Abstract: The spatiotemporal heterogeneity of rainfall is one of the key factors affecting the accuracy of flood forecasting. In this study, a deep learning framework combining rainfall spatial features and flood process features is proposed. Based on the coupled graph convolutional neural network (GCN) and long short-term memory network (LSTM), the adjacency matrix with rainfall stations as nodes is used to improve the LSTM input module, and the runoff process vectorization method (RPV) is introduced to construct the GCN-RPV-LSTM flood process forecasting model. The model was trained and validated by 34 measured rainfall-runoff data in the the control basin of Jialu River at Zhongmou station, and the results were compared with GCN-LSTM and RPV-LSTM models. The results show that: ① The GCN-RPV-LSTM model has the best flood forecasting accuracy. Under the lead time of 6h, the Nash efficiency coefficient (E_NS) in the validation period is 0.014 and 0.076 higher than that of the two comparison models, respectively. ② The model is more accurate for the forecasting of bimodal floods, and the E_NS is above 0.9. It also has significant performance advantages in high-flow flood forecasting. ③ The model has stronger robustness, but there are underestimation of flood peak and delay of peak time. In the future, the model structure can be further improved to provide scientific decision-making basis for flood control safety of the basin.