GCN-Bi-LSTM: A Neural Network Method Based on Spatiotemporal Features for Short-Term Traffic Flow Prediction

Road traffic flow is generally a time series of nonlinear and non-stationary random processes, and the recording time of observation points is generally 5 minutes to 15 minutes, which is greatly affected by random interference factors, with strong uncertainty and no obvious regularity. Nevertheless, better prediction of short-term traffic flow can help with traffic decision-making, road selection, shortening congestion time, etc., greatly improving travel happiness. Therefore, this paper proposes a short-term traffic flow prediction based on the GCN-Bi-LSTM hybrid model, in which the traffic flow time series data is input into the GCN model to extract the spatial features. The early stop method is used to select the best model parameters, and then the data is input into the double-layer Bi-LSTM to learn the time features to obtain the final prediction results. Experimental results show that the prediction errors of the proposed model are 8.7408, 0.0039 and 0.0285, respectively, and the corresponding errors of the GRU model with poor prediction performance in the baseline model are 27.4173, 0.0243 and 0.0816, which fully shows the improvement effect of the proposed model compared with the single neural network model in prediction performance.