Generalized neural network model of alternative refrigerant (R600a) inside a smooth tube

The correct prediction of refrigerant condensation heat transfer performance is important for design of condensers. A generalized neural network correlation for condensation heat transfer coefficient of alternative refrigerant R600a inside horizontal tube has been developed in this paper. Mass flow rate, vapor qualities, saturation temperature, difference value temperature are selected as the input parameters, while the Nusselt number and heat transfer coefficient as the output. Three-layer network is used for predicting the Nusselt number and the heat transfer coefficient. The number of the neurons in the hidden layer was determined by a trial and error process together with cross-validation of the experimental data evaluating the performance of the network and standard sensitivity analysis. The trained network gives the best values over the correlations with less than 4% mean relative error. The experimental data of the heat transfer coefficients of R600a, a hydrocarbon refrigerant, in a horizontal smooth copper tube with an inner diameter of 4 mm and outer diameter of 6 mm are from Agra et al. [O. Agra, “Condensation of refrigerants in a horizontal tube in annular flow regime”, PhD thesis Yildiz Technical University, 2007]. The condensing heat transfer coefficients obtained from the experimental study were seen to be consistent by ± 20% with the correlations developed by Shah [M.M. Shah, A general correlation for heat transfer during film condensation inside pipes, Int. J. Heat Mass Transfer 22 (1979) 547–556], Travis [D.P. Traviss, W.M. Rohsenow, A.B. Baron, Forced convection condensation inside tubes: a heat transfer equation for condenser design, ASHRAE Trans. 79 (1972) 157–165] and Cavallini-Zecchin [A. Cavallini, R. Zecchin, A dimensionless correlation for heat transfer in forced convection condensation, Proceedings of the Fifth International Heat Transfer Conference, vol. 3, 1974, pp. 309–313]. And it is seen that results from the trained network shows good agreement with the experimental data and better results than the correlations given by Shah, Cavallini and Travis.