Roughness-Informed Federated Learning for Energy Consumption Modeling of Battery Electric Vehicles

Battery electric vehicles (BEVs) are central to reducing urban air pollution and require accurate, real-time energy-consumption prediction to optimize itineraries and vehicle systems. Growing privacy concerns motivate on-device learning. We introduce Roughness-Index-based Federated Averaging (RI-FedAvg) in a centralized, cross-device FL setup for BEV energy prediction under non-IID data. RI-FedAvg adapts client regularization using a per-vehicle roughness index to mitigate client drift. In addition, we introduce RI-DP-FedAvg, a differentially private extension of RI-FedAvg that combines roughness-informed regularization with formal privacy guarantees using R´enyi Differential Privacy (RDP)-calibrated Gaussian noise. We compare against FedBEV, FedAvg, FedDyn, FedProx, SCAFFOLD, and a DP-enabled baseline (DP-FedAvg) using MAE. Using real driving logs from VED and simulated heterogeneity (vehicle classes, traffic/weather), RI-FedAvg improves MAE versus baselines and converges in fewer rounds, while RI-DP-FedAvg maintains competitive accuracy under strict privacy levels. We discuss an edge-cloud deployment workflow. These results indicate that RI-FedAvg and RI-DP-FedAvg improve accuracy while keeping data on vehicles in centralized FL architectures.