[Purposes] This paper aims to seek a method with high data utilization efficiency and strong interpretability of the propagation mechanism of traffic flow in a road network to solve the short-term traffic flow prediction problem in urban road networks. [Methods] The paper proposed a short-term traffic flow prediction model for urban road networks (NNLTM-TGAT) that integrated traffic flow propagation characteristics. The model used the link transmission model (LTM) as the road network traffic flow model, converted the traditional non-differentiable traffic flow model into a differentiable computational graph, incorporated the propagation mechanism of traffic flow in the road network into the computational graph model, and introduced the graph attention gated recurrent neural network (GAT-GRU) model to extract feature information in the time and spatial dimensions from urban road network traffic flow data, achieving short-term traffic flow prediction at the intersection flow level in urban road networks. [Findings] The paper uses traffic flow data from a partial road network in the central urban area of Taizhou, Zhejiang Province for model performance testing. The results show that the NNLTM-TGAT model proposed in this paper achieves root mean squared error (RMSE) values of 6.97 vehicle, 6.49 vehicle, and 6.86 vehicle for 5-minute flow prediction in morning peak, off-peak, and evening peak scenarios, respectively, outperforming the comparative models in prediction performance. [Conclusions] The NNLTM-TGAT model proposed in this paper can leverage the powerful high-dimensional spatiotemporal modeling capabilities and dynamic feature learning abilities of deep learning models to extract spatiotemporal characteristics from raw data and utilize the prior knowledge of traffic flow propagation mechanisms to achieve better prediction performance in different scenarios.