Evaluation of Artificial Neural Network (ANN) and Irmak Experimental Models to Predict Daily Solar Net Radiation (Rn) in Cold Semi-arid Climate (Case study: Hamedan)

Solar net radiation (R_n) is one of the most important factors influencing soil heat flux and evapotranspiration rate process. This parameter is derived from the difference between downward and upward radiation fluxes reaching the earth’s surface. Field measurements of R_n are cost effective and difficult to maintain. Therefore, in the most cases, R_n is estimated by empirical, semi- empirical and physical-based models. Recent studies show that the artificial neural network (ANN) is a reliable tool for estimating daily R_n with reasonable performance for the area where lack or shortage of field R_n exists. Using Irmak model and ANN approach, we tried to estimate daily R_n for one of the cold semi-arid sites located in Hamedan. For model evaluations, R_n data were measured in hourly base during December 2011 to December 2012 at Bu-Ali Sina University weather site. In this study, we used 11 daily meteorological parameters as the inputs of ANN to generate the R_n estimates (70% of the data set for training data and 30% for model validation). The results showed that the best model performance of ANN was obtained from a 11-2-1 architecture and the sigmoid function based on the back- propagation training algorithm. The least ANN error was observed by employing 10000 iterations for the training step and two neurons in the hidden layers. The results indicated that the daily net radiation from ANN was more accurate (R²>0.95) than the previously recommended Irmak model.