A Deep Learning Method Fusing Multiscale Temporal Features for Metro Tunnel Deformation Advanced Prediction

During the excavation of foundation pits, accurately predicting the deformation of existing metro tunnel structures plays a crucial role in ensuring the safe operation of the metro system. In this paper, a novel deep learning method for predicting tunnel structural deformation is proposed, namely the time series prediction network based on multiscale temporal feature fusion. This method incorporates the concept of multiscale temporal feature fusion. It extracts time series features at different scales from the input sequence using Conv1d, then reconstructs and fuses these features into new time series data. Subsequently, it integrates long short-term memory layers and fully connected layers to predict future tunnel structural deformation. The proposed deep learning method was validated using deformation monitoring data from the Beijing Metro Line 6 tunnel structure, and a series of ablation experiments were conducted to analyze the impact of multiscale temporal feature fusion on model performance. The experimental results show that the multiscale temporal feature fusion module proposed can effectively improve the prediction accuracy of the model. The R2 of the single-step predictions for different monitoring points can reach up to 0.99, while the multistep prediction accuracy can reach 0.85. It is capable of accurately predicting the development of metro tunnel structural deformation during foundation pit excavation.