Traffic Flow Prediction Based on Spatio-Temporal Aggregated Graph Neural Networks

Because of the increasingly complex spatio-temporal relationships among transportation networks, accurately predicting traffic flow has become a challenging task. Most existing frameworks primarily utilize given spatial adjacency graphs and other complex mechanisms to construct spatio-temporal information of transportation networks. However, the relationships between non-adjacent spatial positions in the road network may affect the effectiveness of these models. Therefore, this paper proposes a new framework called the "spatio-temporal aggregated graph neural network" to enhance the feature relationships in spatio-temporal data. Firstly, a method for generating temporal graphs from spatio-temporal data is introduced to compensate for the spatial graphs' potential inability to reflect temporal correlations. Next, the correlation coefficient matrix is computed to further enhance the spatio-temporal correlations in the traffic road network. Finally, multiple graph structures are overlayed to comprehensively consider the spatio-temporal correlations of the traffic road network. Experimental results on extensive datasets demonstrate the superiority of this model.