铁道科学与工程学报2025，Vol.22Issue(8)：3352-3363,12.DOI:10.19713/j.cnki.43-1423/u.T20241709
顾及导热系数与冻土环境变量的高铁路基冻深预测LSTM模型构建及应用
Construction and application of an LSTM model for predicting the frost depth of high-speed railway subgrades considering thermal conductivity and frozen soil environmental variables
摘要
Abstract
Improving the prediction accuracy of the freezing depth of high-speed railway subgrades in seasonal frozen soil areas is of great significance for ensuring the safe dispatching and smooth operation of high-speed railways in cold regions.To address the lack of utilization of multivariate environmental sequence information in existing freezing depth prediction models for high-speed railway subgrades in seasonal frozen areas,an LSTM model considering thermal conductivity and frozen soil environmental variables was proposed.By taking the three sections DK371+900,DK383+345,and DK391+940 of the Lanxin High-Speed Railway in the Shandanmachang-Minle section as examples,the freezing depth of the subgrade during the rapid growth period from 2015 to 2017 was predicted.The model first used the EMD algorithm to decompose the time series data of thermal conductivity and frozen soil environmental variables,producing a series of data sequences with different characteristic scales that reflect the trends and volatility of the original data and thereby enriching the detail and diversity.It then employed the KPCA algorithm to extract the key factors affecting the subgrade freezing depth,achieving dimensionality reduction and eliminating the data redundancy caused by EMD.Finally,the LSTM network was used to realize multivariable-based subgrade freezing depth prediction.Experimental results show that this model outperforms traditional embankment freezing depth prediction models,EMD-LSTM models,multivariate BP neural network models,and multivariate LSTM models in terms of accuracy.The Mean Absolute Error(fmae)at the three cross-sections is 0.029 m,0.033 m,and 0.060 m,respectively.The Root Mean Square Error(frmse)is 0.036 m,0.042 m,and 0.075 m,respectively.The coefficient of determination(R2)is 0.924,0.949,and 0.906,respectively.The fmae and frmse are lower than those of traditional subgrade freezing depth prediction models by up to 89.1%and 86.8%,respectively;by up to 87.7%and 85.1%as compared with the EMD-LSTM model,respectively;by up to 66.3%and 64.1%as compared with the multivariable BP neural network model,respectively;and by up to 60.2%and 56.7%as compared with the multivariable LSTM model,respectively.The research results can provide a new reference for predicting the freezing depth of high-speed railway subgrades in seasonal frozen soil areas.关键词
高铁路基/经验模态分解/核主成分分析/长短期记忆神经网络/冻结深度预测Key words
high-speed railway subgrade/empirical mode decomposition/kernel principal component analysis/long short-term memory/frost heave depth prediction分类
建筑与水利引用本文复制引用
张超越,魏冠军..顾及导热系数与冻土环境变量的高铁路基冻深预测LSTM模型构建及应用[J].铁道科学与工程学报,2025,22(8):3352-3363,12.基金项目
国家自然科学基金资助项目(42364003,41964008) （42364003,41964008）
