A transferable federated learning approach for wind power prediction based on active privacy clustering and knowledge merge

As wind power continues to develop, advancing wind power prediction becomes more and more crucial. This study focuses on advancing wind power prediction by addressing data privacy and enhancing model applicability at multi-spatial scales, from individual turbines to entire farms. Traditional methods are typically confined to a single scale, lacking flexibility in application, requiring extensive data from farms, which potentially compromises energy data privacy. To tackle these challenges, we introduce an innovative Divide-Merge Federated Learning with Active Private Clustering (D-M APCFed) approach. This approach strategically employs federated learning to train models within privacy-preserving boundary, overcoming the adverse effects of wind power data heterogeneity through a novel APC method and knowledge merge technique. The primary innovation of this study is a scalable and accurate wind power prediction model that operates effectively at multi-spatial scales while safeguarding energy data privacy. In case study of two spatial scales, the D-M APCFed approach achieves an average prediction accuracy of 87.11 % in the twelve federated farms and 81.69 % in the twenty federated turbines. This approach enables a more generalized model through the secure use of data from diverse sources at multi-spatial scales, enhancing prediction accuracy and ensuring the confidentiality of sensitive information.