Can machine learning enhance day-ahead renewable power forecasts? A study on data-driven methods for solar photovoltaic and wind turbines

This study evaluates and compares various data-driven approaches to improve renewable energy production forecasts. In particular, focus is given here to photovoltaic (PV) fields and wind turbines (WTs), which commonly represent the backbone of modern Hybrid Power Stations (HPSs). With the aim of increasing the accuracy of commercial forecasting, the study investigates the potential improvement that machine learning (ML) techniques in correcting case-specific prediction errors. This improvement is particularly valuable for optimizing HPS scheduling and battery management, ultimately boosting system efficiency and reliability. With reference to a small Mediterranean island as the case study, for the PV Linear Regression (LR) achieved notable results, reducing the Normalized Mean Absolute Error (NMAE) from 8.6 % down to 3.59 %. Support Vector Regression (SVR) and Neural Networks (NNs) further enhanced accuracy, with NNs providing the best results (NMAE reduced to 3.34 %). In contrast, the improvement in wind power forecasts was more modest, with SVR performing best (NMAE only reduced from 10.25 % to 9.84 %). While ML successfully eliminated bias errors in the commercial forecast, they were less effective in addressing the variability and oscillations in wind production, highlighting inherent data limitations. Sensitivity analyses also revealed that shorter training windows improved algorithm performance, capturing recent patterns while minimizing noise from older, less relevant data. Results suggest that the introduction of ML correction methods into RES planning can support policy strategies aimed at enhancing renewable integration and improving the resilience of isolated power systems.