A novel transformer-based approach for predicting frost heave of high-speed railway in cold regions

In cold regions, the frost-heave of soil can cause uneven railway subgrades, affecting the safety and efficiency of high-speed railways. This study proposes a novel PCW-iTransformer model for predicting frost heave, which integrates PCHIP data interpolation, CEEMDAN signal decomposition, and WPT denoising to extract sequential features. Compared to existing models like Autoformer, Crossformer, and DLinear, PCW-iTransformer achieves a reduction of 19.1% to 34.5% in error metrics and an improvement of 2.8% to 4.6% in the coefficient of determination. Additionally, a fused parameter model based on normalized moisture and temperature improves prediction accuracy, reducing MSE, MAE, and RMSE by up to 7.6%. The model also demonstrates robustness under data scarcity, maintaining stable performance with 40% continuous or 60% random missing data. Overall, PCW-iTransformer provides a reliable approach for predicting frost heave, offering valuable insights for the maintaining and long-term stability of high-speed railway subgrades in cold regions. • A new deep learning method for predicting the future frost heave of subgrade is proposed. • A model fusing moisture and temperature input parameters is introduced to improve prediction accuracy. • The effect of missing data is investigated in time series forecasting.