MatchEstimate: A Robust Aggregation Method for Federated Learning, Electric Vehicles Case Study

The rise of Connected Electric Vehicles (CEVs) is driving demand for intelligent, data-driven solutions in mobility and energy management. However, using data from each CEV independently fails to generalize to unseen scenarios, while sharing data with a centralized system raises privacy concerns. Federated Learning (FL) enables collaborative model training across distributed EV systems while preserving data privacy. Nevertheless, standard FL algorithms face challenges due to client heterogeneity, including non-IID (non-Independent and non-Identically Distributed) data, varying computational capabilities, and unstable communication links. We propose Match-Estimate (ME), a robust, flexible aggregation method designed for heterogeneous FL. As a drop-in replacement for strategies like Federated Averaging (FedAvg), ME extends our prior method, FedEstimate, to improve performance and training stability across diverse clients. MatchEstimate is a two-step server-side aggregator: (i) it aligns neurons across client models via layer-wise bipartite matching, and (ii) it trains a small regressor to map proxy shard-level updates to each client’s full update, which are then size-weighted and aggregated. This plug-in design mitigates non-IID drift and parameter misalignment. In a CEVs case study, ME significantly enhances their accuracy under realistic conditions, advancing privacy-preserving, decentralized intelligence for next-gen mobility.