Currently, deep learning-based traffic flow prediction methods have deficiencies. Firstly, the prediction model based on graph convolutional network uses a simplified road network topology, ignores the actual traffic organization information, and affects the prediction accuracy. Secondly, the clustering-based prediction model does not consider the regional and temporal similarity of traffic flow and fails to effectively utilize spatio-temporal patterns, resulting in limited enhancement of prediction by clustering results. In addition, overly large training samples increases the training and prediction time, affecting real-time performance. In order to solve the above problems, this paper proposed a deep temporal clustering traffic flow prediction(DTCTFP) model based on deep temporal clustering for short-term traffic flow prediction at urban chokepoints. Firstly, the method constructed the road network topology containing actual traffic organization information and used a graph convolutional network to mine the spatio-temporal characteristics between chokepoints. Secondly, it introduced improved dynamic temporal regularization and shortest path analysis methods to classify similar traffic flow objects into the same cluster, allowing the model to make full use of feature information such as flow rate, time, and location to improve prediction accuracy. Finally, it used a cluster-based recurrent neural network for prediction to enhance the real-time and computational efficiency of the model. Using Chongqing Dadukou traffic data, experiments show that the model reduces MAE, RMSE, and MAPE by 15.02%, 10.72%, and 10.98% on average compared to the latest benchmark. The ablation test also confirms a 14.5% improvement in prediction accuracy with the proposed clustering method.