基于关键变量和SVM交叉验证改进的LSTM大坝变形预测  

Improved LSTM Dam Deformation Prediction Based on Key Variables and SVM Cross Validation

作　　者：杨振亚 王正新 高剑峰 YANG Zhenya;WANG Zhengxin;GAO Jianfeng(Nanjing Water Conservancy Planning and Design Institute Co.Ltd.,Nanjing 210000,China)

机构地区：[1]南京市水利规划设计院股份有限公司,南京210000

出　　处：《河南科学》2024年第9期1307-1314,共8页Henan Science

基　　金：2021年江苏省水利科技项目(2021068);2022年江苏省水利科技项目(2022011)。

摘　　要：为了提升大坝变形预测精度,更好地对未来大坝运行状态进行评估,提出了基于显著变量和自相关分析优化长短期记忆神经网络(LSTM)的大坝变形多步预测方法.该方法首先基于随机森林(RF)确定对变形影响程度高的关键变量,同时基于SVM交叉验证自相关分析提取显著的历史特征,在提升LSTM信息挖掘能力的同时提高预测精度,最后将确定的影响因子与历史数据输入LSTM模型中,通过全连接网络输出变形预测结果.经分析,所提方法大坝变形预测精度高,与LSTM、RF-LSTM、SVM-LSTM方法相比,该方法预测精度最大提升幅度分别为62.15%(MAE)和60.12%(RMSE).In order to improve the accuracy of dam deformation prediction and better evaluate the future operation status of dams,this paper proposes a multi-step prediction method for dam deformation based on significant variables and autocorrelation analysis optimized long short-term memory neural network(LSTM).This method first determines the key variables that have a high impact on deformation based on Random Forest(RF),and extracts significant historical features based on SVM cross validation autocorrelation analysis to improve LSTM information mining ability and prediction accuracy.Finally,the determined influencing factors and historical data are input into the LSTM model,and the deformation prediction results are output through a fully connected network.It is analyzed that the proposed method has high deformation prediction accuracy,and the maximum improvement of the prediction accuracy of this method is 62.15%(MAE)and 60.12%(RMSE)respectively,compared with the LSTM,RF-LSTM,and SVM-LSTM methods.

关 键 词：大坝变形预测 随机森林 自相关分析 交叉验证 LSTM 

分 类 号：TV698.1[水利工程—水利水电工程]