Abstract: In the process of coal mining, the loss of people and property caused by mine water inrush is extremely serious. To prevent the occurrence of water inrush accidents and grasp the law of change of water inrush, the water inrush prediction and forecasting, especially the accurate estimation of mine water inrush, is very important, which is also an important task in the prevention and control of mine water damage. To increase the prediction accuracy of mine water inrush, an efficient time series prediction model combining Variational Mode Decomposition (VMD) and Deep Belief Network (DBN) was proposed for the series of water inrush with no obvious change with time. Firstly, the original data were initially denoised by VMD to break up the original mine water inrush time series into multiple Intrinsic Mode Function (IMF) components, so that each IMF component has the statistical characteristic quantity of the original time series at different time scales, which reduces the strong oscillation and instability of the original time series. Secondly, DBN model was established separately to each IMF component for training and learning, and then the corresponding prediction network model was built. Finally, the predicted values of each component were fused as a result. The results show that the E_MA, E_MAP, E_RMS and R² of VMD-DBN are 9.23, 0.76%, 11.55 and 0.97 respectively, which are compared with the predicted values of GA-BP, LSTM, VMD-LSTM, RBM, VMD-RBM, and DBN models, finding that the mine water inrush prediction with VMD-DBN model has a higher accuracy. Therefore, the VMD-DBN model has relatively obvious advantages under the conditions that the water inrush has no obvious change law over time but with strong oscillation and instability, thus enriching the mine water inrush prediction methods, providing a new technical means for the intelligent mine safety monitoring, with some theoretical value and practical significance.