Scholarlink

<p>The post-construction settlement prediction of high-speed railway subgrades in loess areas was improved by developing a multifactor bi-directional long- and short-term memory neural network optimized by attention mechanism (MBiLSTM-AM) model. Environmental factors and geotechnical properties were incorporated to address the limitations of traditional single-factor models, such as the hyperbolic, exponential curve, Asaoka, and grey theory models, which primarily rely on time-dependent data while neglecting temperature, precipitation, and load effect. Model performance was evaluated using on-site measurements and the rolling cross-validation method, comparing the MBiLSTM-AM model against the hyperbolic, multifactor bi-directional long- and short-term memory neural network (MBiLSTM), and a multifactor long- and short-term memory neural network (MLSTM). Results showed that the MBiLSTM-AM model achieved the highest accuracy, with mean coefficients of determination (<i>R</i><sup>2</sup>) improvements of 7.90%, 4.98%, and 3.25% over the hyperbolic, MBiLSTM, and MLSTM models, respectively. Combining the MBiLSTM-AM model with the hyperbolic model further enhanced prediction accuracy for longer time series. The study demonstrated that integrating multiple influencing factors with an attention mechanism significantly improved settlement prediction in loess subgrades.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Multifactor Settlement Prediction Method for the Loess Subgrade of a High-Speed Railway Based on Deep Learning

  • 期刊论文

  • 发表日期：

    2025年8月5日

查看全文
  • Fei Gao,
  • Wenfang Wang,
  • Xuansheng Cheng,
  • Qingfeng Lv,
  • Qingdong Wang

摘要

The post-construction settlement prediction of high-speed railway subgrades in loess areas was improved by developing a multifactor bi-directional long- and short-term memory neural network optimized by attention mechanism (MBiLSTM-AM) model. Environmental factors and geotechnical properties were incorporated to address the limitations of traditional single-factor models, such as the hyperbolic, exponential curve, Asaoka, and grey theory models, which primarily rely on time-dependent data while neglecting temperature, precipitation, and load effect. Model performance was evaluated using on-site measurements and the rolling cross-validation method, comparing the MBiLSTM-AM model against the hyperbolic, multifactor bi-directional long- and short-term memory neural network (MBiLSTM), and a multifactor long- and short-term memory neural network (MLSTM). Results showed that the MBiLSTM-AM model achieved the highest accuracy, with mean coefficients of determination (R2) improvements of 7.90%, 4.98%, and 3.25% over the hyperbolic, MBiLSTM, and MLSTM models, respectively. Combining the MBiLSTM-AM model with the hyperbolic model further enhanced prediction accuracy for longer time series. The study demonstrated that integrating multiple influencing factors with an attention mechanism significantly improved settlement prediction in loess subgrades.