基于循环神经网络的超大直径盾构掘进地表沉降预测方法研究  

Surface Settlement Prediction Method for Super-large Diameter Shield Tunneling Based on Recurrent Neural Network

作　　者：马川 盛光祖 陈健 李义翔[1] 黄兴 张建勇[1] 文天 MA Chuan;SHENG Guangzu;CHEN Jian;LI Yixiang;HUANG Xing;ZHANG Jianyong;WEN Tian(China Railway 14 Bureau Group Large Shield Engineering Co.Ltd.,Nanjing 210000,China;Wuhan Urban Construction Group,Wuhan 430040,China;CRCC Underwater Tunnel Engineering Laboratory,Jinan 250101,China;Intelligent Construction and Equipment Remanufacturing of Large Shield Tunnel Shandong Engineering Research Center(Engineering Laboratory),Jinan 250101,China;Ocean University of China,Qingdao 266100,Shandong China;State Key Laboratory of Geomechanics and Geotechnical Engineering,Institute of Rock and Soil Mechanics,Chinese Academy of Sciences,Wuhan 430071,China;School of Urban Construction,Wuhan University of Science and Technology,Wuhan 430081,China)

机构地区：[1]中铁十四局集团大盾构工程有限公司,南京210000 [2]武汉城建集团建设管理有限公司,武汉430040 [3]中国铁建水下隧道工程实验室,济南250101 [4]大盾构隧道智能化建造与装备再制造山东省工程研究中心(工程实验室),济南250101 [5]中国海洋大学,山东青岛266100 [6]中国科学院武汉岩土力学研究所岩土力学与工程国家重点实验室,武汉430071 [7]武汉科技大学城市建设学院,武汉430081

出　　处：《河南科学》2024年第4期558-566,共9页Henan Science

基　　金：山东省自然科学基金面上项目(ZR2023ME048);泰山产业领军人才工程专项经费资助(tscx202306015);中国铁建科研开发计划(2023-B04);中铁十四局科技研发计划项目(9137000016305598912021A01)。

摘　　要：为了研究超大直径盾构掘进过程地面沉降规律,以武汉市和平大道南延线盾构工程为研究对象,首先收集了超大直径盾构下穿过程掘进参数和地层地质参数,并使用盾构掘进过程深跨比描述超大直径盾构影响特征;其次,通过收集现场沉降测点数据分析盾构隧道施工阶段地表沉降的影响范围,计算了90%、95%、99%三种置信区间下地表沉降影响范围;最后,选取不同范围内的多元时序数据作为输入参数,分别建立了基于贝叶斯优化算法(BO)的长短期记忆(LSTM)、BP神经网络和随机森林(RF)大直径盾构地面沉降预测模型.模型运行过程中,通过贝叶斯优化算法分别寻找三种不同模型下的最优超参数,并通过四种评价指标对比模型精度.结果如下:①在90%置信水平下三种算法均表现出最高精度,通过区间计算筛选有效输入参数能有效提高模型预测精度;②LSTM对隧道沉降的预测结果优于传统机器学习算法模型,MAPE最低达到8.91%,R^(2)达到90%.In order to study the law of land subsidence in the process of super large diameter shield tunneling,a case study of the shield project of the south extension line of Heping Avenue in Wuhan is carried out.Firstly,the tunneling parameters and stratum geological parameters of the super-large diameter shield tunneling process are collected,and the depth-span ratio is used to describe the influence characteristics of the super-large diameter shield.Secondly,the influence range of surface settlement in shield tunnel construction stage is analyzed by collecting the field settlement measurement data,and the influence range of surface settlement under three confidence intervals of 90%,95% and 99% is calculated.Finally,multivariate time series data in different ranges are selected as input parameters.The long short-term memory(LSTM),BP neural network and random forest(RF)large-diameter shield ground subsidence prediction models based on Bayesian optimization algorithm(BO)are established respectively.During the operation of the model,the Bayesian optimization algorithm is used to find the optimal hyperparameters under three different models,and the accuracy of the model is compared by four evaluation indexes.The results are as follows:①At the 90% confidence level,the three algorithms all show the highest accuracy,and the effective input parameters screened by interval calculation can effectively improve the prediction accuracy of the model;②The prediction result of LSTM for tunnel settlement is better than that of traditional machine learning algorithm model.The minimum MAPE is 8.91% and R^(2) is 90%.

关 键 词：超大直径盾构 地表沉降预测 循环神经网络 深跨比 

分 类 号：TU478[建筑科学—结构工程]