FDSA-STG: Fully Dynamic Self-Attention Spatio-Temporal Graph Networks for Intelligent Traffic Flow Prediction.

With the development of transportation and the ever-improving of vehicular technology, Artificial Intelligence (AI) has been popularized in Intelligent Transportation Systems (ITS), especially in Traffic Flow Prediction (TFP). TFP plays an increasingly important role in alleviating traffic pressure caused by regional emergencies and coordinating resource allocation in advance to deployment decisions. However, existing research can hardly model the original intricate structural relationships of the transportation network (TN) due to the lack of in-depth consideration of the dynamic relevance of spatial, temporal, and periodic characteristics. Motivated by this and combined with deep learning (DL), we propose a novel framework entitled Fully Dynamic Self-Attention Spatio-Temporal Graph Networks (FDSA-STG) by improving the attention mechanism using Graph Attention Networks (GATs). In particular, to dynamically integrate the correlations of spatial dimension, time dimension, and periodic characteristics for highly-accurate prediction, we correspondingly devised three components including the spatial graph attention component (SGAT), the temporal graph attention component (TGAT), and the fusion layer. In this case, three groups of similar structures are designed to extract the flow characteristics of recent periodicity, daily periodicity, and weekly periodicity. Extensive evaluation results show the superiority of FDSA-STG from perspectives of prediction accuracy and prediction stability improvements, which also testifies high model adaptability to the dynamic characteristics of the actual observed traffic flow (TF). [ABSTRACT FROM AUTHOR]