International Communications in Heat and Mass Transfer 38 (2011) 7584

Contents lists available at ScienceDirect

International Communications in Heat and Mass Transfer

j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / i c h m t

Articial neural network techniques for the determination of condensation heat

transfer characteristics during downward annular ow of R134a inside a vertical
smooth tube
M. Balcilar a, A.S. Dalkilic b,, S. Wongwises c,
a
Computer Engineering Department, Yildiz Technical University, Yildiz, Besiktas, Istanbul 34349, Turkey
b
Heat and Thermodynamics Division, Department of Mechanical Engineering, Yildiz Technical University, Yildiz, Besiktas, Istanbul 34349, Turkey
c
Fluid Mechanics, Thermal Engineering and Multiphase Flow Research Lab. (FUTURE), Department of Mechanical Engineering, King Mongkut's University of Technology Thonburi,
Bangmod, Bangkok 10140, Thailand

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

Available online 18 November 2010 In this study, the best articial intelligence method is investigated to estimate the measured convective
heat transfer coefcient and pressure drop of R134a owing downward inside a vertical smooth copper
Keywords: tube having an inner diameter of 8.1 mm and a length of 500 mm during annular ow numerically. R134a
Condensation and water are used as working uids in the tube side and annular side of a double tube heat exchanger,
Heat transfer coefcient
respectively. The ANN training sets have the experimental data of in-tube condensation tests including six
Pressure drop
different mass uxes of R134a such as 260, 300, 340, 400, 456 and 515 kg m 2 s 1, two different saturation
Modeling
Neural network
temperatures of R134a such as 40 and 50 C and heat uxes ranging from 10.16 to 66.61 kW m 2. The
quality of the refrigerant in the test section is calculated considering the temperature and pressure
obtained from the experiment. The pressure drop across the test section is directly measured by a
differential pressure transducer. Input of the ANNs are the measured values of test section such as mass
ux, heat ux, the temperature difference between the tube wall and saturation temperature, average
vapor quality, while the outputs of the ANNs are the experimental condensation heat transfer coefcient
and measured pressure drop in the analysis. Condensation heat transfer characteristics of R134a are
modeled to decide the best approach using several articial neural network (ANN) methods such as
multilayer perceptron (MLP), radial basis networks (RBFN), generalized regression neural network
(GRNN) and adaptive neuro-fuzzy inference system (ANFIS). Elimination process of the ANN methods is
performed by means of 183 data points, divided into two sets randomly, obtained in the experiments. Sets
of test and training/validation include 33 and 120/30 data points respectively for the elimination process.
Validation process, in terms of various experimental conditions, is done by means of 368 experimental data
points having 68 data points for test set and 300 data points for training set. In training phase, 5-fold cross
validation is used to determine the best value of ANNs control parameters. The ANNs performances were
measured by means of relative error criteria with the usage of unknown test sets. The performance of the
method of multi layer perceptron (MLP) with 5-13-1 architecture and radial basis function networks
(RBFN) were found to be in good agreement, predicting the experimental condensation heat transfer
coefcient and pressure drop with their deviations being within the range of 5% for all tested conditions.
Dependency of outputs of the ANNs from input values is also investigated in the paper.
2010 Elsevier Ltd. All rights reserved.

1. Introduction heat transfer systems have higher thermal heat transfer performance
than single-phase heat transfer systems in limited apparatus volumes.
Determination of heat transfer characteristics of compact systems The advantage of high thermal performance in comparison to the
such as heat transfer coefcient, frictional pressure drop, ow type single-phase applications includes smaller, more compact systems
between the uids, heat transfer surface area in heat exchangers regarding tube length, tube diameter, internal tube geometry shape in
should be known accurately for the optimum design. The two-phase a wide variety of applications, e.g. refrigeration and air conditioning
systems, power engineering and other thermal processing plants.
Before doing any numerical work on this subject, validation process of
Communicated by W.J. Minkowycz.
the experimental setup and data is a very signicant for reliability of
Corresponding authors.
E-mail addresses: [email protected] (A.S. Dalkilic), [email protected] the study. The correctness of experimental data should be evaluated
(S. Wongwises). by either graphics regarding the characteristics of trend lines such as

0735-1933/$ see front matter 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.icheatmasstransfer.2010.10.009
76 M. Balcilar et al. / International Communications in Heat and Mass Transfer 38 (2011) 7584

alteration of heat transfer coefcient vs quality, heat ux vs previous publications and for the identication of ow regimes in data
temperature difference between the vapor and inlet wall tempera- corresponding to annular ow downward condensation of R134a in a
ture, pressure drop vs vapor quality etc. or widely accepted vertical smooth copper tube. In addition to this, 12 number of well-
correlations in the literature and desired working conditions should known ow regime correlations from the literature are investigated to
be agreeable with the heat balance correlations and ow type identify the ow regime occurring in the test tube. Dalkilic et al. [22]
observations by means of sight glasses at the different points of determined the average predicted heat transfer coefcient of the
experimental apparatus. refrigerant by means of Kosky and Staub's model [23] and the Von
There have been a number of studies on the in-tube condensation Karman universal velocity distribution correlations using different
heat transfer coefcient and pressure drop. The frictional, acceleration, interfacial shear stress equations valid for annular ow in horizontal
and gravitational components determine the two-phase total pressure and vertical tubes in order to investigate Chen et al.'s annular ow
drop in tubes. Determination of the void fraction is required to compute theory [24]. The effects of heat ux, mass ux and condensation
the acceleration and gravitational components, and in a similar way, temperature on the pressure drop were also discussed. A new
determination of either the two-phase friction factor or the two-phase correlation, based on dimensionless parameters such as the equivalent
frictional multiplier is required to compute the frictional component of Reynolds number, Prandtl number, R number (- ratio), Lockhart and
pressure drop. Generally, empirical methods have been used to compute Martinelli parameter, Bond number and Froude number, using a large
the condensation heat transfer coefcients and pressure drops in tubes. number of data points for the turbulent condensation heat transfer
Most of these proposed models are modications of the DittusBoelter coefcient, was proposed in their paper. Dalkilic et al. [25] presented a
single-phase forced convection correlation [1], as in Akers et al. [2], new experimental approach on the determination of condensation
Cavallini and Zecchin [3], and Shah [4]. heat transfer coefcient using 13 number of frictional pressure drop
Dalkilic et al. [5] compared thirteen well-known two-phase models and 35 number of void fraction models in a vertical tube by
pressure drop models with the experimental results of a condensation means of von Karman's universal velocity distribution and Kosky and
pressure drop of R600a and R134a in horizontal and vertical smooth Staub's annular ow lm thickness model [23]. Dalkilic and Wong-
copper tubes respectively and discussed the main parameters of wises [26] presented a detailed review of research on in-tube
related models and correlations. Dalkilic et al. [6] proposed a new condensation due to its signicance in refrigeration, air conditioning
correlation for the two-phase friction factor of R134a ow based on and heat pump applications. Their review considered heat transfer and
the equivalent Reynolds number model [7] and also showed the pressure drop investigations during in-tube condensation. Almost all
effects of heat ux, mass ux and condensation temperature on the relevant research subjects were summarized, such as condensation
pressure drop. Dalkilic et al. [8] compared eleven well-known heat transfer and pressure drop studies according to tube orientation
correlations for annular ow to each other using a large amount of (horizontal, vertical, inclined tubes) and tube geometry (smooth and
data obtained under various experimental conditions and proposed a enhanced tubes), ow pattern studies of condensation, void fraction
new correlation based on Bellinghausen and Renz's method [9] for the studies, and refrigerants with oil. Dalkilic and Wongwises [27]
condensation heat transfer coefcient of high mass ux ow of R134a. performed a theoretical performance study on a traditional vapor-
They also reported the effects of heat ux, mass ux and condensation compression refrigeration system with refrigerant mixtures based on
temperature on the heat transfer coefcients in their paper. Dalkilic et HFC134a, HFC152a, HFC32, HC290, HC1270, HC600, and HC600a for
al. [10] emphasized that an interfacial shear effect is signicant for the various ratios and their results are compared with CFC12, CFC22, and
laminar condensation heat transfer of R134a using Carey's analysis HFC134a as possible alternative replacements. Dalkilic et al. [28]
[11], which is the improved version of Nusselt's theory [12], and investigated the direct relationship between the measured condensa-
proposed a new correlation based on Bellinghausen and Renz's tion pressure drop and convective heat transfer coefcient of R134a to
method [9] for the condensation heat transfer coefcient during determine the condensation heat transfer coefcient, heat ux, two-
annular ow of R134a at low mass ux in a vertical tube. Dalkilic et al. phase length and pressure drop experimentally. In addition to the
[13] reviewed thirty-three void fraction models and correlations, above, various other papers [2936] support and validate the proposed
determined using relevant data from the available literature. They models and correlations in their papers.
calculated the friction factors, based on the analysis of Ma et al. [14], To the best of the authors' knowledge, there has been only one
obtained from various void fraction models and correlations, and work just for the condensation of R600a in a horizontal tube, recently
compared these with each other and also with those determined from carried out by Demir et al. [37] dealing with numerical solution using
graphical information provided by Bergelin et al. [15]. The effect of neural network analysis of heat transfer coefcient, and no work
void fraction alteration on the momentum pressure drop was also regarding the numerical solution using neural network analysis for the
presented in their paper. Dalkilic et al.'s [16] analysis based on simple condensation heat transfer coefcient and pressure drop of HFC-134a
void fraction models of the annular ow pattern was presented for the either in a horizontal or vertical tube. For these reasons, there remains
forced convection condensation of pure R134a, taking into account room for further research. Especially, it can be noted that all of the
the effect of the different saturation temperatures in high mass ux numerical investigations of neural network analysis do not have any
conditions. The comparisons of calculated lm thickness from void that are concerned with the study. As a consequence, in the present
fraction models and correlations with those from Whalley's annular study, the main concern is to ll the gap in the literature. The results for
ow model [17] were made using their experimental database. the solution of neural network analysis on the two-phase heat transfer
Dalkilic and Wongwises [18] studied a mathematical model, devel- coefcient and pressure drop which have never before appeared in
oped by Barnea et al. [19], and based on the momentum balance of
liquid and vapor phases, in order to determine the condensation lm Table 1
thickness of R134a in their paper. The effects of heat ux, mass ux Operating conditions of the study.
and condensation temperature on the lm thickness and condensa- Refrigerant HFC-134a
tion heat transfer coefcient were also discussed for laminar and
Test tube material, inner diameter d, length L Copper, 8.1 mm, 500 mm
turbulent ow conditions. Six well-known ow regime maps from the
Range of condensing pressure, Psat (bar) 1012
literature were found to be predictive for the annular ow conditions Range of condensing temperature, Tsat (C) 4050
in the test tube in spite of their different operating conditions. Range of mass ux, G (kg m 2 s 1) 260515
Dalkilic et al. [20] reports the experimental investigation of Kosky's Range of heat ux, q (kW m 2) 10.1666.61
Range of temperature difference, T = Tsat Twi , K 1.268.98
model [21] for predicting ow pattern transitions and for the
Range of average vapor quality, x 0.690.96
validation of void fraction models and correlations proposed in their
 M. Balcilar et al. / International Communications in Heat and Mass Transfer 38 (2011) 7584 77

Table 2 Table 4
Uncertainty of experimental parameters. Effect of the experimental data number on the error rates for the determination of
condensation heat transfer coefcient (a) and pressure drop (b) for the mass uxes of
Parameters Uncertainty 300, 400, 515 kg m 2 s 1 and Tsat = 40 and 50 C.
Tref, sat (C) 0.19
Experimental data number 35 68
xi 6.968.24%
T (K) 0.191 error analysis/ANN methods MLP RBFN MLP RBFN
(Tw,o Tw,i)TS (K) 0.045
a)
(Tw,i Tw,o)ph (K) 0.13
h (W m 2 K 1)
mref (g s 1) 0.023
R square 0.999 0.988 0.999 0.99
mw,TS (g s 1) 0.35
Proportional (%) 0.059 1.295 0.061 1.444
mw,ph (g s 1) 0.38
Square law 6.048 144.952 5.551 132.68
qTS (kW m 2) 6.558.93%
qph (kW m2) 1214.81%
b)
href (W m 2 K 1) 7.6410.71%.
P (Pa)
P (kPa) 0.15
R square 0.998 0.997 0.998 0.997
Proportional (%) 1.793 2.28 1.754 2.017
Square law 67.702 78.77 64.581 71.11

open literature are presented. The large amount of measured data

obtained from the experimental study regarding with the convective
heat transfer coefcient and pressure drop is compared with the
results from articial neural networks' methods such as multi layer
perceptron (MLP), radial basis function networks (RBFN), generalized
regression neural network (GRNN) and adaptive neuro-fuzzy infer-
ence system (ANFIS). In addition to this, the effect of input parameters
(mass ux (G), heat ux (q), saturation temperature (Tsat), temper-
ature difference between the condensing temperature and inlet tube
wall temperature (Tsat), vapor quality (x)) on the output parameters
(convective heat transfer coefcient (h), pressure drop (P)) is also
shown and discussed as a rst study on this subject in the literature.

2. Experimental apparatus and method

Detailed descriptions of the experimental apparatus for studying

condensation of R134a inside a vertical tube can be found in the
authors' previous publications.

3. Data reduction and experimental uncertainty

The data reduction of the measured results can be analyzed as follows:

3.1. The inlet vapor quality of the test section (xTS,i)

iTS;i il@TTS;i
xTS;i = 1
ifg@TTS;i

where il@TTS;i is the enthalpy of the saturated liquid based on the

temperature of the test section inlet, ifg@TTS;i is the enthalpy of

Table 3
Elimination of ANN methods according to the error rates for the determination of
condensation heat transfer coefcient (a) and pressure drop (b) for the mass uxes of
260, 340, 456 kg m 2 s 1 and Tsat = 40 and 50 C.

Error analysis/ h (W m 2 K 1)
ANN methods
MLP RBFN ANFIS GRNNN

a)
R square 0.999 0.992 0.994 0.056
Proportional (%) 0.588 1.558 1.182 16.062
Square law 39.452 113.617 103.296 1293.76

P (Pa)

MLP RBFN ANFIS GRNNN

b)
R square 0.997 0.998 0.989 0.432
Fig. 1. Schematic diagrams of articial intelligence models used in the analysis. a) Multi
Proportional (%) 2.127 1.74 4.15 37.608
layer perceptron (MLP) with 5-13-1 architecture. b) Articial neural fuzzy inference
Square law 64.895 55.208 147.85 1148.447
system (ANFIS).
78 M. Balcilar et al. / International Communications in Heat and Mass Transfer 38 (2011) 7584

vaporization based on the temperature of the test section inlet, iTS;i is 3.2. The outlet vapor quality of the test section (xTS,o)
the refrigerant enthalpy at the test section inlet and is given by:
iTS;o il@TTS;o
xTS;o = 4
Qph ifg@TTS;o
iTS;i = iph;i + 2
mref
where iTS;o is the refrigerant enthalpy at the test section outlet, il@TTS;o
where iph;i is the inlet enthalpy of the liquid refrigerant before is the enthalpy of the saturated liquid based on the temperature of the
entering the preheater, mref is the mass ow rate of the refrigerant, test section outlet, and ifg@TTS;o is the enthalpy of vaporization. The
and Qph is the heat transfer rate in the preheater: outlet enthalpy of the refrigerant ow is calculated as follows:


Qph = mw;ph cp;w Tw;i Tw;o 3 QTS
ph
iTS;o = iTS;i 5
mref

where mw,ph is the mass ow rate of the water entering the preheater, where the heat transfer rate, QTS, in the test section is obtained from:


cp,w is the specic heat of water, Tw;i Tw;o is the temperature

ph
QTS = mw;TS cp;w Tw;o Tw;i 6
difference between inlet and outlet positions of the preheater. TS

Fig. 2. Comparisons of experimental convective heat transfer coefcients with the articial neural network methods using 33 number of experimental data points as ANN test group
for the mass uxes of 260, 340, 456 kg m 2 s 1 and Tsat = 40 C (a) and 50 C (b).
 M. Balcilar et al. / International Communications in Heat and Mass Transfer 38 (2011) 7584 79

where mw,TS is the mass ow rate of the water entering the test where d is the inside diameter of the test tube. L is the length of the


section, and Tw;o Tw;i is temperature difference between outlet test tube.
TS
and inlet position of the test section.
3.4. Uncertainties
3.3. The average heat transfer coefcient
The uncertainties of the Nusselt number and condensation heat
QTS transfer coefcient in the test tube varied from 7.64% to 10.71%. It
hexp =
 7 should be noted that similar range of uncertainty with minimum and
Awi Tref;sat Twi
maximum values are exist in the literature. The procedures of Kline and
McClintock [38] were used for the calculation of all uncertainties.
where hexp is the experimental average heat transfer coefcient, QTS is Various uncertainty values of the study for the operating conditions in
the heat transfer rate in the test section, Twi is the average Table 1 can be seen from authors' previous works and in Table 2 in detail.
temperature of the inner wall, Tref;sat is the average temperature of
the refrigerant at the test section inlet and outlet, and Awi is the inside 3.5. Error analysis
surface area of the test section:
Error analysis of the study is performed by means of Eqs. (9), (10)
Awi = dL 8 and (11) and results of them were given in Tables 3 and 4. In these

a) 8000
G=260 kg m-2s-1 MLP
G=260 kg m-2s-1 RBFN
G=260 kg m-2s-1 ANFIS +20%
G=260 kg m-2s-1 GRNN
6000 G=340 kg m-2s-1 MLP
G=340 kg m-2s-1 RBFN
G=340 kg m-2s-1 ANFIS
G=340 kg m-2s-1 GRNN
P (Pa)

G=456 kg m-2s-1 MLP -20%

4000
G=456 kg m-2s-1 RBFN
G=456 kg m-2s-1 ANFIS
G=456 kg m-2s-1 GRNN

2000 Tsat=40 oC
x=0.71-0.94
o
Tsat=1.63-8.48 C
q=10.83-50.89 kW m-2
0
0 2000 4000 6000 8000
Pexp (Pa) -30%

b) 6000
Tsat=50 oC
5000 x=0.76-0.95
Tsat=1.92-8.58 oC
q=11.21-48.24 kW m-2
4000
+20% G=260 kg m-2s-1 MLP
G=260 kg m-2s-1 RBFN
P (Pa)

3000 G=260 kg m-2s-1 ANFIS

G=260 kg m-2s-1 GRNN
G=340 kg m-2s-1 MLP
2000 G=340 kg m-2s-1 RBFN
-20%
G=340 kg m-2s-1 ANFIS
G=340 kg m-2s-1 GRNN
1000 G=456 kg m-2s-1 MLP
G=456 kg m-2s-1 RBFN
G=456 kg m-2s-1 ANFIS
0 G=456 kg m-2s-1 GRNN
0 1000 2000 3000 4000 5000 6000
Pexp (Pa)

Fig. 3. Comparisons of measured two-phase pressure drops with the articial neural network methods using 33 number of experimental data points as ANN test group for the mass
uxes of 260, 340, 456 kg m 2 s 1 and Tsat = 40 C (a) and 50 C (b).
80 M. Balcilar et al. / International Communications in Heat and Mass Transfer 38 (2011) 7584

equations, N is the element number of test group, hiexp and hicalc are the 3.6.2. Radial basis function networks (RBFN)
experimental and calculated results of heat transfer coefcient RBFN architecture has a single hidden layer and some similarities
obtained from i number of test data respectively and hexp and with MLP method. On the other hand, it has different calculation
hcalc are the average values of experimental and calculated results of process for the determination of neuron number, different activation
heat transfer coefcient obtained from the numerical analysis. It function and difference in neuron number in hidden layer. RBFN
should be noted that error analysis of pressure drop was done architecture's of single layer has equal neuron numbers with the
similarly. R square error is determined as follows: element numbers in training group. Bias numbers are the control
parameters of RBFN architecture and should be given by the user.
!2 Training of this architecture, in other words, values of neuron
N


hiexp
hexp hicalc h
calc connections is calculated using training group by means of least
i=1
ERsquare = ! ! 9 square method.

N 2 N
2
hiexp h
exp hicalc h
calc
i=1 i=1 3.6.3. Generalized regression neural networks (GRNN)
A generalized regression neural network (GRNN) is often used for
Proportional error is calculated as follows: function approximation. It has a radial basis layer and a special linear
layer. It performs regression rather than classication tasks. GRNNs
  can only be used for regression problems and trains almost instantly,
 i i  but tends to be large and slow. Like an RBFN network, a GRNN does
100 N hexp hcalc 
Eratio = 10 not extrapolate. Detailed information can be seen from [40].
N i=1 hiexp

3.6.4. Articial neural fuzzy inference system (ANFIS)

Square law error is estimated as follows: ANFIS is an adaptive network which permits the usage of neural
s network topology together with fuzzy logic. It not only includes the
1 N
i 2 characteristics of both methods, but also eliminates some disadvan-
Esquare = : hexp hicalc 11
N i=1 tages of their lonely-used case.

4. Results and discussion

3.6. Calculation procedure for the generalized articial neural networks
(ANNs) model The present study's experiments were performed at the Fluid
Mechanics, Thermal Engineering and Multiphase Flow Research Lab.
An articial neural network is an information-processing system (FUTURE) in KMUTT. Pure HFC-134a was used in the experiments.
that has certain performance characteristics in common with Special attention was paid on the accuracy of the measurements
biological neural networks. Articial neural networks have been during experiments as can be seen from authors' previous publica-
developed as generalizations of mathematical models of human tions on in-tube condensation process.
cognition or neural biology, based on the assumptions that: (1) Numerical analysis of articial intelligence includes 5 dimensional
Information processing occurs at many simple elements called input parameters such as mass ux (G, kg m 2 s 1), heat ux
neurons (2) Signals are passed between neurons over connection (q, kW m 2), saturation temperature (Tsat), temperature difference
links (3) Each connection link has an associated weight, which, in a between the saturation and inlet wall temperatures (Tsat = Tsat
typical neural net, multiplies the signal transmitted (4) Each neuron Twi) and average vapor quality of R134a (x), 2 dimensional output
applies an activation function (usually non-linear) to its net input parameters such as convective heat transfer (h, W m 2 K 1) and
(sum of weighted input signals) to determine its output signal [39]. pressure drop (P, Pa).
As the articial neural networks (ANNs) are one of the commonly
used and developed models to investigate the relationship between Table 5
linear or non-linear inputoutput patterns, they try to generalize the The most predictive combinations of input parameters according to the error rates for
training group and then estimate the test group. Performance of ANNs the determination of condensation heat transfer coefcient (h) and pressure drop (P).
is measured with the success of the prediction. There are many Inputs Input Outputs
described ANN types related with the function approximation in the number
G q Tsat Tsat x R2 h R2 P
literature. The most known methods are multi layer perceptron
(kg m 2 s 1) (kW m 2) (C) (K) (%) (%)
(MLP), radial basis functions networks (RBFN), generalized regression
neural networks (GRNN) and articial neural fuzzy inference system 1 0 0 0 0 1 0.616 0.886
0 1 0 1 0 2 0.999 0.897
(ANFIS) which is the combination of fuzzy logic decision systems and 1 0 0 1 0 2 0.93 0.807
ANNs. Schematic diagrams of some articial intelligence models used 1 0 1 0 0 2 0.656 0.988
in the analysis are shown in Fig. 1. 1 1 0 0 0 2 0.895 0.82
0 1 0 1 1 3 0.999 0.894
0 1 1 1 0 3 0.998 0.945
3.6.1. Multi layer perceptron (MLP)
1 0 0 1 1 3 0.939 0.801
MLP is the member of feed-forward network family. According to 1 0 1 0 1 3 0.696 0.994
this method, number of input and output neurons is equal to the 1 0 1 1 0 3 0.954 0.989
number of input and output parameters respectively and there are 1 1 0 0 1 3 0.893 0.83
some layers including neurons between them. Specication of non- 1 1 0 1 0 3 0.999 0.915
1 1 1 0 0 3 0.914 0.989
linearity and calculation process increase with increasing numbers of 0 1 1 1 1 4 0.999 0.938
hidden layer and neurons. Different architectures of MLP were used 1 0 1 1 1 4 0.969 0.996
for the calculation of heat transfer coefcient and pressure drop. All 1 1 0 1 1 4 0.999 0.92
neurons of MLP are connected with another neurons and the way of 1 1 1 0 1 4 0.95 0.995
1 1 1 1 0 4 0.999 0.994
connections is forward. LevenbergMarquardt algorithm, based on
1 1 1 1 1 5 0.998 0.996
the method of back propagation, was used for the training process.
 M. Balcilar et al. / International Communications in Heat and Mass Transfer 38 (2011) 7584 81

In terms of the elimination process of several ANN methods such best predictive method for the convective heat transfer coefcient (a)
as neural network (ANN) methods such as multilayer perceptron and pressure drop (b) accurately according to the their low average
(MLP), radial basis networks (RBFN), generalized regression neural values of test group of 33 data points' results using several error
network (GRNN) and adaptive neuro-fuzzy inference system (ANFIS), analysis. The performance of the method of multi layer perceptron
random sets of test and training/validation were determined as 33 (MLP) with 5-13-1 architecture and radial basis function networks
and 120/30 among 183 data points by means of Matlab software [40] (RBFN) with the spread coefcient (sp) of 40,000, shown in Figs. 2 and
respectively. The best value of ANNs control parameters is determined 3, were found to be superior to other methods and architectures by
using 5-fold cross validation in training phase. During the calculation means of satisfactory results with their deviations within the range of
process of heat transfer coefcient and pressure drop, the most 0.58% for the estimated condensation heat transfer coefcient and
agreeable hidden layer number value was found as 13 (h and P) 1.74% for the estimated pressure drop respectively. In addition to
among the numbers from 1 to 30 for the method of multi layer this, according to the condensation temperatures of 40 and 50 C,
perceptron (MLP), the most agreeable spread numbers (sp) were comparison of experimental heat transfer coefcients and pressure
found as 12,000 (h) and 40,000 (P) for the radial basis function drops with those from various articial neural network methods of
networks (RBFN). The experimental data includes the mass uxes of test groups are shown in Figs. 2 and 3 in a 20% deviation line for the
260, 340 and 456 kg m 2 s 1 belong to the saturation temperatures mass uxes of 260, 340 and 456 kg m 2 s 1respectively. It can be
of 40 and 50 C in Table 3 and also in Figs. 2 and 3. Table 3 shows the clearly seen from these gures that the generalized regression neural

Fig. 4. Comparisons of experimental convective heat transfer coefcients with the most predictive articial neural network methods using 35 number of experimental data points as
ANN test group for the mass uxes of 300, 400, 515 kg m 2 s 1 and Tsat = 40 C (a) and 50 C (b).
82 M. Balcilar et al. / International Communications in Heat and Mass Transfer 38 (2011) 7584

networks (GRNN) method has the poorest predictability of experi- according to the condensing temperatures of 40 C (a) and 50 C (b)
mental data. respectively with their deviations being within the range of 5% for all
In terms of the increase in data number for the accuracy and tested conditions. Table 5 shows the most predictive combinations of
validation process of the experiments, numerical calculations is done by input parameters, whose deviations are over 0.5%, according to the error
means of 368 experimental data points having 68 data points for test set rates for the determination of condensation heat transfer coefcient (h)
and 300 data points for training set in Tables 4 and 5. Table 4 shows the and pressure drop (P) using the method of multi layer perceptron
effect of the experimental data number on the error rates of the test (MLP) with 5-13-1 architecture including 68 experimental data points.
group of ANN according to the most predictive ANN methods, When the input parameter number is 2, heat ux (q) and temperature
determined from Table 3 and Figs. 2 and 3, for the determination of difference between the condensing temperature and inlet wall
condensation heat transfer coefcient (a) and pressure drop (b) for the temperatures (Tsat) seem to be dominant on the prediction of
mass uxes of 300, 400, 515 kg m 2 s 1 and condensation tempera- convective heat transfer coefcient with the deviation of 0.999% and
tures of 40 and 50 C. According to Table 3, results of error analysis pressure drop with the deviation of 0.897%. When there are 3 known
decrease with increasing data number due to the result of learning input parameters, heat ux (q), temperature difference between the
process of articial intelligence. Figs. 4 and 5 show the success of condensing temperature and inlet wall temperatures (Tsat), and
determined compatible articial neural network methods for the saturation temperature (Tsat) has the best predictability for the
prediction of convective heat transfer coefcient and pressure drop prediction of convective heat transfer coefcient with the deviation of

a)
8000

G=300 kg m-2s-1 MLP

G=300 kg m-2s-1 RBF
G=400 kg m-2s-1 MLP
6000 G=400 kg m-2s-1 RBF +5%
G=515 kg m-2s-1 MLP
G=515 kg m-2s-1 RBF
P (Pa)

4000 -5%

Tsat=40 oC
2000
x=0.77-0.95
T sat=2.47-8.26 oC
q=15.81-52.39 W m-2
0
0 2000 4000 6000 8000
Pexp (Pa)
b)
6000

G=300 kg m-2s-1 MLP

5000 G=300 kg m-2s-1 RBF
G=400 kg m-2s-1 MLP
G=400 kg m-2s-1 RBF +5%

4000 G=515 kg m-2s-1 MLP

G=515 kg m-2s-1 RBF
P (Pa)

-5%
3000

2000
Tsat=50 oC
x=0.76-0.93
1000 Tsat=1.34-8.52 oC
q=12.1-54.22 kW m-2

0
0 1000 2000 3000 4000 5000 6000
Pexp (Pa)

Fig. 5. Comparisons of measured two-phase pressure drops with the most predictive articial neural network methods using 35 number of experimental data points as ANN test
group for the mass uxes of 300, 400, 515 kg m 2 s 1 and Tsat = 40 C (a) and 50 C (b).
 M. Balcilar et al. / International Communications in Heat and Mass Transfer 38 (2011) 7584 83

0.998% and pressure drop with the deviation of 0.945%. It shows that the ph preheater
accuracy of the estimated output values by means of input values o outlet
increases with increasing added proper input parameters. ref refrigerant
A large number of graphics could be generated from the output of sat saturation
the calculations, however, due to space limitation, only typical results TS test section
are shown for limited data of the experimental database of the w water
authors. It should be also noted that detailed information on the heat
transfer characteristics of the study with some additional gures and
different experimental parameters can be seen from the authors'
previous publications. Acknowledgements

5. Conclusion The rst author is indebted to the Department of Mechanical

Engineering, Yildiz Technical University (YTU). The second author
Comparison of the numerical condensation heat transfer coefcient wishes to thank King Mongkut's University of Technology Thonburi
and pressure drop values with the measured values by means of articial (KMUTT) for providing him with a Post-doctoral fellowship. The third
intelligence methods was performed in this paper. There is no research on author wishes to thank the Thailand Research Fund, the Ofce of the
the content of this paper considering its parameters in current literature. Higher Education Commission and the National Research University
The results of this study are expected to ll this gap in the literature. Project for nancial support for this study.
Except from the generalized regression neural network (GRNN),
the articial neural network (ANN) methods of multilayer perceptron References
(MLP), radial basis networks (RBFN) and adaptive neuro-fuzzy
inference system (ANFIS) are found to be capable of predicting the [1] F.W. Dittus, L.M.E. Boelter, Engineering Publication, 2, University of California,
1930, p. 443.
condensation heat transfer coefcient and pressure drop for the high [2] W.W. Akers, H.A. Deans, O.K. Crosser, Condensation heat transfer within
mass region of R134a in a vertical tube within the range of 20% and horizontal tubes, Chemical Engineering Progress Symposium Series 55 (1959)
also within the slightly small error rates. 171176.
[3] A. Cavallini, R. Zecchin, A dimensionless correlation for heat transfer in forced
The performance of the method of multi layer perceptron (MLP)
convection condensation, Proceedings of the Sixth International Heat Transfer
and radial basis function networks (RBFN) were the most predictive Conference 3 (1974) 309313.
methods with their deviations within the range of 5% for the [4] M.M. Shah, A general correlation for heat transfer during lm condensation inside
pipes, International Journal of Heat and Mass Transfer 22 (1979) 547556.
estimated condensation heat transfer coefcient and pressure drop. It
[5] A.S. Dalkilic, O. Agra, I. Teke, S. Wongwises, Comparison of frictional pressure drop
should be noted that these articial intelligence methods have better models during annular ow condensation of R600a in a horizontal tube at low
results on the prediction of condensation heat transfer coefcient and mass ux and of R134a in a vertical tube at high mass ux, International Journal of
pressure drop than the empirical models and correlations proposed in Heat and Mass Transfer 53 (2010) 20522064.
[6] A.S. Dalkilic, S. Laohalertdecha, S. Wongwises, Two-phase friction factor in vertical
the authors' previous publications. downward ow in high mass ux region of refrigerant HFC-134a during
Convective heat transfer coefcient and pressure drop of R134a in condensation, International Communications in Heat and Mass Transfer 35
the paper are found to be dependent to the heat ux and temperature (2008) 11471152.
[7] K. Moser, R.L. Webb, B. Na, A new equivalent Reynolds number model for condensation
difference between the condensing temperature and inlet wall in smooth tubes, International Journal of Heat Transfer 120 (1998) 410417.
temperatures strongly according to the results of analysis. [8] A.S. Dalkilic, S. Laohalertdecha, S. Wongwises, Experimental investigation of heat
transfer coefcient of R134a during condensation in vertical downward ow at
high mass ux in a smooth tube, International Communications in Heat and Mass
Nomenclature Transfer 36 (2009) 10361043.
A surface area, m2 [9] R. Bellinghausen, U. Renz, Heat transfer and lm thickness during condensation of
cp specic heat, J kg 1 K 1 steam owing at high velocity in a vertical pipe, International Journal of Heat and
Mass Transfer 35 (1992) 683689.
d internal tube diameter, m
[10] A.S. Dalkilic, S. Yildiz, S. Wongwises, Experimental investigation of convective
E error heat transfer coefcient during downward laminar ow condensation of R134a in
G mass ux, kg m 2 s 1 a vertical smooth tube, International Journal of Heat and Mass Transfer 52 (2009)
h heat transfer coefcient, W m 2 K 1 142150.
[11] V.P. Carey, LiquidVapor Phase Change Phenomena, Hemisphere Publishing,
i enthalpy, J kg 1 1992.
ifg latent heat of condensation, J kg 1 [12] W. Nusselt, Die Oberachen Kondenastion des Wasserdamfes, Z. Vereines Dtsch.
L length of test tube, m Ingenieure 60 (1916) 541546, 569575.
[13] A.S. Dalkilic, S. Laohalertdecha, S. Wongwises, Effect of void fraction models on the
m mass ow rate, kg s 1 two-phase friction factor of R134a during condensation in vertical downward
N element number ow in a smooth tube, International Communications in Heat and Mass Transfer
P Pressure, N m 2 35 (2008) 921927.
[14] X. Ma, A. Briggs, J.W. Rose, Heat transfer and pressure drop characteristics for
P pressure drop, N m 2 condensation of R113 in a vertical micro-nned tube with wire insert,
Q heat transfer rate, W International Communications in Heat and Mass Transfer 31 (2004) 619627.
q heat ux, kW m 2 [15] O.P. Bergelin, P.K. Kegel, F.G. Carpenter, C. Gazley, Co-current gasliquid ow. II.
Flow in vertical tubes. Meeting at Berkeley, Cal. (USA) June, ASME, Heat Transfer
T temperature, C and Fluid Mechanic Institute (1946) 1928.
T temperature difference, K [16] A.S. Dalkilic, S. Laohalertdecha, S. Wongwises, Effect of void fraction models on the
x average vapor quality lm thickness of R134a during downward condensation in a vertical smooth tube,
International Communications in Heat and Mass Transfer 36 (2009) 172179.
[17] P.B. Whalley, Boiling, Condensation, and GasLiquid Flow, Oxford University
Subscripts Press, 1987.
avg average [18] A.S. Dalkilic, S. Wongwises, An investigation of a model of the ow pattern
calc calculated transition mechanism in relation to the identication of annular ow of R134a in a
vertical tube using various void fraction models and ow regime maps,
cond condensate Experimental Thermal and Fluid Science 34 (2010) 692705.
exp experimental [19] D. Barnea, O. Shoham, Y. Taitel, Flow pattern transition for vertical downward two
g gas/vapor phase ow, Chemical Engineering Science 37 (1982) 741744.
[20] A.S. Dalkilic, S. Laohalertdecha, S. Wongwises, Validation of void fraction models
i inlet and correlations using a ow pattern transition mechanism model in relation to
l liquid the identication of annular vertical downow in-tube condensation of R134a.
84 M. Balcilar et al. / International Communications in Heat and Mass Transfer 38 (2011) 7584

International Communications in Heat and Mass Transfer, International Commu- downward ow at high mass ux, HT2008: Proceedings of the ASME Summer
nications in Heat and Mass Transfer 37 (2010) 827834. Heat Transfer Conference-2008 2, 2009, pp. 193206.
[21] P.G. Kosky, Thin liquid lms under simultaneous shear and gravity forces, [31] A.S. Dalkilic, S. Laohalertdecha, S. Wongwises, Experimental investigation on the
International Journal of Heat Mass Transfer 14 (1971) 12201224. condensation heat transfer and pressure drop characteristics of R134a at high
[22] A.S. Dalkilic, S. Wongwises, Experimental study on the modeling of condensation mass ux conditions during annular ow regime inside a vertical smooth tube,
heat transfer coefcients in high mass ux region of refrigerant HFC-134a inside Summer Heat Transfer Conference, ASME (2009), July 1923, USA, 2009.
the vertical smooth tube during annular ow regime, Heat Transfer Engineering [32] A.S. Dalkilic, O. Agra, Experimental apparatus for the determination of
32 (2011) 112. condensation heat transfer coefcient for R134a and R600a owing inside
[23] P.G. Kosky, F.W. Staub, Local condensing heat transfer coefcients in the annular vertical and horizontal tubes, Summer Heat Transfer Conference, ASME (2009),
ow regime, AIChE Journal 17 (1971) 10371043. July 1923, USA, 2009.
[24] S.L. Chen, F.M. Gerner, C.L. Tien, General lm condensation correlations, [33] A.S. Dalkilic, S. Laohalertdecha, S. Wongwises, Experimental research on the
Experimental Heat Transfer 1 (1987) 93107. similarity of annular ow models and correlations for the condensation of R134a
[25] A.S. Dalkilic, S. Laohalertdecha, S. Wongwises, New experimental approach on the at high mass ux inside vertical and horizontal tubes, International Mechanical
determination of condensation heat transfer coefcient using frictional pressure Engineering Congress and Exposition, ASME (2009), November 1319, USA, 2009.
drop and void fraction models in a vertical tube, Energy Conversion and [34] A.S. Dalkilic, S. Laohalertdecha, S. Wongwises, Comparison of condensation
Management 51 (2010) 25352547. frictional pressure drop models and correlations during annular ow of R134a
[26] A.S. Dalkilic, S. Wongwises, Intensive literature review of condensation inside inside a vertical tube, Thermal and Environmental Issues in Energy Systems,
smooth and enhanced tubes, International Journal of Heat and Mass Transfer 52 ASME-ATI-UTI, (2010), May 1619, Italy, 2010.
(2009) 34093426. [35] A.S. Dalkilic, S. Wongwises, Experimental study on the ow regime identication
[27] A.S. Dalkilic, S. Wongwises, A performance comparison of vapour compression in the case of co-current condensation of R134a in a vertical smooth tube,
refrigeration system using various alternative refrigerants, International Commu- International Heat Transfer Conference, ASME, (2010), August 813, USA, 2010.
nications in Heat and Mass Transfer 37 (2010) 13401349. [36] A.S. Dalkilic, S. Wongwises, Experimental study on the ow regime identication
[28] A.S. Dalkilic, I. Teke, S. Wongwises, Experimental analysis for the determination of in the case of co-current condensation of R134a in a vertical smooth tube,
the convective heat transfer coefcient by measuring pressure drop directly International Heat Transfer Conference, ASME, (2010), August 813, USA, 2010.
during annular condensation ow of R134a in a vertical smooth tube, International [37] H. Demir, . Ara, .. Ataylmaz, Generalized neural network model of
Journal of Heat and Mass Transfer (in press). alternative refrigerant (R600a) inside a smooth tube, International Communica-
[29] A.S. Dalkilic, S. Laohalertdecha, S. Wongwises, A comparison of the void fraction tions in Heat and Mass Transfer 36 (2009) 744749.
correlations of R134a during condensation in vertical downward laminar ow in a [38] S.J. Kline, F.A. McClintock, Describing uncertainties in single sample experiments,
smooth and micron tube, MNHT2008: Proceedings of the Micro/Nanoscale Heat Mechanical Engineering 75 (1953) 38.
Transfer International Conference-2008 Parts AB, 2008, pp. 10291040. [39] L. Fausett, Fundamentals of Neural Networks, Architectures, Algorithms, and
[30] A.S. Dalkilic, S. Laohalertdecha, S. Wongwises, Two-phase friction factor obtained Applications, Prentice-Hall, Inc. Press, 1994.
from various void fraction models of R-134a during condensation in vertical [40] Matlab R2010a, 2010, http://mathworks.com/.