Shadow-Price DRL: A Framework for Online Scheduling of Shared Autonomous EVs Fleets

This paper studies the online scheduling of shared autonomous electric vehicle (SAEV) fleets. The study includes charging management, routing and rebalancing strategies for SAEVs to serve the trip demands in the coupled power and transportation network (PTN). It aims to minimize the total social cost of PTN. The difficulty lies in how to deal with time-varying trip demands with the time-coupled SAEV scheduling and PTN operation considered. To address this challenge, for the first time, we propose a novel framework named the shadow-price deep reinforcement learning (shadow-price DRL), which combines the rigorous PTN operation model and the data-driven model-free DRL-based algorithm. Within the shadow-price DRL, the policy neural network adaptively learns the system dynamics and imposes its actions on the online SAEV scheduling problem as the dynamic shadow prices. By doing so, the SAEV schedule will be determined in the online manner. In addition, the connection between the proposed framework and Lagrangian Relaxation method is discussed, which illustrates the principles and effectiveness of the proposed method. The case studies include the practical Xi'an city which verifies the effectiveness of the shadow-price DRL and illustrates its significant superiority over the safe DRL and model predictive control (MPC) based methods.

Automated summary

This paper focuses on the online scheduling problem of shared autonomous electric vehicle (SAEV) fleets, including charging management, routing, and rebalancing strategies. A novel framework called shadow-price deep reinforcement learning (shadow-price DRL) is proposed to address the challenge of time-varying trip demands. The framework combines a rigorous PTN operation model and a data-driven model-free DRL-based algorithm. Case studies validate the effectiveness of the proposed method.