Deep Learning for Time Series Prediction of Strata Pressure in Coal Mining

Abstract

Hydraulic support plays a vital role in maintaining the structural integrity and safety of underground coal mines. We analyze a six-month dataset (May 1-October 31) of strata pressure from ten hydraulic supports (No. 65-74) in a 5966m $\times 280$ m longwall face, preprocessed into one-minute intervals, to predict strata pressure in underground coal mines, which is critical for ensuring safety and structural integrity. Using Pearson Correlation Coefficient (PCC), Fourier Transform (FT), and change point detection, we uncover strong intra-support correlations (PCC > 0.9), non-periodic patterns, and frequent abrupt shifts (3-5 events/hour). For short-term (one-minute) prediction, we propose a novel CNN-DLinear hybrid model that integrates DLinear’s interpretable trend-residual decomposition, tailored to strata pressure dynamics, with CNN’s localized spike detection for abrupt geological events. For long-term (30-minute) forecasting, we employ a smoothing technique to mitigate abrupt fluctuations and a sliding window approach to capture evolving trends. Experimental results show that our CNN-DLinear model achieves superior performance compared to ARIMA, LSTM, and Transformer models, with average reductions of 67% in MAE, 71% in MAPE, and 62% in RMSE, and an average $R^{2}$ of 0.96 across ten supports. Our approach excels in capturing non-periodic, noisy strata pressure dynamics with lower computational complexity ( $O(L)$ vs. $O(L^{2})$ for Transformers), enabling real-time safety monitoring. This work addresses the urgent need for accurate, efficient strata pressure forecasting in dynamic underground environments, thereby advancing operational safety and decision-making in coal mining.