Spatial-temporal Graph Attention Networks for Traffic Flow Forecasting
Chang Wei, Sheng Jin
Abstract
Abstract In order to accurately forecast the road section traffic volume, in this study, a spatial-temporal graph attention network model(GALSTM), which is based on graph attention architecture and long and short memory network(LSTM), is proposed to predict the traffic volume of road section. LSTM network is used to extract the temporal correlation of traffic flow data, and graph attention network is used to get adaptive adjacency matrix at each time step to capture the spatial correlation of road network. The proposed GALSTM model and other frequently-used traffic flow prediction methods were validated by using dataset collected by the California highway administration PeMS. Experimental results on two traffic datasets indicate that GALSTM model achieves the best prediction accuracy in all three evaluation metrics of mean absolute errors, mean absolute percentage errors, and root mean squared errors. GALSTM model can be used as an effective method to forecast the traffic volume of road section.