主管:中华人民共和国应急管理部
主办:应急管理部天津消防研究所
ISSN 1009-0029  CN 12-1311/TU

消防科学与技术 ›› 2024, Vol. 43 ›› Issue (5): 704-708.

• • 上一篇    下一篇

基于深度学习的氢喷射火事故后果预测方法

何旭, 孔得朋, 杨国栋, 于溪芮   

  1. (中国石油大学(华东) 机电工程学院,山东 青岛 266580)
  • 收稿日期:2023-12-18 修回日期:2024-03-08 出版日期:2024-05-15 发布日期:2024-05-15
  • 作者简介:何 旭(1994- ),中国石油大学(华东)机电工程学院博士研究生,主要从事氢泄漏燃爆事故后果预测方向的研究,山东省青岛市黄岛区长江西路66号,266580。
  • 基金资助:
    国家重点研发计划(2021YFB4000905)

Hydrogen jet fire accident consequences predicting method based on deep learning

He Xu, Kong Depeng, Yang Guodong, Yu Xirui   

  1. (1. Department of Building Environment and Energy Engineering, The Hong Kong Polytechnic University, Hong Kong 999077, China;2. Centre for Offshore Engineering and Safety Technology, China University of Petroleum, Shandong Qingdao 266580, China)
  • Received:2023-12-18 Revised:2024-03-08 Online:2024-05-15 Published:2024-05-15

摘要: 加氢站作为氢能应用产业最重要的配套基础设施之一,其中存在的大量高压氢气伴随着泄漏风险。高压氢气一旦泄漏,极易形成喷射火,对加氢站内构筑物及人员生命财产安全造成严重威胁。为了实现氢喷射火事故后果的快速准确预测,提出了基于神经网络的代理模型事故后果预测方法,相较于传统数值仿真方法具有显著的时效性优势。该方法构建了一种基于生成对抗网络和长短期记忆神经网络的混合代理模型,使用数值仿真生成的训练样本对代理模型进行训练,完成训练的代理模型能够预测加氢站高压氢气泄漏产生喷射火事故后的温度分布情况。使用模糊C均值聚类方法和结构相似度指标定量分析了代理模型预测结果的准确度,结果表明氢喷射火后果预测代理模型能够在保证可接受预测精度的前提下,极大地提高后果预测效率,实现了加氢站氢喷射火后果的时空快速预测。

关键词: 神经网络, 代理模型, 后果预测, 加氢站, 氢喷射火

Abstract: As one of the most important supporting infrastructures for the hydrogen energy application industry, hydrogen refueling station (HRS) is characterized by the presence of large quantities of high-pressure hydrogen accompanied by significant leakage risks. Once the high-pressure hydrogen leaks, it is very easy to form a jet fire, which poses a serious threat to the structures in the HRS as well as to the safety of people's lives and properties. In order to realize the fast and accurate prediction of the consequences of hydrogen jet fire accidents, a neural network-based surrogate model accident consequence prediction method is proposed, which has a significant time-saving advantage over the traditional numerical simulation methods. The method constructs a hybrid surrogate model based on adversarial generative network and long and short-term memory neural network, and the training samples generated by numerical simulation are used to train the surrogate model, and the completed surrogate model can predict the temperature distribution after the jet fire accident caused by the high-pressure hydrogen leakage from HRS. The accuracy of the prediction results of the surrogate model was quantitatively analyzed using the fuzzy C-means and the Structure Similarity Index Measure, and the results showed that the surrogate model for predicting the consequences of hydrogen jet fires can greatly improve the efficiency of consequence prediction under the premise of guaranteeing the acceptable prediction accuracy, realizing the spatio-temporal and fast prediction of the consequences of hydrogen jet fires in HRS.

Key words: neural network, surrogate model, consequence prediction, hydrogen refueling station, hydrogen jet fire