液化场地水平位移预测数据驱动方法研究  

Study of data-driven methods for predicting soil liquefactioninduced lateral displacement

作　　者：张屹蕃 王睿[1] 张建民[1] 张建红[1] ZHANG Yifan;WANG Rui;ZHANG Jianmin;ZHANG Jianhong(State Key Laboratory of Hydroscience and Engineering,Tsinghua University,Beijing 100084,China)

机构地区：[1]清华大学水沙科学与水利水电工程国家重点实验室,北京100084

出　　处：《水力发电学报》2023年第3期103-117,共15页Journal of Hydroelectric Engineering

基　　金：国家自然科学基金项目(52090084,52279101,52038005)。

摘　　要：基于Youd等2002液化变形数据库,结合机器学习方法,本研究建立了基于机器学习的液化场地水平位移预测数据驱动方法,并利用其对新近地震液化场地的水平位移进行了预测。首先,收集了新近几次地震液化场地水平位移案例,在此基础上采用已有常用工程经验方法对新近地震液化场地水平位移进行预测以探究其适用性,发现Youd 2018方法具有良好的表现。为获得最优的机器学习模型,本研究分别讨论了BP神经网络模型(BPNN)、径向基神经网络模型(RBF)、决策树模型(DT)、随机森林模型(RF)、支持向量机模型(SVM)等机器学习模型的适用性。研究发现,随机森林模型(RF)表现优越,该方法计算效率高、数据可扩展性好,同时能够很好地反映已有数据特性。相较于Youd 2018方法,随机森林模型(RF)对新近地震液化场地水平位移的整体预测精度提升18.17%。最后,本研究探讨了随机森林模型(RF)预测液化场地水平位移时不同影响因素的敏感性。A machine learning-based method is developed based on the liquefaction-induced lateral displacement database of Youd et al.,2002,and applied to the simulation of some soil displacement cases in recent earthquakes.We collect the histories of these cases,and then predict them using the existing engineering experience methods to explore the applicability of the model,showing the Youd 2018 method has a good performance.To obtain an optimal machine learning model,this paper discusses the applicability of five models-BP neural network(BPNN),radial basis neural network(RBF),decision tree(DT),random forest(RF),and support vector machine(SVM).We find that the performance of RF is superior to the other machine learning methods.It has high computational efficiency and good data scalability,and can well reflect the characteristics of the data available to this study.Its prediction accuracy is increased by 18.17%relative to the Youd 2018 method.In addition,a sensitivity analysis is carried out of the influencing factors of liquefaction-induced lateral deformation simulation using the RF model.

关 键 词：砂土液化 水平位移 数据驱动 机器学习 敏感性分析 

分 类 号：TU443[建筑科学—岩土工程]