基于CNN-BiLSTM-Attention的三峡库区滑坡地表位移预测研究  

Research on predicting surface displacement of landslides based on CNNBiLSTM-Attention in the Three Gorges reservoir area

作　　者：陈欢[1,2] 冯晓亮 刘一民[3] 赵晗 刘洋 郭浪 张军 CHEN Huan;FENG Xiaoliang;LIU Yimin;ZHAO Han;LIU Yang;GUO Lang;ZHANG Jun(Institute of Exploration Technology,CGS,Chengdu 611734,China;Technology Innovation Center for Risk Prevention and Mitigation of Geohazard,Ministry of Natural Resources,Chengdu 611734,China;School of Intelligent Manufacturing,Chengdu Technological University,Chengdu 611730,China;Geological Environment Monitoring Station of Yunyang County,Chongqing 404500,China;Chongqing 107 Municipal Construction Engineering Co.,Ltd.,Chongqing 401120,China)

机构地区：[1]中国地质科学院探矿工艺研究所,四川成都611734 [2]自然资源部地质灾害风险防控工程技术创新中心,四川成都611734 [3]成都工业学院智能制造学院,四川成都611730 [4]云阳县地质环境监测站,重庆404500 [5]重庆一零七市政建设工程有限公司,重庆401120

出　　处：《沉积与特提斯地质》2024年第3期572-581,共10页Sedimentary Geology and Tethyan Geology

基　　金：国家自然科学青年基金资助项目“断层面库仑应力变化监测方法的力学机理实验研究”(41804089);中国地质调查局项目“地质灾害监测预警与防治支撑(探矿工艺所)”(DD20230447)。

摘　　要：地表位移预测在滑坡监测预警中具有重要意义,建立稳定可靠的滑坡位移预测模型是关键。本文基于卷积神经网络和注意力机制的滑坡位移预测方法,并以三峡库区黄泥巴蹬坎滑坡为例进行了验证。本文综合分析了该滑坡长达8年的降雨量、库水位和地表位移等监测数据,建立了结合卷积神经网络(convolutional neural network,CNN)、双向长短期记忆(bidirectional long short-term memory,BiLSTM)网络和注意力机制(attention)的CNN-BiLSTM-Attention深度学习组合预测模型,采用了适应性学习率和正则化技术进行模型训练,提高了模型的泛化能力同时避免过拟合,并与传统LSTM模型进行对比验证。结果表明:相较于传统的机器学习和神经网络方法,该模型在滑坡位移预测精度上取得了显著提升,预测模型拟合优度(R^(2))达0.989,平均绝对百分比误差(MAPE)仅为0.059。Surface displacement prediction is of great significance in landslide monitoring and early warning,and establishing a stable and reliable landslide displacement prediction model is crucial.This paper utilizes a convolutional neural network(CNN)and attention mechanism to predict landslide displacement,and takes the Huangniba Dengkan landslide in the Three Gorges reservoir area as an example for verification.This paper comprehensively analyzes the landslide's monitoring data on rainfall,reservoir water level,and surface displacement for 8 years.It establishes a CNN-BiLSTM-Attention deep learning combination prediction model,and uses adaptive learning rate and regularization techniques for model training,improving the generalization ability of the model while avoiding overfitting.Additionally,the model is subjected to comparative validation with the traditional long short-term memory(LSTM)model.The results show that the model's landslide displacement prediction accuracy has been significantly enhanced compared to traditional machine learning and neural network methods.The prediction model's goodness of fit(R^(2))reaches 0.989,and the mean absolute percentage error(MAPE)is merely 0.059.

关 键 词：滑坡监测 地表位移 注意力机制 预测模型 三峡库区 

分 类 号：P694[天文地球—地质学]