Enabling Continuous Operation of Shared Autonomous Vehicles With Dynamic Wireless Charging

Shared Autonomous Vehicles (SAVs), serving as an alternative to private cars, are emerging as a pivotal solution to severe traffic problems. Leveraging Dynamic Wireless Charging (DWC) technology, SAVs can potentially operate continuously meet passenger demands, presenting a complex challenge known as the SAV routing and dynamic wireless charging problem (SVRCP). However, traditional heuristic algorithms, which typically reliant on extensive experiential rule settings, fall short in addressing the multiple objectives and complex constraints of SVRCP. Therefore, we propose a Deep Slack Induction by String Removals-based Reinforcement Learning (DSRL) framework, specifically designed to optimise cost-efficiency and energy stabilisation in SAVs operating with DWC conditions. First, a traffic-vehicle cooperation model, grounded in real-world road networks, is constructed to facilitate interactions for the DSRL’s encoder. We then integrate the Slack Induction by String Removals algorithm to adaptively optimise the DSRL parameters, enhancing the decoder’s ability to find reliable solutions. Experiments show that DSRL outperforms heuristic methods in optimising route cost and stabilising the State-of-Charge (SOC) of SAVs, while achieving a more centralised
$\Delta$
SOC distribution, ensuring the continuous operational capability of the SAVs. The results across instance sizes and DWC power demonstrate DSRL’s robust generalisation capabilities and efficient training capabilities facilitated by transfer learning. Additionally, sensitivity analysis from the perspectives of depot location, passenger distribution, and power levels offers insights for DWC road layout.