融合静动态特征的高铁隧道进口段拱顶沉降预测模型研究OA
Research on the Prediction Model for Arch Settlement at the Entrance Section of High-speed Rail Tunnels with Static and Dynamic Features
针对复杂地质条件下高铁隧道进口段拱顶沉降长期预测难题,选取围岩等级、隧道埋深、黏聚力和内摩擦角等静态地质力学参数与开挖初期拱顶沉降序列(A2～A8)等11个影响因子,提出了一种基于卷积神经网络(convolutional neural network,CNN)和长短期记忆神经网络(long short-term memory,LSTM)的预测模型.以某高铁隧道进口段为工程背景,首先对Ⅲ级、Ⅳ级、V级围岩的拱顶沉降与周边收敛进行了监测特征分析,结果表明,变形普遍呈快速增加—缓慢增加—趋于稳定的时程特征,且围岩等级对累计沉降值的影响最为显著.随后,在同一数据集上,将所提出CNN-LSTM模型与LSTM、ANN-LSTM等传统模型进行预测性能对比,结果表明,在决定系数R2、均方误差MSE、平均绝对百分比误差MAPE指标上,CNN-LSTM模型均表现出显著优势.针对4个不同工况下的独立验证测试,CNN-LSTM模型表现最优,其R2均在0.93以上,表明其具有较强的泛化能力.敏感性分析揭示了各影响因子的贡献度排序:围岩等级和黏聚力对最终沉降的贡献最为突出,其次为内摩擦角、隧道埋深和开挖初期拱顶沉降时间序列.研究结果表明,将开挖初期拱顶沉降时间序列与静态地质力学参数在深度网络中协同表征,能够实现隧道进口段拱顶最终沉降的提前预测,从而为工程现场的支护优化设计、监测方案加密与风险预警决策提供定量化理论支撑.
In order to solve the challenge of long-term prediction of arch settlement for the tunnel entrance section under complex geological conditions,static geomechanical parameters such as surrounding rock grade,tunnel burial depth,cohesion,and internal friction angle,and 11 influencing factors including the initial arch settlement se-quence after excavation(A2～A8),a prediction model based on convolutional neural network(CNN)and long short term memory(LSTM)is proposed.Taking the entrance section of a high-speed railway tunnel as the engineering background,the monitoring characteristics of the arch settlement and peripheral convergence of Class Ⅲ,Ⅳ,and V surrounding rocks were first analyzed.The results showed that the deformation generally exhibited a time history of rapid increase-slow increase-and tend to be stable,and the influence of surrounding rock grade on the cumulative settlement value was the most significant.Subsequently,on the same dataset,the proposed CNN-LSTM model was compared with traditional models of LSTM and ANN-LSTM in terms of predictive performance.The results showed that the CNN-LSTM model exhibited significant advantages in the determination coefficient R2,mean square error MSE,and mean absolute percentage error MA PE indicators.The CNN-LSTM model performed the best in indepen-dent validation tests under four different operating conditions,with all R2 values above 0.93,indicating its strong generalization ability.The contribution ranking of various influencing factors were revealed by sensitivity analysis:the rock mass grade and cohesion had the most prominent contribution to the final settlement,followed by the time series of internal friction angle,tunnel burial depth,and initial arch crown settlement during excavation.The re-search results indicate that the time series of initial arch crown settlement and static geomechanical parameters were synergistically characterized in a deep network,which is possible to predict the final settlement of the tunnel en-trance section arch in advance.And then,it can provide quantitative theoretical support for the optimization design of support,the refinement of monitoring schemes,and risk warning decisions for engineering sites.
卢振忠
中铁十八局集团第一工程有限公司,河北涿州 072750
交通工程
高铁隧道拱顶沉降静动态特征CNN-LSTM敏感性分析
high-speed rail tunnelarch settlementstatic and dynamic characteristicsconvolutional neural net-work and long short-term memory(CNN-LSTM)sensitivity analysis
《市政技术》 2026 (1)
193-201,9
中铁十八局集团有限公司2023年度科技发展计划课题(G23-47)
评论