Composite Foundation Settlement Prediction Based on LSTM–Transformer Model for CFG

Abstract

Roadbed construction typically employs layered and staged filling, characterized by a periodic feature of ‘layered filling—filling interval’. The load and settlement histories established during staged construction offer crucial insights into long-term deformation under filling loads. However, models often rely solely on post-construction settlement data, neglecting the rich filling data. To accurately predict composite foundation ground (CFG) settlement, an LSTM–Transformer deep learning model is used. Five factors from the ‘fill height–time–foundation settlement’ curve are extracted as input variables. The first-layer LSTM model’s gate units capture long-term dependencies, while the second-layer Transformer model’s self-attention mechanism focuses on key features, efficiently and accurately predicting ground settlement. The model is trained and analyzed based on the newly constructed Changsha–Zhuzhou–Xiangtan intercity railway section CSLLXZQ-1, which has a CFG pile composite foundation. The research shows that the proposed LSTM–Transformer model for the settlement prediction of composite foundations has an average absolute error, mean absolute percentage error, and root mean square error of 0.224, 0.563%, and 0.274, respectively. Compared to SVM, LSTM, and Transformer neural network models, it demonstrates higher prediction accuracy, indicating better reliability and practicality. This can provide a new approach and method for the settlement prediction of newly constructed CFG composite foundations.