Optimal Graph Information Fused Graph Attention Network for Traffic Flow Forecasting

Abstract

To manage and make decisions about intelligent transportation systems more efficiently, accurate traffic flow forecasting is necessary. Traffic flow forecasting has complex spatial correlation and time dependence. Most current research models are based on a predefined graph structure with a priori knowledge for prediction, which cannot well extract the hidden spatial relationships in traffic data. In this paper, we propose the Optimal Graph Information Fused Graph Attention Network (OGIF-GAT). Specifically, we learn the actual connections between nodes and the hidden spatial relationships through the multigraph feature fusion structure. Next, we design a new graph attention network (GAT), which improves the problem of ignoring edge features in the graph structure in the traditional GAT model and considers their edge features when estimating the correlation of each neighboring node pair: the effect that the distance factor between neighboring nodes has on the spatial correlation. In addition, we use the temporal hybrid transformer (THT) to learn temporal dependencies. Extensive experiments on four public transportation datasets (PeMS04, PeMS08, PeMS-BAY, and METR-LA) demonstrate that our model achieves the optimal level of traffic flow prediction accuracy on all of them and is shown to have strong generalization ability. Compared to STSGCN, the mean absolute error (MAE) decreases by 7.9%, 10.3%, 33.2%, and 19.6%, respectively.