A data-driven spatial-temporal model for prediction of tunnel deformation

Predicting deformation trends in tunnels is crucial for structural damage diagnosis and accident prevention. Due to the deficiency of incomplete influencing factors and rough prediction accuracy, this paper proposes a data-driven spatial-temporal model to predict tunnel deformation behavior. Considering multiple factors including stress, strain, temperature, and deformation, this study combines the Graph Attention Networks with Gated Recurrent Unit (GAT-GRU) model to analyze the spatial and temporal deformation patterns at different monitoring locations. The main innovation of this study lies in the coupling of graph-based attention mechanisms with recurrent neural networks, which allows the model to learn spatial relationships between tunnel nodes based on their spatial proximity, while also capturing the temporal deformation patterns of the tunnel over time. Firstly, the proposed model was discussed to demonstrate its superiority over other spatial and temporal models. The results show that the proposed model significantly outperforms other algorithms, including GRU, TCN, ChebNet, GCN, GAT, and GAT-TCN, in terms of prediction accuracy. The prediction accuracy at different nodes was then compared. It was found that the fitting effect is best at the crown, while the predictive performance at the side is unsatisfactory. Finally, the robustness of the model was analyzed from different time steps and prediction scales. The experiments showed a strong correlation between data from the past 6−7 h and tunnel deformation. The model can effectively predict deformation behavior within the next 5 h.