Robust Ground Settlement Forecasting with Spatiotemporal Transformers and Geotechnical Priors

Abstract

The rapid expansion of urban underground infrastructure, particularly metro systems and utility tunnels, necessitates precise monitoring and forecasting of ground settlement to mitigate risks to existing surface structures. Traditional empirical methods and finite element analyses often struggle to balance computational efficiency with the complex, non-linear dynamics of soil-structure interactions in heterogeneous geological environments. While deep learning has emerged as a viable alternative, purely data-driven models frequently violate physical laws and generalize poorly in data-sparse regimes. This paper presents a novel framework, the Geotechnical Spatiotemporal Transformer (Geo-STT), which integrates geotechnical priors directly into the attention mechanism of a transformer architecture. By embedding static soil parameters—specifically Atterberg limits, void ratios, and shear strength—alongside dynamic time-series data from sensor arrays, the model learns a physics-aware representation of ground deformation. We introduce a novel governing equation-based loss function that penalizes predictions diverging from established settlement trough profiles. Extensive experiments on real-world shield tunneling datasets demonstrate that Geo-STT significantly outperforms state-of-the-art baselines in long-term forecasting accuracy and robustness against sensor noise.