Hybrid Spatial-Temporal Graph Convolutional Network for Long-Term Traffic Flow Forecasting

Recently, short-term traffic flow forecasting has been highly focused on by scholars, and some effective methods and models have been presented. However, few attentions have been taken on long-term traffic flow forecasting. Accurate long-term traffic flow prediction is of great practical significance in alleviating road congestion, reducing potential urban road traffic safety hazards and guiding people’s daily travel planning. Due to the influence of time and space external factors, traffic flow data has highly nonlinear and non-Euclidean structure characteristics, and some traffic flow prediction algorithms have low prediction accuracy. Nevertheless, graph convolutional networks have been widely concerned in various fields in recent years because they can handle data with non-Euclidean structures. In this paper, a longitudinal series slicing method is presented for preprocessing the data to obtain sample data and labeling for model training. In addition, a spatial-temporal graph convolutional network model through mining the potential spatial-temporal dependence in traffic flow dataset is put forward. And experimental results show that the model has good results compared with the benchmark method.