Accurate and real-time short-term traffic flow prediction is critical for the construction of modern traffic management service systems. In order to fully exploit and utilize the spatial-temporal characteristics of traffic flow interaction in different road sections, a two-level screening mechanism composed of the autocorrelation functions, the cross-correlation functions and the KNN algorithm is constructed to evaluate the correlation between the target road section and optimize the combination of road sections, and realize deep mining of spatial information. One of the GCN- GRU combination forecasting model is proposed. The spatial characteristics of short-term traffic flow are captured by using the advantage of the graph convolutional network (GCN) in the global processing of section topology information, and the time characteristics are extracted by using the long-term memory ability of the gated recurrent unit (GRU) for time information, which are verified by the measured short-term traffic flow data of expressway. The results show that using the two-level screening mechanism to effectively screen the road sections and introducing a deep learning combination model, the prediction performance will be significantly improved, which is better than the commonly used models such as the stacked autoencoders network (SAEs) and the temporal convolutional network (TCN).