Spatiotemporal Federated Learning Based Regional Distributed PV Ultra-Short-Term Power Forecasting Method

Accurate distributed photovoltaic power forecasting is crucial for both electricity retailers and distribution network operators. Mining the rich correlations within distributed photovoltaic data has immense potential to boost forecasting accuracy. However, existing correlation modeling approaches often demand centralized aggregation of raw power data, raising privacy concerns. To address this issue, this research proposes a novel spatiotemporal federated learning-based regional distributed photovoltaic ultra-short-term power forecasting method. First, the power forecasting model is trained with federated learning to achieve correlation information sharing by the model interaction. Then, considering that the information shared by model interaction is very limited, a spatiotemporal correlation modeling method based on temporal feature sharing is proposed. Based on this dual information sharing mechanism, effective mining of spatiotemporal correlation information is realized and the accuracy of power forecasting can be enhanced. Under this framework, the central server only generates global models and features without aggregating raw power data, and local users only need to share local model and temporal feature information. Therefore, user data privacy can be protected. Finally, the effect of the proposed method is verified via a China's dataset.