Data-locality-preserving carbon-aware coordination of power and electrified transportation systems using federated reinforcement learning

Abstract

The rapid electrification of transportation systems has intensified the coupling between regional power grids, electric vehicle charging demand, and spatiotemporal traffic dynamics, posing significant coordination challenges under carbon constraints. Conventional centralized optimization and learning-based control approaches often face limitations in scalability, data exposure, and institutional feasibility when applied across geographically distributed regions. Moreover, many existing methods treat traffic dynamics and carbon emissions as exogenous factors, limiting their ability to capture the endogenous environmental impacts of integrated energy–mobility systems. We develop a data-locality-preserving federated reinforcement learning framework for carbon-aware coordination of coupled power and electrified transportation systems across multiple regions. Each region is modeled as an autonomous learning agent that jointly manages energy dispatch, electric vehicle charging behavior, and traffic-related states using a local Proximal Policy Optimization–based actor–critic architecture. The framework reduces raw-data exposure by keeping regional traffic, charging, renewable, and grid observations local and by exchanging encrypted model-update information rather than raw operational records. This design should be interpreted as a data-locality and raw-data non-sharing mechanism, not as a formal differential-privacy or leakage-free guarantee. Explicit physical models of energy balance, grid exchange constraints, vehicle flow conservation, traffic density evolution, charging queue dynamics, and carbon emission formation are embedded into the learning environment, ensuring operational feasibility and interpretability. Carbon emissions emerge endogenously from the joint evolution of power system operation and mobility dynamics rather than being imposed as external penalties. Comprehensive simulation studies involving multiple interconnected regions with heterogeneous traffic intensities, renewable penetration levels, and charging infrastructures show that the proposed federated framework reduces system-wide carbon emissions by approximately 22–28% compared with decentralized local learning baselines, while simultaneously lowering inter-regional policy variance by over 60%. In addition, the federated approach achieves faster convergence, reaching stable performance in around 40% fewer training episodes, and exhibits improved robustness under stochastic demand and renewable variability. The primary contribution of this work is the development of a unified, physically grounded federated learning architecture that enables carbon-aware coordination of energy and transportation systems under realistic operational constraints while avoiding centralized pooling of sensitive regional data.