基于MIC-XGBoost-LSTM模型的边坡位移预测研究  

Slope Displacement Prediction Using MIC-XGBoost-LSTM Model

作　　者：许江波[1] 侯鑫敏 吴雄 刘一凡 孙国政 XU Jiang-bo;HOU Xin-min;WU Xiong;LIU Yi-fan;SUN Guo-zheng(School of Highway,Chang'an University,Xi'an 710064,Shaanxi,China;China Railway First Survey and Design Institute Group Co.Ltd.,Xi'an 710043,Shaanxi,China;Gansu Provincial Transportation Research Institute Group Co.Ltd.,Lanzhou 730030,Gansu,China)

机构地区：[1]长安大学公路学院,陕西西安710064 [2]中铁第一勘察设计院集团有限公司,陕西西安710043 [3]甘肃省交通科学研究院集团有限公司,甘肃兰州730030

出　　处：《中国公路学报》2024年第10期38-48,共11页China Journal of Highway and Transport

基　　金：国家自然科学基金项目(52178310);陕西省交通厅交通科研项目(22-38K&23-39R);陕西省重点研发项目(2024GX-YBXM-372&2024QCY-KXJ-176)。

摘　　要：为实现对边坡位移精确的预测,建立了一种基于最大互信息系数与XGBoost算法的长短期记忆神经网络(MIC-XGBoost-LSTM)边坡位移预测模型。首先,对边坡受到不同降雨条件的影响进行研究,利用最大互信息系数分析不同降雨条件因素与边坡累计位移之间的相关关系,确定相关性显著的降雨影响因素;其次,基于XGBoost算法对与边坡累计位移数据相关性较高的影响因素进行特征构造,将构造特征与原有特征一起进行归一化处理,把归一化后的数据分为训练集与验证集;然后,利用LSTM对G312国道商洛岩质边坡位移进行预测。同时,分别利用XGBoost、LSTM、MIC-XGBoost-LSTM预测模型对边坡的累计位移值进行训练预测,并通过RMSE、MAE、MAPE指标进行预测精度评价,并采用RMSE指标对MIC-XGBoost-LSTM模型最长预测周期及最小训练样本数量进行确定。最后,采用白水河滑坡位移数据对模型进行进一步验证。研究结果表明:日位移增量、蒸散量、净降雨量、累计7 d降雨量与监测点累计位移量的相关性较其他因素更高,且利用4种相关因素构造的特征值与输出的特征值的MIC值为0.97;同时,利用MIC-XGBoost-LSTM模型对G312国道商洛边坡进行预测的结果中均方根误差(Root Mean Squared Error,RMSE)、平均绝对误差(Mean Absolute Error,MAE)、平均绝对百分比误差(Mean Absolute Percentage Error,MAPE)分别为0.25%、0.185%、0.024%,均小于XGBoost、LSTM的RMSE、MAE、MAPE值,并根据RMSE指标获取MIC-XGBoost-LSTM模型最长预测周期与最小训练样本数量分别为56、675;最终采用白水河滑坡位移数据进行验证,其评价指标均小于XGBoost、LSTM模型,这表明MIC-XGBoost-LSTM边坡预测模型具有较高的可靠性。A long short-term memory(LSTM)neural network model for predicting slope displacements based on maximum mutual information coefficients(MICs)and the XGBoost algorithm(MIC-XGBoost LSTM)was established to accurately predict slope displacements.First,the effects of different rainfall conditions on the slope were investigated.The maximum MIC was used to analyze the correlation between different rainfall conditions and the cumulative displacement of the slope,and the rainfall-influencing factors with significant correlations were determined.Next,based on the XGBoost algorithm,feature construction was performed on the influencing factors with high correlation using the cumulative displacement data of the slope,and the construction features were normalized with the original features.The normalized data were divided into training and validation sets.LSTM was used to predict the displacement of the Shangluo rock slope on the G312 National Highway.The XGBoost,LSTM,and MIC-XGBoost-LSTM prediction models were used to train and predict the cumulative displacement value of the slope,and the prediction accuracy was evaluated based on the root-mean-square error(RMSE),mean absolute error(MAE),and mean absolute percentage error(MAPE)indicators.In addition,the RMSE was used to determine the longest prediction cycle and minimum training sample size for the MIC-XGBoost LSTM model.Finally,the displacement data of the Baishui River landslide were used to further validate the model.The results show that the correlations between daily displacement increment,evapotranspiration,net rainfall,cumulative seven-day rainfall,and cumulative displacement at the monitoring point are higher than those of other factors,and the MIC of the feature values constructed using four related factors and the output feature values is 0.97.The RMSE,MAE,and(MAPE)of the predicted results obtained using the MIC-XGBoost-LSTM model are 0.25%,0.185%,and 0.024%,respectively,which are lower than those of XGBoost and LSTM.Based on the RMSE,the longest prediction cycle an

关 键 词：路基工程 预测模型 MIC 岩质边坡 XGBoost模型 LSTM模型 特征构造 

分 类 号：U417.1[交通运输工程—道路与铁道工程]