Privacy-Preserving and Adaptive Federated Deep Learning for Multiparty Wind Power Forecasting

Advanced forecasting tools are essential for modern power systems to mitigate the uncertainty of renewable energy sources. Although data-driven approaches have achieved significant progress in wind power forecasting, their effectiveness is hindered by limited data availability. Strict data regulation rules and competitive interests block neighboring wind farms from combing datasets to learn a more accurate forecasting model. To tackle this challenge, we propose SecFedAProx-LSTM, a novel wind power forecasting method incorporating a deep learning model and a privacy-preserving adaptive federated learning framework. This approach dynamically adjusts local optimization objectives to strike a balance between global convergence performance and exploring individual characteristics, thereby addressing statistical heterogeneity. Additionally, it employs decentralized multiclient functional encryption for secure model aggregation to enhance privacy-preserving guarantees. Extensive experiments conducted on the Wind Integration National Dataset validate the superiority of our method over existing approaches, as it improves forecasting performance and accelerates convergence speed. Furthermore, it efficiently protects user privacy from inference attacks without introducing extra entities.