Abstract:A leakage accident in offshore liquefied natural gas (LNG) transfer systems can lead to severe consequences, including the risk of fire, explosions, and poisoning. These accidents occur rapidly, making it crucial to predict and respond swiftly, particularly for emergency evacuations and equipment protection. In this study, we propose a prediction model for LNG leak diffusion in offshore transfer systems, based on Long Short-Term Memory (LSTM) neural networks. Leveraging fluid dynamics simulations, we gather a substantial dataset. After rigorous training, our model effectively forecasts gas concentration diffusion. The mean square error and average absolute error are both lower than those of the gated recurrent unit (GRU) and backpropagation neural network models.