This action might not be possible to undo. Are you sure you want to continue?

# 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

/ρ

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

,

˙

Q¦

p

are

the set of predictions, and O

e

i,k

= ¦j

s

, j

t

,

˙

Q¦

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