Abstract: To analyze two sub-processes (rainfall-runoff in upstream urban watersheds and river flood evolution), semi-distributed hydrological model of FLOWS-Tank combining both mechanism-driven and data-driven approaches was applied to the small watersheds of Bayi and Douding reservoirs in Fuzhou, and the main channel of Jin’an River． Sensitivity of FLOWS-Tank model parameters and effectiveness of flood simulation in the river channel were studied． Most parameters of the FLOWS-Tank model were found to exhibit low sensitivity． For the Nash-Sutcliffe efficiency coefficient(NSEC) and root mean squared error(RMSE), the model parameters of side orifice height 7 and confluence parameters (nonlinear reservoir 2) showed strong sensitivity in both first-order and total sensitivity analysis． Water level simulation at the Wusi station achieved an mean squared error(MSE) of 0.001, mean absolute error (MAE) of 0.012, mean squared log error (MSLE) of 0.0007, and RMSE of 0.033． The FLOWS-Tank model demonstrated good simulation performance for the Bayi and Douding reservoir catchments, with total runoff increasing gradually as return period increased． In addition, coupling of long short-term memory (LSTM) neural networks and generative adversarial networks (GANs) proved to be well-suited for river flood simulation．