Diesel-electric hybrid trains (DEHTs) are regarded as a transitional but indispensable solution for decarbonizing heavy-haul railways. Their energy-management strategy (EMS) must simultaneously minimize diesel consumption and sustain battery health under highly dynamic operating conditions. This paper proposes a rule-augmented deep reinforcement learning (RDRL) framework that embeds expert knowledge into a Deep Deterministic Policy Gradient (DDPG) agent to achieve real-time optimal power split. Three categories of prior knowledge—(i) the engine’s optimal brake-specific fuel-consumption (BSFC) curve, (ii) charge–discharge characteristics of the traction battery, and (iii) train operation timetables—are encoded as action-space shaping and reward regularization. A high-fidelity backward simulation model of a 1.5 MW DEHTs is established; engine, generator and motor efficiency maps are derived from bench tests, whereas the battery is represented by an Rint model. The RDRL agent is trained with prioritized experience replay and decayed Ornstein–Uhlenbeck noise. Compared with a baseline DDPG and a deterministic rule-based EMS, the proposed method reduces specific fuel consumption by up to 3.1% and accelerates convergence by 38.7%. When benchmarked against dynamic programing, it achieves 94.6% fuel-economy optimality while satisfying state-of-charge constraints. Semi-physical experiments on an autonomous-rail rapid-transit platform further confirm the framework’s adaptability to fuel-cell–battery hybrids. The results demonstrate that integrating domain knowledge with deep reinforcement learning effectively narrows the search space, enhances robustness, and enables real-time deployment for large-scale railway applications.
Li et al. (2026) studied this question.