An Empirical Study of Tree-Based and Instance-Based Models for Short- and Long-Term Solar Energy Forecasting

Abstract

The integration of solar energy into modern power grids necessitates accurate forecasting across multiple time horizons to ensure grid stability and operational efficiency. This paper presents a comprehensive empirical investigation of tree-based ensemble methods and instance-based learning approaches for solar irradiance forecasting. We evaluate random forests, gradient boosting, evolutionary forests, and quantile regression forests against instance-based methods including k-nearest neighbours and regime-dependent artificial neural networks. Experiments are conducted across six climatically diverse locations in Morocco and three sites with varying meteorological variability. Results demonstrate that tree-based ensemble methods consistently outperform instance-based approaches across most forecasting horizons (1–6 hours), with the proposed evolutionary forest model achieving n RMSE values between 4.94% and 18.94% depending on climatic conditions. Hybrid input configurations combining endogenous and exogenous variables yield superior accuracy across all models. Our findings indicate that tree-based methods exhibit particular advantages in high-variability conditions and when training data is limited, while instance-based methods show competitive performance primarily in stable, clear sky conditions.