基于优化变分模态分解和深度学习的滑坡位移预测模型

金属矿山 ›› 2025, Vol. 54 ›› Issue (11): 138-145.

基于优化变分模态分解和深度学习的滑坡位移预测模型

张　研^1,2　叶玉龙¹　王根伟³　荆浩然¹

1.桂林理工大学土木工程学院,广西桂林 541004;2.广西岩土力学与工程重点实验室,广西桂林 541004; 3.广西自然资源职业技术学院自然资源工程系,广西南宁 532100

出版日期:2025-11-15 发布日期:2025-12-01
通讯作者: 张　研(1983—),男,教授,博士,硕士研究生导师。
作者简介:王根伟(1985—),男,副教授,硕士。
基金资助:
国家自然科学基金项目(编号:52068016);广西自然科学基金项目(编号:2020GXNSFAA159118);水利工程岩石力学广西高等学校高水平创新团队及卓越学者计划项目(编号:202006);广西岩土力学与工程重点实验室开放基金项目(编号:桂科能20-Y-XT-01)。

Landslide Displacement Prediction Model Based on Optimal Variational Modal Decomposition and Deep Learning

ZHANG Yan^1,2　YE Yulong¹　WANG Genwei³　JING Haoran¹

1.School of Civil Engineering,Guilin University of Technology,Guilin 541004,China; 2.Guangxi Key Laboratory of Geotechnical Mechanics and Engineering,Guilin 541004,China; 3.Department of Natural Resources Engineering,Guangxi National Resource Vocational and Technical College,Nanning 532100,China

Online:2025-11-15 Published:2025-12-01

摘要/Abstract

摘要： 针对传统方法在预测滑坡位移时准确率不高、泛化性不足等问题,基于“分解—预测—重构”思想,提出了一种变分模态分解和深度学习相结合的滑坡位移预测模型。该模型通过灰色关联度分析确定滑坡位移的影响特征,采用粒子群优化算法(Particle Swarm Optimization,PSO)对变分模态分解(Variational Mode Decomposition,VMD)进行优化,将滑坡位移分解为具有不同物理意义的分量。针对分解后各分量的时序特点,分别采用多项式曲线拟合、长短时记忆网络(Long Short-Term Memory,LSTM)和时间卷积网络(Temporal Convolutional Networks,TCN)进行预测,最终将各分量预测值重构叠加,实现滑坡位移的精准预测。以三峡库区八字门滑坡数据为例,采用决定系数(R2)、平均绝对误差(MAE)、均方根误差(RMSE)等指标对模型进行量化评估。结果表明:所提出的滑坡位移预测模型精度达到 98.6%,能够有效提取滑坡位移数据中隐含的信息特征,对滑坡位移获取具有一定的借鉴意义;在各分量预测中采用不同的模型时,预测效果存在显著差异,因此,针对各分量特征的不同,建立相应的预测模型能够有效提高滑坡位移预测精度;通过参数敏感性分析,得出模型的输入序列长度为12时精度最佳。所提模型预测精度良好,可以为滑坡防灾减灾工程的实际应用提供参考。

关键词: 滑坡位移　变分模态分解　深度学习　粒子群优化　灰色关联度

Abstract: Addressing the issues of low accuracy and insufficient generalization in traditional landslide displacement pre diction methods,this paper proposes a novel model that integrates Variational Mode Decomposition (VMD) and deep learning based on the concept of "decomposition-prediction-reconstruction".The model identifies the influential features of landslide displacement through grey relational analysis and optimizes VMD using the Particle Swarm Optimization (PSO) algorithm, thereby decomposing landslide displacement into components with distinct physical meanings.To predict these components,the model employs polynomial curve fitting,Long Short-term Memory (LSTM) networks,and Temporal Convolutional Networks (TCN),depending on the time series characteristics of each component.The predictions of these components are then recon structed and aggregated to achieve an accurate prediction of landslide displacement.Using the data from the Bazimen landslide in Three Gorges Reservoir Region as an example,the model is quantitatively evaluated using metrics such as R2,MAE,and RMSE.The results demonstrate that the proposed landslide displacement prediction model achieves an accuracy of 98.6%,ef fectively extracting hidden information features from landslide displacement data,and has certain reference significance for get ting landslide displacement.The prediction outcomes vary significantly when different models are used for each component,in dicating that establishing specific prediction models based on the characteristics of each component can effectively improve the accuracy of landslide displacement predictions.Parameter sensitivity analysis reveals that the model achieves optimal accuracy when the input sequence length is 12.The model proposed in this article has good prediction accuracy and can provide a refer ence for the practical application of landslide disaster prevention and reduction projects.

Key words: landslide displacement,variational modal decomposition,deep learning,particle swarm optimization,grey cor relation

中图分类号:

张　研, 　叶玉龙　王根伟　荆浩然. 基于优化变分模态分解和深度学习的滑坡位移预测模型[J]. 金属矿山, 2025, 54(11): 138-145.

ZHANG Yan, 　YE Yulong　WANG Genwei　JING Haoran. Landslide Displacement Prediction Model Based on Optimal Variational Modal Decomposition and Deep Learning[J]. Metal Mine, 2025, 54(11): 138-145.

扫码分享