Privacy-Preserving Graph Inference Network for Multi-Entity Wind Power Forecast: A Federated Learning Approach

Abstract

Data sharing is considered by many wind farm stakeholders as the cause of privacy issues and further financial risks, despite its potential to enhance the accuracy of multi-entity wind power forecasting (MWPF). Federated learning (FL) serves as a possible solution to preserve the privacy in MWPF, while the existing FL-based methods still struggle to obtain accurate prediction due to the intricate spatial dependencies and heterogeneous temporal dependencies. In response to these two challenges, this paper proposes a collaborative privacy-preserving framework (CPLF) for MWPF. Within the CPLF, a graph learning-based local model named graph inference network (GIN) is developed to learn local features and obtain the global ones through aggregation. In terms of the spatial dependencies, a structure-independent dynamic graph inference (SiDGI) block is designed to extract spatial features via learnable directed graph representation. Regarding the heterogeneous temporal dependencies, the GIN, with its encoder-decoder to distill general temporal pattern, is trained following a customized FL procedure to effectively extract entity-specific temporal features. This customization can mitigate the communication burden and reverse-engineer risks while yielding improvements in MWPF accuracy. Finally, the extensive experiments are implemented based on two datasets collected from the Northwest and Southeast of California. The superiority of the proposed privacy-preserving MWPF method is verified compared with some classical methods. Specially, for graph attention, MWPF achieves 6.8% and 14.9% average improvements in mean absolute percentage error (MAPE).