Strategic planning of charging infrastructure for shared electric vehicles: A multi-phase stochastic approach using reinforcement learning

The rapid electrification of ride-hailing mobility requires that fast-charging infrastructure be ready to support the dynamic charging needs of shared electric vehicle (SEV) fleets. Deploying city-wide fast-charging stations requires determining the timing, locations, and number of chargers at each station over multiple planning phases, while managing stochastic charging demand, induced demand from enhanced services, and constraints like grid capacity. In contrast to conventional deterministic and myopic methods that optimise charging stations’ configurations based on static demand and immediate benefits, this study formulates the problem as a stochastic sequential decision problem and applies reinforcement learning (RL) to approximate the optimal deployment plan considering long-term benefits and uncertainties. An agent-based model representing an AI gym environment is developed to train the RL model, simulating SEV interactions with charging stations to capture stochastic charging demand across space and time. RL maximises the expected charging service efficiency with the minimal number of fast chargers over the planning horizon. The model reduces the total number of chargers by 30.5% compared to the myopic benchmark, while achieving a 10.9% higher charging service efficiency. This multi-phase fast charger planning approach offers engineers and planners a novel tool for efficient long-term infrastructure allocation and policy development under stochastic charging demand. • Novel framework formulates charging station planning as a stochastic decision problem. • First application of reinforcement learning (RL) in multi-phase infrastructure planning. • Integrates fleet size decisions & grid constraints in charging infrastructure planning. • Agent-based model simulates stochastic charging demand & interacts with RL. • Model reduces required fast chargers by 30.5% & improves service efficiency by 10.9%. • Model is robust to demand fluctuations & is scalable across diverse urban contexts.