A Comparative Study of Statistical and Machine Learning Methods for Solar Irradiance Forecasting Using the Folsom PLC Dataset

The increasing penetration of photovoltaic solar energy has intensified the need for accurate production forecasting to ensure efficient grid operation. This study presents a critical comparison of traditional statistical methods and machine learning approaches for forecasting solar irradiance using the benchmark Folsom PLC dataset. Two primary research questions are addressed: whether machine learning models outperform traditional techniques, and whether time series modelling improves prediction accuracy. The analysis includes an evaluation of a range of models, including statistical regressions (OLS, LASSO, ridge), regression trees, neural networks, LSTM, and random forests, which are applied to physical modelling and time series approaches. The results reveal that although machine learning methods can outperform statistical models, particularly with the inclusion of exogenous weather features, they are not universally superior across all forecasting horizons. Furthermore, pure time series approach models yield lower performance. However, a hybrid approach in which physical models are integrated with machine learning demonstrates significantly improved accuracy. These findings highlight the value of hybrid models for photovoltaic forecasting and suggest strategic directions for operational implementation.