Performance evaluation of Neural Network and Multivariate Regression Methods for Estimation of Total Solar Radiation at several stations in Arid and Semi-arid Climates

In this study, the capability of multi-layer perceptron (MLP) and multivariate linear regression methods were evaluated to estimate the total solar radiation. For this purpose, the daily weather data of 25 years (1992-2017) including maximum temperature, mean temperature, relative humidity, sunshine hours and solar radiation were used in the five synoptic stations (Bandarabbas, Zanjan, Shiraz, Kerman and Mashhad). The inputs used in the models included various combinations of these variables, and the output was the solar radiation. To evaluate the performance of these models, Determination of Coefficient (R2), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE) and Index of Agreement (IA) were used. In order to train the structure of the ANN, two Bayesian-regularization (Br) and Levenberg-Marquardt (LM) algorithms were compared. Moreover, the training and validation processes were performed. The results of regression model showed that all the input variables are effective on the solar radiation estimation at Bandarabbas, Zanjan and Shiraz, but the effect of relative humidity on radiation at Kerman and Mashhad stations was low. The ANN application with two algorithms showed that Bandarabbas and Kerman stations using the Br algorithm and Zanjan, Shiraz and Mashhad using the LM algorithm give a good result. The lowest values of RMSE, MAE and the highest value of IA and R2 related to Kerman station were 2.799, 0.94, 0.954 and 0.838, respectively. As a main result, the comparison between computation and observation data showed that the ANN model gives better results than the linear regression model for estimation of radiation.