Time Series Prediction of Mine Pressure Using a Dual-Layer GRU with Multi-Feature Enhancement

To address the complex temporal series fluctuations, strong nonlinearity, and limited accuracy of traditional prediction methods in fully mechanized mining face pressure monitoring, this study proposes a dual-layer GRU mine pressure prediction model enhanced by multi-feature fusion. First, the original mine pressure monitoring data undergo fixed-time-interval resampling. Subsequently, Lag features, sliding statistical features, and differential features are constructed to improve the model’s ability to capture historical temporal dependencies and local fluctuation patterns. The dual-layer GRU network is then employed for pressure prediction, with Bayesian optimization used to optimize the model hyperparameters. Experiments were conducted using hydraulic support resistance data from support No. 30, comparing the single-layer GRU, LSTM, and XGBoost models. The results demonstrate superior performance across the MAE, RMSE, and R2 metrics, achieving an RMSE of 0.73 and an R2 of 0.91. Feature ablation experiments confirm that the multi-feature enhancement approach significantly improves prediction accuracy. Cross-support generalization experiments show that the model maintains robust predictive performance across hydraulic support resistance data from the upper, middle, and lower sections of the working face, effectively capturing pressure variation trends and demonstrating strong cross-support generalization capability.