Harvesting the Flexibility of Autonomous Mobility-on-Demand System to Enhance Distribution Network Operation

Considering the potential application of Autonomous Mobility-on-Demand (AMoD) systems in future urban transportation, the strategic charging behavior of AMoD can provide flexibility to promote distribution network efficient operation. In this paper, we develop an incentive-based navigation framework that coordinates fleet routing and charging schedules to harvest the flexibility potential of AMoD fleet. Firstly, we formulate the incentive problem for network operator and the scheduling problem for fleet as a Markov game. We address these problems using a novel multi-agent deep reinforcement learning algorithm, which incorporates a physics-informed neural network model to represent the distribution network and resolve unobservable issues. We then employ an asynchronous update and modified prioritized experience replay mechanism to mitigate non-stationary nature of the environment. Additionally, we propose an expert-based pre-training approach to accelerate the training process. Experimental results demonstrate that the proposed method increases AMoD system profit by 7.83%, reduces DSO operational costs by 29.4%, and eliminates voltage violations, thereby achieving both economic and secure operation of the integrated transportation and power system.