A privacy-preserving personalized federated learning and domain adaptation fusion framework for wind power prediction

Wind power forecasting plays a crucial role in achieving high proportions of renewable energy integration and power system dispatch. In engineering deployment, the scarcity of samples is often mitigated by methods such as transfer learning or generative adversarial networks. However, such methods often rely on cross-site data or model exchange, and are constrained by data privacy and compliance, making it difficult to centralize the original data. Therefore, distributed collaborative paradigms such as federated learning are regarded as more feasible alternatives. However, significant data heterogeneity among different sites can cause distribution drift and negative transfer during the transfer or federated learning process, thereby weakening the practical utility of cross-site knowledge transfer and aggregation. To address these issues, this paper proposes a privacy-preserving personalized federated learning framework: a Temporal Convolutional Network (TCN) is employed as the shared encoder for federated aggregation, while a Bidirectional Gated Recurrent Unit (BiGRU) decoder is retained locally for personalized training, thereby achieving the separation of global representation learning and local adaptation. To balance domain adaptation and privacy protection, clients calculate Gaussian embedding signatures based on encoder outputs and report to the server after applying differential privacy perturbations; the server computes the symmetric KL divergence based on Gaussian embedding signatures to dynamically filter similar clients and then progressively aligns the features of selected clients using Conditional Adversarial Domain Adaptation (CDAN) and Maximum Mean Discrepancy (MMD), generating global representation guidance for encoder optimization to mitigate data distribution shifts. Comparative experiments conducted on multiple wind farm datasets in Australia show that the proposed method significantly improves prediction accuracy and cross-site generalization performance while preserving privacy, demonstrating clear advantages over other models discussed in this study.