Two-Stage Privacy-Preserving Collaborative Forecasting for Wind Generation with Defense Against Malicious Eavesdroppers and Saboteurs

Wind power generation point forecasting involves trade-offs among accuracy, privacy, attack resilience, and training efficiency. Individual forecasting (IF) protects data privacy but often suffers from limited accuracy, whereas collaborative forecasting (CF) exploits cross-farm correlations but may introduce privacy risks, malicious attacks, and high training costs. To address these challenges, this paper proposes an efficient two-stage privacy-preserving CF framework. In Stage 1, each wind farm trains an IF model independently and in parallel. In Stage 2, all farms collaboratively train a collaborative fine-tuning (CFT) model to learn IF residuals, improving forecasting accuracy while reducing the collaborative training burden. Considering two types of malicious attackers in information interaction, namely eavesdroppers and saboteurs, targeted defense mechanisms are further designed. For eavesdroppers, secure multi-party computation is embedded into the CFT process, enabling both forward and backward propagation to be performed over secret shares and preventing privacy leakage. For saboteurs, dropout-embedded self-attention and quantile regression are incorporated to mitigate the impact of tampered collaborative information and constrain abnormal residual corrections. Case studies on real-world data from 10 wind farms in Jiangsu Province, China, and the public GEFCom2014 dataset show that the proposed method achieves higher accuracy and lower collaborative training cost than federated-learning-based CF. Sensitivity analyses further reveal the effects of key defense-related parameters, providing practical guidance for parameter tuning under different security requirements.