Abstract: To address the challenge of limited prediction accuracy for underground gas concentration, we propose a spatio-temporal graph neural network(DASTNN) model incorporating a dual attention mechanism. This model aims to enhance the prediction of gas concentration in coal mine drainage pipe networks. By integrating a graph convolutional network (GCN) with a gated recurrent unit (GRU), and applying both spatial and temporal attention mechanisms, the model improves the extraction of features related to network topology and time series patterns. We evaluated the model on the gasnet-data1 and gasnet-data2 datasets. The results demonstrate that our approach outperforms traditional methods such as HA, SVM, GCN, and GRU. On the gasnet-data1 dataset, the model achieved a mean absolute error (e_MA) of 0.310, a root mean square error (e_RMS) of 1.069, and a coefficient of determination (R²) of 0.975. On gasnet-data2, the e_MA was 0.181, the e_RMS was 0.745, and the R² was 0.990. These findings indicate that the dual attention mechanism effectively captures the spatiotemporal dependencies of gas concentration, and significantly improves prediction accuracy.