The transferability of random forest and support vector machine for estimating daily global solar radiation using sunshine duration over different climate zones

The transferability of random forest (RF) and support vector machine (SVM) for estimating daily global solar radiation using long-term data of measured sunshine duration, extraterrestrial solar radiation, and theoretical sunshine duration was evaluated across different climate zones. Root mean square error (RMSE), Pearson correlation coefficient (R), and Lin’s concordance correlation coefficient (LCCC) were applied to evaluate model transferability performance. Generally, RF and SVM gave better transfer performance in the climate zone where they were developed. On average, RF (RMSE = 0.881 kWh/m2, R = 0.918, LCCC = 0.885) performed better than SVM (RMSE = 0.93 kWh/m2, R = 0.913, LCCC = 0.87) over the study area. RF had narrow ranges of RMSE, R, and LCCC, indicating that RF was more stable for transfer. The transferability performance of RF was mainly affected by the difference in elevation between source and target sites, and SVM was mostly controlled by the distance and difference in elevation between source and target sites. The results indicated that RF might be applied to estimate daily global solar radiation using sunshine duration at the sites within 500 km distance and 1000 m difference in elevation, and SVM within 500 km distance and 500 m difference in elevation between source and target sites.