Assessing new intra-daily temperature-based machine learning models to outperform solar radiation predictions in different conditions

The measure of solar radiation is costly, as well as its maintenance and calibration needs; therefore, reliable datasets are scarce. In this work, several machine learning models to predict solar radiation have been developed and assessed at nine locations (Southern Spain and North Carolina in the USA), representing different geo-climatic conditions (aridity, sea distance, and elevation). As a novelty, due to the ease of providing air temperature measurements, different new input variables from intra-daily temperature datasets were used. According to the results, all the models highly outperformed self-calibrated empirical methods such as Hargreaves-Samani and Bristow-Campbell, with improvements in RMSE ranging from 7.56% in arid climate to 45.65% in humid. Moreover, regarding mean NSE and R2 values, several inland locations obtained values above 0.9. In summer, the highest statistics for all sites (more than a 60% improvement in NSE and R2) were obtained, whereas the worst were given in winter (more than an 18% improvement in NSE and R2). Besides, when assessing the models in different non-used locations with similar climatic characteristics, the reduction in RMSE was from 0.305 W m−2 to 0.252 W m−2 in a semiarid coastal climate and from 0.344 W m−2 to 0.233 W m−2 in dry sub-humid climate, compared to Hargreaves-Samani method. Overall, the MLP obtained the highest performance using the new proposed variables in all locations with medium aridity values, whereas in the aridest and most humid sites, SVM and RF models were preferred. Therefore, the temperature-based models developed in this work can predict solar radiation more accurately than the current ones. This is crucial in locations with no available datasets or missing/low quality and can be used to optimize the determination of the potential locations for solar power plants' construction.