Towards Trajectory-Aware Federated Learning for Electric Vehicle Range Prediction

Abstract

Electric-vehicle (EV) range prediction is a route-conditioned and uncertainty-sensitive task: a useful predictor must estimate the en- ergy required by a planned or likely future trajectory, not only extrapo- late recent consumption. The telemetry and trajectories needed to learn such models are distributed across vehicles and fleets, privacy-sensitive, and costly to centralize. This position paper outlines a trajectory-aware federated learning (FL) agenda for EV range prediction. The proposal keeps raw traces local, learns shared route-energy representations, pre- dicts route energy and destination-arrival risk, and combines probabilis- tic prediction with personalization for driver, vehicle, and geographic heterogeneity. Prior work has shown that probabilistic FL can support energy-demand prediction on planned routes. We extend that direction by treating trajectory representation as a first-class modelling object, by evaluating realistic non-IID and intermittent-connectivity settings, and by making calibration, communication cost, tail-client behaviour, and privacy-relevant safeguards primary metrics. The planned methodology combines real-world EV traces with synthetic SUMO-generated data and compares local, centralized, and federated baselines through temporal and geographic holdouts.