Machine Learning for Wind Power Forecasting Under Data Scarcity: Comparing Spatial Feature Configurations and Selection Strategies

Accurate wind power forecasting is essential for integrating renewable energy into power grids, especially under limited data availability. This work introduces a spatially distributed grid-based approach to construct meteorological features, proposing valid alternatives to previous methods that used province-level data in data-scarce scenarios. Using Random Forest and Least Absolute Shrinkage and Selection Operator (LASSO) models, we evaluate feature importance and demonstrate that including the previous hour's power output notably improves prediction accuracy. Additionally, feature selection significantly reduces computational costs and data requirements while maintaining strong performance.