Benchmarking Adaptive EV Charging Recommendation Under Feasibility Constraints and Non-Stationary Infrastructure

The rapid adoption of electric vehicles (EVs) has intensified the need for charging station recommendation systems that can adapt continuously to changing infrastructure conditions. In practice, charging environments are highly dynamic: station availability fluctuates; prices vary over time, and charging capabilities differ widely across locations. However, many existing approaches assume static infrastructure attributes or optimize a single objective in isolation, limiting their effectiveness in real-world deployments where robust adaptation under uncertainty is essential. In this work, we study EV charging recommendations as a constraint-aware online decision problem and present a unified benchmarking framework for evaluating learning-based recom mendation policies under both static and dynamic regimes. Using a large-scale global charg ing infrastructure dataset (1.5M stations) and a real-world time-series availability and pricing dataset, we compare contextual bandit and reinforcement learning approaches across multiple dimensions, including relevance, feasibility, regret, adaptability, and recovery from exogenous shocks. Our results show that learning-based methods significantly reduce constraint violations and cumulative regret compared to heuristic baselines in static settings, while neural contextual bandits achieve the strongest feasibility–utility trade-off. Under dynamic conditions, we ob serve clear differences in adaptation speed and shock recovery across algorithms, indicating that effective EV charging recommendation requires not only feasibility-aware reward design, but also robustness to non-stationary pricing and availability and the preservation of a meaningful learning signal during disruptive events