Large-scale EV charging coordination: a detailed exploration of mean-field and reinforcement learning approaches

The rapid proliferation of electric vehicles (EVs) presents significant challenges to power-grid stability, especially when thousands of cars charge simultaneously without coordination. This paper investigates two complementary families of scalable control schemes-(i) Mean-Field Control (Mean-Field Games and Mean-Field-Type Games), and (ii) model-free Reinforcement Learning (classical Reinforcement Learning and Deep Reinforcement Learning)-that capture stochastic arrivals, nodal capacity limits, ambient-temperature effects and battery degradation. Analytical mean-field-based formulations yield decentralized charging policies that depend only on population statistics and enjoy ϵ-Nash optimality as fleet size grows, while reinforcement-learning-based agents learn directly from interaction histories and cope naturally with partial observability and non-stationary price signals. A common set of nine key performance indicators is applied to two benchmark scenarios: a one-night, three-node grid and a seven-day, heterogeneous 1 000-EV testbed built on real distribution-network data. Results show that Mean-Field-Type Games minimizes unmet energy ( ≥ 1 %) and queueing delays, Deep Reinforcement Learning maximizes average final State-of-Charge ( ≈ 81 %) under volatile tariffs, and classical Reinforcement Learning provides the most interpretable albeit least efficient baseline. These quantified trade-offs clarify when model-based equilibrium methods suffice and when adaptive, data-driven controllers become indispensable, providing actionable guidance for large-scale, battery-health-aware EV-charging deployments.