2009 World Congress on Computer Science and Information Engineering

Prediction of Refrigerant Mass Flow Rates through Capillary Tubes Using

Adaptive Neuro-Fuzzy Inference System

Hui Xie, Fei Ma, Huifang Fan Yanqiang Di

School of Civil & Environment Engineering, China Academy of Building Research,
University of Science and Technology Beijing, Beijing, P R China
Beijing, P R China
E-mail: [email protected]

Abstract refrigerants are being replaced by ecologically

acceptable alternatives as a consequence of the
A capillary tube is a common expansion device Montreal Protocol. For this reason and because some
widely used in small-scale refrigeration and air- of these new refrigerants used are zeotropic or near-
conditioning systems. Generalized correlation method azeotropic refrigerant mixtures, more accurate and
for refrigerant flow rate through adiabatic capillary general methods for predicting the mass flow rates of
tubes is developed by combining dimensional analysis these refrigerant mixtures are required.
and adaptive neuro-fuzzy inference system (ANFIS). Among the design tools of a capillary tube,
Dimensional analysis is utilized to provide the empirical correlations for mass flow rate have simple
generalized dimensionless parameters and reduce the form and are very convenient for engineering
number of input parameters, while a five-layer applications. Empirical correlations can be classified
feedforward ANFIS is served as a universal into dimensional and dimensionless correlations. In
approximator of the nonlinear multi-input and single- dimensional correlations [1–4], the refrigerant
output function. For ANFIS training and test, properties are not taken into account. Therefore, the
measured data for R134a, R22, R290, R407C, R410A, empirical coefficients in a correlation will be changed
and R600a in the open literature are employed. The for different refrigerants. By comparison, the
most suitable membership function and number of dimensionless correlations [5–7] considering the
membership functions are found as Gauss and two, refrigerant properties are more generalized for different
respectively, for the ANFIS correlation. The statistical refrigerants. Summing up the existing generalized
data can be considered as very promising. This paper dimensionless correlations in above literature, it can be
shows the appropriateness of ANFIS for the prediction seen that all the correlations were assumed as the
of refrigerant mass flow rates through capillary tubes. power-law function, which is a simple nonlinear one
and can be correlated as a linear function when taking
logarithm on both sides of the function. But there is no
1. Introduction evidence to prove the power-law function is the best
model for correlation. In fact, such function will
A capillary tube is a fixed length of small diameter become abnormal in the case of zero subcooling and
tubing lying between the outlet of the condenser and should be improved. Generally speaking, it is
the inlet of the evaporator. It reduces the pressure of impossible to pick out a perfect simple function
the refrigerant from the high-pressure side to the low- because of the complex flow mechanism in a capillary
pressure side of the system and controls the flow of tube. For such unknown nonlinear multi-input and
refrigerant to the evaporator. Capillary tube expansion single-output (multioutput) function, some artificial
devices are simple, reliable and inexpensive, so widely intelligence methods such as artificial neural networks
used in refrigeration equipment, such as household (ANNs) or fuzzy-logic can be used. ANNs have been
refrigerators and freezers and air conditioners. The in- successfully proved as universal approximators
tube flow is complex two-phase critical flow under because they do not require prior knowledge of the
typical operating conditions and the mass flow rate input data distribution [8,9]. Recently, ANNs were
through the capillary tube has a strong influence on the extended for the modeling of energy system [10], the
performance of the whole system. Halogenated refrigerating system [11,12], components [13–16] and

978-0-7695-3507-4/08 $25.00 © 2008 IEEE 774

769
DOI 10.1109/CSIE.2009.543
Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on March 06,2025 at 07:31:34 UTC from IEEE Xplore. Restrictions apply.
 refrigerant properties [17,18]. Meanwhile, the benefits 1
μ A (x1 ) = (3)
of fuzzy systems include rapid calculation times, fast 1 + [(x1 − ai ) / bi ] i
i c

and inexpensive construction, easy to understand,

where ai, bi and ci are membership function parameters.
tolerant of imprecise data, the modelling of non-linear
systems, and can be combined with ANNs to create
adaptive neuro-fuzzy inference system (ANFIS) [19].
ANFIS could be a promising method for modeling the
complex flow mechanism in a capillary tube.
In this study, a generalized correlation method for
refrigerant flow rate through adiabatic capillary tubes
was developed by combining dimensional analysis and
ANFIS. Dimensional analysis was utilized to provide
generalized dimensionless parameters and reduce the
number of input parameters, while ANFIS was served
as a generalized approximator of the nonlinear multi- Figure 1. ANFIS architecture with two rules
input and single-output function. The predictions from
the ANFIS correlation were compared with Layer 2: The nodes in this layer are fixed. These are
experimental data of R22, R134a, R290, R407C, labeled M to indicate that play the role of a simple
R410A, and R600a reported in the literature and the multiplier. The outputs, the firing strengths of the rules,
correlations of other researchers. are given by:
Ni2 = wi = μ A (x1 )μ B (x2 ), i = 1,2 (4)
2. Adaptive neuro-fuzzy inference system i i

and are normalized in the third layer.

(ANFIS) Layer 3: the nodes are also fixed nodes. They are
labelled with P, indicating that they play a
ANFIS is a multilayer feed-forward network where normalization role to the firing strengths from the
each node performs a particular function on incoming previous layer. Outputs can be represented as:
signals. Both square and circle node symbols are used wi
to represent different properties of adaptive learning. N i3 = wi = , i = 1,2 (5)
w1 + w2
To perform desired input–output characteristics,
adaptive learning parameters are updated based on which are the so-called normalized firing strengths.
gradient learning rules [19–21]. Such framework Layer 4: In the fourth layer, the nodes are adaptive
makes the ANFIS modeling more systematic and less nodes. The output of each node in this layer is simply
reliant on expert knowledge. To present the ANFIS the product of the normalized firing strength and a first
architecture, here two inputs and one output are order polynomial (for a first order Sugeno model).
considered, with five layers and two fuzzy if-then rules Thus, the outputs of this layer are given by:
based on a first order Sugeno model: N i4 = wi yi = wi ( pi x1 + qi x2 + ri ), i = 1,2 (6)
Rule 1 : If x 1 is A1 and x 2 is B1 then y 1 = p 1 x1 + q 1 x 2 + r1 Layer 5: In the fifth layer, there is only one single
Rule 2 : If x1 is A2 and x 2 is B 2 then y 2 = p 2 x1 + q 2 x 2 + r2 fixed node labelled with Σ . This node performs the
where x1 and x2 are inputs, Ai and Bi are fuzzy sets, yi is summation of all incoming signals. Hence, the overall
the output within the fuzzy region specified by the output of the model is given by:
fuzzy rule, pi, qi and ri are the design parameters 2

determined during the training process. The ANFIS 2 ∑w y i i

(7)
architecture to implement these two rules is shown in N i5 = ∑ wi yi = i =1
, i = 1,2
i =1 w1 + w2
Figure 1, in which a circle indicates a fixed node,
whereas a square indicates an adaptive node. It can be observed that there are two adaptive layers
Layer 1: In the first layer, each “i” node is a square in this ANFIS architecture, namely the first layer and
node with a node function as the fourth layer. In the first layer, there are three
modifiable parameters (ai, bi and ci), which are related
Ni1 = μ A (x1 ), i = 1,2 (1)
i to the input membership functions. These parameters
Ni1 = μBi−2 (x2 ), i = 3,4 (2) are the so-called premise parameters. In the fourth
where μ A (x1 ) and μ B ( x2 ) can adopt any fuzzy layer, there are also three modifiable parameters (pi, qi
i i −2
and ri), pertaining to the first order polynomial. These
membership function. For example, if we choose them parameters are so-called consequent parameters [19].
to be bell-shaped, such as

770
775

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on March 06,2025 at 07:31:34 UTC from IEEE Xplore. Restrictions apply.
 The task of the training algorithm for this repeat such procedure again and chooses the
architecture is tuning all the modifiable parameters to dimensionless parameters of simple form developed by
make the ANFIS output match the training data. The Choi et al. [6], as shown in Table 1. Two-phase inlet
output of the ANFIS model can be written as: condition is not taken into account because the data in
∑ the literature is not sufficient enough for ANN training
2
wy
i =1 i i w1 w2
yˆ = = y1 + y2 and test. Note that the surface roughness is not
∑ wi w1 + w2 w1 + w2
2
(8) included in the dimensionless parameter. In fact, it was
i =1

= w1 y1 + w2 y 2 found [22] that the friction factor of capillary tube is

Substituting Eqs. (5)–(8), the model output can be mainly dependent on the Reynolds number and not
rearranged by using the fuzzy if-then rules as: sensitive to the surface roughness. Consequently, the
yˆ = w1 x1 p1 + w1 x2 q1 + w1r1 surface roughness was excluded by the existing
(9) generalized correlations [5,6,7,23]. For the ANFIS
+ w2 x1 p 2 + w2 x2 q 2 + w2 r2 correlation, α1 through α7 defined in Table 1 are used
In order to obtain the optimal model parameters, the as the inputs, while α8 is used as the output.
fuzzy rule architecture of the ANFIS was designed by
using different membership functions and various Table 1. Dimensionless parameters for ANFIS
number of membership functions. Hybrid learning rule correlation
was used to train the model according to input/output Parameter Definition
data pairs. Five types of membership functions (Figure
2) and various number of membership functions which α1 ( pin − psat ) / pc Inlet pressure
is ranging 2–5 was used in the ANFIS model. α2 Tsub / Tc Subcooling
α3 L/D Geometry
α4 ρf / ρg Density
α5 (μf − μg )/ μg Friction
α6 σ / Dpin Bubble growth
α7 ρ f hfg / psat Vaporization
α8 m/ D 2
ρf pin Mass flow rate
pc , pin , psat : critical, inlet and saturated (bubble point)
pressure (Pa)
Tsub ,Tc : subcooling and critical temperature (ºC)
L, D: capillary tube length and inner diameter (m)
ρf , ρg : saturated liquid and saturated vapor density (kg
m-3)
Figure 2. Membership functions for ANFIS μ f , μ g : saturated liquid and saturated vapor viscosity (Pa
s)
σ : surface tension (N m-1)
hfg: heat of vaporization enthalpy (J kg-1)
3. ANFIS correlation for refrigerant mass m: mass flow rate (kg s-1)
flow rates
Although the ANFIS is of excellent capability in
A generalized correlation for refrigerant mass flow approximation of the nonlinear function, its
rate through adiabatic capillary tubes should meet two performance strongly depends on the database for
basic requirements. One is that all independent training, especially the parameter range that the
parameters influencing the mass flow rate should be database covers and the reliability of the measured data.
included. Another is that all parameters should be In this study, experimental data of R22, R134a, R290,
correlated with a proper function. The former could be R407C, R410A, and R600a in the literature are
implemented by expertise and then reduced by the employed for correlation, as shown in Table 2. The
dimensional analysis method. For the latter, an ANFIS database is divided into two parts: the training data and
universal approximator could be used instead of the the test data. The training data is used in the training
power-law function. step, while the test data is employed for independent
Dimensional analysis of refrigerant flow in an test of the prediction capability of the trained ANFIS.
adiabatic capillary tube was carried out completely in The parameter limitations of the database on capillary
the literature [5,6]. In this paper, the author does not

771
776

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on March 06,2025 at 07:31:34 UTC from IEEE Xplore. Restrictions apply.
 tube geometry and operating conditions are listed in Table 4. Effects of membership function type on the
Table 3. ANFIS model
Function type RMSE b MRE R2
Table 2. Data sources used for ANFIS correlation Gauss-2 a 0.003767 0.54 0.99999895
Source Refrigerant No. of data Purpose Gauss2-2 0.021842 2.32 0.99990478
Kuehl et al. [22] R22 55 Training Gbell-2 0.008552 0.95 0.99999514
Wijava [24] R134a 9 Training Trap-2 0.165330 7.76 0.99912672
Fiorelli et al. R407C 24 Training Tri-2 0.004987 0.87 0.99999801
[10,11] R410A 12 a
number of the membership function.
Melo et al. [27] R134a 19 Training b
RMSE is the root mean square error
R600a 19
∑ (t − oj )
n 2
Wei et al. [28] R22 31 Training j =1 j
R407C 21 RMSE =
n
Kim et al. [29] R22 11 Test I MRE is the mean relative error
R407C 11
1 n t j − oj
Choi et al. [6]
R410A
R22
10
14 Test II
MRE =
n
∑ j =1
100×
tj
R290 14
R2 is the coefficient of multiple determinations
R407C 14
⎛ ∑n (t j − o j )2 ⎞
⎜ j =1 ⎟
R2 = 1− ⎜
Table 3. Parameter ranges of the present database ⎜ ∑ (o j )2 ⎟⎟
n

Parameter Unit Min. Max. ⎝ j =1 ⎠

Condensing temperature ºC 28.0 73.4 where t is the predicted value, o is the measured
Inner diameter mm 0.601 2 value and n is the number of data.
Inlet pressure MPa 0.703 2.75
Length mm 400 3048
Subcooling temperature ºC 0.9 16.7

4. Results and discussion

In this section, the predictions from the ANFIS
correlation are compared with experimental data and
the correlations of other researchers.
Statistical values such as RMSE, MRE and R2 are
given in Table 4 for the ANFIS correlation with
various membership function types. The number of the
membership function is fixed to 2 for these simulations.
From the data presented in Table 4, for mass flow rates
of refrigerants, Gauss appeared to be most optimal Figure 3. Comparison of Choi et al. [6] and
membership function for the ANFIS model. The worst ANFIS-predicted values for the test data set with
results were obtained with the Trap. In Figure 3, a different membership function
comparison is presented between the experimental and
ANFIS-predicted values for the mass flow rates of R22 Table 5. Effects of number of membership functions
with different membership functions. The results on the ANFIS model
shown in Figure 3 were also supported the above Number RMSE MRE R2
results (Table 4). Gauss-2 0.003767 0.54 0.99999895
The statistical results for different numbers of the Gauss-3 0.080921 3.86 0.99956427
Gauss membership function were presented in Table 5. Gauss-4 0.415346 9.17 0.99402785
The best result is obtained when the number of input Gauss-5 1.297825 20.83 0.91517574
membership function was 2. When the number of
membership function was increased the performance of
the ANFIS was decreased, which also can be observed
from Figure 4.

772
777

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on March 06,2025 at 07:31:34 UTC from IEEE Xplore. Restrictions apply.
 et al. [7] correlation quite over-predict the second test
data. In fact, the correlation developed by Choi et al. [6]
is based on the data used as the second test data set of
ANFIS. Therefore, it is naturally in accordance with
these data. However, in their subsequent paper [7],
Choi et al. did not compare another correlation with
their own data [6].

5. Conclusions
An ANFIS-based generalized correlation method of
refrigerant mass flow rate through adiabatic capillary
tubes has been developed in this paper. Dimensional
analysis is utilized to provide generalized
Figure 4. Comparison of Choi et al. [6] and dimensionless parameters and reduce the number of
ANFIS-predicted values for the test data set with input parameters, while a five-layer perception neural
different number of Gauss membership function network is served as a generalized correlation of the
nonlinear data. For ANFIS training and testing,
Table 6 gives the comparison of ANFIS correlation measured data for R134a, R22, R290, R407C, R410A,
and correlations of other researchers. Three typical and R600a in the open literature are employed.
correlations in the literature are selected for According to statistical results, the ANFIS with two
comparison. One is the ASHRAE correlation [23,24], Gauss membership functions is of the most compact
the other two are developed by Choi et al. [6,7] on architecture and considered as the optimal for the
different database. Compared with other correlations in present input and output.
the literature, the three generalized correlations had In a further study, the ANFIS correlation method
been tested by more refrigerants and data sources. As will be extended for the two-phase inlet condition. It is
shown in Table 6, the ANFIS correlation can predict possible to merge the subcooled and two-phase inlet
better than other correlations for R134a, R22, R290, conditions into a generalized ANFIS correlation since
R407C, R410A, and R600a. All the correlations except the ANFIS can act as a universal approximator for the
Choi et al. [6] give acceptable statistical values of the complex nonlinear function.
ANFIS training data and the first test data set,
meanwhile the ASHRAE correlation [23,24] and Choi

Table 6. Comparison of ANFIS correlation and correlations in the literature

Data set ANFIS ASHRAE [23,24] Choi et al. [6] Choi et al. [7]
RMSE MRE R2 RMSE MRE R2 RMSE MRE R2 RMSE MRE R2
Training 0.0038 0.54 0.999999 0.0102 1.36 0.999988 0.8138 14.25 0.990991 0.0080 0.93 0.999998
Test I 0.0745 3.07 0.999817 0.0231 2.68 0.999891 0.5054 9.45 0.992137 0.1918 6.31 0.999596
Test II 0.2150 6.52 0.997235 1.7970 22.84 0.905624 0.0031 0.28 0.999999 0.9978 17.53 0.990025

Acknowledgements [1] P.K. Bansal, A.S. Rupasinghe, “An empirical model for
sizing capillary tubes”, Int J Refrig, 1996, 19 (8), pp. 497–
505.
The authors would like to acknowledge the
financial support provided by National Major Project [2] D. Jung, C. Park, B. Park, “Capillary tube selection for
of Scientific and Technical Supporting Programs of HCFC22 alternatives”, Int J Refrig, 1999, 22 (8), pp. 604–
China during the 11th Five-year Plan Period (Grant 614.
No. 2006BAJ03A06) and by Beijing Municipality Key
Lab of Heating, Gas Supply, Ventilating and Air [3] S.L. Chen, C.H. Liu, C.S. Cheng, C.S. Jwo, “Simulation
Conditioning Engineering. of refrigerants flowing through adiabatic capillary tubes”,
HVAC&R Res, 2000, 6 (2), pp. 101–115.
References [4] V. Trisaksri, S. Wongwises, “Correlations for sizing
adiabatic capillary tubes”, Int J Energy Res, 2003, 27, pp.
1145–1164.

773
778

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on March 06,2025 at 07:31:34 UTC from IEEE Xplore. Restrictions apply.
 [17] A. Chouai, S. Laugier, D. Richon, “Modeling of
[5] R.R. Bittle, D.A. Wolf, M.B. Pate, “A generalized thermodynamic properties using neural networks: application
performance prediction method for adiabatic capillary tubes”, to refrigerants”, Fluid Phase Equilib, 2002, 199 (1–2), pp.
HVAC&R Res, 1998, 4 (1), pp. 27–43. 53–62.

[6] J. Choi, Y. Kim, H.Y. Kim, “A generalized correlation [18] G. Scalabrin, G. Cristofoli, “The viscosity surfaces of
for refrigerant mass flow rate through adiabatic capillary R152a in the form of multilayer feed forward neural
tubes”, Int J Refrig, 2003, 26 (7), pp. 881–888. networks”, Int J Refrig, 2003, 26 (3), pp. 302–314.

[7] J. Choi, Y. Kim, J.T. Chung, “An empirical correlation [19] Jang J-SR, “ANFIS: Adaptive-network-based fuzzy
and rating charts for the performance of adiabatic capillary inference system”, IEEE Trans Syst Man Cybern, 1993,
tubes with alternative refrigerants”, Appl Therm Eng, 2004, 23(3), pp. 665–85.
24 (1), pp. 29–41.
[20] Jang J-SR, Gulley N, Fuzzy logic toolbox: reference
[8] K. Hornik, M. Stinchcombe, H. White, “Multilayer manual, The Mathworks Inc., Natick, MA, 1996.
feedforward networks are universal approximators”, Neural
Networks, 1989, 2, pp. 359–366. [21] Baylar A, Hanbay D, Ozpolat E, “Modeling aeration
efficiency of stepped cascades by using ANFIS”, Clean-Soil,
[9] M.T. Hagan, H.B. Demuth, M. Beale, Neural network Air, Water, 2007, 35(2), pp. 186–92.
design, PWS Publishing Company, New York, 1996.
[22] S.J. Kuehl, V.W. Goldschmidt, “Modeling of steady
[10] S.A. Kalogirou, “Applications of artificial neural- flows of R-22 through capillary tubes”, ASHRAE Trans,
networks for energy systems”, Appl Energy, 2000, 67, pp. 1991, 97 (1), pp. 139–148.
17–35.
[23] ASHRAE, ASHRAE handbook: refrigeration, American
[11] H. Bechtler, M.W. Browne, P.K. Bansal, V. Kecman, Society of Heating, Refrigerating and Air-Conditioning
“New approach to dynamic modeling of vapor-compression Engineers, Atlanta, 2002.
liquid chillers: artificial neural networks”, Appl Therm Eng,
2001, 21, pp. 941–953. [24] H. Wijaya, “Adiabatic capillary tube test data for HFC-
134a”, Proceedings of the IIR-Purdue Refrigeration
[12] E. Arcaklioglu, A. Erisen, R. Yilmaz, “Artificial neural Conference, West Lafayette, Ind., 1992, vol. 1, pp. 63–71.
network analysis of heat pumps using refrigerant mixtures”,
Energy Conversion Mgmt, 2004, 45, pp. 1917–1929. [25] F.A.S. Fiorelli, A.A.S. Huerta, O.M. Silvares,
“Experimental analysis of refrigerant mixtures flow through
[13] G. Diaz, M. Sen, K.T. Yang, R.L. McClain, “Simulation adiabatic capillary tubes”, Exp Therm Fluid Sci, 2002, 26, pp.
of heat exchanger performance by artificial neural networks”, 499–512.
HVAC&R Res, 1999, 5 (3), pp. 195–208.
[26] F.A.S. Fiorelli, O.M. Silvares, “Refrigerant mixtures
[14] A. Pacheco-Vega, G. Diaz, M. Sen, K.T. Yang, R.L. flow through capillary tubes: a comparison between
McClain, “Heat rate predictions in humid air–water heat homogeneous and separated-flow models”, HVAC&R Res,
exchangers using correlations and neural networks”, ASME J 2003, 9 (1), pp. 33–53.
Heat Transfer, 2001, 123, pp. 348–354.
[27] C. Melo, R.T.S. Ferreira, C.B. Neto, J.M. Goncalves,
[15] C.L. Zhang, G.L. Ding, H. Li, “Thermodynamic M.M. Mezavila, “An experimental analysis of adiabatic
calculation of refrigeration compressors with modified capillary tubes”, Appl Therm Eng, 1999, 19 (6), pp. 669–684.
artificial neural network method”, Chin J Mech Eng, 2001,
37 (1), pp. 75–77 (in Chinese). [28] C.Z. Wei, Y.T. Lin, C.C. Wang, J.S. Leu, “Experimental
study of the performance of capillary tubes for R-407C
[16] Chun-Lu Zhang, “Generalized correlation of refrigerant refrigerant”, ASHRAE Trans, 1999, 105 (2), pp. 634–638.
mass flow rate through adiabatic capillary tubes using
artificial neural network”, International Journal of [29] S.G. Kim, M.S. Kim, S.T. Ro, “Experimental
Refrigeration, 2005, 28, pp. 506–514. investigation of the performance of R22, R407C and R410A
in several capillary tubes for air-conditioners”, Int J Refrig,
2002, 25, pp. 521–531.

774
779

Authorized licensed use limited to: VIT University- Chennai Campus. Downloaded on March 06,2025 at 07:31:34 UTC from IEEE Xplore. Restrictions apply.