New soft computing model for multi-hours forecasting of global solar radiation

The growing development of photovoltaic technology has explored the role of effective solar irradiance forecasting in grid security and stability However, due to its non-stationary nature and complexity, make its estimating extremely difficult. The scope of this work is to deal with this issue by introducing a new machine learning forecasting architecture for multi-hours ahead in multiple-site in Algeria. Specifically, we proposed a new decomposition-based ensemble-forecasting model. The developed forecasting strategy based on a new multi-scale decomposition algorithm named Iterative Filtering (IF) used as a pre-processing stage of the historical solar radiation data combined with Gaussian Process Regression (GPR) as an essence predictor to build an IF-GPR model. Hourly global solar radiation data of two years from different cities with diverse solar radiation profiles are used to validate the full potential of the newly proposed IF-GPR model. The performance of the proposed IF-GPR is rigorously assessed utilizing effective metrics and comparing its performance with the reference model. The forecasting results demonstrate the potential of the hybridization IF-GPR methodology for multi-hour forecasting up to four hours ahead. The forecasting errors in terms of normalized root-mean-square error for four hours ahead are as follows: 0.7, 2.45, 5.496, and 9.76 for the Algiers site; 0.373, 1.34, 2.81 and 11.22 for the Ghardaia region, while the attained results for the Adrar site are equal to: 0.525, 1.36, 2.73, and 4.73. Furthermore, the proposed IF method outperforms the recently introduced decomposition algorithm, complete ensemble empirical mode decomposition with adaptive noise, in boosting the forecasting ability of a stand-alone model.