摘要
为实现软岩隧道施工围岩变形的准确预测,利用长短时记忆(LSTM)神经网络构建软岩隧道拱顶沉降及水平收敛预测模型,采用麻雀搜索算法(SSA)实现LSTM模型超参数寻优,构建SSA-LSTM模型计算流程框架。以梁王山隧道软岩段为例,对围岩大变形进行现场实测与分析,获取拱顶沉降和水平收敛监测数据,代入SSA-LSTM模型进行计算,并与LSTM模型和SSA优化的传统机器学习模型进行误差对比分析。结果表明,SSA-LSTM模型相对误差率为[-1%,2%],R^(2)为0.9986,MAPE为2.3458%,RMSE为0.5298,均为所有模型中最优。为验证SSA-LSTM模型对于未开挖断面沉降变形的预测效果,选取K33+260断面作为研究对象构建未开挖断面沉降变形预测模型,误差分析结果表明,模型预测精度满足指导施工的要求。
In order to accurately predict the surrounding rock deformation during construction of soft rock tunnel,the soft rock tunnel crown settlement and horizontal convergence prediction model has been created by using the long short-term memory(LSTM)neural network,the hyper-parameters of LSTM model have been optimized by using the sparrow search algorithm(SSA),and the calculation process framework for SSA-LSTM has been created.With the soft rock section of Liangwangshan Tunnel being used as an example,field measurement and analysis of large deformation of surrounding rock have been conducted.The crown settlement and horizontal convergence monitoring data is obtained and then substituted into the SSA-LSTM model for calculation.The calculation results are com⁃pared against the results of LSTM model and the SSA optimized traditional machine learning model and the errors are analyzed.As the results indicate:the relative error rate of SSA-LSTM model is[-1%,2%],R^(2) is 0.9986,MAPE is 2.3458%,RMSE is 0.5298,and this model is the best one of all the models.In order to verify the settlement and deformation of unexcavated section predicted by the SSA-LSTM model,the K33+260 section is used as the object of study and the prediction model for settlement and deformation of unexcavated section is created.According to the re⁃sults of error analysis,the prediction accuracy of the model is good enough to guide construction.
作者
王锋
WANG Feng(China Railway Siyuan Survey and Design Group Co.,Ltd.,Wuhan 430063)
出处
《现代隧道技术》
CSCD
北大核心
2024年第1期56-66,共11页
Modern Tunnelling Technology
基金
中铁第四勘察设计院集团有限公司科研课题(2021K062).
关键词
软岩隧道
变形特征
深度学习
智能预测
Soft rock tunnel
Deformation characteristics
Deep learning
Intelligent prediction
