Abstract
Mining-induced subsidence is a complex deformation process influenced by both human activities and natural factors, making accurate prediction essential. Conventional methods face challenges in spatial modeling, deformation interpretation, and computational efficiency. Although deep learning (DL) offers strong nonlinear modeling capabilities and avoids complex physical parameter constraints, its application to mining subsidence prediction remains limited. This study introduces a new prediction framework, Earthformer_MSTSP, and a Gaussian-based dynamic simulation method (MSDS_G) for generating training data. The synthesized dataset is trained using a Cuboid Attention.based hierarchical codec architecture. Using SBASInSAR displacement time series from the Datong mine as input, Earthformer_MSTSP achieves an MSE of 0.45 m and a 3.7% relative error in maximum subsidence, improving prediction accuracy by 74% over the Weibull model and 82% over LSTM. A transferability test in the Changwu mine shows that the model trained with MSDS_G effectively preserves the spatial deformation patterns, as verified by Moran.s I. Additionally, subsidence-rate-based early warning analysis and the impact of mining speed are explored, providing multiple indicators for hazard mitigation.
Funding Information
  • Shaanxi Province Geoscience Big Data and Geohazard Prevention Innovation Team
  • Generic Technical Development Platform of Shaanxi Province for Imaging Geodesy (2024ZG-GXPT-07)
  • ESA-MOST China DRAGON-6 project (95355)
  • Fundamental Research Funds for the Central Universities (300102261108)
  • Fundamental Research Funds for the Central Universities (300102261308)
  • Fundamental Research Funds for the Central Universities (300102264302)

This publication has 22 references indexed in Scilit: