Escolar Documentos
Profissional Documentos
Cultura Documentos
a r t i c l e i n f o a b s t r a c t
Article history: In this paper, an application of artificial neural networks (ANNs) was presented to predict the pressure
Received 20 August 2008 drop and heat transfer characteristics in the plate-fin heat exchangers (PFHEs). First, the thermal perfor-
Accepted 12 November 2008 mances of five different PFHEs were evaluated experimentally. The Colburn factor j and friction factor f to
Available online 27 November 2008
different type fins were obtained under various experimental conditions. Then, a feed-forward neural
network based on back propagation algorithm was developed to model the thermal performance of
Keywords: the PFHEs. The ANNs was trained using the experimental data to predict j and f factors in PFHEs. Different
Plate-fin heat exchanger
network configurations were also examined for searching a better network for prediction. The predicted
Fin
Artificial neural network
values were found to be in good agreement with the actual values from the experiments with mean
Back propagation algorithm squared errors (MSE) less than 1.5% for j factor and 1% for f factor, respectively. This demonstrated that
Colburn factor the neural network presented can help the engineers and manufacturers predict the thermal character-
Friction factor istics of new type fins in PFHEs under various operating conditions.
Ó 2008 Elsevier Ltd. All rights reserved.
1359-4311/$ - see front matter Ó 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.applthermaleng.2008.11.011
2252 H. Peng, X. Ling / Applied Thermal Engineering 29 (2009) 2251–2256
Nomenclature
2. Experimental data
The heat transfer performance and pressure drops of serrated For the air side, the three main parameters: mass flow rate,
fins for a given geometric and flow conditions can be characterized temperature distribution and pressure drop were obtained. And
by a Colburn factor j and a friction factor f, respectively. The flow for steam side, the flow rate of condensed water was measured.
conditions can be characterized by Reynolds number, which can These measurements gave the total heat flux at both the air and
be calculated in terms of dimension parameters such as fin pitch, steam side:
fin height, etc.
Q as ¼ mas cpas Dt as ð9Þ
In PFHEs, the Colburn factor j and friction factor f can be defined
in Eqs. (1) and (2): Q ss ¼ mss css ð10Þ
For all studied cases in this paper, the heat balances between air
j ¼ StPr2=3 ð1Þ and steam were less than 5% by on-line calculations. If the heat bal-
ance were well satisfied, all the corresponding experimental data
4L qi DP q were saved on a computer for data reduction. All thermocouples
f ¼2 ðK c þ 1 r2 Þ þ ð1 r2 ¼ K e Þ i
De qm qum 2 q0 used in the present experiment with an experimental uncertainty
q of ±0.1 K, and the uncertainties of flow rate and pressure drop mea-
þ2 i 1 ð2Þ
q0 surement are less than ±0.5% and ±1%, respectively.
Experimental were performed for Reynolds number ranging
While the Prandtl number is: 600–7000 at the air side. Forty sets of experimental data were ob-
tained and divided into two parts: one part is used for training net-
lc p works (see Table 2); the other is used for testing networks (see
Pr ¼ ð3Þ
k Table 3).
The Stanton number can be evaluated by solving Eq. (4):
3. Prediction of thermal performance using ANNs
qw
St ¼ ð4Þ
cp GðT m T w Þ
One of the ANN models extensively used is the multilayer per-
where qw is the mean heat flux density, G is the mass flux and Tm is ceptron model based on the back propagation (BP) algorithm [16].
the average temperature between the inlet and outlet This BP architecture has proven to be robust and flexible and has
temperatures: been used widely to classify images in the remote sensing field.
This type of neural network generally consist of three or four lay-
Q
qw ¼ ð5Þ ers: an input layer, an output layer, and one or two middle (some-
Ac
times called hidden) layers, as shown in Fig. 4. Every unit in a layer
Ti þ To is connected to every unit in the next layer by weighted connec-
Tm ¼ ð6Þ
2 tions. The connections have variable weights, which are usually
The Reynolds number is defined in terms of hydraulic diameter De: set initially to random values. The data from the input layer are
passed to hidden layer by the weighted connections. Each hidden
GDe
Re ¼ ð7Þ layer unit sums all of its inputs in some way and if the total value
l is above certain threshold, it then produces an output. The outputs
2ðhf df Þðsf df Þ are passed to the output layer, where the process is repeated. Any
De ¼ ð8Þ
ðhf df Þ þ ðsf df Þ outputs from the output layer provide the classification result.
2254 H. Peng, X. Ling / Applied Thermal Engineering 29 (2009) 2251–2256
Table 2 During training, input data for which the correct result is known
Experimental data for training the network. are entered into the ANN. The data pass forward through each of
No. Inputs Outputs layers and the results are compared with the known correct result.
hf (mm) sf (mm) f (mm) Lf (mm) Reas j f The error in the output is propagated back through the network
and is used to adjust the weights of each connection, such that
1 4.7 2.0 0.3 5 727.95 0.01705 0.07630
2 4.7 2.0 0.3 5 1033.6 0.01375 0.06652
the error in the result will be reduced. This training cycle is re-
3 4.7 2.0 0.3 5 1806.4 0.01155 0.05557 peated many thousands of times, until a useful level of accuracy
4 4.7 2.0 0.3 5 3490.7 0.00870 0.04707 is achieved.
5 4.7 2.0 0.3 5 4443.2 0.00756 0.04470 The advantage of ANNs method than conventional regression
6 4.7 2.0 0.3 5 5188 0.00691 0.04331
analysis is: free of linear supposition, have large degrees of free-
7 4.7 2.5 0.3 5 845.55 0.01615 0.08476
8 4.7 2.5 0.3 5 1198.5 0.01396 0.07133 dom, and more effectively deal with nonlinear functional forms
9 4.7 2.5 0.3 5 2106.4 0.01188 0.05736 [17].
10 4.7 2.5 0.3 5 2760.3 0.00997 0.05295 For PFHEs in the present study, five independent parameters
11 4.7 2.5 0.3 5 3974.5 0.00903 0.04855
were fed into the input layer of the network (as shown in Fig. 4):
12 4.7 2.5 0.3 5 5807.2 0.00722 0.04534
13 4.7 2.5 0.3 5 6561.2 0.00684 0.04453
fin height, hf; fin pitch, sf; fin thickness, df; and fin length Lf; Rey-
14 6.5 2.0 0.3 5 1142.3 0.01561 0.06836 nolds number at the air side, Reas; The output layer included two
15 6.5 2.0 0.3 5 1978.8 0.01203 0.05792 parameters: Colburn factor, j; and friction factor, f.
16 6.5 2.0 0.3 5 2555.9 0.01066 0.05526 A total of 40 sets of data were run in the network, of which
17 6.5 2.0 0.3 5 3616.3 0.00904 0.05127
M = 33 sets of experimental data, were used to train the network
18 6.5 2.0 0.3 5 4431.2 0.00821 0.04937
19 6.5 2.0 0.3 5 5118.3 0.00767 0.04785 (Table 2), while the rest of N = 7 data, were used to test the net-
20 6.5 2.0 0.3 5 5724.3 0.00728 0.04671 work (Table 3). The sigmoid function was adopted in ANNs, which
21 6.5 2.5 0.3 5 949.26 0.01546 0.08289 requires the range of both input and output values should between
22 6.5 2.5 0.3 5 1342.5 0.01303 0.07060 0.1 and 0.9. The data evaluated experimentally in this study are
23 6.5 2.5 0.3 5 2326.3 0.00993 0.05759
24 6.5 2.5 0.3 5 3005.7 0.00875 0.05574
normalized in order to have the values. The formula used in the
25 6.5 2.5 0.3 5 4253.3 0.00737 0.05239 following:
26 6.5 2.5 0.3 5 5210.9 0.00666 0.05078
27 6.5 2.5 0.3 5 6730.6 0.00587 0.04756 ActualðuÞ MinðuÞ
NoVðuÞ ¼ ðHigh LowÞ þ Low ð11Þ
28 9.5 2.5 0.2 5 911.2 0.01767 0.07666 MaxðuÞ MinðuÞ
29 9.5 2.5 0.2 5 1288.7 0.01566 0.05602
30 9.5 2.5 0.2 5 2232.7 0.01248 0.04128
where u represents the input and output variables. Min is minimum
31 9.5 2.5 0.2 5 4991 0.00781 0.03302
32 9.5 2.5 0.2 5 5763.1 0.00698 0.03185 data value, Max is the maximum data value, High is the maximum
33 9.5 2.5 0.2 5 6443.3 0.00633 0.03137 normalized data value = 0.9, and low is the minimum normalized
data value = 0.1.
Ve Vp
Error ¼ 100% ð12Þ
Ve
During the training process, the performance of the network
was evaluated by calculating the mean squared error (MSE) values,
which is defined according to following the expression:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
PM e p 2
i¼1 ðV i V i Þ
MSE ¼ ð13Þ
M
In the present work, the learning rate of 0.25 and the momen-
tum rate of 0.9 were determined by the trail and error process,
respectively. For searching a relatively good configuration of ANNs,
six different networks, 5–7–3–2, 5–6–4–2, 5–4–3–2, 5–6–2, 5–4–2
and 5–3–2 were examined. During the training period the devel-
oped ANNs models, had a 5–6–4–2 network (Fig. 4) and 150,000
Fig. 4. Configuration of a 5–6–4–2 neural network for modeling PFHE. training cycles found to have best performance, which has a
H. Peng, X. Ling / Applied Thermal Engineering 29 (2009) 2251–2256 2255
maximum relative error less than 4% and less than 1.5% for the MSE Table 4
value. Fig. 5 shows the MSE values during the training process. It Comparison of experimental data and ANNs results for j and f factors: testing results.
can be seen that the mean squared error asymptotes at about No. Reas Experimental data ANNs testing Error (%)
100,000 cycles. At the end of training process, the relative errors j f j f j f
for training data are shown in Fig. 6. Most of errors of Colburn fac-
1 2381.2 0.00979 0.05157 0.01021 0.05186 4.29 0.57
tor j and friction factor f are within 2% region, and the maximum 2 5814.7 0.00615 0.04236 0.00631 0.04298 2.56 1.47
relative errors is about 4%. 3 4954.1 0.00808 0.04655 0.00805 0.04678 0.39 0.50
Once training is completed, predictions from the testing data 4 807.99 0.01840 0.07975 0.01799 0.08367 2.21 4.91
set were done using the already trained network. Table 4 shows 5 6018 0.00621 0.04914 0.00636 0.04976 2.49 1.26
6 2881.5 0.01100 0.03892 0.01117 0.03713 1.58 4.61
the error values between the output of the network and the testing
7 4075.1 0.00899 0.03509 0.00893 0.03403 0.72 3.01
experimental data. It is obvious that all the values are in the range
of 0%–5%. Fig. 7 shows the predicted j and f factors from trained
and tested ANNs, it can be seen that both predicted results are well
close to the corresponding measured experimental data. The max-
0.018
imum relative errors were approximately 5% for both j and f fac-
tors, and the MRE were 1.19% of j factor and 0.64% of f factor, Training results
0.016
respectively. Testing results +5%
In order to estimate the prediction performance of the ANNs
0.014 Desired line
developed in the present study, also the effects of input parameters
MSE=1.19% -5%
on the outputs were investigated. For this purpose, while one of
0.012
the input parameters was allowed to change from its minimum
j
to maximum values in the range of the available data set with
0.010
equivalent intervals, the others were kept constant, and then they
0.008
10 0.006
0.004
8 0.004 0.006 0.008 0.010 0.012 0.014 0.016 0.018
MSE values (%)
j
6
0.08
Mean squared errors (MSE)
4
Training results
Testing results +5 %
2 0.06 Desired line
MSE=0.64% -5 %
0
f
4
0.02
0.02 0.04 0.06 0.08
f
Relative errors (%)
3
Fig. 7. Thermal performance from training and testing results evaluated using
Colburn factor j ANNs.
Friction factor f
2
were introduced as the inputs to the trained neural network. For
example, Fig. 8 indicates the changes in the predicted values of j
factor with respect to the Reas, which ranges of 600–7000, when
1
other four input parameters (hf, sf, df, Lf) were kept constant at
the values shown in the figure. As expected, j factor decreases with
increasing Reas and this trend is very similar to the existing exper-
0 imental results. That is to say, when the initial conditions are given,
0 5 10 15 20 25 30 35 the ANNs model can predict the output parameter without imple-
Number of training data menting any experiments.
As shown in the above figures, the precision of ANNs is vali-
Fig. 6. Relative errors of training data. dated. It can be seen that the PFHE performance can directly obtain
2256 H. Peng, X. Ling / Applied Thermal Engineering 29 (2009) 2251–2256
0.020 Acknowledgements
References
0.012
[1] G. Diaz, M. Sen, K.T. Yang, R.T. McClain, Simulation of heat exchanger
0.010 performance by artificial neural networks, Int. J. HVAC&R Res. 5 (1999) 195–
208.
[2] G. Diaz, M. Sen, K.T. Yang, R.T. McClain, Dynamic prediction and control of heat
0.008 exchangers using artificial neural networks, Int. J. Heat Mass Transfer. 45
(2001) 1671–1679.
[3] G. Diaz, M. Sen, K.T. Yang, R.T. McClain, Adaptive neuro-control of heat
0.006 exchangers, ASME J. Heat Transfer. 123 (2001) 417–612.
1000 2000 3000 4000 5000 6000 [4] A. Pacheco-Vega, M. Sen, K.T. Yang, R.T. McClain, Neural network analysis of
fin-tube refrigerating heat exchanger with limited experimental data, Int. J.
Reas Heat Mass Transfer. 44 (2000) 763–770.
[5] A. Pacheco-Vega, G. Diaz, M. Sen, K.T. Yang, R.T. McClain, Heat rate predictions
in humid air–water heat exchangers using correlations and neural networks,
Fig. 8. ANNs predictions for j factor with respect to Reas. ASME J. Heat Transfer. 123 (2001) 348–354.
[6] A. Pacheco-Vega, M. Sen, K.T. Yang, Simultaneous determination of in-and-
over-tube heat transfer correlations in heat exchangers by global regression,
from the input information through the ANN network. In other Int. J. Heat Mass Transfer. 46 (2003) 1029–1040.
words, when designing a plate-fin heat exchanger under given in- [7] J. Thibault, B.P. Grandjean, A neural network methodology for heat transfer
let mass flow rate, i.e., the Reynolds number at inlet, the inlet and data analysis, Int. J. Heat Mass Transfer. 34 (1991) 2063–2070.
[8] Y. Islamoglu, A new approach for the prediction of the heat transfer rate of the
outlet temperatures or the temperature difference and the fin wire-on-tube type heat exchanger-use of an artificial neural network model,
geometries at both cold and hot side. The ANN approach is useful Appl. Therm. Eng. 23 (2003) 243–249.
and convenient for engineers or designers to predict the perfor- [9] Y. Islamoglu, A. Kurt, Heat transfers analysis using ANNs with experimental
data with air flow in corrugated channels, Int. J. Heat Mass Transfer. 47 (2004)
mance of PFHE with limited experimental data. It does not require
1361–1365.
understanding the heat transfer and flow characteristics, which is [10] N. Xie, Q.W. Wang, M. Zeng, L.Q. Luo, Heat transfer analysis for shell-and-tube
actually a very complex phenomenon to be expressed by mathe- heat exchangers with experimental data by artificial neural networks
approach, Appl. Therm. Eng. 27 (2007) 1096–1104.
matical formulations. Once the ANN is trained, the weighted con-
[11] Q.W. Wang, G.N. Xie, M. Zeng, L.Q. Luo, Prediction of heat transfer rates for
nections from the network corresponding to a practical heat shell-and-tube heat exchangers by artificial neural networks approach, J.
exchanger can be transferred to engineers or designers. Then engi- Therm. Sci. 15 (2006) 257–262.
neers may simply put the testing data into the trained network and [12] K. Ermis, ANN modeling of compact heat exchangers, Int. J. Energy Res. 32
(2008) 581–594.
therefore quickly make accurate prediction of thermal perfor- [13] S.V. Dudul, Identification of a liquid saturated steam heat exchanger using
mance for the heat exchangers. focused time lagged recurrent neural network model, IETE J. Res. 53 (2007)
69–82.
[14] H. Peng, Z.J. Sun, X. Ling, Numerical and experimental study on flow and heat
5. Conclusion transfer performance of some compact surface recuperators, J. Aerospace
Power 23 (2008) 138–144.
[15] H. Peng, X. Ling, Numerical modeling and experimental verification of flow and
An artificial neural network (ANNs) based on back propagation
heat transfer over serrated fins at low Reynolds number, Exp. Therm. Fluid Sci.
algorithm has been developed to predict the thermal performance 32 (2008) 1039–1048.
of PFHEs. Results indicated that the ANNs model can be trained to [16] Y. Wang, A neural network adaptive control based on rapid learning method
provide satisfactory estimations of both j and f factors in PFHEs. and its application, Adv. Model Anal. 46 (1994) 27–34.
[17] F.F. Farshad, J.D. Garber, J.N. Lorde, Predicting temperature profiles in
The MRE were 1.19% of j factor and 0.64% of f factor, respectively. producing oil wells using artificial neural networks, Eng. Comput. 17 (2000)
Also the versatility of the ANNs modeling was demonstrated by 735–754.
presenting the effects of some input parameters on the outputs [18] G. Chryssolouris, L. Moshin, A. Ramsey, Confidence interval prediction for
neural network models, IEEE Trans. Neural Netw. 7 (1996) 229–232.
using the developed network. It is recommended that the ANNs [19] R. Shao, E.B. Martin, J. Zhang, A.J. Morris, Confidence bounds for neural
prediction model has a high accuracy and reliability for simulating network representations, Comput. Chem. Eng. 21 (1997) 1173–1178.
thermal systems, especially for manufacturers or designers to [20] P. Niyogi, F. Girosi, Generalization bounds for function approximation from
scattered noisy data, Adv. Comput. Math. 10 (1999) 51–80.
model the complicated heat exchangers in engineering applica- [21] M.A. Kramer, J.A. Leonard, Diagnosis using back propagation neural networks-
tions with limited experimental data. analysis and criticism, Comput. Chem. Eng. 14 (1990) 1323–1338.