A Mechanics-Informed Data-Driven Model for Predicting the Lateral Deformation of Retaining Piles in Deep Excavations: Integrating Spatial Graph Convolutional Networks With Transformer-Augmented Temporal Kolmogorov–Arnold Networks

Abstract

Accurate prediction of lateral deformation in retaining piles is critically important for ensuring safety in deep excavations. However, the deformation process of such piles involves complex spatiotemporal coupling effects and nonlinear dynamic characteristics, which are often inadequately captured by traditional methods, thereby limiting prediction accuracy. To address this gap, in this study, a GCN–Former–TKAN prediction framework with a hierarchical architecture is proposed for systematic spatiotemporal information integration. First, a recursive feature elimination (RFE) algorithm is adopted to screen the most deformation-sensitive influencing factors from multisource monitoring data, and parameters with clear physical driving significance are retained. Subsequently, graph convolutional networks (GCNs) are used to explicitly model the spatial topology of monitoring points, whose Spearman correlation-based graph structure implicitly reflects the propagation pattern of excavation disturbance. Next, a Transformer is introduced to conduct deep temporal modeling of spatial feature sequences output by the GCN, thereby capturing long-term dependencies and adaptively weighting features at different time steps via a self-attention mechanism. Finally, the temporal Kolmogorov–Arnold network (TKAN) serves as a nonlinear mapping module that fuses Transformer-enhanced spatiotemporal features with multisource time series to produce predictions. On the basis of a real-world engineering case, a time-window sensitivity analysis reveals that a 7-day historical data window provides the most stable overall performance in this case study. The results of ablation and comparative experiments indicate that the proposed model achieves higher prediction accuracy and more stable performance than the selected baseline models do. This study provides a practical data-driven reference framework for the deformation prediction and health monitoring of deep excavation retaining structures.