International Journal of Heat and Mass Transfer 62 (2013) 481–494

Contents lists available at SciVerse ScienceDirect

International Journal of Heat and Mass Transfer

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

Combination of support vector regression and artificial neural

networks for prediction of critical heat flux
B.T. Jiang, F.Y. Zhao ⇑
Department of Nuclear Science and Technology, Xi’an Jiaotong University, Xi’an City 710049, China

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

Article history: This paper presents a hybrid model that couples m-support vector regression (m-SVR) with radial basis
Received 26 December 2012 function networks (RBFNs) for prediction of critical heat flux (CHF). The hybrid model is achieved in
Received in revised form 8 March 2013 two steps. The first step is to determine the initial architecture and initial weights of the hybrid model
Accepted 9 March 2013
by an m-SVR. The second step is to adjust the initial weights using an annealing robust back propagation
Available online 9 April 2013
(ARBP) learning algorithm. Then the hybrid model is used to predict CHF, which is divided into two parts:
prediction of CHF for water flow in vertical round tubes and prediction of dryout type CHF for deionized
Keywords:
water upflowing through a narrow annular channel with 0.95 mm gap. The dataset used in this paper is
Support vector regression
Critical heat flux
taken from literature. In the first part, prediction of CHF and analysis of parametric trends of CHF are both
Annealing robust back propagation carried out based on three conditions, fixed inlet conditions, local conditions and fixed outlet conditions.
Back-propagation network The predicted results agree better with the corresponding dataset than that of e-SVR. In the second part,
the predicted results are in better agreement with the experimental data than that of back-propagation
network (BPN) employed in the literature. Therefore, the hybrid model presented in this paper is a poten-
tial tool for predicting CHF and has advantages over other methods.
Ó 2013 Elsevier Ltd. All rights reserved.

1. Introduction
Critical heat flux (CHF) has been thought as one of the most
important parameters in the design and operation of high heat flux
systems, such as pressurized water reactors, steam generators and
other boiling heat transfer units. Nuclear reactors are deigned to
receive maximum efficiency under full working power and its effi-
ciency will be improved when the core exit temperature increases.
In this case, the nuclear reactors shall be designed with appropriate
thermal margin to assure that specified acceptable fuel design lim-
its are not exceeded during any condition of normal operation.
Thus, the thermal margin has a vital importance in the design
and safety assessment of nuclear reactors. However, in the thermo
hydraulic design of nuclear reactor, CHF limits the heat flux from
the fuel rods and the power capabilities of nuclear reactors. The
CHF condition is characterized by a sharp reduction of the local
heat transfer coefficient that results from the replacement of liquid
by vapor adjacent to the heat transfer surface [1,2]. For a nuclear
reactor core, exceeding CHF can lead to a sudden large increase
in cladding temperature due to the relatively poor heat transfer
characteristics of vapor, which for most coolants, can lead to a cat-
astrophic failure of nuclear fuel.
⇑ Corresponding author. Tel./fax: +86 29 82668648.
E-mail addresses: jiangburt@gmail.com (B.T. Jiang), fyzhao@mail.xjtu.edu.cn
(F.Y. Zhao).
The ability to predict CHF is therefore a vital issue for the per-
formance and the safety of nuclear reactors. CHF prediction in nu-
clear reactors could be useful to know the real causes of the failure,
like the burnout of tubes or leaks that appear as consequences of
an accelerated process of corrosion caused by the high temperature
reached in the material. A considerable amount of significant
experimental and theoretical research concerning CHF has been
performed over the last five decades with the development of
water cooled nuclear reactors. As a result of these efforts, several
prediction methodologies have been developed to predict CHF.
They can be categorized as three principal approaches: look-up ta-
bles, empirical correlations and phenomenological models.
Although many analytical and experimental studies on CHF
have been presented in the literature, it would be incorrect to
say that a commonly shared apprehension of this problem has
been established [3] because CHF is a complex phenomenon and
is influenced by many parameters. Even these three principle ap-
proaches have their own limitations, which are described below.
Firstly, Tong and Tang [4] evaluated look-up tables and empirical
correlations and summarized the limitations of these two methods
as follows. On the one hand, the Groeneveld CHF look-up table
[5,6] does not provide much more convenience when it is applied
to reactor design, since in this case proper adjustments for detailed
geometrical effects of a prototype rod bundle would be needed.
Such adjustments are usually expressed in empirical formulae that
are as complicated as the existing CHF design correlations. On the

0017-9310/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.ijheatmasstransfer.2013.03.025
482 B.T. Jiang, F.Y. Zhao / International Journal of Heat and Mass Transfer 62 (2013) 481–494
Nomenclature
b bias of the final regression function
D tube diameter (m)
G mass flow rate (kg/m2 s)
ID inner diameter (mm)
OD outer diameter (mm)
QCHF critical heat flux (kW/m2)
qo heat flux of outside tube (kW/m2)
w support vector weight
X exit quality
C regularization parameter
dH hydraulic diameter (m)
Hfg latent heat of vaporization (kJ/kg)
L/dH length to hydraulic to diameter ratio
P pressure (MPa)
qi heat flux of inside tube (kW/m2)
other hand, a CHF correlation is accurate only in the particular flow
regimes within the ranges of the operating parameters in which it
was developed, thus the application of CHF correlations should be
limited to within these ranges of parameters. Secondly, the limita-
tions of phenomenological models are also summarized as follows.
Phenomenological models were initiated in the 1970s [7] and fined
tuned in the 1990s [8]. Although the phenomenological approach
is the most reliable as it is a mechanistic model, it does require a
lot of empirical input in the form of rates of entrainment and rede-
position, etc. [9]. Therefore, there has not yet developed a universal
correlation for CHF prediction up to now.
Alternatively, advanced information processing approaches and
numerical optimization techniques have been applied to predict
CHF. In the past two decades, artificial neural networks (ANNs),
as one of artificial intelligence (AI) techniques, have been used to
predict CHF. There have been many studies on this topic. Most of
them were summarized in Table 4 of [10] and the others were re-
ported in the references [11,12]. Besides, genetic algorithm (GA), as
one of the optimization techniques, has also been used to predict
CHF [13]. Generally, these above studies can be divided into two
categories: one is that ANNs alone were used to predict CHF and
the other is that ANNs were combined with other techniques to
predict CHF. In the first category, one commonly used type of ANNs
was back-propagation network (BPN), since BPN has the character-
istics of simple structure and easy implementation. Besides, two
other types of ANNs, radial basis function networks (RBFNs) and
high order neural network (HONN), were also used in this category.
In the second category, ANNs were combined with wavelet trans-
form [11], fuzzy theory [12] and GA [14] for CHF prediction and
approximation. Additionally, as can be seen from Table 4 of [10],
some other researchers proposed adaptive network-based fuzzy
inference systems (ANFIS) and genetic neural network (GNN) to
predict CHF. Both ANFIS and GNN methods were also a combina-
tion of ANNs and fuzzy theory, GA. Compared with using ANNs
alone, those methods combined with the advantages of ANNs
and other techniques. In the GNN method, GA was used to opti-
mize the weight and threshold of BPN. In the proposed nonpara-
metric model of [14], GA was used to find useful input features
in BPN. Thus, an accurate degree of CHF prediction and approxima-
tion was obtained by these methods. Compared with conventional
modeling approaches, ANNs do not require a deep knowledge of
CHF phenomena or their best-fit correlations. However, ANNs have
some limitations due to the algorithm itself, such as depending on
researchers experience or knowledge to select structure parame-
ters, difficulty in coming up with a reasonable interpretation of
u weight of the network
Xe equilibrium quality
Xc critical quality
Greek symbols
v tunable parameter
ni slack variable
b deterministic annealing schedule
g learning constant
Dhi inlet subcooling (kJ/kg)
e insensitivity zone
a lagrangian multiplier
k weight of the network
u influence function
the overall structure of the network [15], and easily getting stuck
in a local minimum.
Recently, support vector regression (SVR) is considered to be a
promising technique that can overcome the drawbacks of ANNs.
SVR is the application of support vector machine (SVM) to the gen-
eral regression problem. SVM, originally developed by Vapnik and
his colleagues [16], is a machine learning method based on statis-
tical learning theory (SLT). SVR implements the structural risk min-
imization principle (SRM) principle, which has been shown to be
superior to the traditional empirical risk minimization (ERM) prin-
ciple employed by conventional ANNs. It is the difference that
equips SVR with many attractive features and good generalization
performance, which is the goal in statistical learning. In the last
few years, SVR has been used as an alternative method to ANNs
in many nuclear engineering applications [17–23]. In [17,18],
SVR alone was applied to two different prediction problems. In
[19,20], SVR was combined with GA, in which GA was used to opti-
mize the parameters of SVR. In [21–23], SVR was used to predict
departure from nucleate boiling ratio (DNBR), which is an impor-
tant design parameter for water-cooled reactors. As we all know,
departure from nucleate boiling (DNB) is one type of CHF. Although
these studies have been carried out in various areas of nuclear
engineering, few studies have been conducted on the use of SVR
for prediction of CHF with the exception of studies by Cai
[24,25]. In [24,25], SVM was applied to predict CHF in concen-
tric-tube open thermosiphon. The parameters of SVM were opti-
mized using a stepwise searching method [24] and chaotic
particle swarm optimization (CPSO) [25]. However, the effect of
various input parameters on CHF, which was important to develop
reliable prediction models, was not investigated in [24,25].
In this paper, a hybrid model that combines m-SVR and RBFNs is
presented to predict CHF. The proposed hybrid model is based on
the following two considerations. First, there are two common
types of SVR based on different loss functions, e-SVR and its mod-
ified version, m-SVR. All SVR used in [17–25] were e-SVR. However,
a shortcoming of e-SVR is that it is difficult to choose an appropri-
ate value of e- in practice. To avoid this difficulty, Schölkopf et al.
[26] proposed m-SVR, which introduces a new parameter v in the
primal problem to trade-off the tube size against model complexity
and empirical risk. We thus choose m-SVR in our hybrid model. Sec-
ond, because the kernels of the m-SVR are similar to the basis func-
tions of the RBFNs with scatter partitioning [27], we use RBFNs as
another component of our hybrid model. However, since RBFNs are
a kind of ANNs, they also have the disadvantages of ANNs. One dis-
advantage is that it is difficult to determine the number of neurons
 B.T. Jiang, F.Y. Zhao / International Journal of Heat and Mass Transfer 62 (2013) 481–494 483

Fig. 1. The structure of the hybrid model based on SVs.

Table 1
Experimental range of CHF dataset.

Parameters Ranges
Fixed inlet and outlet conditions [29] Local conditions [28]
Tube diameter, D (mm) 6, 8, 10, 12 8
Heated length, L (m) 0.3–1.77 0.64–19.88
Pressure, P (MPa) 0.106–0.951 0.1–21
Mass flow rate, G (kg/m2 s) 20–277 0–8000
Input Inlet subcooling, Dhi (kJ/kg) 50–654 –
Variables Critical quality, Xe (–) – 0.5–1.0
Exit quality, X (–) 0.323–1.251 –
Output Critical heat flux 0.108–1.598 0–44.338
Variable QCHF (MW/m2)
Working fluid Water Water
No. of data 513 24781

Fig. 2. The whole process of CHF prediction.

Table 2
Optimal values of three parameters and SVs.

C r v SVs Initial structure

Fixed inlet conditions 51.2 0.76 0.18 378 5–378–1
Local conditions 68.4 18.62 0.37 4202 3–4202–1
Fixed outlet conditions 65.6 1.68 0.24 381 5–381–1
484 B.T. Jiang, F.Y. Zhao / International Journal of Heat and Mass Transfer 62 (2013) 481–494

Fig. 3. Comparisons between the hybrid model and e-SVR. (a) Fixed inlet conditions (b) Local conditions (c) Fixed outlet conditions.

Table 3
Comparison of prediction errors between the hybrid model and e-SVR.

Fixed inlet conditions Local conditions Fixed outlet conditions

v-SVR + RBFNs e-SVR v-SVR + RBFNs e-SVR v-SVR + RBFNs e-SVR
Training data Test data Training data Test data Training data Test data Training data Test data Training data Test data Training data Test data
RMSE (%) 1.013 1.189 2.478 2.341 1.172 1.109 2.095 2.532 1.69 1.118 2.528 2.192
MAPE (%) 1.298 1.097 2.424 2.212 1.734 1.591 2.471 2.181 1.343 1.252 2.167 2.233
 B.T. Jiang, F.Y. Zhao / International Journal of Heat and Mass Transfer 62 (2013) 481–494 485

Fig. 4. Comparisons between the hybrid model and empirical correlations.

in the hidden layer. To overcome this disadvantage, we use the subject to

support vectors (SVs) as the hidden layer neurons of RBFNs. Then 8
> yi  wT uðxÞ  b 6 e þ ni
the resulting RBFNs can be viewed as the hybrid model. It should >
<
be noted that there are two differences between the initial struc- wT uðxÞ þ b  yi 6 e þ ni ð1Þ
ture of the hybrid model and that of traditional RBFNs. The first >
>
:
is that the center of each basis function in the initial structure of ni ; ni P 0; i ¼ 1; 2; . . . ; N; e P 0
the hybrid model corresponds to one support vector of m-SVR. where 0 6 v 6 1, C is a regularization parameter. It should be noted
The second is that the SVs and output weights are determined that here e is not a priori but a variable to optimize. However, e has
automatically by m-SVR algorithm itself. These two differences the reverse effect on e-SVR. The advantage of Eq. (1) lies in the prop-
make the hybrid model more advantageous since it not only re- erties of the constant v. It is proved that the parameter v is an upper
tains the advantage of m-SVR, but also overcomes the shortcoming bound on the fraction of margin errors and lower bound of the frac-
of traditional RBFNs. tion of SVs [31] and with probability 1, asymptotically, v is equal to
Predictions of CHF for two special cases are carried out in this both fractions, as shown in the following:
paper: CHF for water flow in vertical round tubes and dryout  NNE 6 v ; where NE is the number of margin error vectors
CHF for deionized water in a vertical narrow annular channel with
(jyi  ðwT uðxÞ þ bÞj > e) in the training set.
0.95 mm gap. In the first case, CHF is predicted based on three con- NS
 P v ; where N is the number of SVs (jyi  ðwT uðxÞ þ bÞj 6 e) in
ditions: fixed inlet conditions, local conditions and fixed outlet N

conditions. We use the dataset of 2006 CHF look-up table [28]
for local conditions and the dataset of Kim et al. [29] for fixed inlet
and outlet conditions. The predicted results are compared with
those of e-SVR and empirical correlations. In the second case, the
dataset is collected from the literature [30]. In [30], BPN was used
to predict the dryout type CHF. To further validate the effectiveness
of the hybrid model, the predicted results are compared with those
of the BPN.
the training set.
Although this primal formulation is different from that of e-SVR,
it is still a linearly constrained quadratic programming problem.
Thus, it can be solved by using multiplier techniques and applying
the Karush–Kuhn–Tucker (KKT) conditions to its corresponding
dual problem. Then the regression function, which has the identi-
cal form to that of e-SVR, is given by

X
N
f ðxÞ ¼ ðai  ai Þkðxi ; xÞ þ b ð2Þ
2. The proposed hybrid model
i¼1

2.1. m-support vector regression (v-SVR) where kðxi ; xj Þ ¼ huðxi Þ; uðxj Þi, where the notation h,i represents
the inner product, is the kernel function. The commonly used kernel
As described in the introduction, since it is difficult to select an functions consist of the linear kernel, the Gaussian kernel function
appropriate e for e-SVR, m-SVR is introduced as a modified version and the polynomial kernel with different degrees. In this paper,
of e-SVR. However, because we use m-SVR in our hybrid model, we the following Gaussian kernel is used as a kernel function due to
only give a brief description about it. More detailed about the the- its robustness and symmetry properties.
ory of e-SVR can be found in [31]. Given a set of data points
!
G ¼ ðxi ; yi ÞNi¼1 ; where xi e Rm is the input vector; yi e R is the actual
kxi  xj k2
value; N is the total number of data points. The primal problem of kðxi ; xj Þ ¼ exp  ð3Þ
2r2
m-SVR is given as follows:
!
From Eq. (2), it can be seen obviously that only the training data
1 1X N
min kwk2 þ C  ve þ ðn þ nk Þ points xi with jai  ai j – 0 have an impact on the regression func-
w;b;e;nk ;nk 2 N k¼1 k
tion f(x). All the other data points could be deleted from the data
486 B.T. Jiang, F.Y. Zhao / International Journal of Heat and Mass Transfer 62 (2013) 481–494

Fig. 5. Effect of system pressure on CHF. (a) Fixed inlet conditions, (b) Local conditions, (c) Fixed outlet conditions.

set and the same solution would be obtained. Thus, these training 2.2. Hybrid model
data points are called SVs. Let SV be the support vector set, then
Eq. (2) can be reformulated as: The hybrid model is implemented in two stages, initializing and
adjusting. In the first stage, the initial structure and weights of the
X hybrid model are determined by m-SVR. To facilitate the further
f ðxÞ ¼ ðai  ai Þkðxi ; xÞ þ b ð4Þ
xi 2SV discussion, we rewrite Eq. (4) as follows:
 B.T. Jiang, F.Y. Zhao / International Journal of Heat and Mass Transfer 62 (2013) 481–494 487

X
P on the above considerations, we believe that it is a promising strat-
^¼
y ki kðui ; xÞ þ b ð5Þ egy to use the SVs as the neurons in the hidden layer, because this
i¼1
strategy not only maintains the advantage of m-SVR, but also over-
comes the shortcoming of traditional RBFNs.
where y ^i is the output of m-SVR network. P is the number of kernel
In the second stage, the weights of the hybrid model are ad-
functions, which is equivalent to the number of SVs. The decision
justed using the ARBP learning algorithm proposed in [32]. In this
function given by Eq. (5) can be viewed as a three-layer neural net-
learning algorithm, a robust cost function of the hybrid model is
work, as shown in Fig. 1. From Eq. (5) and Fig. 1, it is intuitive to
defined as
consider the network as RBFNs using normalization basis function.
We give a more detailed mathematical derivation in the Appendix.  
X
P X
P
bðtÞ e2 ðtÞ
From Fig. 1, we can see that the hidden layer of the network is com- ER ¼ r½ei ðtÞ; bðtÞ ¼ ln 1 þ i ð6Þ
posed of kernel functions of m-SVR. The number of neurons in the i¼1 i¼1
2 bðtÞ
hidden layer corresponds to the number of SVs. Discussions on
^i ðtÞ is the error between
where t is the epoch number. ei ðtÞ ¼ yi  y
the difference between the network and traditional RBFNs are de-
the ith desired output and the ith output of the hybrid model at
scribed in the introduction.
epoch t. b(t) is a deterministic annealing schedule acting like the
In this paper, the choice of the SVs as the neurons in the hidden
scale estimator in robust learning algorithms. r() is the so-called
layer is based on two considerations. First, SVs are automatically
logistic loss function. As we can see from Fig. 1, there are two
selected from the training data during the learning process in m-
weights, k and u, that need to be adjusted for the obtained network.
SVR and they contain all the essential information needed to define
In ARBP algorithm, the weights k and u can be updated based on the
the final decision function. Second, SVs are used to provide sparse
following gradient descent method:
representation. In supervised learning based on kernel functions
like SVR, sparseness is viewed as a desirable property since it can
lead to a structural simplification of the learned model by remov- ki ðt þ 1Þ ¼ ki ðtÞ þ Dki ðtÞ ð7Þ
ing its irrelevant or redundant components. Accordingly the gener-
alization capability of the learned model can be improved. Based ui ðt þ 1Þ ¼ ui ðtÞ þ Dui ðtÞ ð8Þ

Fig. 6. Effect of mass flux on CHF. (a) Fixed inlet conditions, (b) Local conditions for different pressures under constant critical qualities, (c) Local conditions for different
critical qualities under constant pressure, (d) Fixed outlet conditions
488 B.T. Jiang, F.Y. Zhao / International Journal of Heat and Mass Transfer 62 (2013) 481–494

Fig. 6. (continued)

with performance, we have to preprocess the CHF data before applying

the hybrid model. These data are standardized to zero mean and
@ER @ei ðtÞ
Dki ¼ g ¼ guðei ; bÞ ¼ guðei ; bÞkðui ; xÞ ð9Þ unit variance.
@ki @ki

@ER @ei ðtÞ ui  x 3.2. Parameters initialization

Du i ¼  g ¼ guðei ; bÞ ¼ guðei ; bÞki kðui ; xÞ
@ui @ui r2
where g is learning constant. u() is usually called the influence
@ rðei ðtÞ;bðtÞÞ
function and is defined as uðei ; bÞ ¼ @ei ðtÞ
¼ 1þe2i ðtÞ=bðtÞ
3. Prediction of CHF in vertical round tubes
3.1. Dataset description and preprocessing
In this section, we use the hybrid model to predict CHF for
water flow in vertical round tubes. The dataset used in this section
is collected from the published literature [28,29]. The data of 2006
CHF look-up table [28] is used for local conditions and the data of
Kim et al. [29] is for fixed inlet and outlet conditions. The experi-
mental range of the dataset is given in Table 1. From Table 1, we
can see that D, L, P, G, Dhi, Xe and X are the input variables, which
denote tube diameter, heated length, pressure, mass flow rate, inlet
subcooling, equilibrium quality and exit quality; QCHF is the output
variable, which represents CHF value. To increase the prediction
ð10Þ
The CHF data is partitioned into the training data and test
ei ðtÞ data. The training data are selected using the SC scheme [33] in
.
order to obtain more informative data. The SC scheme has also
been used an efficient data-based modeling method to select
training data in [21–23]. The rest data is used as the test data,
which does not include the training data. As can be seen from Ta-
ble 1, CHF is predicted under three different conditions, thus
three separate hybrid models are established for the correspond-
ing conditions. We summarize the whole process of CHF predic-
tion in Fig. 2. As shown from Fig. 2, to establish the hybrid
model, three parameters C, r and m should be optimized firstly.
In this paper, we employ particle swarm optimization (PSO)
[34] technique to obtain the optimal C and r. The optimal value
of m is obtained by the method mentioned in [26]. The optimal
parameters and initial structures of the hybrid models for three
conditions are shown in Table 2. The parameters used in the ARBP
algorithm are set as follows: the learning rate g = 0.04; the
momentum coefficient is 0.6 in this paper.
 B.T. Jiang, F.Y. Zhao / International Journal of Heat and Mass Transfer 62 (2013) 481–494
Fig. 7. Effect of inlet subcooling or critical quality on CHF. (a) Fixed inlet conditions, (b) Local conditions.
3.3. Prediction of CHF
In this section, CHF for water flow in vertical tubes is predicted
by the hybrid model. The predicted results are compared with that
of e-SVR. For comparison, both methods are trained and tested
using the same data. Comparisons of the predicted results with
the experimental data are shown in Fig. 3. As can be seen from
Fig. 3(a) and (c), the ratio of predicted results obtained by the hy-
brid model to experimental data is closer to 1 than that of pre-
dicted results obtained by the e-SVR to experimental data. From
Fig. 3(b), it can be seen that although the two figures are both pro-
vided with a ±3% error band based on error analysis and a diagonal
straight line indicating the perfect mapping, the predicted results
of the hybrid model are more consistent with the experimental
data than those of the e-SVR. Additionally, we summarize the com-
parison of performance evaluation in Table 3. In Table 3, two statis-
tical criteria are considered to evaluate the performance of the
hybrid model, namely, root mean squared error (RMSE) and mean
absolute percentage error (MAPE). RMSE is a commonly used per-
formance measure of determining estimation accuracy. MAPE
gives the mean ratio between the errors and the actual values, as
given by Eq. (11)
N  
1X ^
yi  yi   100
MAPE ¼
N i¼1  yi 
^i is the predicted value and N
where yi is the experimental value, y
denotes the number of the data points. As shown in Table 3, the
489
RMSE and MAPE values of the training and test data for the hybrid
model are both smaller than those obtained by the e-SVR. From
Fig. 3 and Table 3, we can see that the proposed hybrid model is
suitable and has superior performance over the e-SVR.
Finally, we also compare the hybrid model with two empirical
correlations. Kottowski et al. [35] empirical correlation is derived
based on 170 data points. These data points include their own data
and others. The correlation is given as follows:
Q CHF ¼ 0:216ðdH=LÞ0:8 ð1  2xÞG0:807 Hfg ð12Þ
The data points have the following ranges of variables: G = 50–
800 kg/m2 s, x = 0.4–0, L/dH = 30–125. B&W-2 correlation is de-
rived based on 207 experimental data points and is described as
the following equation::
Q CHF ¼ 3:155  103 ð1:155  16:025dHÞ  107 ð4:3604
 104 GÞB  48:21GxHfg =½12:71ð2:252  103 GA Þ ð13Þ
where A = 0.712 + 3.006  105(P  13790), B = 0.834 + 9.93  105
(P  13790). The corresponding ranges of parameters are as fol-
lows: P = 13.79–16.55 MPa, G = 1024–5425 kg/m2 s, Xc = 0.3–0.2,
dH = 0.0051–0.0127 m. The comparison of the predicted results be-
tween the hybrid model and the two correlations is shown in Fig. 4.
ð11Þ
From Fig. 4, it can be seen that the hybrid model has a better agree-
ment with the experimental data than those of the two correlations.
Therefore, the hybrid model has been applied successfully to predict
CHF for water flow in vertical tubes.
490 B.T. Jiang, F.Y. Zhao / International Journal of Heat and Mass Transfer 62 (2013) 481–494
4. Parametric trends of CHF
In this section, we investigate the influence of the principal
parameters on CHF using the hybrid model.
4.1. Effect of system pressure
The effect of system pressure on CHF is shown in Fig. 5. Fig. 5(a)
and (c) show the results for fixed inlet and outlet conditions,
respectively. We can see that the results in Fig. 5(a) and (c) show
the similar trends. CHF slowly increases with the increasing pres-
sure at high mass flux, but it approximately does not change with
the pressure increase at low mass flux and low pressure conditions.
Fig. 5(b) shows the results for local conditions. From Fig. 5(b), we
can clearly see that CHF increases with pressure at low pressures,
passes through a maximum, and then gradually decreases at high
pressures. This trend agrees with the general understanding.
4.2. Effect of mass flux
Fig. 6 illustrates the mass flux effect on CHF. Fig. 6(a) shows the
results for fixed inlet conditions. From Fig. 6(a), we can see that CHF
increases with mass flux increase. Fig. 6(a) also shows that CHF data
for the large L/D value lie nearly on a single line. Both Fig. 6(b) and
Fig. 8. Effect of heated length to diameter ratio on CHF. (a) Fixed inlet conditions, (b) Fixed outlet conditions.
(c) shows the results for local conditions. From both figures, we can
see that CHF generally increases with increasing mass flux for
Xe < 0. However, at low mass flux, CHF decreases with the increase
of mass flux, as shown in curve 4 of Fig. 6(b). From both figures, we
can also see that the mass effect on CHF appears to be complex for
Xe > 0. Under the condition of Xe > 0, CHF firstly decreases with mass
flux at low mass flux, goes through a minimum, and then increases
gradually at higher mass flux. Fig. 6(d) shows the results for fixed
outlet conditions. As can be seen from Fig. 6(d), the trends are gen-
erally similar to that in Fig. 6(a). However, there are some complex-
ities in the low mass flux and high exit quality regions
(P = 0.9 MPa, X = 0.8). Under these conditions, CHF increases with
mass flux at first, sharply decreases to a minimum, and then in-
creases gradually, as shown in curve 2 of Fig. 6(d).
4.3. Effect of inlet subcooling or equilibrium quality
The inlet subcooling effect on CHF is indicated in Fig. 7. The re-
sults for fixed inlet conditions are shown in Fig. 7(a). From Fig. 7(a),
we can see that CHF increases with increasing inlet subcooling. The
inlet subcooling effect decreases as the mass flux decreases and the
effect becomes small as the heated length increases. Fig. 7(b)
shows the results for local conditions. As can be seen from
Fig. 7(b), CHF decreases with increasing quality for both Xe < 0
 B.T. Jiang, F.Y. Zhao / International Journal of Heat and Mass Transfer 62 (2013) 481–494 491

Fig. 9. Schematic diagram of experimental test loop.

and Xe > 0. It also shows that CHF increases with increasing mass
flux at Xe < 0, but decreases with increasing mass flux at Xe > 0.
However, at high qualities, CHF remains almost constant regard-
less of the quality increase.
4.4. Effect of heated length to diameter ratio
Fig. 8 represents the effect of heated length to diameter ratio on
CHF. As can be seen from Fig. 8, the results obtained under fixed
inlet conditions (Fig. 8(a)) have similar trends to those obtained
under fixed outlet conditions (Fig. 8(b)). CHF decreases with
increasing L/D under a fixed diameter D, which means that the
longer tube results in the lower CHF. The effect of L/D becomes
smaller as the length increases. Moreover, CHF decreases with
decreasing mass flux for a given length.
In conclusion, it is important to assess the reliability of the pre-
diction model according to the analysis of parametric trends of
CHF. However, in some previous studies, there are several inade-
quacies in the section of the parametric trends analysis of CHF.
The inadequacies mainly are lack of comprehensive comparison
between different CHF prediction methods. To overcome the inad-
equacies, we provide a more comprehensive comparison and eval-
492 B.T. Jiang, F.Y. Zhao / International Journal of Heat and Mass Transfer 62 (2013) 481–494

Fig. 10. Comparisons between the hybrid model and BPN. (a) Inner tube, (b) Outer tube.

Table 4
Comparison of prediction errors between the hybrid model and BPN.

Inner tube dryout Outer tube dryout

v-SVR + RBFNs BPN v-SVR + RBFNs BPN
Training data Test data Training data Test data Training data Test data Training data Test data
RMSE (%) 1.773 1.518 9.368 8.597 1.568 1.953 7.238 8.876
MAPE (%) 1.637 1.251 7.484 7.627 2.229 2.072 8.622 6.213

uation of the performance of the hybrid model using the experi-
mental data as a reference. In this comprehensive comparison,
the prediction results of the hybrid model are compared not only
with that of the e-SVR, but also with the experimental data. As
can be seen from Figs. 5–8, the results obtained from the e-SVR
have a good agreement only with some of the experimental data
under given conditions. However, by comparison, the results ob-
tained from the hybrid model are consistent with all experimental
data over the full range of parameters. Additionally, the observed
parametric trends based on the hybrid model are found to agree
with the previous understanding. Therefore, the analysis of para-
metric trends of CHF shows the better accuracy of the hybrid mod-
el with regard to variations of pressure, mass flux, inlet subcooling
or critical quality and heated length to diameter ratio, as compared
with the e-SVR method.
5. Prediction of CHF in narrow annuli
5.1. A brief introduction to experimental setup
In this section, we briefly describe the experimental setup and
more details about it can be found in [30]. The schematic diagram
of the dryout experimental test loop is shown in Fig. 9. As shown
in Fig. 9(a), the experimental system consists of two main loops,
the primary loop and secondary loop. The primary loop is pro-
vided with a circulating water pump, pressurizer, serpentine pre-
heater, calibrated flow water, U-shaped preheater, test section,
condenser, valves and connected pipes. The secondary loop is
connected to the primary loop through a condenser, which is
used to condense the fluid in the primary loop. The working fluid
is deionized water.
Fig. 9(b) shows the schematic diagram of the test section. The
test section is composed of two vertical concentric stainless steel
tubes (outside and inside tubes) with the fluid flowing through
the annular space between the two tubes. The geometrical diam-
eters of the test section are given as follows: length in 835 mm,
10 mm inner diameter and 2 mm thickness of outside tube,
8.1 mm outer diameter and 1 mm thickness of inside tube. Thus,
the corresponding annular gap size of the test section is
0.95 mm.
112 groups of experimental data are obtained and ranges of
experimental parameters are summarized as the following:
P = 0.67–4.0 Mpa, G = 34–140 kg/m2 s, Xc = 0.7–0.99, qo = 8.3–
81.5 kW/m2, qi = 14.9–40.5 kW/m2.
 B.T. Jiang, F.Y. Zhao / International Journal of Heat and Mass Transfer 62 (2013) 481–494 493

hP ihP i hP i
P P P
5.2. Prediction of CHF X
P
i¼1 kðui ; xÞ i¼1 ki kðui ; xÞ þ i¼1 kðui ;xÞ b
^ðxÞ ¼
y ki kðui ;xÞ þ b ¼ PP
i¼1 i¼1 kðui ;xÞ
0 1 0 1 0 1
This section presents that the dryout type CHF for deionized 1 k1 1
water upflowing in a vertical narrow annular channel with B .. C B . C B.C
B C
½kðu1 ; xÞ;... kðuP ;xÞ@ . A½kðu1 ; xÞ;.. .kðuP ;xÞ@ . A þ ½kðu1 ;xÞ;.. .kðuP ;xÞ@ .. A  b
.
0.95 mm gap is predicted using the hybrid model. 112 groups of 1 1
kP
experimental data are obtained from [30]. The experimental data ¼ PP
i¼1 kðui ;xÞ
consist of inner tube dryout type CHF and outer tube dryout type 0 1 0 1
k1 1
CHF. The CHF data is first normalized into zero mean and unit var- B . C
C þ ½kðu1 ;xÞ;. ..kðuP ; xÞB .C
½kðu1 ; xÞ;... kðuP ;xÞT B .
@ . A @ .. A  b
iance, and then is split into two parts: a training set with 75% of the
kP 1
data and a test set with the 25% of the data. The training set and ¼ PP
i¼1 kðui ; xÞ
test set are selected randomly, both of which have the same size
ð16Þ
as that in [30]. The predicted results are compared with those of
0 1 0 1
BPN method, as shown in Fig. 10. For a fair comparison, the param- 1 kðu1 ; xÞ ... kðuP ; xÞ
. B .. .. .. C
eters of the BPN are kept identical to that used in [30]. From Fig. 10, where T ¼ @ .. A½kðu1 ; xÞ; . . . kðuP ; xÞ ¼ @ . . . A.
we can clearly see that the predicted results have a better agree- 1 kðu1 ; xÞ ... kðuP ; xÞ
ment with the experimental data than those of the BPN. Table 4 Rearranging Eqs. (16) and (5) can be rewritten as
shows the comparison of performance measures. As shown in 2 0
1 0 1 3
Table 4, the RMSE and MAPE values of the hybrid model are smal- k1 1
6 B . C B.C 7
ler than that of the BPN. From Fig. 10 and Table 4, it can be seen ½kðu1 ; xÞ; . . . kðuP ; xÞ4T @ . A þ @ .. A  b7
6 B . C
5
that the predicted performance of the hybrid model is superior
to that of the BPN. Thus the hybrid model can be used as an effec- kP 1
^ðxÞ ¼
y PP
tive tool for prediction of CHF. i¼1 kðui ; xÞ
0 1
w1
B . C
6. Conclusions ½kðu1 ; xÞ; . . . kðuP ; xÞB @ .. A
C

wP
In this paper, a hybrid model is presented to predict CHF. The ¼¼ PP ð17Þ
prediction of CHF for two different cases are provided: water i¼1 kðui ; xÞ

flow in vertical round tubes and deionized water flow upward where the weights w1, . . . wP can be expressed in terms of
in a vertical narrow annular channel with 0.95 mm gap. In the fk1 ; . . . kP ; bg as the following
first case, CHF is predicted using the hybrid model based on fixed 0 1 0 1 0 1
inlet conditions, local conditions and fixed outlet conditions. The
w1 k1 1
B . C B . C B.C
hybrid model yields smaller RMSE and MAPE values than the B . C ¼ T B . C þ @ .. A  b ð18Þ
@ . A @ . A
standard e-SVR method. In addition, the hybrid model gives bet-
wP kP 1
ter performance than the empirical correlations. The effects of
the input parameters on CHF are analyzed using the hybrid mod- From Eqs. (17) and (18), the RBFNs using normalization a basis
el. The results obtained by the hybrid model not only show a function is obtained as follows
general agreement with the previous studies, but also fit the PP XP
experimental data better than those obtained by the standard ^ðxÞ ¼ Pi¼1 wi kðui ; xÞ
y P
¼ wi Nðui ; xÞ ð19Þ
e-SVR method. In the second case, the hybrid model is used to i¼1 kðui ; xÞ i¼1
predict the dryout type CHF. It is found that the predicted results
agree well with the experimental data. The results also show that where Nðui ; xÞ ¼ Pkðu
P
i ;xÞ
is the normalized basis functions and the
kðui ;xÞ
i¼1
the hybrid model, with narrower error bands and smaller RMSE network given by Eq. (19) is referred to as the m-SVR + RBFNs.
and MAPE values, outperforms the BP-ANNs method. The pro-
posed methodology allows us to achieve accurate results, thus References
it is suitable for the thermal–hydraulic and heat transfer data
processing like CHF. [1] J.G. Collier, J.R. Thome, Convective Boiling and Condensation, third ed., Oxford
University Press, New York, 1994.
[2] S.M. Ghiaasiaan, Two-Phase Flow, Boiling and Condensation in Conventional
and Miniature Systems, Cambridge University Press, 2008.
Appendix A. [3] A.V. Dedov, Critical heat flowrates in subcooled flow boiling, Therm. Eng. 57 (3)
(2010) 185–192.
Derivation. of the hybrid model [4] L.S. Tong, Y.S. Tang, Boiling Heat Transfer and Two Phase Flow, second ed.,
Taylor and Francis, Washington, DC, 1997.
[5] D.C. Groeneveld, S.C. Cheng, T. Doan, 1986 AECL-UO critical heat flux lookup
The RBFNs using normalization basis functions is given by table, Heat Transfer Eng. 7 (1–2) (1986) 46–62.
Pn [6] D.C. Groeneveld, L.K.H. Leung, P.L. Kirillov, V.P. Bobkov, I.P. Smogalev, V.N.
hi lðXðkÞÞ Vinogradov, X.C. Huang, E. Royer, The, 1995 look-up table for critical heat flux
yðkÞ ¼ Pi¼1
n ð14Þ
i¼1 lðXðkÞÞ
in tubes, Nucl. Eng. Des. 163 (1996) 1–23.
[7] P.B. Whalley, P. Hutchinson, G.F. Hewitt, The calculation of critical heat flux in
where hi is the ith weight. X(k) = [x1(k),. . . xN(k)]T (k = 1, 2. . . m) is the forced convective boiling, in: Fifth International Heat Transfer conference,
Tokyo, Paper No. B6. 11, 1974.
input data. li() is the ith Gaussian basis function given by Eq. (3) [8] G.F. Hewitt, A.H. Govan, Phenomenological modelling of non-equilibrium
! flows with phase change, Int. J. Heat Mass Transfer 33 (1990) 229–242.
jXðkÞ  X0 ðiÞj2 [9] B.R. Vijayarangan, S. Jayanti, A.R. Balakrishnan, Studies on critical heat flux in
li ðXðkÞÞ ¼ exp  ð15Þ flow boiling at near critical pressures, Int. J. Heat Mass Transfer 49 (2006) 259–
2r 2
268.
[10] T.L. Cong, G.H. Su, S.Z. Qiu, W.X. Tian, Applications of ANNs in flow and heat
For derivation convenience, we reformulate the Eq. (6) and give the transfer problems in nuclear engineering: a review work, Prog. Nucl. Energy 62
following derivation (2013) 54–71.
494 B.T. Jiang, F.Y. Zhao / International Journal of Heat and Mass Transfer 62 (2013) 481–494
[11] S.H. Kim, I.C. Bang, W.P. Baek, S.H. Chang, S.K. Moon, CHF detection using
spationtemporal neural network and wavelet transform, Int. Commun. Heat
Mass 27 (2) (2000) 285–292.
[12] S.K. Moon, S.H. Chang, Classification and prediction of the critical heat flux
using fuzzy theory and artificial neural networks, Nucl. Eng. Des. 150 (1)
(1994) 151–161.
[13] D.G. Lee, H.G. Kim, W.P. Baek, S.H. Chang, Critical heat flux prediction using
genetic programming for water flow in vertical round tubes, Int. Commun.
Heat Mass 24 (7) (1997) 919–929.
[14] Y.K. Kwon, B.R. Moon, S.D. Hong, Critical heat flux function approximation
using genetic algorithms, IEEE Trans. Nucl. Sci. 52 (2) (2005) 535–545.
[15] V. Kecman, Learning and Soft Computing: Support Vector Machines, Neural
Networks, and Fuzzy Logic Models, MIT Press, Cambridge, MA, London,
England, 2001.
[16] V.N. Vapnik, The Nature of Statistical Learning Theory, Springer-Verlag, New
York, 2000.
[17] I.H. Bae, M.G. Na, Y.J. Lee, G.C. Park, Calculation of the power peaking factor in a
nuclear reactor using support vector regression models, Ann. Nucl. Energy 35
(12) (2008) 2200–2205.
[18] M.G. Na, J.W. Kim, D.H. Lim, Y.J. Kang, Residual stress prediction of dissimilar
metals welding at NPPs using support vector regression, Nucl. Eng. Des. 238
(7) (2008) 1503–1510.
[19] M.G. Na, H.Y. Yang, Power reconstruction of fuel rods by support vector
regression for CANDU reactors, IEEE Trans. Nucl. Sci. 54 (4) (2007) 1302–1310.
[20] M.G. Na, B.R. Upadhyaya, Model predictive control of an SP-100 space reactor
using support vector regression and genetic optimization, IEEE Trans. Nucl. Sci.
53 (4) (2006) 2318–2327.
[21] D.H. Lim, H.Y. Yang, M.G. Na, Monitoring minimum DNBR using a support
vector regression model, IEEE Trans. Nucl. Sci. 56 (1) (2009) 286–293.
[22] D.S. Kim, S.W. Lee, M.G. Na, Prediction of axial DNBR distribution in a hot fuel
rod using support vector regression models, IEEE Trans. Nucl. Sci. 58 (4) (2011)
2084–2090.
[23] S.W. Lee, D.S. Kim, M.G. Na, Prediction of DNBR using fuzzy support vector
regression and uncertainty analysis, IEEE Trans. Nucl. Sci. 57 (3) (2007) 1595–
1601.
[24] J.J. Cai, Applying support vector machine to predict the critical heat flux in
concentric-tube open thermosiphon, Ann. Nucl. Energy 43 (2012) 114–122.
[25] J.J. Cai, Predicting the critical heat flux in concentric-tube open thermosiphon:
a method based on support vector machine optimized by chaotic particle
swarm optimization algorithm, Heat Mass Transfer 48 (8) (2012) 1425–1435.
[26] B. Schölkopf, A. Smola, P. Bartlett, R. Williamson, P. Bartlett, New support
vector algorithms, Neural Comput. 12 (2000) 1207–1245.
[27] B. Schölkopf, K. Sung, C.J.C. Burges, F. Girosi, P. Niyogi, T. Poggio, V. Vapnik,
Comparing support vector machines with Gaussian kernels to radial basis
function classifiers, IEEE Trans. Signal Proces. 45 (11) (1997) 2758–2765.
[28] D.C. Groeneveld, J.Q. Shan, A.Z. Vasic, L.K.H. Leung, A. Durmayaz, J. Yang, S.C.
Cheng, A. Tanase, The, 2006 CHF look-up table, Nucl. Eng. Des. 237 (2007)
1909–1922.
[29] H.C. Kim, W.P. Baek, S.H. Chang, Critical heat flux of water in vertical round
tubes at low pressure and low flow conditions, Nucl. Eng. Des. 99 (2000) 49–
73.
[30] F. Duan, Numerical simulation on dryout point of annular upward flow in
narrow annuli with bilateral heating, M.S. thesis, Xi’an Jiaotong University,
Xi’an, 2007.
[31] V. Vapnik, Statistical Learning Theory, Wiley, New York, 1998.
[32] C.C. Chuang, S.F. Su, C.C. Hsiao, The annealing robust back propagation (ARBP)
learning algorithm, IEEE Trans. Neural Network 11 (2000) 1067–1077.
[33] S.L. Chiu, Fuzzy model identification based on cluster estimation, J. Intell.
Fuzzy. Syst. 2 (3) (1994) 267–278.
[34] J. Kennedy, R.C. Eberhart, Particle swarm optimization, in: Proceedings of IEEE
International Conference on Neutral Networks, Perth, Australia, 1995,
pp.1942–1948.
[35] H.M. Kottowski, C. Savatteri, W. Hufschmidt, A new critical heat flux
correlation for boiling liquid metals, Nucl. Eng. Des. 108 (1991) 396–413.