A Hybrid Fuzzy Dynamic Grammar Graph Attention Network for Traffic Flow Prediction

Abstract

As an indispensable part of the intelligent transportation system, traffic flow prediction can provide managers with decision-making reference data for maintaining traffic order and help individuals choose better travel plans. Because the data of large-scale traffic networks are highly nonlinear and have spatio-temporal fluctuation, it is challenging to establish an accurate and effective prediction model. For this regard, this paper proposes a hybrid fuzzy dynamic spatio-temporal grammar graph attention network for traffic flow prediction, which uses a grammar structure to simultaneously capture the influence of the spatio-temporal dependence of the three historical traffic parameters and their coupling relationship on the prediction values and uses the fuzzy inference structure to capture the influence of other uncertain factors on the prediction values. The prediction values are obtained by using the fusion prediction structure to fuse the output features of the grammar structure and the fuzzy inference structure in an attention mode. Simulation results on two real data sets show that the prediction accuracy of this model is better than the existing prediction methods in different traffic networks.

Keywords: traffic flow prediction, grammar structure, spatio-temporal dependence, fuzzy inference structure, fusion prediction structure, attention mode