Construction and verification of real-time prediction model for coal mine gas concentration based on deep learning

China’s coal industry is facing severe safety challenges due to the frequent occurrence of gas accidents. Traditional prediction methods such as autoregressive （AR） models and ARIMA models are difficult to meet the requirements of precise early warning due to problems such as low prediction accuracy and difficulty in handling nonlinear relationships. Based on deep learning technology, this study constructs a real-time prediction model for coal mine gas concentration that integrates the attention mechanism and graph convolution. By processing the monitoring data of Shendong Coal Mine of CHN Energy, adopting means such as data preprocessing, feature selection and standardization, combined with the spatio-temporal encoder-decoder architecture, efficient prediction of gas concentration has been achieved. The model takes the mean square error （MSE） as the loss function, is trained with the Adam optimizer, and its performance is verified through indicators such as MAE, MAPE, and RMSE. The results show that this model outperforms the benchmark model under different prediction Windows and has good real-time performance and reliability, providing a new method for the prediction and early warning of coal mine gas concentration.