Forecasting global solar irradiance for various resolutions using time series models-case study: Algeria

The objective of this research is to build models for various time resolutions to predict global solar irradiation using data mining and statistical techniques. The time resolutions analyzed are 5 min, 1 hour and one-day horizon ahead. The models tested herein are three supervised machine learning (ML) techniques: nonlinear autoregressive neural network (NAR), support vector regression (SVR) and random forest (RF). A linear autoregressive (AR) model and the naive persistence (PER) model have also been included. The datasets come from two sites situated in Algeria: Algiers and Ghardaia that have different climatic conditions during the year corresponding to two types of climate, Mediterranean and Arid. One important contribution of this research to global irradiance forecasting is the benchmarking of the ML used, taking into account the lack of practical results and the needs detected in the literature, especially for the case of RF model; according to our best knowledge, the random forest method has never been tested as it has been done in our study: it is just based on past values of the same variable without exogenous data to forecast the future ones. The results of this research show that RF is the best technique with a slight difference in performance, specially for hourly forecasts ahead. The proposed models appear to be less outstanding both in the case of unstable sky conditions (Algiers) and when the resolution is 1 day, due to the fact that time series become significantly less correlated by including more randomness characteristic.