Solar Irradiance Forecasting Using Ensemble Machine Learning Methods: A Sustainable Energy Perspective

Abstract

Accurate solar irradiance forecasting is critical for reliable power-grid management. This study evaluates six ensemble machine learning architectures — Random Forest (RF), Gradient Boosting Machine (GBM), XGBoost, AdaBoost, LightGBM, and a stacked heterogeneous ensemble — for global horizontal irradiance (GHI) prediction. Three publicly available datasets were employed: the National Solar Radiation Database (NSRDB v3.2.2), the PVGIS Typical Meteorological Year (TMY v5.2), and ERA5 reanalysis fields. The stacked ensemble achieved a normalised root mean square error (nRMSE) of 7.3% across a 12-month hold-out evaluation window, surpassing the best single-model baseline (XGBoost, nRMSE 9.8%) by 25.5% in relative error reduction. Model sensitivity to cloud-cover variability, aerosol loading, and seasonal transitions was systematically characterised. The findings indicate that no single tree-based model generalises robustly across all irradiance regimes; stacking, however, compensates for individual model biases in a measurable and reproducible manner. A critical synthesis of 37 prior studies (2015–2024), a full experimental pipeline, and an operationally grounded evaluation metric framework are provided as a practitioner-oriented reference.