Transforming driving-scenario priors into adaptive reward modulation for reinforcement learning-based energy management of heavy-duty fuel cell vehicles

Deep reinforcement learning (DRL)-based energy management strategies (EMSs) reduce the lifetime operating cost of heavy-duty fuel cell vehicles (HDFCVs) by jointly considering hydrogen consumption, fuel cell degradation, and battery aging. However, most multi-objective DRL-based EMSs use fixed reward weights, imposing the same optimization preference across changing driving scenarios. Such fixed preferences can perform well on a specific training cycle but may limit optimization performance when driving conditions change. Existing scenario-aware methods mainly use scenario priors as state inputs or policy-switching signals, which either provide context without explicitly guiding objective adaptation or require multiple scenario-specific policies, increasing training and deployment complexity. To address this gap, this study proposes a scenario-adaptive multi-objective DRL framework that transforms driving-scenario priors into adaptive reward modulation. An online soft scenario cognition model provides continuous scenario memberships, while dynamic programming (DP)-based weight perturbation identifies scenario-specific cost priorities. Based on these priorities, confidence-gated reward modulation adapts the emphasis among cost objectives and suppresses modulation when scenario dominance is ambiguous. For a 49-ton HDFCV, training-domain ablation shows that A-SAC achieves the lowest comprehensive operating cost among the evaluated DRL strategies and reaches 98.57% of the DP economy benchmark. Across 40 test cycles excluded from DRL policy training and hyperparameter tuning, A-SAC reduces mean comprehensive operating cost by 10.37% relative to SAC and achieves lower costs in 38 cases, demonstrating improved multi-objective EMS optimization and cross-cycle generalization.