In order to improve the prediction accuracy of expressway traffic flow, combined with S-G filter, BiLSTM, GRU and other methods, a BiLSTM-GRU combined model based on S-G filter is proposed to predict the short-term traffic flow of expressway. The combined model first uses the S-G filter to reasonably remove noise and abnormal fluctuations in the data, and solves the problem that the model is difficult to capture and learn the large fluctuation of high-speed traffic flow. Secondly, by using BiLSTM, the long-term dependence and time pattern of traffic flow data are captured according to its two-way advantages. Furthermore, GRU efficiently extracts and converts the output features of BiLSTM to output predictive values by automatically learning features. Finally, the mean absolute error ( MAE ), mean square error ( MSE ) and mean absolute percentage error ( MAPE ) were used to evaluate the prediction results. The case analysis selects the traffic flow data of the 104KM main line bayonet Shenyang direction of the Beijing-Harbin Expressway in Tianjin ( Tangshan section ), and analyzes the prediction performance of the combined model. The results show that the BiLSTM-GRU combined model based on S-G filter improves MSE by 0.75 to 51.28, MAE by 0.14 to 3.82, and MAPE by 0.95 % to 20.84 % compared with the traditional single model and other combined models. It can be seen that the BiLSTM-GRU combined model based on S-G filter can reasonably remove noise and abnormal fluctuations in the data, improve the prediction accuracy of short-term traffic flow on expressways, and provide theoretical basis for traffic flow control on expressways.