Deep reinforcement learning-based eco-driving strategy for hybrid heavy-duty trucks in long-haul highway scenarios

Hybrid heavy-duty truck eco-driving strategies are an effective means to reduce fuel consumption and improve vehicle fuel efficiency. However, in long-distance highway driving scenarios, existing strategies lack road prediction capabilities and are greatly affected by changes in the vehicle's total mass, resulting in poor fuel efficiency. To address these limitations, we propose a deep reinforcement learning (DRL)-based eco-driving strategy tailored for long-haul highway operations of hybrid heavy-duty trucks. First, we develop a hierarchical field-of-view and multidimensional feature fusion method to identify longitudinal road gradients, converting previewed slope information into structured feature vectors that enable robust recognition of diverse gradient segments. Second, we introduce a dual-timescale adaptive Kalman filter-based mass observer that accounts for both the slow-varying nature of truck mass and fast-varying internal disturbances, achieving high-precision and robust mass estimation. Third, we design a DRL-driven eco-driving strategy that integrates road gradient and vehicle mass dynamics, incorporating these states as critical inputs into a soft actor-critic (SAC) framework. This model balances cumulative energy-efficiency rewards with policy entropy to generate optimal velocity profiles for hybrid heavy-duty trucks across varied long-haul highway scenarios, thereby minimizing energy consumption under differing gradients and mass conditions. Finally, validation on randomly sampled road segments excluded from training demonstrates that the proposed strategy significantly outperforms existing approaches in both total fuel consumption and equivalent energy consumption, delivering superior energy savings while exhibiting strong robustness to variations in road topology and payload.