Hybrid fuzzy grammar dynamic graph diffusing attention network for traffic flow prediction

Accurate and real-time traffic flow prediction is an indispensable part of the intelligent transportation system and is essential in improving traffic planning capability. However, due to the highly nonlinear and spatiotemporal fluctuation characteristics of the large-scale traffic network data, it is a challenging issue to establish an accurate and effective prediction model. In this regard, a hybrid fuzzy grammar dynamic graph diffusing attention network is proposed for traffic flow prediction. Firstly, the network utilizes the grammar network structure composed of grammar rules to synchronously capture the interactive information of observable traffic parameters and the dynamic spatio-temporal correlation of each node. Secondly, the network utilizes an improved graph attention network for spatio-temporal node aggregation and dynamic edge information extraction, effectively mitigating over-smoothing. Finally, the network combines hidden features captured by the grammar structure with the change rate of the traffic flow through the fuzzy network to deduce the blend of hidden features of observable and unobservable information. Simulation results on three real datasets show that the proposed model outperforms existing prediction methods under traffic networks.