摘要
边坡稳定性研究对于地质灾害的防治尤为重要,为了研究不同机器学习模型的边坡稳定性预测效果,以边坡稳定性系数为预测对象,选择重度、黏聚力、内摩擦角、边坡高度、坡角、孔隙压力比为输入变量,建立了基于支持向量机(SVR)的边坡稳定性系数预测模型。研究了不同核函数对SVR模型预测结果的影响,对比了SVR模型与极限学习机(ELM)模型、BP神经网络模型的预测效果,分析了各输入变量对SVR模型与ELM模型预测性能的影响。结果表明:基于SVR的边坡稳定性系数预测模型的预测效果较好,其中采用了RBF核函数的SVR模型预测效果较好,平均绝对误差为8.10%,均方根误差为0.034;输入变量对SVR模型与ELM模型的影响较大,剔除各变量后SVR模型预测结果的均方根误差变化小于0.02,因此本文建立的SVR模型具有更好的稳定性。
Slope stability is a very important issue for the prevention and control of geological disasters. In order to analyze the prediction effect of slope stability of different machine learning models, the prediction model of slope stability coefficient based on support vector machine(SVR) is established by taking the slope stability coefficient as the prediction object and selecting the weight, cohesion, internal friction angle, slope height, slope angle and pore pressure ratio as the input variables. The influence of different kernel functions on the prediction results of SVR model is compared, the prediction effects of SVR model, ELM model and BP neural network model are compared, and the influence of each input variable on the prediction performance of SVR model and ELM model is analyzed. The results show that the SVR model using RBF kernel function has a good prediction effect, with an average absolute error of 8.10% and a root mean square error of 0.034. The input variables have a great impact on SVR model and ELM model. The root-mean-square error of SVR model prediction results is less than 0.02 after removing all variables, so the SVR model established in this paper has better stability.
作者
刘相龙
刘仲武
马灵会
任凯
王晓东
LIU Xianglong;LIU Zhongwu;MA Linghui;REN Kai;WANG Xiaodong(Zhonglan Railway Passenger Dedicated Line Co.,Ltd.,Lanzhou,Gansu 730000,China;Gansu Academy of Civil Engineering Science Co.,Ltd.,Lanzhou,Gansu 730000,China)
出处
《水利与建筑工程学报》
2023年第1期172-178,共7页
Journal of Water Resources and Architectural Engineering
基金
甘肃省住房和城乡建设厅建设科技项目(JK2021-46,JK2021-55)。
关键词
边坡稳定性系数
支持向量机
变量分析
机器学习
slope safety factor
support vector machine
variable analysis
machine learning
