Estimation of Global Solar Radiation using Regression and Artificial Neural Networks Models

Artificial neural networks (ANNs) was used to estimate daily global solar radiation at a weather station lacking any measured Rs values based on the measured Rs values at another station with similar climate. The accuracy of ANNs was compared with that of six other models developed for estimating Rs including FAO-56, Hargreaves-Samani, Mahmood-Hubard, Bahel, Annandale, and Bristow-Campbell models. The weather data was selected from Karaj and Shiraz weather stations having arid and semi-arid climates, respectively. The weather data of Karaj station, where daily global solar radiation was measured, was used to train ANNs and Shiraz data was used for validation. ANNs generated daily global solar radiation estimates with higher degree of accuracy as compared with all the other models used with the input parameters of maximum possible sunshine hours and daily extraterrestrial solar radiation, which both depend on latitude and day of the year, and actual sunshine hours with root mean square error (RMSE) of 2.34 Mj m-2 day-1 and correlation coefficient (R) of 0.94 (at 1 percent significant level). In case actual sunshine hours was not available, Annandel and Hargreaves-Samani models with locally calibrated empirical parameters and ANNs with minimum and maximum air temperatures and extraterrestrial radiation as input parameters gave the best results.