基于多元算法融合的软岩隧道围岩变形预测模型及应用研究  

Study on the Prediction Model of Surrounding Rock Deformation in Soft Rock Tunnel Based on Multivariate Algorithm Fusion and Its Application

作　　者：侯守江 HOU Shoujiang(China Railway 18th Bureau Group Co.,Ltd.,Tianjin 300350)

机构地区：[1]中铁十八局集团有限公司,天津300350

出　　处：《现代隧道技术》2023年第6期151-164,共14页Modern Tunnelling Technology

基　　金：中铁十八局集团有限公司科技研究计划课题(G18-07).

摘　　要：针对传统软岩隧道围岩变形预测方法适用性和泛化能力不足的问题,提出一种基于改进D-S证据理论的多元算法融合模型,以天桥山隧道为工程依托,构建相应的软岩隧道围岩变形预测流程框架。首先采用果蝇优化算法(FOA)优化随机森林(RF)模型,采用鲸鱼优化算法(WOA)优化长短时记忆神经网络(LSTM)模型,结合偏最小二乘回归(PLSR)预测模型,采用改进D-S证据理论计算3个模型的融合权重并进行决策级融合,得到基于多元算法融合的围岩拱顶沉降及水平收敛预测模型;然后以天桥山隧道DK110+600断面G、S1、S2和S3测点的监测数据作为训练和测试样本,对比PLSR模型、RF模型、LSTM模型、FOA-RF模型、WOA-LSTM模型和多元算法融合模型的预测效果;最后以DK110+605断面为例对多元算法融合模型进行工程应用。对比结果表明,多元算法融合模型预测精度最高,相对误差在[-1.5 mm,1.5 mm]以内,平均R2值为0.9985,平均MAPE值为3.09%,可以实现对软岩隧道围岩变形的准确预测;工程应用表明,模型总体预测误差在[-3 mm,2 mm]以内,平均R2值为0.9956,平均MAPE值为5.65%,具有较高预测精度,满足指导施工的要求。Since the traditional methods for predicting surrounding rock deformation in soft rock tunnels are lacking in suitability and generalization,a multivariate algorithm fusion model based on the improved D-S evidence theory has been proposed,and the process framework for prediction of surrounding rock deformation in soft rock tunnel has been established by using the Tianqiaoshan Tunnel Project as an engineering example.First,the fruit fly optimiza⁃tion algorithm(FOA)is used to optimize the random forest(RF)model and the whale optimization algorithm(WOA)is used to optimize the long short-term memory neural network(LSTM)model.The partial least squares regression(PLSR)prediction model and the improved D-S evidence theory are used to calculate the fusion weight of the 3 models and realize decision-level fusion,which generates the surrounding rock crown settlement and horizontal conver⁃gence prediction model based on the multivariate algorithm fusion.Then the data from the measuring points G,S1,S2 and S3 of the DK110+600 cross section of Tianqiaoshan Tunnel are used as training and test samples,to com⁃pare the prediction results of PLSR model,RF model,LSTM model,FOA-RF model,WOA-LSTM model and multi⁃variate algorithm fusion model.At last,the multivariate algorithm fusion model is put to engineering application at the DK110+605 cross section.According to comparison of the models,the multivariate algorithm fusion model has the highest prediction accuracy,its relative error is within the range of[-1.5 mm,1.5 mm],its average R2 value is 0.9985 and average MAPE value 3.09%,and it allows accurate prediction of the surrounding rock deformation pat⁃tern of soft rock tunnel.Accounting to its engineering application,the overall prediction error of the model is within the range of[-3 mm,2 mm],its average R2 value is 0.9956 and average MAPE value 5.65%,and it has high enough prediction accuracy to guide construction activities.

关 键 词：软岩隧道 变形预测 多元算法融合 改进D-S证据理论 

分 类 号：U451.2[建筑科学—桥梁与隧道工程]