基于随机森林算法和物理模型试验的软土滑坡运动特性预测研究

周书东; 丁其乐; 王奕仁; 张彤炜; 张益; 王凤阳

doi:10.13928/j.cnki.wrahe.2025.07.018

水利水电技术（中英文） ›› 2025, Vol. 56 ›› Issue (7) : 239 -248. DOI: 10.13928/j.cnki.wrahe.2025.07.018

工程地质

基于随机森林算法和物理模型试验的软土滑坡运动特性预测研究

周书东 ¹ ,
丁其乐 ²^,³ ,
王奕仁 ²^,³ ,
张彤炜 ¹ ,
张益 ¹ ,
王凤阳 ²^,³

作者信息 +

Predicting the dynamic characteristics of soft soil landslides based on the random forest algorithm and physical model experiments

Shudong ZHOU ¹ ,
Qile DING ²^,³ ,
Yiren WANG ²^,³ ,
Tongwei ZHANG ¹ ,
Yi ZHANG ¹ ,
Fengyang WANG ²^,³

Author information +

文章历史 +

PDF

摘要

【目的】软土滑坡因其复杂的塑性变形和流变特性，破坏机理不同于硬质岩土中的普通滑坡，传统的滑坡研究方法难以有效捕捉其运动特性非线性和不确定性特征。【方法】为量化评估软土滑坡的运动特性，采用软土的黏塑性流体假设，利用理想黏塑性材料Carbopol作为试验材料开展物理模型试验，并选取滑坡到达指定距离的厚度和速度作为代表性参数，通过随机森林算法训练试验数据构建预测模型。【结果】结果显示：厚度训练数据和测试数据的决定系数分别为0.941和0.923,速度训练数据和测试数据的决定系数分别为0.936和0.917;厚度和速度残差主要分布范围分别为(-0.02,0.02)和(-0.1,0.075),且呈正态分布。此外，自变量重要性分析显示屈服应力为对模型影响最大的输入变量，其重要程度特征值为0.35。【结论】结果表明：该模型具有较高的预测精度和良好的泛化能力，在处理软土滑坡动力学的高维度数据和复杂非线性关系方面具有较好的表现，可为软土滑坡灾害防治提供科学依据。

Abstract

[Objective] Soft soil landslides, due to their complex plastic deformation and rheological properties, have failure mechanisms distinct from ordinary landslides in hard rock and soil. Traditional method of landslide research are ineffective at capturing the nonlinear and uncertain characteristics of their movement. [Methods] To quantitatively evaluate the movement characteristics of soft soil landslides, a viscoplastic fluid assumption of soft soil was adopted, using ideal viscoplastic material Carbopol as the test material in physical model experiments. Thickness and velocity at a specified distance were chosen as representative parameters, and a prediction model was built by training the experimental data with the random forest algorithm. [Results] The results show that the determination coefficients of the thickness training and testing data were 0.941 and 0.923, respectively, while those of the velocity training and testing data were 0.936 and 0.917, respectively. The residuals for thickness and velocity were mainly distributed within the ranges of(-0.02, 0.02) and(-0.1, 0.075), respectively, and showed a normal distribution. Furthermore, an analysis of feature importance indicated that yield stress had the most significant impact on the model, with an importance value of 0.35. [Conclusion] The results demonstrate that the model has high predictive accuracy and good generalization ability, and it performs well in handling high-dimensional data and complex nonlinear relationships in the dynamics of soft soil landslides. This provides a scientific basis for the prevention and control of soft soil landslide disasters.