Federated Learning with Mixed Encryption for Distributed Energy Resources: Forecasting Electricity Charging Demand at Multiple Electrical Vehicle Charging Stations

Sharing information or data between actors is crucial for forecasting distributed energy resources. However, collaboration involving the exchange of confidential data requires privacy-preserving models. Current models rely solely on fully encrypted frameworks, neglecting publicly available data or enforcing encryption on both private and public data. Furthermore, it is unknown whether a combination of public and private data types is beneficial to federated learning on spatiotemporal energy applications, in particular forecasting. In this work, we introduce a mixed encryption setting incorporating public data (e.g. weather, calendar) and private data relative to energy measurements. In this case study of day-ahead electric consumption of neighboring Electric Vehicle charging stations, we show that combining data increases the accuracy by up to 6% in RMSE and that mixed encryption reduces the computation time by up to 2 times compared to full encryption.