Abstract
Abstract Accurate prediction of the volume of coal rock gas is a key prerequisite for the low carbon transformation of the coal industry. Based on this foundation, strengthening the extraction and utilization of gas, and through the substitution of methane for fossil fuels, coupling CO₂ displacement for increased production and geological storage technologies, it is possible to jointly achieve methane emission control, resource efficiency improvement, and carbon sequestration, precisely meeting the industry’s demands for low carbon development. This study takes the 21,605 working face of Guizhou Qinglong Coal Mine as the research object. Employ COMSOL software for multi-physics field simulations and integrate machine learning models to predict gas extraction volumes, further refining the prediction accuracy through optimization algorithms. The results demonstrate that the combination of numerical simulation and machine learning methods markedly boosts the precision and reliability of gas extraction volume predictions. Evaluation using the Entropy Weight Method shows that the eXtreme Gradient Boosting (XGBoost)- Long Short-Term Memory (LSTM) hybrid model has the highest weight. Compared to the base XGBoost model, MAE, MAPE, MSE, and RMSE decreased by 72.64%, 46.89%, 95.77%, and 79.44%, respectively. Compared to the base LSTM model, these metrics dropped by 79.53%, 33.39%, 98.65%, and 88.37%, respectively. The R² value improved to 0.9998. This study identifies that gas extraction volumes are influenced by multiple factors, with coal seam permeability and temperature being particularly significant. This research not only provides a scientific basis for the low carbon production technology strategy of coal rocks, but also offers new approaches and tools for accurately predicting the volume of gas extraction from coal rocks.