Application of robust deep learning models to predict mine water inflow: Implication for groundwater environment...

Area of Science:

• Hydrology
• Data Science
• Mining Engineering

Background:

• Traditional mine water inflow prediction faces challenges due to parameter uncertainty and complex mechanisms.
• Existing data-driven models struggle with nonlinearity and non-stationarity in water inflow processes.

Purpose of the Study:

• To propose robust DIFF-TCN and DIFF-LSTM models for accurate daily average mine water inflow prediction.
• To overcome the limitations of existing models in handling nonlinear and non-stationary data.

Main Methods:

• Developed two models: Difference Method (DIFF) combined with Temporal Convolutional Neural Network (TCN) and Long Short-Term Memory Neural Network (LSTM).
• Applied and validated the models on data from Tingnan Coal Mine, Shanxi Province, China.
• Utilized SHAP values to interpret the contribution of input features to predictions.

Main Results:

• The DIFF-TCN model outperformed other deep learning and traditional time series models, achieving MAE of 5.88 m³/h, RMSE of 6.85 m³/h, and R² of 0.96.
• Historical water inflow data were identified as the most significant input feature.
• Advance distance and groundwater level data also influenced predictions, though past groundwater levels sometimes had a negative impact.

Conclusions:

• The proposed DIFF-TCN model offers a robust and accurate solution for mine water inflow forecasting.
• Findings highlight the potential of advanced deep learning for smart hydrological forecasting in mining contexts.
• The study provides valuable technical guidance for ensuring mining safety and protecting surrounding water resources.