次固结系数的集成机器学习预测模型  

A Ensemble Machine Learning Prediction Model for Sub-consolidation Coefficient

作　　者：刘延昭 汪洋 张明 苏秉强 许明秀 苏明明 LIU Yanzhao;WANG Yang;ZHANG Ming;SU Bingqiang;XU Mingxiu;SU Mingming(Zhejiang Jiaogong Underground Construction Co.,Ltd.Hangzhou 311421,China;Zhejiang Communications Construction Group Co,Ltd.Hangzhou 310051,China;Research Center of Coastal and Urban Geotechnical Engineering,Zhejiang University Hangzhou 310058,China)

机构地区：[1]浙江交工地下工程有限公司,杭州311421 [2]浙江交工集团股份有限公司,杭州310051 [3]浙江大学滨海和城市岩土工程研究中心,杭州310058

出　　处：《广东土木与建筑》2025年第1期45-50,共6页Guangdong Architecture Civil Engineering

基　　金：浙江省建设厅项目(2022K073);浙江大学-浙江交工协同创新联合研究中心项目(ZDJG2021001)。

摘　　要：工程实践中的长期沉降问题由次固结系数控制,但是现有的次固结系数经验模型存在着考虑因素不周、研究对象不足等诸多问题,不够可靠。与以往的相关性研究不同,研究提出了一种预测次固结系数的集成机器学习模型。新的模型采用了随机森林、自适应提升法、梯度提升回归树、极致梯度提升法4种集成学习方法,并采取网格搜索法和k折交叉验证法优化模型的超参数。文章建立起了4个输入变量(液限、塑性指数、孔隙比、黏土含量)和1个输出变量(次固结系数)之间的复杂关系。此外,分析了特征的相对重要性,提高了模型的可解释性。研究结果发现,极致梯度提升法的预测效果最好,随机森林的预测效果最差。孔隙比是4个特征中最重要的,黏土含量的相对特征重要性最低。The long-term settlement problem in engineering practice is controlled by the sub-consolidation coefficient,but the existing empirical models of sub-consolidation coefficient have many problems such as inadequate consideration of factors and insufficient research objects,which are not reliable enough.Unlike previous correlation studies,this study proposes an ensemble machine learning model for predicting secondary consolidation coefficients.The new model uses four integrated learning methods:random forest,adaptive boosting method,gradient boosting regression tree,and extreme gradient boosting method,and adopts grid search method and k-fold cross-validation method to optimize the hyperparameters of the model.This study establishes a complex relationship between four input variables(liquid limit,plasticity index,void ratio,clay content)and one output variable(sub-consolidation coefficient).And the relative importance of features is analyzed,which improves the interpretability of the model.The research results found that the prediction effect of the extreme gradient boosting method is the best,while the prediction effect of random forests is the worst.The void ratio is the most important of the four characteristics,while the relative characteristic of clay content is the least important.

关 键 词：软黏土 次固结系数 蠕变 集成学习 机器学习 

分 类 号：TU433[建筑科学—岩土工程]