DCSGN: A knowledge-data driven method for predicting fine-grained deformation of newly constructed tunnels

Accurate prediction of deformation in newly constructed tunnels is crucial for early risk detection and ensuring structural stability. However, due to the limited sample size and high variability of data in these tunnels, it becomes difficult to achieve both high prediction accuracy and strong generalization performance when relying solely on data-driven approaches. To address these challenges, a knowledge-data driven method is proposed to predict tunnel deformation. Specifically, a distance-correlation coupled symmetric graph network (DCSGN) framework is presented, which incorporates prior knowledge of node positions and symmetry constraints of the tunnel structure. A data-driven model based on a Chebyshev graph convolutional network with a gated recurrent unit (ChebNet-GRU) is employed to predict deformation behavior across the entire tunnel section. A typical newly constructed tunnel located in Eastern China was selected as a case study. Sensitivity analysis and component comparisons demonstrated the DCSGN model’s high prediction accuracy and robustness, with optimal values of RMSE at 0.11 mm, MAE at 0.087 mm, and R² at 0.73. Furthermore, the model’s predictive accuracy was validated by comparison against state-of-the-art (SOTA) hybrid neural network baselines. This proposed model effectively forecasts fine-grained tunnel deformation, offering valuable decision support for deformation prediction in newly constructed tunnels.