Reducing Residential Grid Congestion from Electric Vehicles Using Privacy-Preserving Federated Smart Charging

electric vehicles, federated learning, smart charging, differential privacy, secure aggregation, residential feeder.

Authors

September 25, 2026

Downloads

Residential electric vehicle charging can create short feeder peaks during already crowded evening hours. Utilities need coordinated charging, yet household charging records reveal daily routines and travel habits. This paper develops a privacy-preserving federated smart charging method for residential feeders. Each client trains an energy predictor without sending raw sessions to a server. Clipped updates receive distributed Gaussian noise before pairwise-mask secure aggregation. A feeder-aware scheduler then places predicted energy within each vehicle connection window. The experiment uses 6,163 measured charging sessions from 74 apartment users. It also uses measured household imports from Open Power System Data. The corrected workflow converts cumulative meter readings into hourly demand increments. Testing covers 958 unseen sessions on a modeled feeder serving 100 homes. The proposed private model uses a noise multiplier of 1.0. Its privacy budget is 14.91 at a delta value of 0.00001. Private smart charging lowers the combined peak from 160.07 kW to 100.55 kW. This reduction equals 37.18 percent and removes ten modeled overload hours. However, the schedule leaves 10.57 percent of requested energy unmet. Its mean charging delay reaches 8.37 hours. The private predictor records 12.80 kWh test RMSE. The centralized model performs better, with 8.60 kWh RMSE. Results show that privacy noise adds little error beyond ordinary federated training. Client data differences remain the main accuracy barrier. The method protects data location and bounds shared updates. It exposes the service cost behind congestion relief. These findings support stronger local models and service constraints before deployment.