Applied Thermal Engineering 27 (2007) 10961104 www.elsevier.

com/locate/apthermeng

Heat transfer analysis for shell-and-tube heat exchangers with experimental data by articial neural networks approach
G.N. Xie, Q.W. Wang *, M. Zeng, L.Q. Luo
State Key Laboratory of Multiphase Flow in Power Engineering, Xian Jiaotong University, Shaanxi, Xian 710049, China Received 21 March 2006; accepted 11 July 2006 Available online 24 October 2006

Abstract This work applied Articial Neural Network (ANN) for heat transfer analysis of shell-and-tube heat exchangers with segmental bafes or continuous helical baes. Three heat exchangers were experimentally investigated. Limited experimental data was obtained for training and testing neural network congurations. The commonly used Back Propagation (BP) algorithm was used to train and test networks. Prediction of the outlet temperature dierences in each side and overall heat transfer rates were performed. Dierent network congurations were also studied by the aid of searching a relatively better network for prediction. The maximum deviation between the predicted results and experimental data was less than 2%. Comparison with correlation for prediction shows superiority of ANN. It is recommended that ANN can be used to predict the performances of thermal systems in engineering applications, such as modeling heat exchangers for heat transfer analysis. 2006 Elsevier Ltd. All rights reserved.
Keywords: Heat transfer rate; Outlet temperature dierence; Articial neural network; Shell-and-tube heat exchanger; Segmental baes; Continuous helical baes

1. Introduction The Computational Intelligence (CI) techniques, such as Articial Neural Networks (ANNs), Genetic Algorithms (GAs), Fuzzy Logic (FL), have been successfully applied in many scientic researches and engineering practices. ANNs have been developed for about two decades and now widely used in various application areas such as pattern recognition, system identication, dynamic control and so on. ANN oers a new way to simulate nonlinear, or uncertain, or unknown complex system without requiring any explicit knowledge about input/output relationship. ANN has more attractive advantages. It can approximate any continuous or nonlinear function by using certain network conguration. It can be used to learn complex nonlinear relationship from a set of associated input/output vectors. It can be implemented to dynami*

Corresponding author. Tel./fax: +86 29 82663502. E-mail address: wangqw@mail.xjtu.edu.cn (Q.W. Wang).

cally simulate and control unknown or uncertain process. In recent years, ANNs have been used in thermal systems for heat transfer analysis, performance prediction and dynamic control. For example, Thibault and Grandjean [1] earlier used a Neural Network (NN) for heat transfer data analysis, Jambunathan et al. [2] evaluated heat transfer coecients from experimental data by using a NN, Bittanti and Piroddi [3] used a NN to identify and control heat exchangers, Yang and Sen [4,5] reviewed works in dynamic modeling and controlling of heat exchangers using ANNs and GAs, Diaz et al. [610] did lots of works in steady/ dynamic simulation and control heat exchangers using ANNs, Parcheco-Vega et al. [1114] also did many works in analysis for n-and-tube heat exchangers with limited experimental data using soft computing and global regression, Islamoglu et al. [15,16] predicted heat transfer rate for a wire-on-tube heat exchanger and made heat transfer analysis for air owing in corrugated channels. Other researches about heat exchangers control by means of ANNs were reported in references [1719]. From

1359-4311/$ - see front matter 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.applthermaleng.2006.07.036

G.N. Xie et al. / Applied Thermal Engineering 27 (2007) 10961104

1097

Nomenclature Dc diameter of center blocked tube (mm) Do outside diameter of tube (mm) Er relative error M number of sets of data for training network N number of sets of data for testing network Nb number of bae Nt number of tube Pr Prandtl number U heat transfer rate (W) R evaluation factor for average accuracy, Eq. (3) Rew, Reo Reynolds number in water-side, oil-side rms Root-mean-squares error Sb bae pitch (mm) C coecient of heat transfer correlation m exponent of heat transfer correlation Nu Nusselt number Pr Prandtl number Tw,in, To,in inlet temperature in water-side, oil-side (K) umax velocity at minimum cross-section (m/s) DTo,out temperature dierence in oil-side (K) DTw,out temperature dierence in water-side (K) r evaluation factor for scatter accuracy, Eq. (4) m kinematic viscosity Superscripts e, p experimental, predicted Subscripts o, w oil, water in, out Inlet, outlet

aforementioned successful applications, it is shown that ANNs are well suitable to thermal analysis in engineering systems, especially in heat exchangers. In many experimental studies and engineering applications of thermal science, researchers and engineers expect to reduce experimental data into one or more simple and compact dimensionless heat transfer correlations. The disadvantages of the correlation methods are that heat transfer coecients strongly depend on their denitions and temperature dierences, and inevitably need iterative method to obtain correlations when uid properties are dependent on uid temperatures [20]. However, ANN does not need denition of correlations and iterative method, only needs input/output samples for training a special neural network, in turn, obtaining output results as test samples fed into trained network. In the above-mentioned literature, most works were done in thermal analysis for n-tube heat exchangers, while for shell-and-tube heat exchangers only few works were done in open literature. For this reason, the objective of this paper is that, setting up experimental system for investigation on three shelland-tube heat exchangers, and applying ANN for heat transfer analysis of heat exchangers with experimental data based on back propagation algorithm to train the network. The predicted outputs of ANN are temperature dierences of two sides and heat transfer rate. Dierent network congurations were studied for searching an optimal network. In addition, the predicted results by ANN were compared with those by correlations from references. 2. Physical model and experimental data 2.1. Experimental system The experimental loop is shown in Fig. 1, which was designed by our research group and built at school of Energy and Power Engineering, Xian JiaoTong Univer-

sity. There are three sub-loops (an oil loop, a cold water loop and a cooling water loop) for achieving the heat exchange of the experimental loop in the present study. In Fig. 1, 41 is the tested heat exchanger. We can carry out oilwater (by oil loop and cold water loop) or water water (by hot water loop and cold water loop) heat exchanger on the experimental loop. The cooling water loop is used to cool the heated water of the cold water loop. More detailed description of the experimental system and tested heat exchangers can be found in Ref. [21]. 2.2. Data acquisition Three main parameters, mass ow rate, temperature and pressure drop are obtained, for both hot and cold working medium of the tested heat exchangers. The heat balances between water and oil are less than 8% by on-line calculations for all studied cases. If the heat balance were well satised, all the corresponding experimental data were saved and averaged on a computer for data reduction. The uncertainties of obtained temperature dierence, ow rate, heat transfer rate and heat transfer coecient are less than 2%, 0.15%, 2.5% and 4%, respectively. The three tested heat exchangers are shown in Fig. 2. Fig. 2(a) is a heat exchanger with segmental baes (hereafter, the heat exchanger is called HX1), the other two are heat exchangers with continuous helical baes, as shown in Fig. 2(b) and Fig. 2(c). The only dierence between the latter two helical heat exchangers is the inletoutlet ow manner of shell-side uid. One is middle-in-middle-out (so called HX2, Fig. 2(b)) and the other one is side-inside-out (so called HX3, Fig. 2(c)). The cores of HX2 and HX3 are same (i.e., the layouts of tubes and baes are identical), and the only dierence between the HX2 and HX3 is the position/location of inlets and outlets of shell-side ows. The positions of inlet and outlet of HX2 are on the middle of the shell side, which is normal to shell,

1098

G.N. Xie et al. / Applied Thermal Engineering 27 (2007) 10961104

Fig. 1. Experimental loop.

while the positions of inlet and outlet of HX3 are on the side of the shell side, which is tangential to shell. It should be noted that for helical heat exchangers (HX2, HX3) there is a blocked center tube. The heat exchangers are 12 type, with hot oil owing in shell-side and cold water owing in tube-side. The detail parameters for the three tested heat exchangers are shown in Table 1. Note that the helical characteristics are determined by helical pitch (herein is bae pitch, Sb) and bae diameter, instead of helix angle, because the helical baes are continuous. One cycle of continuous helical bae is rstly produced, and each cycle occupies one circle of the exchanger, and spires for one helical pitch. Experiments were performed for Reynolds number ranging from 300 to 7000 in the shell-side, 3000 to 4000 in the tube-side. Heat transfer rate varied from 20 kW to 50 kW. Thirty nine sets of experimental data were obtained and divided into two parts: one part is used for training networks (as listed in Table 2), the other is used for testing networks (as listed in Table 3). It should be noted that the diameter of center tube in HX1 is zero since there is no center tube. In this study, the shell-side Reynolds number is dened by Reo umax Do m 1

where Do is the outer diameter of tube, umax is the velocity at minimum cross-section of shell-side, m is kinematic viscosity of oil. It should be noted that the properties of oil and water are determined at the mean temperature (averaged by inlet and outlet temperature). 3. Neural network conguration ANNs comprise of a great number of interconnected neurons. Fig. 3 illustrates a typical full-connected network conguration. Such an ANN consists of a series of layers with a number of nodes. The nodes (circle points in Fig. 3) sometimes called neuron are the basic processors of neural network. Each connection between two nodes with a real value is called weight. Nodes are gathered together into a column called layer. For each node, there exist an activation and a bias associated it. Among various types of ANNs, the feedforward or multilayer perception neural network is widely used in engineering applications. The input information is propagated forward through the network, while the output error is back propagated through the network for updating the weights. As shown in Fig. 3, the rst layer with eight nodes and last layer with three nodes are called input layer and output layer respectively, while the others in the middle are called hidden layers. The conguration in Fig. 3 has two hidden layers,

G.N. Xie et al. / Applied Thermal Engineering 27 (2007) 10961104

1099

Oil

Oil

Water

(a) HX1

Blocked center tube

Oil

Oil Water

(b) HX2

Blocked center tube

Oil Oil Water

(c) HX3
Fig. 2. Shell-and-tube heat exchangers. Table 1 Geometrical parameters of heat exchangers Parameters Inner diameter of shell Outer diameter of tube Inner diameter of tube Arrangement of tube Eective length of tube Number of tubes (helical) Number of tubes (segmental) Outer diameter of inlet tube in water-side Inner diameter of inlet tube in water-side Outer diameter of outlet tube in water-side Inner diameter of outlet tube in water-side Outer diameter of inlet tube in oil-side Inner diameter of inlet tube in oil-side Outer diameter of outlet tube in oil-side Inner diameter of outlet tube in oil-side Length of heat exchanger Unit mm mm mm mm mm mm mm mm mm mm mm mm mm Value 207 10 8 Triangle 620 158 176 57 50 57 50 57 50 57 50 670

which has six and ve nodes respectively. There are many ways to implement ANNs. It is dicult to nd an optimal network, considering the uniqueness of a real problem. Thus, a priori choice, such as selection of network topol-

ogy, training algorithm and network size should be made based on experience in order to keep the task to a manageable proportion. The back propagation (BP) algorithm is widely used to train the networks. The main idea of this algorithm is to minimize cost function by steepest descent method to add small changes in the direction of minimization. It simply consists of back-propagating the output errors to the network by modifying the weight matrices. More descriptions of BP algorithm can be found in Refs. [1,5]. The drawback of BP algorithm is that it may get stuck in a local minimum and it needs a great of time to arrive at convergence. Varying the learning rate dynamically or using momentum terms can improve the convergence speed. The mathematical background, the procedures for training and testing the ANN, and description of BP algorithm can be found in references [22,23]. Although the BP algorithm needs long time to converge, the algorithm has gained a remarkable popularity in neural network community, because it is relatively easy to be implemented in engineering applications. For the heat exchangers at hand, eight independent parameters were fed into the input layer of the network

1100

G.N. Xie et al. / Applied Thermal Engineering 27 (2007) 10961104

Table 2 Experimental data for training the network No. 1 2a 3c 4a 5c 6a 7c 8a 9b 10a 11a 12b 13a 14a 15a 16c 17c 18c 19c 20b 21c 22b 23c 24b 25c 26b 27b 28b 29b 30c
a

Reo 296 525 571 697 745 821 981 1102 1148 1253 1399 1413 1486 1693 1825 1950 2565 2591 3045 3121 3507 4365 4949 4979 5536 5669 5843 6702 6996 7018

To,in (C) 59.6 58.7 58.4 60.3 59.1 60.5 58.5 59.6 61 59.4 58.9 61 59.4 59.1 59.8 59.3 59 59.4 59.2 60.8 59.6 60.1 60.7 61 61.4 61.6 60.9 62 61.8 62.2

Rew 3010 3014 3358 2942 3706 3033 3814 3121 3887 2954 3168 3391 3022 3165 3094 3369 3488 3742 3607 3705 3811 3832 3862 3997 3820 3907 4203 4103 4091 4000

Tw,in (C) 30.2 28.1 30.7 27.7 32.9 29.3 34.3 30.2 33.6 26.7 30.8 26 27.9 30.6 29.3 26.7 27.9 31.6 29.4 29.7 32.7 31.3 33.1 33.4 32.2 32 33.9 34.8 34.5 34.6

Nt 176 176 158 176 158 176 158 176 158 176 176 158 176 176 176 158 158 158 158 158 158 158 158 158 158 158 158 158 158 158

Nb 7 7 9 7 9 7 9 7 9 7 7 9 7 7 7 9 9 9 9 9 9 9 9 9 9 9 9 9 9 9

Dc (mm) 0 0 48 0 48 0 48 0 48 0 0 48 0 0 0 48 48 48 48 48 48 48 48 48 48 48 48 48 48 48

Sb (mm) 70 70 48 70 48 70 48 70 48 70 70 48 70 70 70 48 48 48 48 48 48 48 48 48 48 48 48 48 48 48

Note: Dc = 0 indicates that there exists no center tube in the heat exchanger. Superscripts a, b, c refer to the data from HX1, HX2, HX3 respectively.

Table 3 Experimental data for testing the network No. 1a 2a 3c 4b 5b 6b 7c 8c 9c Reo 378 912 1371 1978 2610 3480 4251 5761 6625 To,in (C) 59.9 59.8 58 59.8 60.2 59.1 60.7 61 61.5 Rew 2911 2794 3319 3537 3596 3923 3721 3936 4015 Tw,in (C) 26.5 24.1 26.1 27.9 28.5 33.3 31.1 34.1 35.3 Nt 176 176 158 158 158 158 158 158 158 Nb 7 7 9 9 9 9 9 9 9 Dc (mm) 0 0 48 48 48 48 48 48 48 Sb (mm) 70 70 48 48 48 48 48 48 48

Note: Dc = 0 indicates that there exists no center tube in the heat exchanger. Superscripts a, b, c refer to the data from HX1, HX2, HX3 respectively.

(as shown in Fig. 3): Reynolds numbers and inlet temperature in each side Rew, Reo, Tw,in, To,in, total number of tubes Nt, diameter of center tube Dc, total number of bafes Nb and bae pitch Sb. The main reason for selection of these input variables is that, heat transfer rate as well as outlet temperature are aected by inlet mass ow rate, inlet temperature on each side, and structure of heat exchangers core due to the aforementioned dierences between the three heat exchangers. The eects of tube and bae

arrangements can be considered into the hydraulic diameter, which is included in Reynolds number. The output layer contains three parameters: heat transfer rate, Up, temperature dierences in each side, DT p , DT p . It is o;out w;out noticed that the ANN prediction is o-line carried out after all of the dynamic parameters, including Re, thermal properties, heat transfer rate, have been post-handled. Thus, once the ow rate has been measured, Re can be obtained through its denition, so can the heat transfer rate and

G.N. Xie et al. / Applied Thermal Engineering 27 (2007) 10961104

Reo
Rew

1101

To ,in
p T o,out

Tw ,in Twp,out Nt

p
Nb Sb Dc
Fig. 3. 8-6-5-3 neural network conguration used for modeling heat exchangers.

temperature dierences. In other words, when all experimental data have been reduced, the ANN predictions can be o-line conducted. A total of 39 set of data were run in the network, of which M = 30 sets of experimental data, as listed in Table 2, were used to train the network, while the rest of N = 9 data, as listed in Table 3, were used to test the network. Note that 77% of the experimental data were used for training the network. The selection of test data from each heat exchanger may be somewhat arbitrary, however these data are based on approximately uniform variation of Reo from 300 to 7000 and based on total number of data from each heat exchanger. In the present study, the popular used sigmoid function was adopted in hidden layers and output layer. It should be noted that the sigmoid function has the asymptotic limits of [0, 1]. It is desirable to normalize all the input/output data with the largest and smallest values of each of the data sets, since the variables of input/output data have dierent physical units and range sizes. Thus, to avoid any computational diculty, all of the input/output pairs were normalized in range of [0.15, 0.85] based on previous experience [713]. 4. Results and discussion As aforementioned, drawback of BP algorithm is that it may get stuck in a local minimum, therefore the learning rate was being changed during the training process of the network. In the present study, the learning rate was nally set to 0.4 based on previous tested experience [713]. The training of the neural network was terminated when the maximum number of training cycles was reached. Note that the selection of the number is a trail-and-error process in which it may be changed if the performance of neural network during the training is not good enough. In this

study, after a series of trail tests the number of training cycles was chosen to be 1,000,000, where the maximum relative error between the output of the network and the target output was less than 2%. The relative error of every predicted output was dened by Er jAe Ap j=Ae 2

where Ap is the predicted results (that is, the output of ANN), Ae is the experimental data (that is, the target output. The maximum error was determined by maximum value of the maximum relative errors of the three output variables. During the training process of neural network, the performance of the network was evaluated by calculating the root-mean-square (rms) values of the output errors v u M u 1 X Ae Ap 2 rms t 3 M i1 Ae Then rms error was determined by maximum value of the rms error of the three output variables. As an example, the errors during training 8-6-5-3 network conguration (as shown in Fig. 3), with two hidden layers with 6 and 5 nodes respectively, are shown in Fig. 4(a). It can be seen that the maximum error asymptotes at about 800,000 cycles, while the rms error is reached at 100,000 cycles. At the end of training process, the relative errors for training data are shown in Fig. 4(b). Most of errors are within 1% region, and the maximum relative error is about 1.5%. Generalization is an important quality of ANN. It is the ability to provide accurate output results when the input data that have never been used for training were fed into the trained network. The network topology and size, such as selection of number of hidden layer and number of hidden node, will aect the predicted performance. The performance of the trained network is evaluated by comparing its

1102
10

G.N. Xie et al. / Applied Thermal Engineering 27 (2007) 10961104

N N 1 X 1 X Ae Ri N i1 N i1 Ap v u N uX R Ri 2 rt N i1

Error (%)

maximum error
2

rms

0 0 200,000 400,000 600,000 800,000 1,000,000

Cycles Training process

1.6 1.4 1.2 1.0

Error (%)

0.8 0.6 0.4 0.2 0.0 0 5 10 15 20 25 30

Note that in Table 4, R and r is the maximum value which were determined from R and r of the three output variables respectively. R reects the average accuracy of the prediction, while r reects the scatter of the prediction. Both quantities are important for an assessment of the relative success of the ANN analysis [5]. For three layers, when the number of hidden nodes is increased to 5, R is much closer to unity. This indicates that adding more hidden nodes may not improve the predicted results. From Table 4, network with R close to unity is 8-6-5-3-3, however, its r is larger than that of 8-6-5-3, in this sense, the congurations with four layers have higher accuracy of prediction that those with ve layers. It is also noted that adding more hidden layers may not make the prediction better. Thus, in this case, conguration 8-6-5-3 is selected for testing, with
16 14 12
+1 0%

(C)

10 8 6 4 2 0

-1

0%

No Relative error
Fig. 4. Training error for conguration 8-6-5-3 ANN: (a) Training process and (b) relative error DTo,out (j), DTw,out (m) and Q (w).

T o,out

Table 4 Comparison of errors by dierent ANN congurations Conguration Train error Er (%) 8-4-3 8-5-3 8-6-3 8-7-3 8-8-3 8-6-4-3 8-6-5-3 8-7-5-3 8-6-5-3-3 8-7-5-3-3 5.1347 3.8703 4.2666 4.0109 2.3681 1.3852 1.4471 1.0657 2.1113 1.8007 rms (%) 1.5891 1.0828 0.8151 0.8442 0.5448 0.5306 0.5276 0.4831 0.8686 0.7028 Test error R 1.1282 1.1084 1.1914 1.1904 1.1226 1.0893 1.0890 1.0958 1.0595 1.0925 r 0.12920 0.12272 0.19488 0.19391 0.12947 0.13259 0.13872 0.13531 0.14540 0.13764
10

8 e T o,out

10

12

14

16

(C)

Tpw,out (C)

+1

0%

-1

0%

Note: the error is the maximum value among from the error of the three output variables.

0 0 2 4 6 8 10

prediction with the data set aside for testing. Thus, in this study, by the aid of searching a relatively good conguration for prediction, ten dierent ANN congurations were studied, as shown in Table 4. R and r are dened by

Tew,out

(C)

Fig. 5. Predicted temperature dierence vs. experimental data (C): (a) predicted temperature dierence in oil-side and (b) predicted temperature dierence in water-side.

G.N. Xie et al. / Applied Thermal Engineering 27 (2007) 10961104 Table 5 Constants of C and m in correlations for tested HX Heat exchanger HX1 HX3 HX2 C 0.16442 0.06801 0.34571 m 0.65582 0.57861 0.62296 Error (%) 3.528 2.28 1.76

1103

40,000

+1

0%

50,000

p (W)

-1

0%

30,000

20,000

Correlations ANN

10,000

0 0 10,000 20,000 30,000 40,000 50,000

e (W)
Fig. 6. Comparison of 8-6-5-3 ANN and correlation for heat transfer rate (W).

smallest R = 1.089 and r = 0.1387 and the maximum relative error is less than 1.5%.This selection agrees well with that in Ref. [24]. The predicted temperature dierences for two sides of heat exchangers from trained and tested ANN plotted against experimental data is shown in Fig. 5. It can be seen that both predicted results are well close to the corresponding measured variables. In this study, the predicted heat transfer rates obtained by conguration 8-6-5-3 ANN, and those by dimensionless correlations from Peng [21], were compared. It is noted that heat transfer correlations in shell-side could be shown as ln(Nu/Pr1/3) = C + m ln(Re), and the constants C and m are listed in Table 5 [21]. The comparison results are shown in Fig. 6. For most of data, the ANN error is within 2% while the correlation error is within 8%. The ANN predictions are much better than those of correlations. Thus, for the tested heat exchangers at hand, ANN is superior to correlation for prediction. Prediction performance of heat exchanger is one of important objective to a designer or engineer so as to understand the performance before performing the experimental investigations. There are many approaches, which can be used. As usual, for example, the data information obtained by experiments can be compressed as a compact form in correlation such as Nusselt number vs. Reynolds number and Prandtl numbers, Nu = f (Re, Pr), sometimes including geometrical factors. However, there exist some assumptions in deriving the correlation, which generally are not quite valid for real problem. For example, it is very

often to assume the heat coecient along heat transfer wall to be constant and the temperature dierence between hot and cold uid to be constant, and uid properties are often independent on uid temperature. In fact, these assumptions do not always hold for an actual heat exchanger. As shown in the above gures, the precision of ANN is much better than that of simplied correlation. It is can be seen that we can directly obtain the heat transfer rates from the input information through the network, instead of using them to get Nusselt numbers from correlations, and then in turn indirectly obtain the heat transfer rates. On the other hand, when designing a heat exchanger under given inlet mass ow rate and temperature, the outlet temperature or sometimes temperature dierence should be evaluated. Under this case, the heat transfer rate needs to be predicted. The ANN approach is useful and convenient for engineers or researchers to predict the performance of a given heat exchanger with limited experimental data. It does not need to provide accurate and detailed mathematical formulations as well as compact form of correlations. Once the ANN is trained, the weights and biases from the network corresponding to a practical heat exchanger can be transferred to engineers or researchers who are going to use the test data for prediction. Then engineers may simply feed these data into the trained network and therefore quickly make accurate prediction of thermal performance for the practical heat exchanger. However, some limitation should be considered for ANNs, since they do not provide any knowledge about physical phenomena.

5. Conclusions In the present study, an experimental system for investigation on performance of shell-and-tube heat exchangers is set up, and limited experimental data is obtained. The ANN is applied to predict temperature dierences and heat transfer rate for heat exchangers. BP algorithm is used to train and test the network. It is shown that the predicted results are close to experimental data by ANN approach. Comparison with correlation for prediction heat transfer rate shows ANN is superior to correlation, indicating that ANN technique is a suitable tool for use in the prediction of heat transfer rates than empirical correlations. It is recommended that ANNs can be applied to simulate thermal systems, especially for engineers to model the complicated heat exchangers in engineering applications.

Acknowledgements This work was supported by National Defense Science and Technology Key Laboratory Foundation of Chnia (Grant No. 51482100204JW0801) and Program for New Century Excellent Talents in University of China (Grant No. NCET-04-0938).

1104

G.N. Xie et al. / Applied Thermal Engineering 27 (2007) 10961104 tions and neural networks, ASME J. Heat Transfer 123 (2001) 348 354. A. Pacheco-Vega, Simulation of compact heat exchangers using global regression and soft computing, Ph.D. thesis, University of Notre Dame, 2002. A. Pacheco-Vega, M. Sen, K.T. Yang, Simultaneous determination of in-and-over-tube heat transfer correlations in heat exchangers by global regression, Int. J. Heat Mass Transfer 46 (2003) 1029 1040. Y. Islamoglu, A new approach for the prediction of the heat transfer rate of the wire-on-tube type heat exchanger-use of an articial neural network model, Appl. Thermal Eng. 23 (2003) 243249. Y. Islamoglu, A. Kurt, Heat transfers analysis using ANNs with experimental data with air ow in corrugated channels, Int. J. Heat Mass Transfer 47 (2004) 13611365. M. Ayoubi, Dynamic multi-layer perception networks: application to the nonlinear identication and predictive control of a heat exchanger, World Sci. Ser. Rob. Intell. Syst. 17 (1997) 205230. M. Jalili-Kharaajoo, B.N. Araabi, Neuro-predictive control of a heat exchanger: comparison with generalized predictive control, IEEE Trans. (2003) 675678. K. Varshney, P.K. Panigrahi, Articial neural network control of a heat exchanger in a closed ow air circuit, Appl. Soft Comput. 5 (2005) 441465. X. Zhao, Performance of a single-row heat exchanger at low in-tube ow rates, Master thesis, University of Notre Dame, 1995. B.T. Peng, Experimental study of heat transfer and pressure drop for shell-and-tube heat exchangers with continuous helical baes, Master thesis, Department of Energy and Power Engineering, Xian JiaoTong University, 2005 (in Chinese). S. Haykin, Neural Networks: A Comprehensive Foundation, Prentice Hall, Upper Saddle River, NJ, 1999. M.T. Hagan, H.B. Demuth, M. Beale, Neural Network Design, China Machine Press, Beijing, 2002. Q.W. Wang, G.N. Xie, M. Zeng, L.Q. Luo, Prediction of heat transfer rates for shell-and-tube heat exchangers by articial neural networks approach, J. Thermal Science 15 (2006) 257262.

References
[1] J. Thibault, B.P. Grandjean, A neural network methodology for heat transfer data analysis, Int. J. Heat Mass Transfer 34 (1991) 2063 2070. [2] K. Jambunathan, S.L. Hartle, S. Ashforth-Frost, V. Fontama, Evaluating heat transfer coecients using neural networks, Int. J. Heat Mass Transfer 39 (1996) 23292332. [3] S. Bittanti, L. Piroddi, Nonlinear identication and control of a heat exchanger-a neural network approach, J. Franklin Inst. 334 (1997) 135153. [4] K.T. Yang, M. Sen, Articial neural network-based dynamic modeling thermal systems and their control, in: Wang Bu-xuan (Ed.), Heat Transfer Science and Technology, Higher Education Press, Beijing, 2000. [5] M. Sen, K.T. Yang, Applications of articial neural networks and genetic algorithms in thermal engineering, in: F. Kreith (Ed.), The CRC Handbook of Thermal Engineering, CRC Press, Boca Raton, FL, 2000. [6] G. Diaz, M. Sen, K.T. Yang, R.T. McClain, Simulation of heat exchanger performance by articial neural networks, Int. J. HVAC R Res. 5 (1999) 195208. [7] G. Diaz, Simulation and control of heat exchangers using articial neural networks, Ph.D. thesis, University of Notre Dame, 2000. [8] G. Diaz, M. Sen, K.T. Yang, R.T. McClain, Dynamic prediction and control of heat exchangers using articial neural networks, Int. J. Heat Mass Transfer 45 (2001) 16711679. [9] G. Diaz, M. Sen, K.T. Yang, R.T. McClain, Adaptive neuro-control of heat exchangers, ASME J. Heat Transfer 123 (2001) 417 612. [10] G. Diaz, M. Sen, K.T. Yang, R.T. McClain, Stabilization of thermal neuro-controllers, Appl. Artif. Intell. 18 (2004) 447466. [11] A. Pacheco-Vega, M. Sen, K.T. Yang, R.T. McClain, Neural network analysis of n-tube refrigerating heat exchanger with limited experimental data, Int. J. Heat Mass Transfer 44 (2000) 763770. [12] A. Pacheco-Vega, G. Diaz, M. Sen, K.T. Yang, R.T. McClain, Heat rate predictions in humid airwater heat exchangers using correla[13]

[14]

[15]

[16]

[17]

[18]

[19]

[20] [21]

[22] [23] [24]