With the rapid development of urbanization in China, the number of motor vehicles is also increasing rapidly, and at the same time it has also caused a series of traffic problems. These include serious road traffic congestion, rising traffic accident rates, and waste of resources caused by inefficient transportation, most notably the severe congestion of road traffic. Therefore, forecasting traffic flow is the key to ease traffic congestion. To solve the above problems, this paper proposes an urban short-term traffic flow prediction model based on stacked self-encoder and support vector machine. This model utilizes the characteristics of Stacked Auto Encoder (SAE) and Support Vector Machine (SVM): SAE is used to extract features from the traffic Flow and the extracted features are input into the SVM. Then, Traffic congestion is predicted. In order to verify the validity of the model, data on buses and taxis in Wuhan are processed to extract important traffic flow parameters in traffic congestion prediction, including traffic volume, average speed, and average density. For the preprocessed data, SAE and SVM were used for model training and prediction. Then the model was modified by BP back-propagation neural network to make the model prediction error converge. The prediction results show that the SAE-SVM model has higher accuracy than the traditional prediction model, and the mean square error (MSE) and average absolute error (MAE) are low, so the SAE-SVM model is more accurate in traffic congestion prediction.