Evaluating different types of artificial neural network structures for performance prediction of compact heat exchanger

In the present work, the performance of an air-to-refrigerant laminated type evaporator is predicted using a genetic algorithm (GA)-integrated feed-forward neural network (FFNN) and recurrent neural network (RNN). The obtained results are compared with the results of the FFNN with back-propagation learning algorithm, as the most recommended algorithm in the literature. The considered evaporator consists of single-phase and two-phase regions in the refrigerant side which makes the ANN-based methods so suitable for its modeling. To train the mentioned neural networks, the steady-state experimental data of the evaporator performance include capacity, outlet refrigerant pressure and temperature and outlet air dry- and wet-bulb temperatures is collected with varying input parameters. The results show a good agreement with experimental data, and it is observed that RNN-based method has the best average root-mean-square error (1.169 against 5.017, 4.791 and 2.286 for FFNN, GA-trained FFNN and numerical modeling, respectively). In fact, using GA to optimize FFNN structure makes better results than conventional FFNN, but the RNN method provides the best results because of using suitable intelligent configuration. Also, in contrary to numerical method, it is much faster and calculation processing load is lower. Therefore, RNN is proposed as a substitute for FFNN and the GA-trained FFNN. Finally, a sensitivity analysis determined the inlet refrigerant pressure as the most important parameter in predicting the evaporator capacity.