Optimization of Mobility Incentives in Electric Vehicle Car Sharing Systems: A Reinforcement Learning Framework

Abstract

This work introduces a novel reinforcement learning (RL) framework for optimizing mobility incentives in free-floating electric vehicle carsharing systems. User participation in relocation and rebalancing activities is facilitated by a centralized control agent that assigns dynamic and space-heterogeneous tariff discounts and mobility credits. These mobility incentives can adapt to various operational conditions and are learned through direct interaction with a new simulation environment. This environment integrates a probabilistic demand model, historical reservation data, and a transportation mode choice model, and can simulate reservations, relocation, and charging activities affected by the incentives. The efficacy of the RL framework is demonstrated on a simulated fleet in Tel Aviv, using real-world data from the AutoTel company. Four RL agents are trained via standard PPO, TD3, DDPG, and SAC algorithms designed to handle continuous action spaces. Efficacy is evaluated against baseline relocation strategies (crew-based and hybrid relocation) and the results demonstrate higher revenues and lower charging and relocation costs achieved by the RL agents. To test the scalability of the approach, two observation spaces of increasing dimensionality are investigated. Overall, this study highlights the potential of RL-based mobility incentivization strategies, offering benefits for transportation and energy grid operators. A powerful tool that can support sustainability and electrified carsharing operations.

Keywords: Reinforcement Learning, FFEVCS, Mobility Incentives, Relocation, Charging Stations, Acceptable Walking Distance