Abstract: Current study incorporates both the watershed and urban areas into a unified spatial context in order to address the problem of coordinated flood forecasting in watershed-urban compound systems.Based on the proposed framework for distinguishing easily generated runoff patterns, a hybrid forecasting model, called GRGM-LSTM, and is developed by coupling the Grid-based Runoff Generation Model (GRGM) with Long Short-Term Memory neural networks (LSTM).The model is tested using 18 observed flood events in the control basin of the Jialu River at Zhongmou station.In addition, the forecast results are compared and analyzed against the Storm Water Management Model (SWMM) and GRGM-SWMM model.The study reveals that: ① The relative error and coefficient of determination obtained from the GRGM for simulating runoff are 8.41% and 0.976, respectively.This indicates that considering the spatial distribution of runoff patterns results in more accurate runoff calculations.② For forest period of less than 6 hours, the GRGM-LSTM hybrid model outperforms physical mechanism models such as GRGM-SWMM and SWMM, yielding Nash-Sutcliffe efficiency coefficients greater than 0.8, indicating superior simulation performance.③ However, for a forest period exceeding 6 hours, the GRGM-LSTM hybrid model experiences some accuracy loss, and when the forest period increases to 12 hours, the simulation accuracy of GRGM-SWMM surpasses that of GRGM-LSTM.The research findings can serve as a scientific basis for coordinated management of flood prevention and disaster reduction in watershed-urban areas.