Abstract: The underground water supply system of coal mine was the lifeline of coal mine safety production. The prediction of water flow of water supply network was the basis of water supply system optimization dispatching. The importance of prediction has important influence on water supply dispatching. This paper proposes a method of flow prediction of underground coal mine water supply network that integrates multi-modal data features. Different from previous methods, deep learning method realizes the integration of multiple data related to spatial topology structure of underground pipe network, historical time dependence, actual underground production conditions and cycle. Specifically, The graph convolutional neural network adding spatial attention mechanism is used to obtain the spatial topological relationship of the monitoring points of underground pipe network, and then the gated circulation unit in the cyclic neural network is used to obtain the time dependence of the monitoring points, and the final prediction results are formed by integrating the production rules of coal mine with the flow data of different periods. Through the actual data experiment of a mine in Shaanxi Province, The results show that compared with SVM, LSTM, STGCN and other methods, the proposed method can more accurately predict the future trend of downhole flow, and the prediction deviation is reduced by 9.3%, 6.84% and 3.65%, respectively.