Advancing Privacy-Preserving Wind Generation Forecasts With Selective Spatial-Temporal Dependencies Extraction, Encryption and Sharing

The adoption of renewable energy sources (RES) is essential for accelerating the low-carbon transition. Accurate forecasting of renewable energy generation is vital for its effective integration into power systems, allowing for mitigation on variability and intermittency of RES and facilitating economic and secure grid operation. In this context, collaborative wind generation forecasting task faces multiple challenges, including secure information sharing, spatial-temporal dependencies characterization and utilization, as well as uncertainty quantification via probabilistic forecasting. To this end, this paper proposes a novel collaborative forecasting framework which features 4 core functionalities: a temporal convolutional network-autoencoder based secure data encryption and temporal dependencies extraction module; a self-attention based secure data sharing and selective spatial-temporal dependencies extraction module; a privacy-preserving collaborative point/probabilistic forecasting module based on improved pinball loss function; and an adaptive fine-tuning procedure to further leverage feature extraction performance and prediction accuracy through joint training. Case studies using real-world data from 10 wind farms located across Jiangsu province, China, validated superior performance of the proposed method while highlighting the value of the developed functionalities. The proposed collaborative forecasting method achieves 18.21%, 15.23%, 8.69% improvement in RMSE compared to the worst-performing baseline method, in benchmarking against individual forecasting, pre-training only and statistical pre-screening based methods. Furthermore, a maximum 26.62% improvement in Coverage Width-based Criterion is attained with respect to parametric and non-parametric baseline methods.