Fei Ma ^1,2* Qingbin Zhang ³ Lichun Sui ²

• ¹ Changzhi University, Changzhi, China
• ² Chang’an University, Xi'an, Shaanxi, China
• ³ Zhongjin Environmental Technology Co., LTD,, Taiyuan, China

In China, the Loess Plateau's fragile geological structure leads to complex and variable surface subsidence in old gob areas following coal mining activities. Accurately predicting this residual subsidence remains a significant scientific challenge. In this study, a method for residual subsidence prediction using an Exponential Smoothing Long Short-Term Memory (EsLSTM) model is proposed. The investigation centers on the 18001# old goaf area of the Yangquan Coal Mine in Shanxi Province. Using Sentinel-1A imagery, continuous SAR data from 98 periods were acquired and processed via Enhanced Distributed Scatter InSAR technology. The EsLSTM model was then developed to capture the subsidence time-series characteristics of all surface scatter points and predict future ground subsidence. The analysis reveals that the EsLSTM model delivered excellent accuracy, achieving an R 2 value of 0.975. It also outperformed SVR and traditional LSTM models, with a Mean Absolute Error of 2.2 mm and a Root Mean Square Error of 7.9 mm. Predicted results indicate that by October 2023, the maximum cumulative subsidence at the 18001# working face of the Yangquan Coal Mine will reach 204 mm. The subsidence trend is expected to become more gradual and stable, suggesting a low likelihood of geological disasters in the area.