Forecasting adjacent-construction-induced radial displacement of MRT shield tunnels using dynamic deep learning

Rapid urbanization is driving the expansion of mass transit systems in densely populated urban areas worldwide. The weak geological conditions of the Songshan Formation, which underlies the Taipei Basin in northern Taiwan, elevate the risk of shield tunnel deformation from nearby construction, as illustrated by a recent incident near Taipei Dome that caused excessive displacement in the MRT Bannan Line. This study proposes the Neural Network-Revolving Gate Fourier Transformation (NN-RGFT) as a novel hybrid deep learning (DL) model for predicting adjacent-construction-induced radial displacement in MRT shield tunnels. The NN-RGFT integrates a non-sequential Neural Network (NN) and Revolving Gate Fourier Transform (RGFT) to capture periodic and oscillatory deformation patterns in sequential monitoring data using iterative frequency-domain transformations and cross-attention. The proposed model was developed and validated on a database covering six monitoring points and was benchmarked against conventional machine learning (ML), hybrid ML, and DL models. The NN-RGFT consistently achieved the lowest error and highest correlation metrics of all compared models. Applied to an independent real-world case in New Taipei City, the model achieved MAPE values below 6.5% at all six observation points, confirming its reliability. The proposed model offers a practical, proactive tool for tunnel deformation monitoring and construction risk management.