Hourly Global Solar Forecasting Models Based on a Supervised Machine Learning Algorithm and Time Series Principle

In this paper, predictive models of hourly global solar radiation (HGSR) at one-hour step ahead have been developed by adopting a new methodology. It consists of the association of a supervised machine learning algorithm named Support Vector Machine (SVM) with time series principle. A dataset of HGSR, collected at Ghardaia city in the south of Algeria, has been used. In order to evaluate the accuracy of the developed models, we have used the Kolmogorov-Smirnov Integral (KSI) in addition to the conventional metrics (RMSE, NRMSE, NMBE, MAPE, and R). The results showed that accurate forecasts of HGSR at one-hour step ahead have been obtained by considering only its previous values at two-step forward. The predictive performances of the selected SVM model have been compared with those of some models available in the literature. This work proved the ability of the investigated machine learning algorithm with time series principle in predicting HGSR with good accuracy.