Privacy-Preserving, Scalable EV Charging Optimization

The growing penetration of electric vehicles
(EVs) leads to non-uniform and time-dependent charging
demand at the distribution level, which complicates feeder
loading and voltage regulation. To manage this variability, a federated learning–based coordination architecture is
developed in which EV charging stations deployed at different network locations train LSTM-based load forecasting models and reinforcement learning (RL) charging controllers using locally measured power and usage data. Model training is performed without transferring individual charging records, relying instead on aggregated parameter updates. Secure aggregation and differential privacy mechanisms are incorporated to bound information leakage, while decentralized policy optimization is adopted to avoid centralized control
dependence. Communication and update synchronization are
designed with explicit bandwidth and latency constraints to
ensure feasibility under operational charging-network
conditions. The proposed architecture is evaluated using
MATLAB simulations that represent stochastic EV arrival
behavior, bounded charging time flexibility, and time-varying
utility tariff structures. Simulation outcomes show lower
forecasting error, reduced training effort for RL policy
convergence, observable feeder-level peak load reduction, and
improved privacy protection compared with centralized and
non-federated learning approaches.