预订
计算机科学
电池(电)
调度(生产过程)
计算机网络
工程类
生产(经济)
实时计算
汽车工程
需求响应
作者
Bo Zhang,Yanchao Zhang,Tao Wang,Zhihua Chen,Guoliang Shi,Rui Miao,Zhibin Jiang
标识
DOI:10.1080/00207543.2026.2709056
摘要
Battery swapping stations (BSSs) are an important component of electric vehicle (EV) energy replenishment. Their operation is subject to stochastic variability and abrupt demand regime shifts, which appear as persistent changes in swapping demand relative to available charging capacity. Under such non-stationarity, a single policy trained for one regime can become persistently mismatched after a transition, leading to sustained vehicle queuing or repeated unused charging-bay reservations. This paper studies sequential charging-bay reservation for a single BSS and proposes Resilient Hierarchical Proximal Policy Optimization (RHPPO) as a multi-level deep reinforcement learning decision framework. RHPPO maintains a small policy library of regime-specialised PPO reservation policies. A meta-learning procedure provides a transferable meta-policy initialisation, and each regime policy is obtained with limited fine-tuning to reduce training cost. During operation, a lightweight neural selection model monitors recent operational state sequences to decide whether to switch policies and which policy to activate, enabling timely adaptation under regime transitions. Simulation results show that RHPPO outperforms rule-based, static, and single-policy DRL baselines in cumulative operational cost by reducing charging cost, alleviating EV queuing, and lowering unused reserved capacity. Additional experiments under varying EV fleet sizes indicate stable performance and improved scalability in larger single-BSS settings.
科研通智能强力驱动
Strongly Powered by AbleSci AI