Abstract: In the complex environment of coal mines, the prediction precision of traditional gas concentration prediction models is relatively low. Although the traditional gas concentration prediction model is optimized by introducing various optimization algorithms to improve the gas concentration prediction accuracy. But modeling only from the time dimension ignores the spatial features of gas concentration. This can easily lead to the loss of important prior knowledge and affect the prediction effect. In order to solve the above problems, a gas concentration spatiotemporal distribution prediction model based on graph convolutional networks (GCN) and gated recurrent unit (GRU) is proposed. Firstly, the historical data of gas concentration is preprocessed. A gas concentration spatial node graph is constructed based on the spatial distance between each collection node. The graph is used to model the complex dependency relationships between nodes. Secondly, at each sampling time point, the gas concentration and distance weight parameters between nodes are used as inputs to obtain the spatial node graph structure of gas. After that, GCN is used for spatial feature adaptive learning and graph convolution operation to obtain the spatial features of gas concentration. Then, the spatial feature information of gas concentration is transformed into sequence data and input to GRU. Finally, GRU processes the gas spatial feature information composed of each time under the time series. Through sequence-to-sequence based models and autoencoders, GRU generates model prediction results. The experimental results show the following points. ① The GCN-GRU model can accurately predict the overall trend of gas concentration changes. The fit between the predicted results and actual data is better than the historical average (HA) model and support vector regression (SVR) model. ② The root mean square error of GCN-GRU model is reduced by 0.5%, 71.4% and 37.9% respectively compared with HA model, SVR model and autoregressive integrated moving average model (ARIMA) model. The average absolute error of GCN-GRU model is reduced by 10.5%, 82.4% and 82.4% respectively compared with HA model, SVR model and ARIMA model. The accuracy of GCN-GRU model is improved by 0.06%, 17.7% and 13.8% respectively compared with HA model, SVR model and ARIMA model. The results indicate that GCN-GRU model has strong robustness, and the generalization performance is good. ③ The GCN-GRU model pays more attention to the influence of important features in the preorder than HA model, SVR model, and ARIMA model. This is mainly because the two gates of GRU focus on the temporal features of the data. While retaining the gating function, GRU reduces training parameters, improves model training efficiency to a certain extent, and reduces training duration.