Hydraulic support pressure prediction via deep learning with multilevel temporal feature integration.

Area of Science:

• Mining Engineering
• Artificial Intelligence
• Time Series Analysis

Background:

• Accurate hydraulic support pressure prediction is vital for coal mine safety, especially with increasing mining depth and complexity.
• Existing methods face challenges in precisely predicting pressure under dynamic and complex underground conditions.

Purpose of the Study:

• To develop an advanced prediction method for hydraulic support pressure by fusing multi-dimensional features.
• To enhance the accuracy and reliability of pressure prediction in challenging mining environments.

Main Methods:

• A Long Short-Term Memory-Patch Temporal Fusion Transformer (LSTM-PatchTST) model was developed, integrating Pearson correlation analysis and Gaussian moving average filtering.
• The LSTM network captures temporal dynamics, while the PatchTST module models local and global dependencies through self-attention mechanisms.
• The model preprocesses time-series data, extracts dynamic features, and fuses them for deep learning-based prediction.

Main Results:

• The proposed LSTM-PatchTST model demonstrated superior performance compared to pure PatchTST and Transformer+LSTM models.
• Significant reductions in Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) were observed, with improvements up to 58.7%.
• The model showed strong generalization ability when applied to data from different coal mines.

Conclusions:

• The developed LSTM-PatchTST method offers a robust and accurate approach for hydraulic support pressure prediction.
• This advancement contributes to improved safety measures and operational efficiency in deep and complex coal mining operations.
• The fusion of LSTM and PatchTST effectively addresses the challenges of predicting dynamic pressure changes in underground environments.