长短期记忆网络在隧道火灾实时致灾态势预测中应用研究  

Application of long short-term memory networks for real-time tunnel fire disaster prediction

作　　者：贾进章[1,2] 陈佳琦 陈怡诺 JIA Jinzhang;CHEN Jiaqi;CHEN Yinuo(School of Safety Science and Engineering,Liaoning Technical University,Fuxin 123000,Liaoning,China;Key Laboratory of Mine Thermodynamic Disaster and Prevention,Ministry of Education,Liaoning Technical University,Huludao 125105,Liaoning,China;School of Civil Engineering,Liaoning Technical University,Fuxin 123000,Liaoning,China)

机构地区：[1]辽宁工程技术大学安全科学与工程学院,辽宁阜新123000 [2]辽宁工程技术大学矿山热动力灾害与防治教育部重点实验室,辽宁葫芦岛125105 [3]辽宁工程技术大学土木工程学院,辽宁阜新123000

出　　处：《安全与环境学报》2025年第4期1298-1309,共12页Journal of Safety and Environment

基　　金：国家自然科学基金项目(52174183,52374203)。

摘　　要：针对隧道火灾过程中高温烟气对人员避灾的威胁,为实现隧道火灾有效控制,及时提供隧道火灾实时救援决策,提出了一种试验测量和人工智能相结合的方法,基于温度传感器和长短期记忆(Long Short-Term Memory,LSTM)网络对烟气温度进行实时预测。首先,通过1∶20小面积火灾试验收集不同工况下的温度数据,然后,采用LSTM模型从试验火灾数据库中学习、训练,并进行不同火源类型测试,发现该算法模型可以很好地预测隧道内温度分布。对模型的预测能力进行测试,测试结果表明,预测结果精度高,相对误差在±10%内。与反向传播神经网络(Back Propagation Neural Network,BPNN)模型进行比较,测试误差均值降低3.85百分点,对比效果明显,满足隧道火灾实时态势检测需要,为隧道火灾事故的应急救援建立了较为新颖的智能预测方法。To achieve effective control of tunnel fires and facilitate real-time rescue decisions,this paper presents a method that integrates experimental measurements with artificial intelligence.Specifically,it utilizes temperature sensors and Long Short-Term Memory(LSTM)networks to predict flue gas temperatures in real time.Firstly,temperature data were collected under various working conditions through a 1:20 small-scale fire experiment.This included different advance times(10 s,20 s,30 s),kerosene sizes(0.1 m×0.1 m,0.13 m×0.13 m,0.16 m×0.16 m),and ventilation wind speeds(0.15 m/s,0.51 m/s,0.78 m/s).The main tunnel and bifurcation tunnel areas of the bifurcated branch tunnel were divided for longitudinal temperature distribution experiments,and an analysis of measurement errors was conducted.Subsequently,the LSTM model is employed to learn from and train on the experimental fire database.Various fire source types are tested,and the results indicate that the algorithm effectively predicts the temperature distribution within the tunnel.The model s predictive capability was evaluated,and the results demonstrated high accuracy,with a relative error within±10%.Compared with the BPNN model,although the training error of the LSTM model is 0.5 percentage points higher,the average value of its test error is 3.85 percentage points lower.Ultimately,it was concluded that as the pyrolysis rate of the fire source increases,longitudinal ventilation results in a decrease in the temperature of the tunnel roof,provided that the wind speed remains constant.Additionally,the ventilation system not only suppresses the reverse flow of smoke but also lowers the temperature of the tunnel roof during the process.By integrating temperature sensor data with the LSTM algorithm,a real-time prediction method for the fire temperature distribution in multi-branch tunnels is proposed.The comparison results are significant,effectively addressing the requirements for real-time detection of tunnel fire situations.This establishes a relatively novel in

关 键 词：安全工程 隧道火灾 长短期记忆网络 烟气温度 实时预测 

分 类 号：X932[环境科学与工程—安全科学]