A Spatiotemporal Federated Learning Based Distributed Photovoltaic Ultra-short-term Power Forecasting Method

Abstract

Accurate distributed photovoltaic power prediction plays a crucial role in supporting power retailers in making optimal trading strategies and distribution network operators in making reasonable scheduling plans. The mining and utilization of distributed photovoltaic power correlation is an effective way to improve the forecasting accuracy of the target station. However, the current photovoltaic power forecasting methods must rely on data sharing in correlation mining, which may lead to serious data privacy problems. In order to protect the privacy of data during power modeling while ensuring the effective utilization of spatiotemporal correlations, the spatiotemporal federated learning based distributed photovoltaic ultra-short-term power forecasting method is proposed in this paper. In this method, the local server takes the Gated Recurrent Unit-based autoencoder as local model to extract local temporal features and forecast the local power. The central server uses Federated Averaging to aggregate and update the parameters of each local forecasting model and uses t-distributed Stochastic Neighbor Embedding to process the temporal features into the global feature which implies the spatiotemporal correlations between stations. Subsequently, the local model combines local temporal features and global spatial features to achieve power prediction. Through the doublelayer sharing of local model parameters and local temporal features, spatiotemporal correlations are effectively utilized while ensuring the privacy of data and the forecasting accuracy of each distributed photovoltaic is improved. Finally, the case study via real-world data is conducted to verify the effectiveness of the proposed method.