Abstract: Gas hazards restrict the development level of coal mine safety production. Gas control is an important part of high gas coal mine mining. Effective prediction of gas concentration in the next period and making reasonable safety protection measures can provide certain reference basis for decision-making of coal mine gas control. This paper builds a prediction model of gas concentration in coal mine working face based on recurrent neural network (RNN) to process continuous time series samples. The model preliminarily determines the network structure parameters of the prediction model based on the broad strategy, and selects the gas concentration time series with larger data volume and longer time span as the training samples. Firstly, the outliers and missing values in training samples are processed by proximity mean method and interpolation method. At the same time, MinMaxScala method is used to normalize the data. Secondly, MSE and running time are used as evaluation indexes, Adam optimizer is used to optimize the weight of the model, Relu is selected as activation function, and Dropout layer is added to the hidden layer. By adjusting the parameters of batch size and network layers, the optimal RNN gas prediction model is finally obtained. The results show that compared with BP neural network prediction model and Bidirection RNN prediction model, the training error of gas concentration prediction model based on RNN is reduced to 0.003, the prediction error is reduced to 0.006, and the prediction error fluctuation range is 0.001～0.024. The results indicate that the RNN gas prediction model has better accuracy, stability and robustness. The model can effectively predict the trend of gas concentration in the next period. Therefore, reasonable protection measures can be taken in advance to ensure coal mine safety production.