The invention discloses a method for dynamically predicting the drainage flow of an urban rainwater system drainage outlet. In step (1), the rainstorm and flood management model is used to simulate the rainfall-runoff, and the drainage flow process lines of multiple sets of drainage pipe network outlets are used as training samples. Step (2), set up RBF neural network and train, carry out the optimization of network hidden layer node and center width Spread in the training process; Step (3), set up NARX neural network and train; Step (4), will finish training The NARX neural network and the RBF neural network are coupled to obtain the coupling network, and then predict, calculate the mean square error of the coupling network and the sample, return the flow value with the smallest mean square error as the optimized coupling point, and randomly select the rainfall data to input the coupling network. Obtain the predicted drainage flow hydrograph. The invention organically combines the advantages and characteristics of different neural networks, the prediction result is in good agreement with SWMM simulation, the mean square error of the curve is 0.000458, and has good prediction accuracy.