基于LSTM的海上LNG转驳系统泄漏事故预测方法研究  

RESEARCH ON LEAKAGE ACCIDENT PREDICTION METHOD OF OFFSHORE LNG TRANSFER SYSTEM BASED ON LSTM

作　　者：邓陈辉 张纪涵 DENG Chenhui;ZHANG Jihan(China State Shipbuilding Corporation,Shanghai 200011,China;China University of Petroleum,Beijing,Beijing 102249,China)

机构地区：[1]中国船舶集团有限公司,上海200011 [2]中国石油大学(北京),北京102249

出　　处：《力学与实践》2024年第3期500-510,共11页Mechanics in Engineering

摘　　要：在海上液化天然气(liquefied natural gas,LNG)转驳系统中,一旦发生泄漏事故,其后果将极其严重,可能引发火灾、爆炸、中毒等危害。液化天然气泄漏事故发生速度迅猛,因此如何快速进行泄漏扩散的预测对于应对突发事件下的人员疏散和设备保护至关重要。本研究构建了一种基于长短期记忆神经网络(long short-term memory,LSTM)的海上液化天然气转驳系统泄漏扩散预测模型,利用流体动力学仿真计算,获取了大量的基础数据集,然后通过训练,能够有效地对气体扩散浓度进行准确预测,所得结果的均方差和平均绝对误差均低于门控循环单元(gated recurrent unit,GRU)神经网络模型和反向传播神经网络模型。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.

关 键 词：海上液化天然气转驳系统 泄漏事故 长短期记忆神经网络 门控循环单元 反向传播 

分 类 号：U698.6[交通运输工程—港口、海岸及近海工程]