A privacy-preserving federated learning framework for load forecasting in electrified intelligent transportation systems

Abstract

The electrification transition of intelligent transportation systems (ITSs) is coupling mobility operations with distribution- grid scheduling and producing highly heterogeneous charging loads whose forecasting must not expose sensitive vehicle, passenger, or operator data. This paper presents a privacy-preserving federated learning (FL) framework for multi-region load forecasting in electrified ITSs, including urban fast-charging corridors, electric-bus depots, logistics fleet parks, commercial mobility hubs, and distributed-energy-assisted suburban charging communities. Each regional gateway trains a temporal encoder locally, keeps raw electric vehicle (EV) charging records and personalized prediction heads within the regional privacy boundary, and releases only protected encoder updates. Differential privacy (DP) clipping and Gaussian noise, secure aggregation, and 8-bit update quantization are embedded jointly in the learning loop. A heterogeneous transportation-energy benchmark is constructed from public EV-charging traces and calibrated regional demand profiles. Compared with Federated Averaging (FedAvg) and Federated Proximal (FedProx), the proposed method reduces mean absolute percentage error (MAPE) from 4.88% and 4.72% to 4.18%, respectively, while lowering the communication payload from 11.2 MB to 4.9 MB per protected round. The results show that privacy-preserving federated forecasting can support day-ahead charging coordination, depot energy management, and edge-cloud automation during the electrification of intelligent transportation systems.