@Article{sdhm.2026.086014,
AUTHOR = {Yimin Jin, Lei Zhang, Jun Tan, Dongyang Geng, Yanhui Zhang, Ripeng Cong},
TITLE = {Spatiotemporal Prediction Model for Surface Settlement in PBA Construction Using Adaptive Decoupled TCN-LSTM},
JOURNAL = {Structural Durability \& Health Monitoring},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/sdhm/online/detail/28221},
ISSN = {1930-2991},
ABSTRACT = {Pile-Beam-Arch (PBA) metro station construction involves complex, multi-stage sequential load redistributions and frequent stress transitions. Accurately predicting surface settlement under these conditions is critical for safety but remains challenging due to the scarcity of fence monitoring data and extreme non-linear soil deformations. This study aims to develop a robust spatiotemporal prediction framework that addresses data sparsity, stage-wise magnitude disparities, and complex geotechnical dependencies. First, targeted Gaussian noise, simulating actual physical instrument measurement errors, is injected exclusively into the training dataset to mitigate small-sample overfitting without data leakage. Second, a Stage-wise Adaptive Normalization (SAN) strategy is implemented to allocate independent feature mapping spaces for distinct construction steps. This aligns with physical stress release mechanisms, effectively preventing the numerical vanishing of early-stage minor deformation features. Third, a decoupled spatiotemporal architecture combining a One-Dimensional Temporal Convolutional Network (1D-TCN) and a Long Short-Term Memory (LSTM) network is employed. The 1D-TCN captures transverse spatial synergistic deformations, while the LSTM retains longitudinal historical stress paths. The framework was rigorously validated using an equidistant blind-zone partitioning method on actual field data. Results indicate that the proposed model effectively mitigates the over-smoothing and peak-shaving errors frequently encountered by standard parameter-heavy networks during abrupt stress transitions. Quantitatively, it accurately reconstructs the macroscopic settlement trough morphology in unobserved spatial zones, achieving a mean absolute error of 0.30 mm and a coefficient of determination (<i>R</i><sup>2</sup>) of 0.98, significantly outperforming benchmark algorithms. By explicitly decoupling spatial feature extraction from temporal memory and incorporating engineering mechanism-guided preprocessing, the proposed framework demonstrates superior numerical stability and generalization, providing a robust and physics-compatible early-warning approach for complex underground engineering projects.},
DOI = {10.32604/sdhm.2026.086014}
}