Prediction-informed distributional deep reinforcement learning for energy management of fuel cell tracked vehicles: A hybrid control framework

Energy management strategies (EMSs) play a pivotal role in promoting energy conservation of vehicles with hybrid energy storage systems, particularly for fuel cell tracked vehicles (FCTVs) operating under complex driving scenarios. While optimization-based and learning-based EMSs have gained significant research traction, fundamental challenges remain in achieving deeper methodological integration for enhanced control performance. To bridge this gap, this article develops a novel hybrid control framework that synergistically combines prediction-informed distributional deep reinforcement learning (DRL) with the equivalent consumption minimization strategy (ECMS) for FCTVs, delivering an intelligent and integrated energy management solution. Firstly, this integrated framework employs DRL for equivalent factor (EF) selection, while ECMS delivers optimization-grounded power allocation, collectively advancing fuel economy. Further elevating this approach, we employ Truncated Quantile Critics (TQC), an advanced distributional DRL algorithm engineered to refine EF precision and energy utilization efficiency. Moreover, a variational mode decomposition (VMD)-Transformer-based power demand prediction module is seamlessly integrated to empower prospective decision-making within the DRL paradigm. Experimental validation demonstrates that this predictor achieves 27.45%–36.07% higher power demand prediction accuracy than conventional baselines. The proposed framework reduces hydrogen consumption by 6.15% compared with the soft actor-critic-based EMS benchmark under the testing cycle. These results demonstrate the effectiveness of integrating prediction-informed DRL with ECMS for energy management of FCTVs.