Large-scale electric bus network transition planning via deep reinforcement learning

Abstract

Urban bus electrification is gaining global interest, playing a crucial role in reducing emissions. This study defines and addresses the electric bus network transition problem (EBNTP), jointly optimizing battery electric bus (BEB) fleet transitions and charging facility planning over a multi-period horizon. Existing research often neglects this interdependent long-term planning and lacks scalable solutions for large systems. This study proposes a deep reinforcement learning (DRL) approach, formulating EBNTP as a Markov Decision Process modeling sequential planning decisions, and introduces the DRL-HetGNN method, integrating heterogeneous graph neural networks (HetGNN) to capture network effects and enhance efficiency in large-scale applications. Using Hong Kong's franchised bus system as a case study, DRL-HetGNN demonstrates superior performance and generalizability compared to benchmark methods. Scenario analyses explore budget allocations, independent operators, BEB subsidies, and price fluctuations, while examining policy-incentive mechanisms to accelerate electrification. The findings will support policymakers in planning sustainable public transportation systems.