Assessment of a Temperature-Based Artificial Neural Network Designed for Global Solar Radiation Estimation in Regions with Sparse Experimental Data

The aim is to evaluate a model of monthly mean global solar radiation based on a simple ANN that uses geographic and temperature data as input variables and is designed for estimations in regions with few radiometric stations. Using data from 414 Spanish stations, the performance of the model is evaluated when both the number and the percentage of data collected for training the network are significantly modified while maintaining the clustering algorithms. The statistical indicators obtained show a compromise between achieving a lower mean error for all stations and limiting the maximum error at each station. In the worst case, the average error is less than 10% for all stations, and the maximum local error only exceeds 20% in less than 2% of the estimates. The least accurate predictions seem to be related to climate types where the clearness index tends to be higher in winter than in summer, which is the case in some locations on the northern Spanish coast. The results are consistent with estimates obtained for 16 non-Spanish stations, selected within the same input data range, suggesting that the variation of the clearness index over the year could be an important factor for local climate characterization.