Implementation Issues in Artificial Neural Network Based Thermal Analysis

It is now known the generally it can be demonstrated that artificial neural network (ANN), particularly the fully-connected feedforward configuration with backward propagation error-correction routine, can be a rather effective and accurate tool to correlate performance data of thermal devices such as heat exchangers (Sen and Yang, 2000; Kalogirou, 1999). Good examples are the recent demonstrations for the compact fin-tube heat exchangers (Diaz et al., 1999a; Yang et al., 2000; Pacheco-Vega et al., 1999) including those with complex geometries and also two-phase evaporators (Pacheco-Vega et al., 2000) as well as the dynamic modeling of such heat exchangers and their adaptive control (Diaz et al., 1999b; Diaz et al., 2000). Unfortunately, despite such successes, there are still implementation issues of the ANN analysis which lead to uncertainties in its applications and the achieved results. The present paper discusses such issues and the current practices in dealing with them. Those that will be discussed include the number of hidden layers, the number of nodes in each hidden layer, the range within which the input-output data are normalized, the initial assignment of weights and biases, the selection of training data sets, and the training rate. As will be shown, the specific choices are by no means trivial, and yet are rather important in achieving good ANN results in any given application. Since there are no general sound theoretical basis for such choices at the present time, past experience and numerical experimentation are often the best guides. However, many of these choices and issues relating to them involve optimization. As a result. Some of the existing optimization algorithms may prove to be useful and highly desirable in this regard. The current on-going research to provide some rational basis in these issues will also be discussed. Finally, it will also be mentioned that successfully implemented ANNs have many additional uses in practice. Examples include parameter sensitivity analysis, training, design of new experiments, and clustering of data sets.