A Reinforcement Learning-based framework for optimizing mobile charging pod operations

Abstract

The rise of autonomous electric vehicles (AEVs) presents new challenges and opportunities for an efficient and flexible charging infrastructure. This study proposes a reinforcement learning (RL) based framework for optimizing the control and operation of mobile autonomous charging pods (MAPs) for maintaining the operation of AEVs through dynamic charging. We formulate a time and energy aware Markov Decision Process (MDP) to maximize the energy delivered, and the number of AEVs serviced, while also minimizing energy consumed and increasing efficiency. We integrate this framework with SUMO to enable realistic MAP-AEV interactions. A Proximal Policy Optimization (PPO) algorithm was used to train this MDP and identify the optimal control strategies for initiating, terminating, and balancing the network. The results show that the PPO agent can service around 175 AEVs, with an efficiency of 91.5%, representing a 25% improvement over baseline greedy heuristics. Moreover, the battery capacities of AEVs can also be reduced by up to 26%, without compromising the performance. The simulation results show the potential of the proposed method in providing a flexible, and scalable charging for future transport.