Deep Reinforcement Learning-Based Cooperation Scheduling Strategy for Single-Route Battery Electric Bus Fleets Considering Battery Capacity Decay and Nonlinear Charging

Abstract

The scheduling problem of battery electric buses (BEBs) has been widely studied. However, most existing work focuses only on either operation or charging. In reality, these two aspects are interdependent and should be considered from a lifecycle perspective. This study proposes a cooperative scheduling strategy for BEB fleets on a single bus route. The strategy addresses opportunity charging under a limited charging infrastructure. It dynamically generates operation and charging decisions based on real-time system states. The goal is to reduce the impact of uneven terminal arrival times and mitigate battery aging. Both the nonlinear charging behavior and battery aging are considered in the charging plan. The BEB work cycle is modeled as two interrelated Markov Decision Processes (MDPs): one for operation and one for charging. A deep reinforcement learning framework, called the Double-SAC-Combination Framework, is used to solve the model. It integrates the Soft Actor-Critic (SAC) and Discrete SAC algorithms. The model is validated using real-world data from Beijing Bus Line 44 (inner loop). Results show that the proposed strategy reduces daily battery capacity degradation by 8.5% to 33.3% compared to a no-control baseline. The best performance is achieved when the headway is 6 minutes, speed is 25 km/h, fleet size is 35, and there are 4 charging piles. Sensitivity analyses are conducted on BEB load differences, driving speed, headway, number of charging piles, fleet size, route length, and battery capacity.