Abstract: As coal mining extends deeper, the geological conditions underground become increasingly complex. The issue of inaccurate prediction of mine gas content due to complex spatiotemporal correlations has become increasingly prominent. Traditional prediction methods cannot effectively overcome the technical bottleneck of “separation between macro trends and local details, and disconnection between spatial correlation and temporal dynamics”. To effectively enhance the accuracy and reliability of gas content prediction, research was conducted on a gas content distribution prediction method based on spatiotemporal characteristics, accompanied by engineering verification tests. Taking the grid nodes divided in coal seam mining areas of high-gas mines as the research carrier, a multi-dimensional grid node feature system was constructed based on node basic features, adjacency matrices, and edge feature matrices. The Universal Kriging (UK) spatial interpolation algorithm was introduced, to establish a baseline model for the macro spatial variation trend of gas content. A UK-Graph Transformer-LSTM fusion prediction model was proposed. By deeply mining the spatial topological associations of grid nodes under the control of fault barriers, lithological differences, and other factors, LSTM was utilized to capture the dynamic temporal dependencies of extraction flow rates. A residual correction module was designed to learn the prediction biases of the UK baseline model, achieving a deep fusion of macro trends with spatiotemporal details. Experimental comparative analysis results showed that while the UK baseline model could depict the macro distribution trend of gas content, its prediction accuracy had significant shortcomings. Its root mean squared error (RMSE) and mean absolute percentage error (MAPE) were as high as 0.24 m³/t and 2.51%, respectively. Single interpolation algorithms and single deep learning models struggled to adapt to the dynamic conditions of gas extraction due to inherent flaws. The former lacked the dimension of temporal evolution, while the latter lacked a spatial constraint mechanism, both unable to achieve spatiotemporal collaborative prediction. In contrast, the fusion model, relying on a multi-feature collaborative driving mechanism, achieved comprehensive improvement in prediction accuracy. Its RMSE was only 0.19 m³/t, the mean absolute error (MAE) was as low as 0.16 m³/t, and the MAPE was controlled at 1.60%. Compared to the UK baseline model, the fusion model reduced MAE by 30.43%, MAPE by 36.30%, maximum error by 9.38%, and residual sum of squares by 76.44%. Based on the prediction results of this model, high-risk areas such as the core zones of geological structures can be accurately identified, providing important theoretical basis and technical support for the prevention and control of mine gas disasters and the optimization of extraction parameters.