Abstract: In order to make early warning of gas overlimit and coal and gas outburst accidents, the model of data preprocessing and gas concentration prediction were established, and an early warning system with three-level of identification criteria and procedures was developed. Through the training of gas concentration data based on LSTM algorithm, the gradient disappearance and gradient explosion in deep learning were solved. The temporal and spatial correlation of gas concentration was reflected through the analysis of gas characteristics, and the optimal forecasting method was obtained by studying the relationship among sample duration, forecasting duration and forecasting accuracy. The results show that when the time of sub-sample is 1.0 h, the prediction accuracy of 5 min in advance is the highest, which meets the evaluation standard of 1%. When the time of sub-sample is 1.5 h, the prediction accuracy of 10 min and 15 min in advance is the highest, which meets the evaluation criteria. When the time of sub-sample is 2.0 h, it takes 20 min in advance to meet the standard. The prediction and analysis of gas concentration and the prewarning system can reflect the variation law of gas concentration in the process of coal drift driving, and carry out gas concentration prediction and real-time anti-outburst prewarning, providing technical support for gas prevention and control safely and efficiently.