FedWindT: Federated learning assisted transformer architecture for collaborative and secure wind power forecasting in diverse conditions

Abstract

Accurate wind power forecasting is crucial for efficient grid management and maximizing the utilization of wind energy. This study introduces the FedWindT, an innovative model that combines transformer neural architectures with federated learning, specifically designed to enhance wind power prediction. The transformer’s self-attention mechanism adeptly captures the temporal dynamics of wind data, while federated learning facilitates a decentralized, privacy-preserving training process. Our comprehensive empirical analysis across multiple wind farm datasets demonstrates that the FedWindT consistently outperforms traditional state-of-the-art centralized approaches. Specifically, the FedWindT achieved an average Normalized Mean Squared Error (NMSE) of 0.0109, Mean Absolute Error (MAE) of 0.0243, and Root Mean Squared Error (RMSE) of 0.0288, with R-squared ($R^2$) values consistently closer to 1, indicating high predictive accuracy. These results validate the effectiveness of combining federated learning with advanced neural architectures and highlight a promising direction for future decentralized energy forecasting solutions.