Robust optimization-based energy management for dual-APUs heavy-duty hybrid electric vehicles using intention-aware prediction and curiosity-driven control

Given the power limitations of pure electric systems in heavy-duty vehicles, developing heavy-duty hybrid electric vehicles has become one of the optimal solutions for improving energy efficiency and reducing emissions. Among them, dual auxiliary power unit configurations have gained prominence due to their enhanced power and flexibility. However, this system demands higher robustness and precise control of the two diesel range extenders and battery pack to ensure stability and efficient energy distribution. Severe environmental disturbances—such as the uncertainty of unknown driving cycles and non-negligible sensor noise can dramatically degrade control accuracy and energy distribution stability. To address these challenges, we propose an improved Twin Delayed Deep Deterministic Policy Gradient (TD3)-based energy management system (EMS) coupling a driving-intention-aware Transformer (DI-Transformer) velocity predictor, a Denoising AutoEncoder for noise suppression, and a diversity curiosity mechanism for enhanced exploration. Against a conventional Transformer model, the DI-Transformer cuts forecast error by at least 12.5 % over 3s, 5s, 8s and 10s horizons. This enhanced foresight, combined with curiosity-driven exploration, accelerates TD3 convergence by 60.43 % relative to the soft actor-critic (SAC) and TD3. Additionality, the proposed EMS attains a fuel consumption outperforming SAC and standard TD3 by 4.63 % and 2.69 %, respectively.