A deep learning-based combination method of spatio-temporal prediction for regional mining surface subsidence.

Area of Science:

• Geotechnical Engineering
• Mining Engineering
• Artificial Intelligence

Background:

• Surface subsidence from coal mining presents significant risks to infrastructure and safety.
• Current deep learning (DL) methods for subsidence prediction often struggle with spatial correlations and temporal nonlinearities.
• Accurate prediction of surface subsidence is crucial for effective risk management in mining operations.

Purpose of the Study:

• To develop a novel deep learning (DL) approach for enhanced prediction of surface subsidence caused by coal mining.
• To address the limitations of existing DL models in capturing spatial and temporal complexities in subsidence data.
• To improve the accuracy and reliability of surface subsidence forecasting.

Main Methods:

• Spatial data partitioning using K-means clustering.
• Application of a gate recurrent unit (GRU) model to analyze nonlinear time-series subsidence data within partitions.
• Global model accuracy enhancement through snake optimization (SO).

Main Results:

• The proposed DL method significantly outperforms traditional Long Short-Term Memory (LSTM) and GRU models.
• Achieved high accuracy, with 99.1% of sample pixels exhibiting less than 8 mm absolute error.
• Demonstrated superior capability in capturing spatial correlations and temporal nonlinearities in subsidence data.

Conclusions:

• The novel DL approach, combining K-means, GRU, and SO, offers a more accurate and reliable method for predicting coal mining surface subsidence.
• This technique effectively addresses the limitations of existing models by better handling spatial and temporal data characteristics.
• The findings provide a valuable tool for improving safety and mitigating risks associated with mining-induced surface subsidence.