Adaptive spatial–temporal fusion transformer for multi-step geological condition prediction ahead of TBM

Predicting geological conditions ahead of the tunnel boring machine (TBM) face remains challenging due to limited survey resolution and inherent subsurface heterogeneity. This paper investigates how spatial–temporal dependencies among TBM operational parameters can be exploited for reliable multi-step, ring-level rock mass class forecasting up to 20 rings ahead of the excavation face. An Adaptive Spatial–Temporal Fusion Transformer (ASTFT) model is developed that integrates graph convolutional networks encoding multi-source inter-parameter relationships with a Transformer decoder through a context-sensitive gating mechanism. Validated on real-world tunneling data from the Bukit Timah Granite Zone, Singapore, the model achieves over 98% accuracy and weighted F1-score across all 20 prediction steps, consistently outperforming five baseline architectures. The results demonstrate that coupling graph-based spatial modeling with adaptive temporal fusion enables accurate, interpretable geological forecasting for proactive decision-making in TBM tunneling. Future research should pursue cross-project validation across diverse geological settings and real-time field deployment.