LSTM-augmented DRL for generalisable energy management of hydrogen-hybrid ship propulsion systems

Enhancing the generalisation of energy management strategies is crucial for hybrid ship power systems to adapt to unknown navigation conditions effectively. A long short-term memory (LSTM)-based data augmentation method is employed to mitigate uncertainty in propulsion power, thereby enhancing the generalisation of energy management strategies based on deep reinforcement learning (DRL). Simulations using a hybrid propulsion model and operational data from “Three Gorges Hydrogen Boat No.1” compared DQN and DDPG algorithms with and without LSTM integration. By evaluating the DRL strategy's performance in reducing fuel cell operating pressure and energy consumption before and after data augmentation, the quality of generalisation performance is characterised. Results show that optimisation target weights affect training convergence and performance under unknown test conditions. Data enhancement via the LSTM model improves DRL generalisation in unknown navigation conditions. Compared to original DDPG, LSTM-DDPG reduces FC operating pressure by 5.82 % and 1.86 %, and cuts hydrogen consumption by 0.80 % and 2.13 % under two days of unknown conditions. This research offers guidance for designing energy management strategies with high generalisation, addressing adaptability issues with real-world data uncertainty.