ACAR Researe Vou. 5,No.3 HVACR Researc Juty 1999 ‘ing sufficient 1 Simulation of Heat Exchanger Performance by dono ene of Artificial Neural Networks E's imerest in ‘model shown ' eno Gerardo Diaz Mihir Sen, Ph.D. to extrapolate 1 behavior, it K.T. Yang, Ph.D. Rodney L. McClain, PE. 1 comfort and ‘Member ASHRAE The artificial neural network technique was applied to heat transfer through a series of prob- lems of increasing complexity. For the simplest problem of one-dimensional heat conduction with linear activation function, itis possible to give physical meaning to the synaptic weights of the network. A network with sigmoid activation function was used for non-linear representation of convection problems where identification of the weights with physical variables was not pos- sible. Two cases of convective heat transfer with one and two heat transfer coefficients and arti- Jicially generated data were examined. Finally, the method was applied to the analysis of data ‘obtained in the laboratory for a single-row, fin-tube heat exchanger. It is shown that a better prediction with smaller scatter is obtained in comparison to a conventional power-law correla- tion for the heat transfer coefficients. INTRODUCTION In the design of thermal systems, predicting the heat transfer rate of heat exchangers under prescribed operating conditions is necessary. For a given device exchanging heat between two fluids, the heat transfer rate depends on the flow rates and the inlet temperatures of each fluid. Currently, most calculations are done on the basis of manufacturers’ data for specific fluids that ‘give the heat transfer rate asa function of the two flow rates and the two inlet temperatures. This is a four-variable function and difficult to represent completely. In principle the functional rela- tion depends on the geometry of the heat exchanger, the materials with which itis made, the sur- face conditions, the fluids used, etc. The relationship completely characterizes the heat ‘exchanger and isthe information that must be transferred in some form from the manufacturer to the design engineer. It would be advantageous to be able to compress the information contained in the heat trans- fer rate function so that it can later be accurately recovered. For instance, if the internal and external heat transfer coefficients are provided, the heat transfer for any flow rate, inlet tem- perature, or fluid can be easily determined. The situation is complicated by the fact that the heat transfer coefficients vary considerably with flow rates and fluid properties. Dimensional analysis can reduce the number of variables to the internal and extemal Nusselt (or Stanton) numbers as functions of the corresponding Reynolds and Prandtl numbers. If such correlations could be determined from experimental measurements, an acceptable procedure can be devised. In practice, without accurate tube wall temperature measurements, itis difficult to separate the experimentally determined overall thermal resistance into its internal and external ‘components. Furthermore, property variations, especially the variation of liquid viscosity, ‘Gerardo Diaz is radar student, Mite Seni profesor, KwangeTzu Vang is Viol D. Hank professor mari, and Rodney L. McClain is an associate profesional specialist inthe Deparment of Asrospao and Mechanical Eng ‘neering, University of Nae Dane, Note Dare, Indiana 195196 HVACER Reseazcr make any correlation obtained highly dependent on fluid temperatures. Procedures that take this variation into account, become very complex and potentially lose generality. In any case the user of the information, the system designer, is usually interested only in the heat transfer rate, and not in intermediate variables such as the heat transfer coefficients. For this purpose a straightforward interpolation of the original experimental data would probably be more useful, ‘but would be inconvenient ‘The artificial neural network (ANN) technique offers an alternative approsch tothe problem of information compression for heat exchangers. Its a procedure that is usually used for predicting the response of a physical system that cannot be easily modeled mathematically. The network is first “trained” by experimentally obtained input-output sets of data, after which it can be used for prediction, The manufacturer can train a network using the experimental date: the constants or Parameters of the trained network can then be transferred to the user who can calculate the per- formance of the heat exchanger under any other flow rate or inlet temperature conditions. ANN are an established technique; see, for example, Haykin (1994) for an account of the history and mathematical background. There have been a few applications to heat transfer prob- Jems. One isin the area of liquid crystal thermography to determine heat transfer coefficients (lambunathan et al. 1996). The generalization capacity of the ANN has been used to design a finned heat exchanger (Lavric et al. 1994, Lavrc etal. 1995). Huang and Nelson (1994) also applied this technique to determine the delay time for a HVAC plaat to respond to control actions. Ding and Wong (1990) controlled a simulated hydronic system using an ANN. In a pre vious paper Diaz etal (1996) applied ANNs to analyze experimental heat-exchanger data, ‘The goal of the present study isto represent heat exchangers using ANNs. The procedure used to set up and train the network is described fist. Then a series of problems of increasing com plexity are formulated to facilitate understanding. These problems are: one-dimensional conduc- tion, convection with one heat transfer coefficient, convection with two heat transfer coefficients, and single-row plate-fin heat exchanger. Artificial data bases are generated forthe firs three problems. Finally, an experimental date base will be used forthe fourth problem and the results ofthe ANN analysis presented, ANN ANALYSIS ‘An ANN, schematically shown in Figure 1, consists of a series of layers, each with a number of nodes. The first and last layers are the input and output layers, respectively. In a fully con- nected network, all nodes are connected to all nodes of the previous and following layers. The available data consist of M runs. Of these Mj will be used for taining and M, for testing. Each run is a single experiment providing a number of values of the physical variables for that run, AI physical variables used by the ANN are normalized to confine them in the range {0.15,0.85}. The following notation is inrogyced: (i) isthe jth node in the ith layer, x, is ts input, yi its output, 8, is its bias, and w? 4 is the synaptic weight between nodes (— 1,8) and (.j). Te total number of layers, including input and output layers, isn, and the number of nodes inthe ft layer is m. Te input information is propagated forward through the nework; ‘m, values enter the network and m, leave. The relation between the output of node (i 1,8) in ‘one layer andthe input of node (jn the following ayer is Hg = 8+ Die o“aR Resear Yount 5, NUMBER 3, JY 1988 rc +s that take {in any case eat transfer ' purpose a vore useful, node number problem of predicting network is be used for onstants or ate the per- ions. ount of the to design a 1994) also to control N.Inapre- data, Figure 1. Schematic of neural network 1 Furthermore, the relation between the input and output of the same node (i,j) is yi, = g¢x) cedure used wwhere aC) isthe sigmoidal activation fntion asing com- ral conduc at wansfer 3) = ated for the coblem and @ For the first layer the function g(x) = x is used. ‘Training of the network consists of changing the weights until the output values differ litle from known target values. This is done here through a backpropagation method (Rumethart et al, 1986; Thibault and Grandjean, 1991). Fist, the error is quantified atthe output layer by hanumber 5 fully con- ng = Ya nj Yay) @ Iyer, The oe sting. Each where fis the target output forthe node (nj). After calculation ofthe 6, we move back to the that run, (n= 1th layer. There isnot a target outpat to compare with for this layer, so the error is defined the range for the nodes inthis layer as 1. is its es (i-1,4) number of network; @-1bin may nD LE a Man ty @ i Having computed the errors forall the nodes ofthe (n~ I)th layer, (n~ 2)th layer is evaluated using an expression similar to the above, and so forth all the way to the second. Next the change in the weights and the biases are computed using a anil iy = AB i143 48, ©HVACER Reseact here 2s the learning rate that is used to scale down the degree of change made to the connec- ons. The larger the learning rate, the faster the net will learn, but the chances of the ANN being sable to reach the desired output are also greater. For the current work a value of = O.4is cho- ‘a which appears to work well. At the end of one cycle ofthe backpropagation, using the cor- cctions above, an updated set of values of the weights is obtained. Its best if the training data contain the minimum and the maximum of the entire data base. te backpropagation algorithm is repeatedly applied to this set of training data until a certain iterion is reached. The percentage error between the heat rate predicted by the ANN, Qanw, 1d the, ong supplied by the training set of data, Q is used. The relative heat flow is defined as 1 = Q/Qaww for i = 1,....My, where Q is the dimensional heat rate obtained expesimen- lly or by means of a known relation, and Qawy is the dimensional heat rate predieted by the NN. The standard deviation of Rj is © vere Ris the average of R, over all the nuns used for training. ‘The network 4-5-5-1 (four layers with 4,5, 5, and I nodes respectively) is analyzed as a typi- | configuration. The different starting conditions for the weights and biases were obtained in a dom fashion. Figure 2 shows @ as a function of the number of training cycles N. Each curve for a different sct of starting conditions. tis seen that the starting conditions do not make a wificant difference afier about 50,000 cycles, and that for ¢~IN-1/2 for large N. It was wumed that a training period of 100,000 cycles was adequate. It is important to note that the al weights obtained after training are not unique and different starting values of the weights ry lead to different outcomes. ‘The trained network was tested using weights and biases found from the traiping process with test data. The evaluation consists of the agreement between Qawiv and Q and the ratio of trates R; as a function of the run number i. NE-DIMENSIONAL CONDUCTION This is simple heat transfer problem that is chosen to illustrate the approach. The governing sation for one-dimensional conduction through a wall is o-"a-r o or situations in which KAYL is constant, function Q(T, 7) from Equation (7) represents a ne in three-dimensional space. Since the problem is linear, a linear activation function g(x) is appropriate. For the simplest work of the form 2-1-1, the functional relation is 3 21 = 05 + 7'1(82,1 + W137] + 4279) ®) Somparing Equations (7) and (8), established the following relations 3 851 +2121 ®snnemee Voune 5 Nome 3.17598 8 Hoxie ANN tng esiseho. ing the cor + data bas Gla genain IN Baam fied as ited bythe © sdasa type N Beaned in Earhcune Figure 2 Ect of ditferent starting conditions on standard deviation routes for configuration 45. For ference sri ie has sop af V2 PNT Diterent gil correspond to ierent Starting condos: viet te Se weighs Th eit arnt onique hough ere ga cela relonhp been hem. Fanbemore cn king, wi! = 1 te weight 99 and ware elated tothe stem parameter £4. ee sn ANN with more noes wil similar resus. Fr example, 222-1 newot within eat etvation fon the neigh re nommiqu ad shew physi sigan fra Spa cate The depen of feedom forthe cole of oye constants, however, cease wih he Sie of te ork. Teeth larger ANNG conta the aller es pial ewes tt fortis + governing Ter problem, dono beter in epresenng he uncon Q(T, 7) ‘The linear activation function is successful because Q(T,, 7) is a plane. Modification of the weights in soch an ANN produce rotation andlor transition of the plane. The function o (7, T,) is nota plane ifthe conductivity isa funetion of temperature, andthe activation func- tion will not accurately represent the surface, For example ifthe thermal conductivity i linear function ofthe temperature, Q(T), 7) is then a quadratic function. It was found tat, even with Spteeiea ‘quadratic activation functions, itis not possible to find physical meanings for the weights of the network. For nonlinear problems, the constants inthe ANN do not have a physical meaning. For he simplest this reason, the sigmoid activation function, which works well with nonlinear problems was used. ® CONVECTION WITH ONE HEAT TRANSFER COEFFICIENT In the situation of forced convection due to flow in a duct with constant wall temperature, the heat transfer is controlled by the convection heat transfer coefficient between the fluid and oy the wall, k. For the situation that h does not vary along the length of the duct, the heat flow is given byHVACER Researci 9 = telly ~Teand[ ero APL) ] a0) For the situation in which only Tip, Tyup and rt are varied and the other parameters being 1 ldconstant, O(TiqsTyqi)8) is a curved surface in four-dimensional space. An artificial data base was generated using equation (10) and known heat transfer coefficients ‘duct flow. The heat transfer for laminar flow, in which Nu = 3.66, and for turbulent flow, Nu 2.023 ReM’SPr?* were used; where Nu, Re, and Prare the Nusselt, Reynolds, and Prandtl num- 's, respectively. The critical Reynolds number for transition was taken to be 2300, the diame- of the duct was 0.05 m, and the fluid properties were those of air. Data were generated in the ge 230 < Re < 23044, For laminar flow the training and testing data sets were My = 142 and = 46, respectively, while for turbulent flow they were My = 277 and My = 91. The ANN was ined using artificial data and then the predictions of the trained network were compared with analytical results, The results for laminar flow data alone are presented first, then for turbu- \tflow alone, and finally for both cases combined. ‘configuration 3 1 was used for which the inputs were the mass flow rate, the inlet tem- ‘ature, and the wall femperature,and the output was the heat transfer rate, Figure 3 shows the 1 tionship between Qanw and G using two different networks trained for laminar and turbu- 14 flow, It was foun thatthe errors for laminar flow are in general, less than 2% with a maxi- 1 am ertor of 6.8%, For turbulent flow, the results were very similar. The maximum error larger Wn 5.8%, and in general the error is less than 1.5%, Most ofthe larger errors were for runs in ich the heat transfer rate was small, which makes the percentage error large. For both flow 1 imes combined, the errors were larger as shown in Figure 4. Large errors were observed for Kk ¢ values of the heat transfer rate, although good agreement was obtained for turbulent flow am Figure 3. Laminar (*) and turbulent flow (x) convection with one heat transfercoefficient. Straight line is Qaww = 9“AR Resoace oun, Neves 3, Ju 1958 21 0) T neters being | 2") covfficients | :nt flow, Na Eat “rand num- - whe diame- rated in the = = 142 and y ie ANN was | y pared with 1 for turbu- ve inlet tem em 3 shows the ined lam heb Figure 4, Combined laminar and turbulent flow with one Sree heat transfer coefficient. Straight line is Qaww = @ error larger . + for nuns in with large values of Q . Thus, this simple exercise showed that separating laminar and turbulent 'r both flow flow data yielded better heat transfer predictions. sbserved for ‘bulent flow CONVECTION WITH TWO HEAT TRANSFER COEFFICIENTS ‘The situation of foreed convection ina duct with extemal fins and a known outside tempera- ture was considered next. There are two heat tansfer coefficients one inside and another ouside the duct. The data for waining the ANN and testing its predictions were generated artificially. For the situation with water the in-tube and sir the over-tube fluid, the heat transfer rate is a funetion of four variables: the two inlet temperatures 7," and 7,2, and the two mass flow rates, ‘iy and mig. For the case with the nner and outer heat transfer coeficiets, hand hi, respec: tively, constant along the length of the heat exchanger tube, the heat flow is given by o= Pie Te- TH) er =) ap where a2) Artificial data were generated using the inner and outer power-law correlations of Zhao (4995) for the same heat exchanger reported in the following section. These are 0585173 Na, = 0.1368Re)*Pr)/? for 200 Re, $700 3)HVACER Researex B88 fs eg Quy (WI eam igure 5. Two heat transfer coefficients, Straightline is Qanw = @ 078, Nu, = 0.01854Re).”"*Pr> for 800mputerized data acquisition system collected the experimental data. A nominal 457 mm byACaR Reseanen 2 (a4) as) he configura: 25. The max- rage ratio of, by means of acteriste pat. imental data snnel facility, as the in-tube blower that is downstream heater raised ‘Type T ther- air is close to average read- 4 upstream of air flow rate, 11457 mm by Vouwne 5, Noser 3, Jay 1999 203 @ €)- * © Figure 6. Experimental setup: (a) wind tunnel, (b) heat exchanger 610 mm (18 in. by 24 in.) water coil single-row, plate-in proprietary heat exchanger was the heat exchanger used. Additional experimental deals are given by Zhao (1995). A total of M = 259 runs were made of which Af, = 197 were for taining and M = 62 for testing The experimental data obtained were rt, and ry, the inlet temperatures, 7% and Ti, outlet temperatures, T3,, and 7", for both fluids. The heat transfer rate Q was determined from energy balances ‘on each fuid. Since the air and water sides give slightly different Q (thin 10%), an average value was used. Data from 75% of the runs were randomly selected for training the ANN and the remaining 25% of the rans were used for testing the predictions of the trained network. ‘The configuration 4-5-2-1-1 was chosen first. The input variables tothe network were Hi. ‘yy, Thy and 77% the output is Oawy. The performance ofthe trained network was evaluated by comparing the heat flow prediction with the dataset aide for testing. The average value of R is found to be R= 1.00212 with a standard deviation of 6 = 0.017 and a maximum error of 17.88%, To examine the effect of network configuration, 14 different configurations using the same seis of data for training and testing were evaluated. The values of P and the standard de ation 6 are shown in Table 1. The network configuration with R closest to unity is 4-1-1-1, while208 HVACER Reseanct 188] ap 70 Figure 7. Ratio of heat transfer rates Rj for run i. x 4-545-1; +4-5-1-1, Table 1. Comparison of Heat Transfer Rates Predicted by Different Configurations (Configuration R © TO2373 0266 0.98732 0.084 0.99796 ois, 1.00065 0.265 0.96579 0.089 1.00075 003s 1.00800 0018 1.00288 ois 095743 0288 0.99481 0.032 45201 1.00212 017 45-5411 Loo2a 0.016 45521 1.00397 019 455. 0.022 a. 5-5-1 is the one with smallest 0. Thus there are atleast two different criteria for the selection £ an optimal configuration, though in either case the erro and the scatter are small. The config ‘ation 4-1-1-1 has a much larger standard deviation with respect tothe others so the configura on 4-S-1-1 was selected, which has a value of R also very close to unity but with a smaller andard deviation. The performance of the 4-5-5-1 and 4-5-1-1 configurations are also com- ved in more detail. The values of R; for individual runs are shown in Figure 7. Although ‘5-1-1 has the second best R, there are some points in which the prediction differed from the ‘periment by more than 14%. The 4-5-5-1 network, on the other hand, has etrors confined to ss than 3.7%, It was observed that no matter how many layers the network has, the predictionAcCAR ReseaRc gurations 266 084 O18 265 089 035 ols ols 258 032 a7 016 a9 om the selection ‘The config- ve configura- ith a smaller re also com- 7. Although red from the s confined to he prediction oun 5, Numger 3, JuLy 1999 205 hhas a large standard deviation if all hidden layers contained just one node, For all but the three configurations of hidden layers with just one node, the different configurations overpredict the hheat transfer rates with an average error of only 0.33%. ‘The effect of the normalization used for the variables was also studied. For the network 45.5-1, which has the smallest standard deviation, a computation using a different normaliza- tion range of (0.05,0.95] was run. The number of cycles used was the same as before, ie. N 100,000. The results showed that R = 1.00063 and = 0.016, which can be compared to the val- ues in Table 1. There was only a 6.6% difference in c due to changing the normalization range, Once the network has been trained for a given heat exchanger, as by the manufacturer, the information can be transferred to the user as a set of weights and biases. For example, the data are put in columnar form for the configuration 4-2-1 and presented in Tables 2 and 3. A com- puter code that will ead this datafile and be able to predict the performance of the device for any other flow rate or inlet temperature within the range tested could easily be written, Table 2. Values of the Weights for Configuration 4-2-1 wy “Ema 401 1321 120 o7n 1.356 -0303 3 1441 856 k Table 3. Values of the iases for Configuration 4-2-1 ey T1314 2aT4 1.848 DISCUSSION AND CONCLUSIONS ‘The problem of prediction of heat-exchanger performance is one of transfer of manufacturer's test data to a designer in suitably compact form so as to be accurate enough for calculations. ‘There are many ways in which this can be done. At one extreme, the information can be con- tained in two fundamental heat transfer correlations, one forthe inside and another forthe out- side, Another extreme is to transfer the entire test data for interpolation. The technique using ‘ANN lies somewhere in the middle of these two approaches. The technique has the accuracy comparable to the latter while the information being transferred consists only of the synaptic ‘weights and nodal biases of the trained network. A comparison between the ANN prediction aww, and that from the correlations Qcor. is shown in Figure 8. The network is seen to give better predictions with a smaller scatter. The correlation was specifically developed from exper- imental data for this heat exchanger and the methodology used is given by Zhao (1995).206, HVACAR Reseanct Figure 8. Comparison of ANN and correlation predictions; correlations (*)s ANN (+) (Straightline is Gann = @ and Qeor = @ dotted lines represent +10% deviations.) ‘The results of tests with artificial data for both one and two heat transfer coefficients indicated that the ANNs predict heat flow accurately and just as well as Equations (10) and (11). These equations assume that the heat transfer coefficient is constant along the length of the heat exchanger tubes. There are several reasons why the assumption does not hold for an actual heat exchanger, There arc hydrodynamic and thermal entrance regions, secondary flows in the tube bends, complex vortex structures in the neighborhood ofthe tube-fn junctions, heat conduction along tube walls, natural convection within the tubes and between fins, temperature dependence of fluid properties like the viscosity. There are far too many nondimensional fluid, flow, thermal and geometrical parameters inthe problem for existing correlations to be accurate. The relation- ships between the Nusselt, Reynolds, and Pranddl numbers only oversimplifie the problem so thatthe resulting predictions have a large uncertainty that are not due to measurement error but 10 information compaction through corelations Correlations in terms of Reynolds and Pranddl numbers have the advantage that predictions ‘ean be made for fluids other than those used in the original experiments. The price for this is a lack of precision inthe predictions. One reason is that any correlation chosen, e.g., power-laws ‘may not accurately represent the actuel phenomena due tothe use of limited number of parame- ters. Another isthe uncertainty due to the property variations with temperature that must be con- sidered twice, once when obtaining the correlations from experimental data and next when using them for predictions for a particular application. This is an important source of error especially when the fluids used for the former are different from the later. ‘Under these circumstances, the ANN approach isan attractive alternative that has been shown to be satisfactory for heat exchangers using data both artificially generated and for experimen- tally obizined from a compact heat exchanger. The errors in prediction using a trained ANN are comparable to measurement errors. The precision is much better than from simplified correla- tions and is comparable toa direct transfer of the entire test data, The usefulness ofthe ANN approach is that it basically allows prediction of the behavior of a given heat exchanger without need for an accurate mathematical model of the details of the pro-ACAR Reseanes 1 AINN (+) fons) cents indicated vd (11). These hh of the heat ‘an actual heat ws in the tube sat conduction ce dependence flow, thermal The relation xe problem so ‘ment error but rat predictions 2e for this is 2 1. power-laws 2er of parame- {must be con- xt when using stor especially as been shown or experimen- ined ANN are ified correla- behavior of a ails of the pro- Vouue , NuMBER 3, JULY 1999 2 ‘cess. With processes involved that cannot be modeled exactly using first principles, the ANN is ‘an effective means for the transfer of information from the manufacturers’ laboratory to the design engineer who would like to use test data for application. All that would have to be trans- ferred are the weights and biases corresponding to a particular heat exchanger. The designer simply reads these values into his or her own network and is quickly in a position to make accu- rate predictions of the thermal behavior of the heat exchanger, ACKNOWLEDGMENTS ‘We gratefully acknowledge the support of Mr. D.K, Dorini of BRDG-TNDR for this and related projects in the Hydronics Laboratory. G.D. also thanks the Organization of American States for a PRA Fellowship, NOMENCLATURE A ‘transverse area for conduction, m? R average of R; over all test runs. Aj Ag inner and outer heat transfer areas, m? Re Reynolds number De te ine daneter m a ncerenn Ty © seaichea G1) temperature, °C tip jnowerta lene fag: Opec 5G) activation function U overall heat transfer coefficient, Wim?-K) b ba ee arene eS ich V,, average velocity of uid, mis . Seer wraps ween as sm mass flow rate, kgs iy imputto node (ij) * woe mcm frm pe 5 My number of raining runs: &y ere Eocene) My mamberet estrus a facets n ‘total number of ANN layers a N —almnterofeelefortaning «Yet wieosyo ui, wt No Nossl mer By ins of nde ,p > See oY standard devin overall runs cease Subscripts and Sapererpts, 3 emtatrratebeveen fu, Wasi Qanw heat transfer rate predicted by ANN, W me inlet Gor eaten rat peiced by powerinw ou! ot torelatons W war ie mS 0/Ban forrai sll REFERENCES Diaz, G, J. Yanes, M. Sen, K-T. Yang, and R-L. McClain, 1996, Analysis of data from single-row heat exchanger experiments using an artificial neural network. In Proceedings of the ASME Fluids Engi- neering Division FED-242: 45.52, Ding, Y., and K.V. Wong. 1990. Control of 2 simulated dual-temperature hydronic system using a neural network approach. ASHRAE Transactions 96(2): 727-32. Haykin,$. 1994. Neural Nenworks, A Comprehensive Foundation. New York: Macmillan College Publish- ing Co, Huang, 8-H. and RM, Nelson,. 1994, Delay time determination using an artificial neural network. ASHRAE Transactions 100(1): 831-40 Jambunathan, K., S.L. Hartle, 8. Ashforth-Frost, and V.N. Fontama, 1996. Evaluating convective heat transfer coefficients using nevral networks. Intemational Jounal of Heas end Mass Transfer 39(11): 2329.32,208 HVACAR Reseancst Lavric, D.,V. Lavrie, 0, Muntean, and E. Danciu, 1994, Auto-orgenizing algorithm for design of fin heat ‘exchanger. Revue Roumaine de Chimie 3911): 1241-56. Lavri,D., V. Lavric, and A. Woinaroschy. 1995, Designing fin heat exchanger with a neural network. Rerue Rouraine de Chimie 106): 561-65. Rumelhar, D.E., GE. Hinton, and RJ. Williams. 1986. Learning internal representations by error propags- tion, In Parallel Distributed Processing: Explorations inthe Microstructure of Cognition. Cambridge, MA: MIT Press. Thibault, J, and B. Grandjean. 1991. A neural network methodology for heat transfer data analysis. Iter- national Jounal of Heat and Mass Transfer 34(8): 2063-70. Zhao, X. 1995. Performance of a Single-Row Heat Exchanger at Low In-Tube Flow Rates. MS. Thesis, Department of Aerospace and Mechanical Engineering, University of Notre Dame, Notre Dame, IN. | | |