H-FLTN: A Privacy-Preserving Hierarchical Framework for Electric Vehicle Spatio-Temporal Charge Prediction

Accurately forecasting electric vehicle (EV) charging demand is critical for energy providers seeking to ensure grid reliability and efficient resource allocation; however, doing so often requires access to highly privacy-sensitive mobility data. This paper introduces the Hierarchical Federated Learning Transformer Network (H-FLTN), a privacy-preserving framework for modelling EV charging events across both location (where) and time (when). H-FLTN adopts a three-tier hierarchical architecture comprising EV clients, community-level Distributed Energy Resource Management Systems (DERMS), and a central Energy Provider Data Centre (EPDC). Within this architecture, transformer-based learning captures long-range spatio-temporal dependencies in EV charging behaviour, while federated learning (FL) enables collaborative model training without exposing raw mobility data beyond individual EVs. Privacy is enforced through additive secret sharing, peer-to-peer (P2P) share augmentation, and secure aggregation. To support scalability, H-FLTN integrates a Dynamic Client Capping Mechanism (DCCM) and Client Rotation Management (CRM), limiting computational cost while mitigating participation bias. Extensive simulation results using large-scale mobility data demonstrate strong location prediction accuracy and stable temporal prediction behaviour under non-independent and non-identically distributed (non-IID) client distributions, while maintaining efficient and scalable training performance. The proposed framework provides a deployable foundation for privacy-preserving EV charging analytics in future smart grid environments.