Abstract: With the increase of coal mining depth, coal production process is faced with complex water inrush mechanism and variable water inrush main control factors, and the uncertainties among the factors make the prediction of floor water inrush more difficult. In order to accurately predict the risk of floor water inrush, aiming at the small sample and non-linear problem of floor water inrush, firstly, genetic Algorithm is used to optimize the initial weights and thresholds of network random assignment, and then Sparrow Search Algorithm with strong search ability and good stability is selected to optimize the weights and thresholds for the second time, so as to establish the SSA-GA-BP neural network floor water inrush prediction model. Based on the analysis of geological and hydrological data of Binhu Coal Mine in Shandong Province, 8 factors including water pressure of aquifer, aquifer thickness, aquiclude thickness, fault density, fractal dimension value of fault, permeability coefficient, unit water inflow and floor failure depth are selected as the main control factors to predict floor water inrush, mapping the main controlling factors of 3D surface map projection. The Kriging interpolation method in surfer software is used to extract 50 data points as the input samples of the model(including 40 training sets and 10 test sets). The model is trained and studied. After the training error accuracy meets the requirements, the water inrush risk of 12 data points of 3 unmined working faces in Binhu Coal Mine is predicted. To verify the accuracy of the model, BP, GA-BP and SSA-GA-BP models are used to predict the test set; to avoid the one-sideness of comparing the model only with the prediction of BP network, the Fuzzy Comprehensive Evaluation Method, which determines the weight by Entropy Weight Method, is selected to predict the test set. The prediction results of each network model and method are compared with the actual values for analysis. The results show that the water inrush prediction error of GA-BP neural network model optimized by sparrow search algorithm is smaller, and the prediction accuracy is higher, which provides a scientific theoretical basis for mine water disaster prediction.