To improve the prediction performance of temporal-spatial characteristics by the existing short-term traffic flow prediction model, a deep neural network prediction model combining convolutional neural network (CNN) and gated recurrent unit (GRU) neural network is proposed. In the model, the CNN is firstly used to extract the spatial characteristics of short-term traffic flow data, and then the results are input into the GRU to mine the temporal features. Taking the traffic flow data in the California Department of Transportation’s performance evaluation system as an example, the model is trained to verify its accuracy. The results show that compared with existing models, the proposed prediction model has better prediction performance. The average absolute percentage error is significantly reduced, which can provide an effective basis for traffic management and control.