The deep-sea mining ships are the dominant contributor to both energy consumption and carbon emissions within the overall deep-sea mining systems. To achieve optimal energy management for the integrated energy system of such ships, this study proposes a Physical-Assisted Safe Reinforcement Learning (PA-SRL) framework. The proposed method fully exploits the model-free nature of reinforcement learning, while incorporating reward shaping based on ship operational constraints, thereby effectively mitigating the operational risks of the agent. Furthermore, the proposed method embeds physics-informed priors into the Soft Actor–Critic (SAC) algorithm, which significantly enhances both the training efficiency and the performance effectiveness of the agent. The real-world case study conducted in the South China Sea demonstrates the effectiveness of the proposed method in balancing economic efficiency with operational robustness under various uncertainties.
Wang et al. (Sun,) studied this question.