Real-time Scheduling of Electric Bus Flash Charging at Intermediate Stops: A Deep Reinforcement Learning Approach

The flash charging of electric buses (EBs) refers to the charging of EBs with pantograph chargers at intermediate stops. By "charging less but more often", flash charging enables EBs to use small batteries, thus improving fuel economy while meeting mileage requirements. However, in real-time operation, flash charging can be susceptible to uncertainties such as passenger demand and electrical load – the former determines how long EB dwells at stops, beyond which charging would delay the transit service, while the latter together with charging loads could put distribution networks at risk. To address the above uncertainties, this paper proposes a deep reinforcement learning (DRL) approach for the real-time scheduling of EB flash charging in terms of location, timing, and duration. Numerical results show that: 1) the proposed DRL approach can find efficient and reliable scheduling policies that outperform benchmarks such as the real-world "uniform" policy by making a better use of EBs' layover at stops based on real-time information; 2) our approach remains effective when applied to flash charging systems with renewable energy resource integration or different scales; 3) pantograph chargers should have sufficiently high power rating to support an efficient transit service whilst without risking the distribution network, and an "adequate" charger setup can be designated for improved utilisation based on our approach.