while minimizing energy consump-
tion. Sensitivity analyses may also be included in the modeling by ANNs in order
to improve even more their accuracy. The current research has opened possibilities
of further application of clustering of experimental data. Cluster analysis of exper-
imental data from different phenomena, e.g. laminar and turbulent flows or heat
transfer with and without condensation, can be applied to thermal components to
classify the data before using a technique for modeling. A further application in-
cludes fuzzy clustering to determine the adequate number of regimes in which heat
exchangers may be working. The analysis of some heat transfer phenomena, such as
spatially inhomogeneous or unsteady problems, can also be pursued with the ANN
approach if the information from previous spatial or temporal locations are supplied
to the network such that the values at new points in space or time can be generated
from it. To investigate the possibility of capturing the physics of several phenomena,
CFD calculations may also be used to produce training data for the neural network
technique.
114
As mentioned in Chapter 5, many different correlating forms, among them power
laws, are commonly used in the heat transfer literature, but there are numerous other
possibilities. Hence, the natural extension to the study of the correlation approach is
to search for the function that best fits the heat exchanger experimental data without
assuming a priori the form of the correlation. The search algorithm chooses from a
variety of possible forms of correlation functions within a restricted functional space,
compares them and picks out the one, along with the corresponding constants, that
provides the closest fit to the experiments. In principle, the above can be carried
out by means of genetic programming (Koza, 1992), which is an extension of the
genetic algorithm. The simulated annealing algorithm may be also suitable in this
regard.
In regards to the techniques that belong to soft computing, the ANN is one
of which several improvements can be attempted. The selection of the network
configuration, i.e. number of layers and number of hidden nodes, is a un-resolved
issue in ANNs. From a set of configurations both genetic algorithms and simulated
annealing can search for the one that provides the best accuracy with the smallest
architecture, i.e. number of layers and hidden nodes. Another issue that can be pur-
sued is in regards to the learning algorithm. This may be based on interval Newton
methods instead of a steepest descent to find the optimum values of the weights and
biases of a given network architecture. One problem that deserves consideration is
the selection of the activation function in ANNs. It would be interesting to explore
if a different function, instead of the common sigmoid, hyperbolic tangent, etc., may
be used to perform the nonlinear transformations between inputs and outputs.
Interval analysis is a powerful tool which can be applied to several areas of
thermal engineering. Control of thermal devices is one of them. For example, the
constants of a PID may be optimized to find the best controller. On the other
115
hand, its application has been so far to algebraic systems. It would be interesting
to expand its application to the solutions of differential equations.
116
APPENDIX A
GENETIC ALGORITHMS
An outline of the procedure, for finding a maximum of a function f(x) in a
specific domain a ≤ x ≤ b, also shown schematically in Figure. A.1, is:
• An initial population of M members x
1
, ..x
M
∈ [a, b] is randomly generated.
• The value of the objective function f(x) which correspond to the fitness, is
calculated for each member of the population.
• The probability distribution for the next generation, on the basis of which
parents are selected, is calculated from the fitness values.
• Crossover is applied to the parents based on a preselected probability p
c
. The
crossover point in both is the same, and bits corresponding to the rest of the
strings at that point are interchanged.
• Mutation is performed on a bit by bit basis with a preselected probability p
m
,
so that a mutated bit is changed from 0 to 1 or vice versa.
• Once crossover and mutation have been applied to the complete population a
new population is created. The process is continued until the largest fitness
in a generation does not change much any more.
117
The numerical implementation of the genetic algorithm was done in MATLAB.
In Figure A.1 the index j refers to a member in the population of size M, where M
is even, and the variable G indicates the number of the current generation.
A.1 Optimization of a function of one variable
Let us now consider the application of GA to a two-dimensional problem to show
how the technique works for finding the global maximum of a multimodal function.
The function of one variable given by
f(x) =
sin
2
(10x)
(x + 1)
(A.1)
being x in the domain [0, 1] is studied as an example. This function has three maxima
and four minima. All minima are global, while only one maximum is a global
extremum as shown in Figure A.2. In order to solve the problem, the fundamental
parameters for the GA were chosen as follows: the population size M = 30, the
length of chromosome n
b
= 18 bits (which gives a resolution of 3.8 10
−6
), the
probability of crossover p
c
= 1, the probability of mutation p
m
= 0.03, and the
maximum number of generations G = 100.
The evolution of the GA algorithm for this function is presented in Figures
A.2(a), (b) and (c). In the figures, the symbol “∗” represent the individuals of
the population plotted in the function domain. Figure A.2(a) illustrates the initial
population randomly distributed. These are the candidate solutions to the problem.
After 20 generations the evolution achieved by natural selection, crossover and mu-
tation tends to eliminate the individuals with less fitness values. Due to crossover,
the best individuals tend to concentrate at the peaks of the function as shown in
Figure A.2(b), but it is mutation which prevents the population from being trapped
in local extrema. The final results at G = 100 are shown in Figure A.2(c) where it
can be observed that most of the population is clustered near the maximum value
118
Evaluate fitness of each
individual in population
+ 1 G = G
Create initial
random population
Designate result
End
Yes
Evaluate fitness of each
individual in population
G = 0
Set initial parameters
No
Termination
criterion satisfied ?
or G ≥ Gmax ?
Select M individuals
based on fitness
j
= 1
Select two individuals
Insert two offspring
into new population
j = j + 2
Perform crossover
with probability Pc
No
j M ≤ ?
Select one individual
Insert individual into
new population
j = j + 1
Perform mutation
with probability Pm
No
j M ≤ ?
j
= 1
Yes
Yes
Figure A.1. Flow chart for the genetic algorithm.
119
of the function. Several runs were performed, the average value of the independent
variable being x = 0.1526 and the function f(x) = 0.86458 which is very close to
the maximum. It is to be noted that the values of the parameters influence the
performance of the algorithm. Goldberg (1989) indicates that p
c
should be a value
close to 1 while p
m
should be chosen very small, e.g. 0.01.
A.2 Optimization of a function of two variables
Let us consider the “Peaks” function given by
f(x
1
, x
2
) = 3 (1 −x
1
)
2
e
−(x
2
1
+(x
2
+1)
2
)
−10
_
x
1
5
−x
3
1
−x
5
2
_
e
−(x
2
1
+x
2
2
)

1
3
e
−((x
1
+1)
2
+x
2
2
)
(A.2)
in the domain x
1
, x
2
∈ [−2.5, 2.5], which is shown in Figure A.3. This function has
several maxima and minima. However, there is only one global minimum within
the domain that is located at (x
1
, x
2
) = (0.0230, −1.6250). At this point, the value
of the function is −6.5511. Let us now employ GAs to find the global minimum of
this function. The parameters used to carry out the calculation are: population size
M = 30, length of chromosome for each independent variable 64 bits (so that the GA
uses a string of n
b
= 128 bits to represent both variables), probability of crossover
p
c
= 1, probability of mutation p
m
= 0.03, and the maximum number of generations
G = 100. The average results over ten runs are x
1
= 0.2282, x
2
= −1.6256, and
f(x
1
, x
2
) = −6.5511 which are very close the “exact” values. This indicates that
the GA is capable of finding the global minimum of this multimodal function.
A.3 Properties of crossover
The crossover operator in the genetic algorithm allows new individuals to be
created. Given a pair of numbers, called parents, which are represented in terms of
strings of zeros and ones, the crossover operator takes these parents and produces
120
x
(a)
x
(b)
0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1
0 0.2 0.4 0.6 0.8 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
x
f
(
x
)
f
(
x
)
f
(
x
)
(c)
Figure A.2. Genetic algorithm for function f(x) =
sin
2
(10x)
(x+1)
.
121
−2
−1
0
1
2
−2
−1
0
1
2
−10
−5
0
5
10
x
2
x
1
f
(
x
,
x
1
2
)
Figure A.3. Plot of function “Peaks.”
two offspring. The crossover process is shown schematically in Figure A.4. It is
to be noticed that the length of the vector corresponds to the number of bits n
b
used in the representation. The crossover point pc is the same for both parents and
bits corresponding to the rest of the strings are interchanged. The position of the
crossover point is given by the value of pc counted from left to right. Thus, from
Figure A.4 it can be seen that the crossover point is at the position 7 (pc = 7) while
the string-length is n
b
= 10.
In order to study the properties of one-position or one-point crossover in the
genetic algorithm, let us consider the variables x and y of a two-dimensional space.
These variables are encoded to produce the variables x
b
and y
b
in binary form using
n
b
bits, with n
b
∈ ^. The crossover operation can be seen as a map between the
122
Parents
1 1 0 1 1 0 1 0 1 0 0 1 0 1 1 1 1 1 0 1 Offspring
Crossover point: pc
1 1 0 1 1 0 1 1 0 1 0 1 0 1 1 1 1 0 1 0
String-length: n
b
Figure A.4. Crossover operation.
space represented by (x, y) to the space given by (x

, y

) when bits of x
b
and y
b
are
interchanged, as shown in Figure A.4, to get x

b
and y

b
which are decoded later to
generate the corresponding values of x

and y

. Therefore, in order to analyze all
possible maps between (x, y) and (x

, y

), the complete set of feasible values of y
b
for each probable value of x
b
has to be considered. From these, it will be possible
to obtain all the combinations between them. The total number of combinations is
a function of n
b
and it is given by N
cb
= 4
n
b
.
Table A.1 shows all the possible mappings of (x, y) → (x

, y

) with x
b
and y
b
being each represented using n
b
= 2 bits. It also illustrates the binary-values of
x

b
and y

b
after crossover between x
b
and y
b
is performed. Note that, in this case,
only one crossover point exists. The corresponding mappings given in Table A.1 are
shown graphically in Figure A.5. Such (x, y) → (x

, y

) mappings are joined by a
line. Here it can observed that sometimes the mapping is to the same point while
other times the mapping is to a different one. It is also seen that there is symmetry
with respect to both diagonals.
123
Table A.1. REPRESENTATION OF ALL POSSIBLE MAPPINGS OF (x, y) TO
(x

, y

) OBTAINED BY CROSSOVER USING n
b
= 2 BITS PER VARIABLE.
x
b
y
b
(x, y) x

b
y

b
(x

, y

)
0 0 0 0 (0, 0) 0 0 0 0 (0, 0)
0 0 0 1 (0, 1) 0 1 0 0 (1, 0)
0 0 1 0 (0, 2) 0 0 1 0 (0, 2)
0 0 1 1 (0, 3) 0 1 1 0 (1, 2)
0 1 0 0 (1, 0) 0 0 0 1 (0, 1)
0 1 0 1 (1, 1) 0 1 0 1 (1, 1)
0 1 1 0 (1, 2) 0 0 1 1 (0, 3)
0 1 1 1 (1, 3) 0 1 1 1 (1, 3)
1 0 0 0 (2, 0) 1 0 0 0 (2, 0)
1 0 0 1 (2, 1) 1 1 0 0 (3, 0)
1 0 1 0 (2, 2) 1 0 1 0 (2, 2)
1 0 1 1 (2, 3) 1 1 1 0 (3, 2)
1 1 0 0 (3, 0) 1 0 0 1 (2, 1)
1 1 0 1 (3, 1) 1 1 0 1 (3, 1)
1 1 1 0 (3, 2) 1 0 1 1 (2, 3)
1 1 1 1 (3, 3) 1 1 1 1 (3, 3)
Let us now consider the case when the representation of x and y is with n
b
= 3
bits per variable, so that the number of possible combinations of their binary values
is N
cb
= 64, and the corresponding number of all possible crossover positions is
N
cp
= 2. The value of N
cp
establishes the number of possible mapping sets which
may be generated by crossover. A mapping set is a map generated by fixing the
crossover point and performing the crossover operation on all combinations of the
parents. For N
cp
= 2 the possible mapping sets are also 2, one for each value of cp.
Figure A.6 shows the mapping set (x, y) → (x

, y

) when cp = 1. In this figure
four sub-mapping blocks and one avenue can be seen in each of the x- and y-
directions. An avenue appears when there is no connection between points of a
block to points of another block. On the other hand, the domain of the mapping set
is bounded in [0, 7] while the sub-domains have, each, a size of [3,3]. From Figure
124
A.6, the average of the Euclidean distance, l
s
, between the mapping points, (x, y)
and (x

, y

), can be easily computed. For the present case, l
s
is found to be 1.77.
The mapping set for cp = 2 is shown in Figure A.7. It is comprised of 16 sub-
mapping blocks and 3 avenues in each direction. The mapping set is also bounded in
[0, 7] but the sub-domains are now of size [1,1] each. The above provides an average
mapping distance between points l
s
= 0.71. It is clear from Figure A.6 that the
lower bound of the sub-domains are determined by the values of the elements of x
b
and y
b
that are located at the left of the crossover point once they have been decoded
when elements at the right of it have all zero values. The size of the sub-domains
is determined by the minimum and maximum values of the elements located at the
right of the crossover point when fixing those elements situated at the left of it.
In the case shown in Figure A.7, for cp = 2, the lower bounds are given by all
the combinatory values of the vector elements located at the left of the crossover
point. This property can be generalized to the case of n
b
bits and (n
b
−1) crossover
points which is shown for n
b
= 4 in Figures A.8–A.10. Therefore one can indicate
that, as the crossover point moves towards the left, the average mapping distance
between points, l
s
, increases generating sub-mapping blocks which are larger in size.
The number of sub-mapping blocks N
smb
generated in each mapping set depends
on the crossover point. Figures A.5, A.6 and A.8 are map sets for cp = 1 with n
b
= 2,
n
b
= 3 and n
b
= 4 respectively in which, for each set, the number of blocks is four.
In the case of cp = 2, Figure A.7 and Figure A.9 show sets of maps with 16 blocks;
while Figure A.10 presents the case for cp = 3 with 64 sub-maps. It can be observed
that the number of sub-mapping blocks follows the rule N
smb
= 4
cp
which indicates
all possible combinations of the values of the elements at the left of the crossover
point.
125
The number of avenues of separation in each direction N
ave
, which can be seen
in Figures A.5–A.10 as examples for n
b
= 2, n
b
= 3 and n
b
= 4 respectively, is also
a function of the crossover point. These are shown in tabular form in Table A.2 for
each value of the crossover point and different values of the string-length used. It is
seen that for cp = 1 the mapping set will always have one avenue, while for cp = 7
which is only valid if n
b
= 8 is used, N
ave
= 127.
Table A.2. NUMBER OF AVENUES N
ave
AS A FUNCTION OF THE NUMBER
OF BITS n
b
AND CROSSOVER VALUES cp.
String-length Number of Number of avenues for each cp-value; N
ave
(cp)
crossover points cp = 1 cp = 2 cp = 3 cp = (n
b
−1)
2 1 1
3 2 1 3
4 3 1 3 7
5 4 1 3 7
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
n
b
(n
b
−1) 2
0
2
1
+ 2
0
2
(cp−1)
+ + 2
1
+ 2
0
Taking into account all the mapping sets that can be generated with a given
representation using n
b
bits, it is possible to compute the mean value of the Eu-
clidean distance between points for all possible maps. This value is calculated in
the following way: Let X = (x, y) and X

= (x

, y

) in 1
2
, then
l
t
=
1
N
cb
N
cb

j=1
[[X
j
−X’
j
[[ (A.3)
where all values of x and y are normalized in the [0, 1] range, and N
cb
is the total
number of combinations between their binary representations, x
b
and y
b
, that gen-
erate x

b
and y

b
. Figure A.11 shows l
t
as a function of n
b
for n
b
= 1, ..., 9. From the
figure it is clear that as n
b
increases, the expected value of l
t
diminishes as l
t
∼ 1/n
b
,
126
which indicates that the use of n
b
+1 bits instead of n
b
bits as n
b
→ ∞ does not im-
prove the representation capability of a particular variable because the probability
of changing the bits in a particular location (or crossover point) is smaller. Although
the accuracy in the representation is increased by a factor of 2, when n
b
+1 instead
of n
b
as n
b
→ ∞ is considered, the size of a possible jump decreases by the same
factor.
The properties of the crossover operator can be summarized as follows:
• For each mapping set there is symmetry with respect to the main diagonals.
The same type of symmetry appears inside each sub-mapping block.
• As the position of crossover point cp moves towards the left, the average dis-
tance l
s
between the mapping points increases. This agrees with the fact
that, as the number of bits used to represent a particular variable increases,
the change in the value of such variable may become larger when the binary
number 0 or 1 of the first position, counted from left to right, changes. On
the other hand, under the same circumstances, if the change is in the binary
number of the last position smaller changes in the values of the variable will
occur.
• The number of avenues appearing in each mapping set is given by the sequence
N
ave
(cp) =
cp

j=1
2
(j−1)
for cp = 1, ..., (n
b
−1) (A.4)
where N
ave
(cp) is the number of avenues for a particular value of cp.
• As the number of bits n
b
increases the mean value of the Euclidean distance
of all possible maps l
t
decreases at a rate given by
l
t

1
n
b
. (A.5)
127
−1 0 1 2 3 4
−1
−0.5
0
0.5
1
1.5
2
2.5
3
3.5
4
x, x’
y
,

y

Figure A.5. Mapping in a two-dimensional space using a representation of string-
length of 2.
−1 0 1 2 3 4 5 6 7 8
−1
0
1
2
3
4
5
6
7
8
x, x’
y
,

y

0
0
1
1
0
1
0
1 y =
x = b
b
Figure A.6. Mapping in a two-dimensional space using a representation of string-
length of 3 for cp = 1.
128
−1 0 1 2 3 4 5 6 7 8
−1
0
1
2
3
4
5
6
7
8
x, x’
y
,

y


y =
x = b
b
Figure A.7. Mapping in a two-dimensional space using a representation of string-
length of 3 for cp = 2.
0 2 4 6 8 10 12 14 16
0
2
4
6
8
10
12
14
16
x, x’
y
,

y

Figure A.8. Mapping in a two-dimensional space using a representation of string-
length of 4 for cp = 1.
129
0 2 4 6 8 10 12 14 16
0
2
4
6
8
10
12
14
16
x, x’
y
,

y

Figure A.9. Mapping in a two-dimensional space using a representation of string-
length of 4 for cp = 2.
0 2 4 6 8 10 12 14 16
0
2
4
6
8
10
12
14
16
x, x’
y
,

y

Figure A.10. Mapping in a two-dimensional space using a representation of string-
length of 4 for cp = 3.
130
10
0
10
1
10
−2
10
−1
10
0
Slope = -1
l
t
n
b
Figure A.11. Mean distance l
t
as a function of the number of bits n
b
.
131
APPENDIX B
SIMULATED ANNEALING
The present section uses the algorithm of the simulated annealing technique de-
veloped in Corana et al. (1987), which is schematically shown in Figure B.1. In the
figure, G indicates the number of the current iteration. The numerical implementa-
tion was made in MATLAB. An outline of the procedure is given next.
Assume, for example, that it is desired to find the minimum of a function f(x)
in a given domain a ≤ x ≤ b. In outline the steps of the procedure are the following.
• First, the temperature T

∈ [0, ∞] is initialized to a high value T

0
, an initial
point x
0
∈ [a, b] is randomly generated and the function f(x
0
) is evaluated.
• Then, a sequence of candidate points are generated around the current point
x
0
by applying random moves uniformly distributed in intervals. The size of
these intervals is chosen so that the candidate point lies inside the domain
[a, b].
• The candidate point is accepted if it produces an improvement in the min-
imization of the function or may be accepted according to the Metropolis
criteria. In mathematical terms this is written as
p [f(x)] =
_
¸
¸
_
¸
¸
_
1.0 f(x) < f(x
0
)
e

_
f(x)−f(x
0
)
T

_
f(x) ≥ f(x
0
)
(B.1)
where p is the probability of the candidate solution for being accepted.
132



G = G + 1
G = 0
Set initial parameters
Randomly generate starting point
Designate result
End
Termination
criterion satisfied ?
Perform a cycle of random moves, each along
one coordinate direction. Accept or reject each
point according to the Metrololis criterion.
Record the optimum reached so far.
Adjust step vector V.
j = 0
j ≥ Ns
Reduce temperature.
Set current point to the optimum.
m = 0
m ≥ Nt
No
Yes
No
No
Yes
Yes
Figure B.1. Flow chart for the simulated annealing.
133
• The process is continued until a sort of “equilibrium” is approached; that is,
when no improvement in the minimization of f is achieved. After this “ther-
mal equilibration,” the temperature-like parameter T

is reduced employing
a prescribed reduction schedule and a new sequence of moves is performed
starting from the optimum value found so far.
• The process described above is continued for each temperature-like parameter
until its value is low enough that no more useful improvement can be expected.
B.1 Optimization of a function of one variable
Let us now consider the application of SA to a two-dimensional problem to show
how the technique works for finding the global optimum. As an example, the same
function of one independent variable used before for the GA case (Section A.1) given
by Eq. (A.1) is now analyzed.
After some initial trials, the chosen parameters for the SA were: the initial tem-
perature T

0
= 50, the temperature reduction factor used in the reduction schedule
r
t
= 0.85, the tolerance = 110
−6
, the maximum number of iterations N
max
= 100,
the number of operations before temperature adjustment N
t
= 30, and the number
of cycles for a given temperature N
s
= 30. With these values of the parameters
one achieves a good schedule in the temperature reduction and, therefore, a good
chance to find the global minimum.
Figure B.2 shows a plot of the function along with the places in the function
domain visited during the process (65 iterations). During the first stages of the
run, shown as (a), (b) and (c) in Figure B.2, the algorithm performs random jumps
accepting almost every function evaluation since T

is initially very high and p → 1
as can be seen from Eq. (B.1). After some temperature reductions have been done,
SA still accepts some movements in the opposite direction of optimization (d) and
134
a very few short jumps as the size of the intervals are reduced. Finally, when the
temperature is reduced to a very low value the algorithm rejects most of the moves,
accepting only moves in the direction of the global optimum (e) until reaching
it. When the algorithm finally converges, it gives as a result x = 0.15275 and
f(x) = 0.86586 which are the actual values of the independent variable and the
function obtained analytically and very close to those obtained by the GA in a
previous section. The number of function evaluations was 2, 114.
0 0.2 0.4 0.6 0.8 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
x
f
(
x
)
(a)
(b)
(c)
(d)
(e)
Figure B.2. Simulated Annealing for function f(x) =
sin
2
(10x)
(x+1)
. Jumps are marked
as (a), (b), (c), (d), (e).
135
B.2 Optimization of a function of two variables
The “Peaks” function given by Eq. (A.2) has also been considered here. Again,
the SA technique is used to find the global minimum of this function. The selected
parameters were: T

0
= 50, r
t
= 0.85, = 1 10
−6
, N
max
= 100, N
t
= 30, and
N
s
= 20. Ten runs were performed and the global minimum was always obtained.
Such minimum corrrespond to x
1
= 0.2282, x
2
= −1.6255 and f(x
1
, x
2
) = −6.5511,
which is the same obtained by means of the genetic algorithm. The simulated
annealing performed 11, 473 function evaluations and 96 iterations.
136
APPENDIX C
INTERVAL ANALYSIS
C.1 Description
Interval methods are deterministic techniques based on interval analysis that are
capable of solving global optimization problems with mathematical and computa-
tional certainty. Interval analysis is discussed in detail by Moore (1966), Hansen
(1992), Kearfott (1996), and Neumaier (1990). The idea of bounding round-off er-
rors during real number computations generated the interval analysis (Moore, 1966),
which relies on the interval mathematics methodology. Interval mathematics is a
generalization in which interval numbers replace real numbers, interval arithmetic
replaces real arithmetic, and interval analysis replaces real analysis. Introduced in
its modern form by Moore (1966), real interval arithmetic is based on arithmetic
within the set of closed intervals of real numbers. An interval X is a closed bounded
set of real numbers
X = [x
L
, x
U
] = ¦x ∈ 1 [ x
L
≤ x ≤ x
U
¦, x
L
< x
U
(C.1)
where x
L
≤ x
U
, and x
L
and x
U
are its lower and an upper bounds respectively.
In order to understand the interval method (IM) some background is provided
next. For clarity, only a one-dimensional approach will be used. An extension to
multiple dimensions is given in Section 4.2.3. In what follows, upper-case symbols
denote intervals while lower-case quantities are real numbers. The superscripts
L and U will indicate the lower and the upper bound of an interval respectively,
137
and a hat will designate its middle point. It is to be noted that a real number x is
equivalent to a degenerate interval, i.e. [x, x]. Hence, when expressing a real number
as an interval, for purposes of applying interval operations, the simpler noninterval
notation will be retained.
If X = [x
L
, x
U
] and Y = [y
L
, y
U
] then the elementary interval arithmetic opera-
tions are
X + Y = [x
L
+ y
L
, x
U
+ y
U
]
X −Y = [x
L
−y
U
, x
U
−y
L
]
X Y =
_
min¦x
L
y
L
, x
L
y
U
, x
U
y
L
, x
U
y
U
¦, max¦x
L
y
L
, x
L
y
U
, x
U
y
L
, x
U
y
U
¦
¸
1
Y
=
_
1
y
U
,
1
y
L
_
, 0 / ∈ [y
L
, y
U
]
X/Y = X
1
Y
.
The midpoint of an interval X is defined as
ˆ x = m(X) =
x
L
+ x
U
2
(C.2)
while its width is given by
w(X) = x
U
−x
L
. (C.3)
Other arithmetic definitions are presented in Hansen (1992), and Schnepper and
Stadtherr (1996).
Real-valued functions of intervals can be also evaluated. That is, if f(x) is
evaluated using interval arithmetic in the interval X = [x
L
, x
U
], the interval valued
function is given by f
I
= f([x
L
, x
U
]). Although most of the time the range of the
function is computed directly using interval analysis, sometimes it is convenient to
use a Taylor series to compute the function over the interval in order to get sharper
bounds on its range. Thus f(X) can be expanded in Taylor series about x as
f
I
= f(X) = f(x) + (X −x)f

(x) +
(X −x)
2
2
f

(X) + (C.4)
138
where the operation (X −x) is defined as
X −x = [x
L
, x
U
] −[x, x] = [x
L
−x, x
U
−x] (C.5)
and here x is generally taken, although not always, as the midpoint of X.
In general the implementation of interval arithmetic on a computer is needed
since the methods based on it are computationally intensive. There are several
methods based on interval analysis which have been used to solve both optimization
problems and systems of nonlinear equations, among them interval Newton methods
and Gauss-Seidel interval procedures which are discussed in Hansen (1992), Kearfott
(1996), and Neumaier (1990).
The interval Newton method (INM) was derived by Moore (1966) using the mean
value theorem and it is an excellent method for finding all the zeros of a continuously
differentiable function. If X is an interval containing both a point x and a zero x

of a function f(x) where f(x

) = 0, then x

∈ N(x, X) where
N(x, X) = x −
f(x)
f

(X)
, 0 / ∈ f

(X) (C.6)
and N(x, X) is a finite interval. Since any zero of f in X is also in N(x, X), then
it should be in the intersection X ∩N(x, X). For the case when 0 ∈ f

(X), Hansen
(1979), and Hansen (1992) give extended interval arithmetic rules.
Thus at any iteration k, the algorithm is
ˆ x
k
= m(X
k
) (C.7)
N(ˆ x
k
, X
k
) = ˆ x
k

f(ˆ x
k
)
f

(X
k
)
(C.8)
X
k+1
= X
k
∩ N(ˆ x
k
, X
k
) (C.9)
where Eqs. (C.8) and (C.9) are key steps of the method. Mathematical proofs of
the existence and uniqueness of a solution are provided in the references (Neumaier,
139
1990; Hansen, 1992; Kearfott, 1996). It is to be noted that the first derivative must
be evaluated using interval arithmetic as well as other operations such as subtraction,
division and intersection. In the context of optimization, the INM is used to find
the roots of the gradient of the objective function. Figure C.1 shows a flow chart of
the algorithm presented in Hansen (1979), which uses the interval Newton method
to find the global minimum of a function of one variable. In the flow chart, w
X
and
w
f
are the widths of X and f(X), while
X
and
f
are tolerances on them that the
user must set.
Assume, for example, that one wants to find the minimum of a function f(x),
which at least has to be twice differentiable, in a given domain a ≤ x ≤ b. In outline
the steps of the procedure given by Hansen (1979) are the following.
• First, the initial interval X
0
is taken from the list L of intervals which need
to be analyzed and the function f(X
0
) is evaluated. Also, for this initial
iteration, the best function so far is taken as f

= min[f(x
L
), f(x
U
)], where
x
L
and x
U
are the lower and upper bounds of X
0
. In subsequent steps, when
updating f

, f

= max [f(X)] if max [f(X)] < f

.
• Then the interval is divided into subintervals. Subintervals for which f(x) ≥
f

, f

(x) ,= 0 or f

(x) < 0, for all x ∈ X, are excluded since they cannot have
the global minimum. All these conditions are checked using interval values of
f, f

and f

. Subintervals that are candidates for having global minima are
considered and put on the list L.
• Use the interval Newton method to exclude subintervals of X, and find the
zeros of f

where possible global minima may exist.
• The procedure is continued until all the intervals in L have been analyzed and
either accepted or rejected as a solution.
140



L = [ ] ?
Set initial parameters
Set initial interval
Termination
criterion satisfied ?
Get current interval X from L and process it
Attempt to improve
interval X
Delete part of interval X
Such that f(y) > f*, where y ∈ X
No
Designate result
End
Yes No
Yes
f* < f(X)
No
Yes
No
Yes
w (X) > ε and w (X) > ε
X X f f
No
f’’(X) < 0
Apply interval Newton method
to find zeros of f’(X) and or
eliminate some intervals
Discard X
Discard X
If new Xs
then add
them to L
Yes
Figure C.1. Flow chart for the interval method.
141
• The solution interval is determined by achieving a tolerance in the minimum
function value found.
It is to be noted that such algorithm is based upon the branch and bound prin-
ciple. This means that the whole domain X is not searched uniformly, but some
parts (branches) are preferred. The interval method was implemented by a program
in MATLAB.
C.2 Optimization of a function of one variable
Let us now consider the application of IM to a function of only one variable
in order to illustrate how the technique works for finding the global optimum of
multimodal functions. As an example, the same function, namely Eq. (A.1) used
before in both GA and SA will be used. Its first derivative is given by
f

(x) = −
sin(10x)
x + 1
_
20 cos(10x) −
sin(10x)
x + 1
_
(C.10)
while the second derivative is
f

(x) =
200
(x + 1)
2
_
2 sin
2
(10x) −1
¸
+
2 sin
2
(10x)
(x + 1)
2
_
20 cos(10x) −
sin(10x)
x + 1
_
. (C.11)
The characteristics of the function, first derivative and second derivative are
plotted in Figure C.2. As can be noticed from the figure, within the domain (0, 1),
the function f(x) has four minima and three maxima. The first derivative, f

(x),
has seven roots, four maxima and three minima. Furthermore f

(x) is also multi-
modal, being negative in some parts of the domain and positive in others. Note that
although the interval method looks for the global minimum, within the context of
optimization, maximizing f(x) is equivalent to minimizing −f(x).
The initial interval, which comprises the whole domain, is chosen to be X
0
=
[0, 1]. The tolerance given for the width of the intervals in x is
X
= 1 10
−4
, while
142
0 0.2 0.4 0.6 0.8 1
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
x
f (x)
f ’’(x)
100
f ’(x)
50
Figure C.2. Plot of function f(x) =
sin
2
(10x)
(x+1)
, and its first and second derivatives.
that of the function is
f
= 1 10
−6
. The tolerance in the width of the intervals
in the Newton method iteration was set to
X
k = 1 10
−3
. Figure C.3 displays the
evolution of the process to find the maximum of f(x) in terms of how the intervals
are either split, reduced or rejected. In a) the initial interval X
0
=[0,1] is considered
first. Later, b), this interval is split in two since no improvement can be obtained;
one of them is put in the list of intervals to be checked. In c) the interval on the left
side was then split, and later part of it was rejected due to the monotonicity test. A
similar procedure as in c) is done in d) for the left-side interval of the remaining two.
In e) the left-side interval has been rejected due to the convexity test and the interval
at the right extremum is then considered. The latter interval is rejected in f) after
applying the objective-function test. The process is continued until the interval that
143
contains the global optimum is the only one remaining in the list. This interval is
then reduced until the specified tolerance is satisfied. Thus, the interval that encloses
the global maximum is found to be [0.15274487654700, 0.15274487654751] while the
corresponding enclosure of the function is [0.86586562719408, 0.86586562719485].
By taking the mid-point of X and f(X), the corresponding point values are x =
0.15274487654726 and f(x) = 0.86586562719447.
C.3 Optimization of a function of two variables
As a second example let us consider finding the global minimum of the so called
“Three-hump camel” function discussed by Hansen (1992), which is given by
f(x
1
, x
2
) = 2x
2
1
−1.05x
4
1
+
1
6
x
6
1
−x
1
x
2
+ x
2
2
(C.12)
whose gradient components g
1
and g
2
are given by
g
1
(x
1
, x
2
) = 4x
1
−4.2x
3
1
+ x
5
1
−x
2
(C.13)
g
2
(x
1
, x
2
) = 2x
2
−x
1
(C.14)
and the Hessian matrix is
h(x
1
, x
2
) =
_
_
_
4 −12.6x
2
1
+ 5x
4
1
−1
−1 2
_
_
_
. (C.15)
Note that in multiple dimensions, it is necessary to have information about the
gradient and the Hessian matrix if the interval method is going to be used. The
function, shown in Figure C.4, has several extrema with a global minimum at x
1
=
x
2
= 0. The initial interval is considered for both x
1
and x
2
to be [-2,2]. The
tolerances are set to
X
= 10
−4
and
F
= 10
−6
for the width of the interval and
the range of the function respectively. After the process is carried out, the final
result given by the method is x
1
= [−0.03630691361532, 0.47254377188262] 10
−6
144
0 0.2 0.4 0.6 0.8 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
x
f
(
x
)
a)
b)
c)
d)
e)
f)
g)
h)
i)
j)
k)
Figure C.3. Interval Newton method for function f(x) =
sin
2
(10x)
(x+1)
.
145
and x
2
= [−0.01817423098431, 0.23627188594152] 10
−6
with the analytic result
being x
1
= x
2
= 0. The above demonstrates that the interval method is capable
of enclosing the global solution. Note that it could be possible to choose a more
restrictive tolerance. Thus, the method would provide a sharper enclosure of the
solution.
−2
−1
0
1
2 −2
−1
0
1
2
0
2
4
6
8
10
x
2
x
1
f
(
x
,
x
1
2
)
Figure C.4. Plot of function “Three-hump camel”.
146
APPENDIX D
EXTRAPOLATION IN NEURAL NETWORKS
In the current section two numerical tests are presented to show the lack of good
extrapolation capabilities offered by artificial neural networks (ANNs). In the first
case, data from a two-dimensional nonlinear function will be generated to train and
test an ANN, while in the second such analysis will be done using the experimental
data of Sieder and Tate (1936).
D.1 Neural network analysis of a two-dimensional function
Let us consider the two-dimensional function
f(x) = 1 + 2x −3x
2
+ e
x
−cos(2πx) (D.1)
within the domain x ∈ [−2.5, 2.5]. This nonlinear function is shown graphically
using a solid line in Figure D.1. In order to test the accuracy of estimations from
neural networks outside the region of training, 21 data sets were generated in the
region of x ∈ [−1, 1], and 30 more within the remaining domain. The ANN archi-
tecture was chosen to be a 1-5-3-1. All the procedures for training followed here,
e.g. data normalization, selection of number of cycles, etc., are the same as those
used in Chapters 6 and 7. The network was trained using the data obtained within
the domain [−1, 1] and tested in the complete range, i.e. x ∈ [−2.5, 2.5]. The cor-
responding results are illustrated in Figure D.1. Also shown are the results from
a polynomial of degree six that was fitted to the same data. From the figure it
147
can be seen that inside the region of available training data, both the ANN and the
polynomial follow the curve accurately. Their rms errors with respect to the original
data are 13.5% and 29.9% respectively. Outside the range [−1, 1], however, both the
ANN and the polynomial provide wrong answers. While the polynomial predictions
blow up immediately outside the interpolation region, the ANN estimations tend to
level off at some values that are independent of x.
−2.5 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5
−8
−6
−4
−2
0
2
4
6
8
x
f
(
x
)
Figure D.1. Neural network and polynomial approximations of function f(x).
− f(x); ∗ ANN; – – polynomial
148
D.2 Neural network analysis of the Sieder and Tate (1936) data
As a second example, let us now consider the heat exchanger experimental data
developed by Sieder and Tate (1936). These data and the corresponding correlation
are widely used to represent the heat transfer coefficients in laminar flow in a pipe.
A total of 67 experimental runs, using three different oils as working fluids, were
reported by Sieder and Tate (1936). These included the Nusselt Nu, Reynolds
Re, and Prandtl Pr numbers, and the viscosity ratio µ/µ
wall
. Here, µ and µ
wall
are the fluid dynamic viscosities calculated at the average and wall temperatures
respectively. The tube diameter D, and its length L were also given.
The correlation developed by Sieder and Tate (1936) is
Nu = 1.86Re
1/3
D
Pr
1/3
_
µ
µ
wall
_
0.14
_
D
L
_
1/3
. (D.2)
In Eq. (D.2), the Nusselt, Reynolds and Prandtl numbers are defined as
Nu =
hD
k
, Re
D
=
DG
µ
, Pr =
c
p
µ
k
(D.3)
where h is the heat transfer coefficient, G is the mass velocity, k is the thermal
conductivity, and c
p
is the specific heat of the fluids used. The corresponding range
of applicability is given as
3.63 < Re
D
< 2110
151 < Pr < 16700
0.0042 <
_
µ
µ
wall
_
< 9.75 . (D.4)
Figure D.2 displays the Sieder and Tate (1936) data in the parameter space. In
the figure, the lines projecting the data to the Re
D
-Pr plane are provided for clarity.
Also shown is the range of applicability of their correlation for each parameter. This
range, given in Eq. (D.4), corresponds to a box in the parameter space. As Figure
D.2 shows, the data do not cover uniformly the complete range in the parameter
149
space but only a small region of it. It is seen that most of the experiments were con-
ducted for small values of µ/µ
wall
and only for particular values of the parameters;
sometimes only one. On the other hand, a more restrictive range of applicability
could be given by the convex hull of the data.
0
500
1000
1500
2000
2500
0
5000
10000
15000
20000
0
2
4
6
8
10
Re
D
Pr
µ

/

µ
w
a
l
l
Figure D.2. Region in parameter space of the Sieder and Tate (1936) experimental
data.
In order to investigate the extrapolation capability of the ANN, the measure-
ments of Re
D
, Pr, µ/µ
wall
, and Nu were used to train the network. The network
configuration is a 3-5-5-1. The number of training cycles was chosen to be 150,000.
The testing was carried out inside and outside the region (cloud) where experiments
were conducted, but always within the box. The Sieder and Tate (1936) correlation
150
was also used for comparison purposes. Thus, the prediction of Nu by means of
correlations and artificial neural networks is
Nu = Nu(Re
D
, Pr, µ/µ
wall
).
0 50 100 150 200
0
10
20
30
40
50
60
Nu
c
N
u
A
N
N
Figure D.3. Comparison in Nu-estimations between neural networks and the Sieder
and Tate (1936) correlation. ◦ Estimations inside the cloud of experimental data;
+ Estimations outside the cloud of experimental data.
A qualitative comparison between the Nu-predictions given by the neural net-
work and the correlation is presented in Figure D.3. The solid line indicates the
equality between their estimations. The symbol “◦” denotes estimations within the
151
region where experiments were performed whereas the symbol “+” corresponds to
predictions outside such region. It is to be noted that, inside the cloud, the pre-
dictions between the ANN and the correlation are very close to each other. The
rms errors between predictions and experiments are 5% for the ANN, and 13.8%
for the correlation. Outside of the cloud, their estimations are very different. As
in the previous example, most of the ANN estimations level off regardless of the
values of the independent parameters. On the other hand, the correlation follows
a certain trend in which Nu increases with Re
D
, Pr, µ/µ
wall
. The examples given
above demonstrate that the neural network does not have good capabilities for ex-
trapolation. Therefore, it is extremely important to both determine the range of
applicability of the ANN and use the neural network within such region in order to
avoid extrapolation.
152
BIBLIOGRAPHY
Aarts, E.H.L., and Korst, J.H.M., 1989, Simulated Annealing and Boltzman Ma-
chines, Wiley, Tiptree Essex, Great Britian.
Abdel-Magid, Y.L., and Dawoud, M.M., 1995, “Tuning of AGC of interconnected
reheat thermal systems with genetic algorithms,” in: Proceedings of the IEEE In-
ternational Conference on Systems, Man and Cybernetics, Vol. 3, pp. 2622–2627.
Abu Madi, M., Johns, R.A., and Heikal, M.R., 1998, “Performance characteristics
correlation for round tube and plate finned heat exchangers,” International Jour-
nal of Refrigeration-Revue Internationale du Froid, Vol. 21, No. 7, pp. 507–517.
Alvarez, E., Correa, J.M., Riverol, C., and Navaza, J.M., 2000, “Model based in neu-
ral networks for the prediction of the mass transfer coefficients in bubble columns.
Study in newtonian and non-newtonian fluids,” International Communications
of Heat and Mass Transfer, Vol. 27, No. 1, pp. 93–98.
Anderberg, M.R., 1973, Cluster Analysis for Applications, Academic Press, New
York, NY.
Bastani, A., Fiebig, M., and Mitra, N.K., 1992, “Numerical studies of a compact
fin-tube heat exchanger,” in: Design and Operation of Heat Exchangers, Roetzel,
W., Heggs, P.J., and Butterworth, D., (Eds.), Springer-Verlag, Berlin, pp. 154–
163.
Beecher, D., and Fagan, T., 1987,“Effects of fin pattern on the air-side heat transfer
coefficient in plate finned-tube heat exchangers,” in: ASHRAE Transactions, Vol.
93, part 2, pp. 1961–1985.
Benelmir, R., Khalfi, M., and Feidt, M., 1997, “Identification of the heat transfer
coefficient over the external area of a finned tubes heat exchanger with respect to
the moisture content of the air without condensation,” Revue G´enerale de Ther-
mique, Vol. 36, No. 4, pp. 289–301.
Bonabeau, E., Dorigo, M., and Th´eraulaz, G., 1999, Swarm Intelligence: From Nat-
ural to Artificial Systems, Oxford University Press, England.
Bowman, R.A., Mueller, A.C., and Nagle, W.M., 1940, “Mean temperature differ-
ence in design,” Transactions of the ASME, Vol. 62, pp. 283–294.
Briggs, D.E., and Young, E.H., 1969, “Modified Wilson plot techniques for obtaining
heat transfer correlations for shell and tube heat exchangers,” in: Chemical En-
gineering Progress Symposium Series, Bandy, W.R., and Martini, W.R., (Eds.),
AIChE, New York, NY., pp. 35–45.
153
Campbell, M.I., Amon, C.H., and Cagan, J., 1997, “Optimal three-dimensional
placement of heat generating electronic components,” Journal of Electronic Pack-
aging, Vol. 119, No. 2, pp. 106–113.
Castellano, G., Fanelli, A.M., and Pelillo, M., 1997, “An iterative pruning algorithm
for feedforward neural networks,” IEEE Transactions on Neural Networks, Vol.
8, No. 3, pp. 519–531.
Castrogiovanni, A., and Sforza, P.M., 1997, “A genetic algorithm model for high
heat flux flow boiling,” Experimental Thermal and Fluid Science, Vol. 15, No. 3,
pp. 193–201.
Ceranic, B., Fryer, C., and Baines, R.W., 2001, “An application of simulated anneal-
ing to the optimum design of reinforced concrete retaining structures,” Computers
& Structures, Vol. 79, No. 17, pp. 1569–1581.
˘
Cerny, V., 1985, “Thermodynamical Approach to the Traveling Salesman Problem:
An Efficient Simulation Algorithm,” Journal of Optimization Theory and Appli-
cations, Vol. 45, pp. 41–51.
Chang, Y.-P., Tsai, R., and Hwang, J.-W., 1997, “Condensing heat transfer charac-
teristics of aluminum flat tube,” Applied Thermal Engineering, Vol. 17, No. 11,
pp. 1055–1065.
Chatterjee, S., and Laudatto, M., 1997, “Genetic algorithms in statistics: procedures
and applications,” Communications in Statistics-Simulation and Computation,
Vol. 26, No. 4, pp. 1617–1630.
Chaudhuri, P.D., Diwekar, U.M., and Logsdon, J.S., 1997, “An automated approach
for the optimal design of heat exchangers,” Industrial and Engineering Chemistry
Research, Vol. 36 No. 9, pp. 3685–3693.
Chinnam, R.B., and Ding, J., 1998, “Prediction limit estimation for neural network
models,” IEEE Transactions on Neural Networks, Vol. 9, No. 6, pp. 1515–1522.
Chryssolouris, G., Moshin, L., and Ramsey, A., 1996, “Confidence interval predic-
tion for neural network models,” IEEE Transactions on Neural Networks, Vol. 7,
No. 1, pp. 229–232.
Corana, A., Marchesi, M., Martini, C., and Ridella, S., 1987, “Minimizing multi-
modal functions of continuous variables with the simulated annealing algorithm,”
ACM Transactions on Mathematical Software, Vol. 13, No. 3, 262–280.
Corberan, J.M., and Melon, M.G., 1998, “Modeling of plate finned tube evapora-
tors and condensers working with R134A,” International Journal of Refrigeration,
Vol. 21, No. 4, pp. 273–284.
Courrieu, P., 1994, “Three algorithms for estimating the domain of validity of feed-
forward neural networks,” Neural Networks, Vol. 7, No. 1, pp. 169–174.
Davalos, R.V., and Rubinsky, B., 1996, “An evolutionary-genetic approach to heat
transfer analysis,” Journal of Heat Transfer, Vol. 118, No. 3, pp. 528–531.
154
Davis, L.D., (Ed.), 1991, Handbook of Genetic Algorithms, Van Nostrand, New York.
Deb, K., 1999, “An Introduction to Genetic Algorithms,” S¯adhan¯ a: Academy Pro-
ceedings in Engineering Sciences, Vol. 24, Parts. 4–5, pp. 293–315.
D´ıaz, G., 2000, Simulation and Control of Heat Exchangers Using Artificial Neural
Networks, Ph.D. Thesis, Department of Aerospace and Mechanical Engineering,
University of Notre Dame, Notre Dame, Indiana.
D´ıaz, G., Sen, M., Yang, K.T., and McClain, R.L., 1999, “Simulation of heat
exchanger performance by artificial neural networks,” International Journal of
HVAC & R Research, Vol. 5, No. 3, pp. 195–208.
Fabbri, G., 1997, “A genetic algorithm for fin profile optimization,” International
Journal of Heat and Mass Transfer, Vol. 40, No. 9, pp. 2165–2172.
Femia, N., Spagnuolo, G., and Tucci, V., 1995, “Interval analysis in power electron-
ics,” Journal of Circuits Systems and Computers, Vol. 5, No. 3, pp. 317–336.
Floudas, C.A., 1999, “Recent advances in global optimization for process synthesis,
design and control: enclosure of all solutions,” Computers & Chemical Engineer-
ing, Vol. 23, suppl. s, pp. s963–s973.
Fourligkas, N., and Doumanidis, C., 1999, “Temperature field regulation in thermal
cutting for layered manufacturing,” ASME Journal of Manufacturing Science and
Engineering, Vol. 121, No. 3, pp. 440–447.
Fovell, R.G., 1997, “Consensus clustering of U.S. temperature and precipitation
data,” Journal of Climate, Vol. 10, No. 6, pp. 1405–1427.
Fowler, M.M., Klett, D.E., Moreno, J.B., and Heermann, P.D., 1997, “Using arti-
ficial neural networks to predict the performance of a liquid sodium reflux pool
boiler solar receiver,” in Proceedings of the International Solar Energy Confer-
ence, pp. 93–104.
Gau, C.-Y., Brennecke, J.F., and Stadtherr, M.A., 2000, “Reliable nonlinear pa-
rameter estimation in VLE modeling,” Fluid Phase Equilibria, Vol. 168, No. 1,
pp. 1–18.
Gobovic, D., and Zaghloul, M.E., 1993, “Design of Locally connected CMOS neural
cells to solve the steady-state heat flow problem,” in: Proceedings of the 36th
Midwest Symposium on Circuits and Systems, Vol. 1, pp. 755–757.
Goering, D.J., Humphrey, J.A.C., and Greif, R., 1997, “The dual influence of curva-
ture and buoyancy in fully developed tube flows,” International Journal of Heat
and Mass Transfer, Vol. 40, No. 9, pp. 2187–2199.
Goldberg, D.E., 1989, Genetic Algorithms in Search, Optimization, and Machine
Learning, Addison-Wesley, Reading, MA.
155
Gong, G., 1986, “Cross-validation, the jackinfe, and the bootstrap: Excess error
estimation in forward logistic regression,” Journal of the American Statistical
Association, Vol. 81, pp. 108–113.
Gonzalez-Monroy, L.I., and Cordoba, A., 2000, “Optimization of energy supply sys-
tems with simulated annealing, continuous and discrete descriptions,” Physica A,
Vol. 284, No. 1–4, pp. 433–447.
Gray, D.L., and Webb, R.L., 1986, “Heat transfer and friction correlations for plate
finned-tube heat exchangers having plain fins,” in: Proceedings of the Eighth In-
ternational Heat Transfer Conference, Vol. 6, pp. 2745–2750.
Haji-Sheikh A., Mashena, M., and Haji-Sheikh M.J., 1983, “Heat transfer coefficient
in ducts with constant wall temperature,” ASME Journal of Heat Transfer, Vol.
105, pp. 878–883.
Hansen, E.R., 1979, “Global optimization using interval analysis: the one-dimensional
case,” Journal of Optimization Theory and Applications, Vol. 29, No. 3, pp. 331–
344.
Hansen, E.R., 1980, “Global optimization using interval analysis: the multi-dimensional
case,” Numerische Mathematik, Vol. 34, pp. 247–270.
Hansen, E.R., 1992, Global Optimization using Interval Analysis, Marcel Dekkar,
New York, NY.
Hansen, E., and Sengupta, S., 1981, “Bounding solutions of systems of equations
using interval Analysis,” BIT, Vol. 21, No. 2, pp. 203–211.
Haykin, S., 1999, Neural Networks, A Comprehensive Foundation, 2nd Ed., Prentice-
Hall, Upper Saddle River.
Holland, J.H., 1975, Adaptation in Natural and Artificial Systems, University of
Michigan Press, Ann Arbor, MI.
Hua, J.Z., Brennecke, J.F., and Stadtherr, M.A., 1998, “Enhanced interval analysis
for phase stability: cubic equation of state models,” Industrial and Engineering
Chemistry Research, Vol. 37, No. 4, pp. 1519–1527.
Hussain, M.A., 1999, “Statistical Pattern Recognition: A Review,” IEEE Transac-
tions on Pattern Analysis and Machine Intelligence, Vol. 22, No. 1, pp. 4–37.
Idem, S.A., Jacobi, A.M., and Goldschmidt, V.W., 1990, “Heat transfer characteri-
zation of a finned-tube heat exchanger (with and without condensation),” ASME
Journal of Heat Transfer, Vol. 112 No. 1, pp. 64–70.
Ingber, L., 1993, “Simulated annealing: practice versus theory,” Journal of Mathe-
matical and Computational Modeling, Vol. 18 No. 11, pp. 29–57.
Ingber, L., 1996, “Statistical mechanics of nonlinear nonequilibrium financial mar-
kets: applications to optimized trading,” Mathematical and Computer Modeling,
Vol. 23, No. 7, pp. 101–121.
156
Jacobi, A.M., and Goldschmidt, V.W., 1990, “Low Reynolds number heat and mass
transfer measurements of an overall counterflow, baffled, finned-tube, condensing
heat exchanger,” International Journal of Heat and Mass Transfer, Vol. 33, No.
4, pp. 755–765.
Jain, A.K., Duin, R.P.W., and Mao, J.C., 2000, “Statistical pattern recognition: a
review,” IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol.
22, No. 1, pp. 4–37.
Jain, A.K., Murty, M.N., and Flynn, P.J., 1999, “Data clustering: a review,” ACM
Computing Surveys, Vol. 31, No. 3, pp. 264–323.
Jakob, M., 1957, Heat Transfer, Volume II, Wiley and Sons, New York, NY.
Jang, J.Y., Lai, J.T., and Liu, L.C., 1998, “The thermal-hydraulic characteristics
of staggered circular finned-tube heat exchangers under dry and dehumidifying
conditions,” International Journal of Heat and Mass Transfer, Vol. 41, No. 21,
pp. 3321–3337.
Jambunathan, K., Hartle, S.L., Ashforth-Frost, S., and Fontama, V.N., 1996, “Eval-
uating convective heat transfer coefficients using neural networks,” International
Journal of Heat and Mass Transfer, Vol. 39, No. 11, pp. 2329–2332.
Jeannette, E., Assawamartbunlue, K., Curtiss, P.S., and Kreider, J.F., 1998, “Ex-
perimental results of a predictive neural network HVAC controller,” in: ASHRAE
Transactions, Vol. 104, part 2, pp. 192–197.
Johnson J.M., and RahmatSamii Y., 1997,“Genetic algorithms in engineering elec-
tromagnetics,” IEEE Antennas and Propagation Magazine, Vol. 39, No. 4, pp.
7–25.
Kaka¸ c, S., Bergles, A.E., and Mayinger, F., (Eds.), 1981, Heat Exchangers: Thermal-
Hydraulic Fundamentals and Design, Hemisphere Publ. Corp., Washington, DC.
Kalogirou, S.A., 1999, “Applications of artificial neural networks in energy systems–
a review,” Energy Conversion and Management, Vol. 40, No. 10, pp. 1073–1087.
Kalogirou, S.A., Panteliou, S., and Dentsoras, A., 1999, “Artificial Neural Networks
used for the Performance Prediction of a Thermosyphon Solar Water Heater,”
Renewable Energy, Vol. 18, No. 1, pp. 87–99.
Kandlikar, S.G., 1990, “Thermal design theory for compact evaporators,” in: Com-
pact Heat Exchangers, Shah, R.K., Kraus, A.D., and Metzger, D., (Eds.), Hemi-
sphere Publishing, New York, NY, pp. 245–286.
Kandlikar, S.G., 1991, “A model for correlating flow boiling heat-transfer in aug-
mented tubes and compact evaporators,” ASME Journal of Heat Transfer, Vol.
113, No. 4, pp. 966–972.
Kaufman, L., and Rousseeuw, P.J., 1990, Finding Groups in Data: An Introduction
to Cluster Analysis, Wiley and Sons, New York, NY.
157
Kayansayan, N., 1994, “Heat transfer characterization of plate fin-tube heat ex-
changers,” International Journal of Refrigeration-Revue Internationale du Froid,
Vol. 17, No. 1, pp. 49–57.
Kays, W.M., and London, A.L., 1984, Compact Heat Exchangers, 3rd Ed., McGraw-
Hill, New York, NY.
Kearfott, R.B., 1996, Rigorous Global Search: Continuous Problems, Kluwer, Dor-
drecht, Netherlands.
Kearfott, R.B., and Kreinovich, V., 1996, Applications of Interval Computations,
Kluwer, Dordrecht, Netherlands.
Kedzierski, M.A., and Kim, M.S., 1996, “Single-phase heat transfer and pressure
drop characteristics of an integral-spine fin within an annulus,” Journal of En-
hanced Heat Transfer, Vol. 3, No. 3, pp. 201–210.
Khandelwal, D., and Kesavadas, T., 1999, “A computational approach to multi-
material solid freeform fabrication using simulated annealing,” Journal of Mate-
rials Processing & Manufacturing Science, Vol. 7, No. 4, pp. 405–415.
Khartabil, H.F., and Christensen, R.N., 1992, “An improved scheme for determin-
ing heat transfer correlations from heat exchanger regression models with three
unknowns,” Experimental Thermal and Fluid Science, Vol. 5, No. 6, pp. 808–819.
Kim, N.H., Youn, B., and Webb, R.L., 1998, “Heat transfer and friction correlations
for plain fin-and-tube heat exchangers,” in: Proceedings of the Eleventh Interna-
tional Heat Transfer Conference, Vol. 6, pp. 209–213.
Kirkpatrick, S., Gelatt, C.D., Jr., and Vecchi, M.P., 1983, “Optimization by Simu-
lated Annealing,” Science, Vol. 220 (4598), pp. 671–680.
Koza, J.R., 1992, Genetic Programming: On the Programming of Computers by
Means of Natural Selection , MIT-Press, Cambridge, MA.
Kramer, M.A., and Leonard, J.A., 1990, “Diagnosis using backpropagation neural
networks–analysis and criticism,” Computers and Chemical Engineering, Vol. 14,
No. 12, pp. 1323–1338.
Kulkarni, S.R., Lugosi, G., and Venkatesh, S.S., 1998, “Learning pattern classification–
a survey,” IEEE Transactions on Information Theory, Vol. 44, No. 6, pp. 2178-
2206.
Kumar, R., Varma, H.K., Agrawal, K.N., and Mohanty, B., 2001, “A comprehensive
study of modified Wilson plot technique to determine the heat transfer coefficient
during condensation of steam and R-134a over single horizontal plain and finned
tubes,” Heat Transfer Engineering, Vol. 22, No. 2, pp. 3–12.
Kuroe, Y., and Kimura, I., 1995, “Modeling of unsteady heat conduction field by
using composite recurrent neural networks,” in: Proceedings of the IEEE Inter-
national Conference on Neural Networks, Vol. 1, pp. 323–328.
158
Kushida, G., Yamashita, H., and Izumi, R., 1986, “Fluid flow and heat transfer in
a plate-fin and tube heat exchanger: analysis of heat transfer around a square
cylinder situated between parallel plates,” Bulletin of the JSME-Japan Society
of Mechanical Engineers, Vol. 29, No. 258, pp. 4185–4191.
Leonard, J.A., Kramer, M.A., and Ungar, L.H., 1992, “A neural network architec-
ture that computes its own reliability,” Computers and Chemical Engineering,
Vol. 16, No. 9, pp. 819–835.
Li, H.Y., and Yang, C.Y., 1997, “A genetic algorithm for inverse radiation prob-
lems,” International Journal of Heat and Mass Transfer, Vol. 40, No. 7, pp.
1545–1549.
Mahajan, R.L., and Wang, X.A., 1993, “Neural network models for thermally based
microelectronic manufacturing processes,” Journal of the Electrochemical Soci-
ety, Vol. 140, No. 8, pp. 2287–2293.
Maier, R.W., Brennecke, J.F., and Stadtherr, M.A., 2000, “Reliable computation
of reactive azeotropes,” Computers & Chemical Engineering, Vol. 24, No. 8, pp.
1851–1858
Maier, R.W., and Stadtherr, M.A., 2001, “Reliable density-functional-theory cal-
culations of adsorption in nanoscale pores,” AIChE Journal, Vol. 47, No. 8, pp.
1874–1884.
Marwah, M., Li, Y., and Mahajan, R.L., 1996, “Integrated neural network modeling
for electronic manufacturing,” Journal of Electronics Manufacturing, Vol. 6, No.
2, pp. 79–91.
Mavridou, T.D., and Pardalos, P.M., 1997, “Simulated annealing and genetic algo-
rithms for the facility layout problem: A survey,” Computational Optimization
and Applications, Vol. 7, No. 1, pp. 111–126.
Mazzola, A., 1997, “Integrating artificial neural networks and empirical correlations
for the prediction of water-subcooled critical heat flux,” Revue Generale de Ther-
mique, Vol. 36, No. 11, pp. 799-806.
McQuiston, F.C., 1976, “Heat, mass and momentum transfer in a parallel plate de-
humidifying exchanger,” in ASHRAE Transactions, Vol. 82, part 2, pp. 87–106.
McQuiston, F.C., 1978a, “Heat, mass and momentum transfer data for five plate-
fin-tube heat transfer surfaces,” in: ASHRAE Transactions, Vol. 84, part 1, pp.
266–293.
McQuiston, F.C., 1978b, “Correlation of heat, mass and momentum transport coeffi-
cients for plate-fin-tube heat transfer surfaces with staggered tubes,” in: ASHRAE
Transactions, Vol. 84, part 1, pp. 294–309.
Metropolis, N., Rosenbluth A.W., Rosenbluth M.N., Teller, A.H., and Teller, E.,
1953, “Equation of state calculations by fast computing machines,” Journal of
Chemical Physics, Vol. 21, No. 6, pp. 1087–1092.
159
Michalewicz, Z., 1992, Genetic Algorithms + Data Structures = Evolution Programs,
2nd. Extended Edition, Springer-Verlag, Berlin Heidelberg, Germany.
Mirth, D.R., and Ramadhyani, S., 1995, “Prediction of cooling-coil performance
under condensing conditions,” International Journal of Heat and Mass Transfer,
Vol. 14, No. 4, pp. 391–400.
Mistry, S.I., and Nair, S.S., 1993, “Nonlinear HVAC computations using artificial
neural networks,” in: ASHRAE Transactions, Vol. 99, Part 1, pp. 775–784.
Mitchell, M., 1996, Introduction to Genetic Algorithms, MIT-Press, Cambridge, MA.
Montgomery, D.C., 1991, Design and Analysis of Experiments, 2nd Ed., Wiley, New
York, NY.
Moon, S.K., and Chang, S.H, 1994, “Classification and Prediction of the Critical
Heat Flux Using Fuzzy Theory and Artificial Neural Networks,” Nuclear Engi-
neering and Design, Vol. 150, No. 1, pp. 151–161.
Moore, R.E., 1966, Interval Analysis, Prentice-Hall, Engelwood Cliffs, NJ.
Nakagiri, S., and Suzuki, K., 1997, “Interval estimation of eigenvalue problem based
on finite element sensitivity analysis and convex model,” JSME International
Journal Series A-Solid Mechanics and Material Engineering, Vol. 40, No. 3, pp.
228–233.
Neumaier, A., 1990, Interval Analysis for Systems of Equations, Cambridge Univer-
sity Press, Cambridge, MA.
Nilar, S.H., 1991, “Applications of the simulated annealing method to intermolecular
interactions,” Journal of Computational Chemistry, Vol. 12, No. 8, pp. 1008–1013.
Nirwan, A., and Hou, E., (Eds.), 1997, Computational intelligence for optimization,
Kluwer, Norwell, MA.
Niyogi, P., and Girosi, F., 1999, “Generalization bounds for function approximation
from scattered noisy data,” Advances in Computational Mathematics, Vol. 10,
No. 1, pp. 51–80.
Okamoto, K., Tsuru, D., and Fumizawa, M., 1996, “Reconstruction of three– di-
mensional density distribution from the limited projection images with genetic
algorithm,” Journal of Flow Visualization and Image Processing, Vol. 3, No. 4,
pp. 253–263.
Pacheco-Vega, A., Sen M., Yang, K.T., and McClain, R.L., 1998, “Genetic-algorithm-
based predictions of fin-tube heat exchanger performance,” in: Proceedings of the
Eleventh International Heat Transfer Conference, Vol. 6, pp. 137–142.
Pacheco-Vega, A., Sen, M., Yang, K.T., and McClain, R.L., 1999, “Prediction of
humid air heat exchanger performance using artificial neural networks,” in: Pro-
ceedings of the ASME Heat Transfer Division-1999, 1999 ASME International
Mechanical Engineering Congress and Exposition, HTD Vol. 364-3, pp. 307–314.
160
Pacheco-Vega, A., Sen, M., Yang, K.T., and McClain, R.L., 2001a, “Neural network
analysis of fin-tube refrigerating heat exchanger with limited experimental data,”
International Journal of Heat and Mass Transfer, Vol. 44, No. 4, pp. 763–770.
Pacheco-Vega, A., D´ıaz, G., Sen, M., Yang, K.T., and McClain, R.L., 2001b, “Heat
rate predictions in humid air-water heat exchangers using correlations and neural
networks,” ASME Journal Heat of Transfer, Vol. 123, No. 2, pp. 348-354.
Panchal, C.B., and Rabas, T.J., 1993, “Thermal performance of advance heat ex-
changers for ammonia refrigeration systems,” Heat Transfer Engineering, Vol. 14,
No. 4, pp. 42–57.
Prechelt, L., 1998, “Automatic early stopping using cross validation: quantifying
the criteria,” Neural Networks, Vol. 11, pp. 761–767.
Queipo, N., Devarakonda, R., and Humphrey, J.A.C., 1994, “Genetic algorithms for
thermosciences research: application to the optimized cooling of electronic com-
ponents,” International Journal of Heat and Mass Transfer, Vol. 37, No. 6, pp.
893–908.
Ramadhyani, S., 1998, “Calculation of air-side heat transfer in compact heat ex-
changers under condensing conditions,” in: Computer Simulations in Compact
Heat Exchangers, Vol. 1, Computational Mechanics Publications, Sunden, B.,
and Faghri, M., (Eds.), South Hampton, UK., pp. 151–168.
Randall, M., McMahon, G., and Sugden, S., 2002, “A simulated annealing approach
to communication network design,” Journal of Combinatorial optimization, Vol.
6, No. 1, pp. 55–65.
Ranganayakulu, C., and Seetharamu, K.N., 1999, “The combined effects of wall
longitudinal heat conduction, inlet fluid flow nonuniformity and temperature
nonuniformity in compact tube-fin heat exchangers: a finite element method,”
International Journal of Heat and Mass Transfer, Vol. 42, No. 2, pp. 263–273.
Ratschek, H., and Rokne, J., 1993, “Experiments using interval-analysis for solv-
ing a circuit-design problem,” Journal of Global Optimization, Vol. 3, No. 4, pp.
501–518.
Raudensky, M., Woodbury, K.A., and Brezina, T., 1995, “Genetic algorithm in so-
lution of inverse heat conduction problems,” Numerical Heat Transfer. Part B,
Fundamentals, Vol. 28, No. 3, pp. 293–306.
Refenes, A.N., (Ed.), 1995, “Neural Networks in the Capital Markets,” Wiley and
Sons, Chichester, England.
Rich, D.G., 1973, “The effect of fin spacing on the heat transfer and friction per-
formance of multi-row, smooth plate fin-and-tube heat exchangers,” in ASHRAE
Transactions, Vol. 79, part 2, pp. 137–145.
Rumelhart, D.E., Hinton, G.E., and Williams, R.J., 1986, “Learning Internal Rep-
resentations by Error Propagation,” in: Parallel Distributed Processing: Explo-
rations in the Microstructure of Cognition, MIT Press, Cambridge, MA.
161
Sahnoun, A., and Webb, R.L., 1992, “Prediction of heat transfer and friction for
the louver fin geometry,” ASME Journal of Heat Transfer, Vol. 114, No. 4, pp.
893–900.
Sastry, U.A., 1964, “Solution of the heat transfer of laminar forced-convection in
non-circular pipes,” Applied Science and Research, Vol. A13, pp. 269–280.
Sastry, U.A., 1965, “Heat transfer and laminar forced convection in multiple con-
nected cross-section,” Indian Journal of Pure and Applied Physics, Vol. 3, pp.
113–116.
Schmit, T.S., Dhingra, A.K., Landis, F., and Kojasoy, G., 1996, “A genetic algo-
rithm optimization technique for compact high intensity cooler design,” Journal
of Enhanced Heat Transfer, Vol. 3, No. 4, pp. 281–290.
Schnepper, C.A., and Stadtherr, M.A., 1996, “Robust Process Simulation using
Interval methods,” Computers and Chemical Engineering, Vol. 20, No. 2, pp.
187–199.
Sen, M., and Yang, K.T., 1999, “Applications of artificial neural networks and ge-
netic algorithms in thermal engineering,” CRC Handbook of Thermal Engineer-
ing, Kreith F., (Ed.), section 4.24 pp. 620–661.
Shah, R.K., 1985, “Compact heat exchangers,” Handbook of Heat Transfer Appli-
cations, Rohsenow, W.M., Hartnett, J.P., and Gani´c, E.N., (Ed.), second Ed.,
McGraw-Hill, NY, pp. 4-174–4-312.
Shah, R.K., 1990, “Assessment of modified Wilson plot techniques for obtaining heat
exchanger design data,” in: Proceedings of the Ninth International Heat Transfer
Conference, Vol. 5, pp. 51–56.
Shah, R.K., Kraus, A.D., and Metzger, D., (Eds.), 1990, Compact Heat Exchangers,
Hemisphere Publishing, New York, NY.
Shah, R.K., and London, A.L., 1978, “Laminar Forced Convection in Ducts,” in:
Advances in Heat Transfer, Suppl. 1, Academic Press, New York, NY.
Shah, R.K., and London, A.L., 1980, “Effects of Nonuniform Passages on Compact
Heat Exchanger Performance,” Journal of Engineering for Power–Transactions
of the ASME, Vol. 102, No. 3, pp. 653–659.
Shao, R., Martin, E.B., Zhang, J., and Morris, A.J., 1997, “Confidence bounds for
neural network representations,” Computers and Chemical Engineering, Vol. 21,
suppl. s, pp. s1173–s1178.
Sieder, E.N., and Tate, G.E., 1936, “Heat transfer and pressure drop of liquids in
tubes,” Industrial and Engineering Chemistry, Vol. 28, No. 12, pp. 1429–1435.
Smith, K.A., 1999, “Neural networks for combinatorial optimization: a review of
more than a decade of research,” Informs Journal on Computing, Vol. 11, No. 1,
pp. 15–34.
162
Srinivasan V., and Shah R.K., 1997, “Condensation in compact heat exchangers,”
Journal of Enhanced Heat Transfer, Vol. 4, No. 4, pp. 237–256.
Stone, M., 1974, “Cross-validatory choice and assessment of statistical predictions,”
Journal of the Royal Statistical Society, Vol. B36 pp. 111–133.
Sumpter B.G., Getino, C., and Noid, D.W., 1994, “Theory and applications of neu-
ral computing in chemical science,” Annual Review of Physical Chemistry, Vol.
45, pp. 439–481.
Svergun, D.I., 1999, “Restoring low resolution structure of biological macromolecules
from solution scattering using simulated annealing,” Biophysical Journal, Vol. 76,
No. 6, pp. 2879–2886.
Tayal, M.C., Fu, Y., and Diwekar, U.M., 1999, “Optimal design of heat exchangers:
a genetic algorithm framework,” Industrial and Engineering Chemistry Research,
Vol. 38, No. 2, pp. 456–467.
Theodoridis, S., and Koutroumbas, K., 1999, Pattern Recognition, Academic Press,
San Diego, CA.
Thibault, J., and Grandjean, B.P.A., 1991, “A neural network methodology for heat
transfer data analysis,” International Journal of Heat and Mass Transfer, Vol.
34, No. 8, pp. 2063–2070.
Torikoshi, K., Xi, G., Nakazawa, Y., and Asano, H., 1994, “Flow and heat trans-
fer performance of a plate-fin and tube heat exchanger, first report: effect of fin
pitch,” in: Proceedings of the Tenth International Heat Transfer Conference, pp.
411–416.
Vanderplaats, G.N., 1984, Numerical Optimization Techniques for Engineering De-
sign: With Applications, McGraw-Hill, Mineola, NY.
van Laarhoven, P.J.M., and Aarts, E.H.L., 1987, Simulated annealing: theory and
applications, Kluwer, Norwell, MA.
Vehi, J., Luo, N., Rodellar, J., and Armengol, J., 2001, “Digital control via interval
analysis,” Nonlinear Analysis-Theory Methods & Applications, Vol. 47, No. 1, pp.
203–212.
von Wolfersdorf, J., Achermann, E., and Weigand, B., 1997, “Shape optimization of
cooling channels using genetic algorithms,” Journal of Heat Transfer, Vol. 119,
No. 2, pp. 380–388.
Wagner, Y., 1998, “A moving grid algorithm for a heat exchanger with phase
change,” Applied Numerical Mathematics, Vol. 28, No. 2–4, pp. 477–491.
Wang, C.C., Chiou, C.B., and Lu, D.C., 1996, “Single-phase heat transfer and flow
friction correlations for microfin tubes,” International Journal of Heat and Fluid
Flow, Vol. 17, No. 5, pp. 500–508.
Webb, R.L., 1980, “Air-side heat transfer in finned tube heat exchangers,” Heat
Transfer Engineering, Vol. 1, No. 3, pp. 33–49.
163
Wilson, E.E., 1915, “A basis for rational design of heat transfer apparatus,” Trans-
actions of the ASME, Vol. 37, pp. 47–70.
W´ojs, K., and Tietze, T., 1997, “Effects of the temperature interference on the re-
sults obtained using the Wilson plot technique,” Heat and Mass Transfer, Vol.
33, No. 3, pp. 241–245.
Yan, W.M., and Sheen, P.J., 2000, “Heat transfer and friction characteristics of
fin-and-tube heat exchangers,” International Journal of Heat and Mass Transfer,
Vol. 43, No. 9, pp. 1651–1659.
Yan, Y.Y., and Lin, T.F., 1999, “Evaporation heat transfer and pressure drop of
refrigerant R-134a in a plate heat exchanger,” ASME Journal of Heat Transfer,
Vol. 118, No. 1, pp. 118–127.
Yang, R.L., 2000, “Convergence of the simulated annealing algorithm for continuous
global optimization,” Journal of Optimization Theory and Applications, Vol. 104,
No. 3, pp. 691–716.
Yao, X., and Liu, Y., 1997, “A new evolutionary system for evolving artificial neu-
ral networks,” IEEE Transactions on Neural Networks, Vol. 8, No. 3, pp. 694–713.
Zhang, G.Q., Patuwo B.E., and Hu, M.Y., 1998, “Forecasting with artificial neural
networks: the state of the art,” International Journal of Forecasting, Vol. 14, No.
1, pp. 35–62.
Zhao, X., 1995, Performance of a Single-Row Heat Exchanger at Low In-Tube Flow
Rates, M.S. Thesis, Department of Aerospace and Mechanical Engineering, Uni-
versity of Notre Dame, Notre Dame, Indiana.
Zeng, P., 1998, “Neural computing in mechanics,” Applied Mechanics Review, Vol.
51, No. 2, pp. 173–197.
164

Sign up to vote on this title
UsefulNot useful