Imitation learning-based energy management for connected hybrid electric vehicles under computational constraints

Traditional learning-based energy management strategies (EMSs) often suffer from low training efficiency and implicitly assume that large neural networks can be deployed on onboard controllers to pursue optimal performance, overlooking the challenges of real-world deployment. To address these issues, this paper first proposes an imitation learning-based EMS using Generative Adversarial Imitation Learning (GAIL) to efficiently approximate expert performance using demonstration data. Then, based on this GAIL-based EMS, a V2X-assisted strategy selection mechanism is proposed to maintain high fuel economy while reducing the network size and computational costs. Specifically, real-world traffic data are clustered into several categories with reduced complexity, allowing dedicated GAIL-based EMSs to be trained for each category. When the hybrid electric bus stops at a bus station, upcoming traffic conditions are obtained via V2X, and the corresponding EMS is selected. Simulation results show that, compared to traditional learning-based strategies, the proposed GAIL-based EMS improves training efficiency by 66.13% and enhances fuel economy by 2.02%. Furthermore, under two real-world driving conditions, the V2X-based adaptive EMS further improves fuel economy by 2.38% and 2.44%, respectively, compared to a static GAIL-based EMS with the same network structure.