Research on Spatio-temporally Consistent Traffic Flow Prediction Based on Adaptive Dynamic Correlation Matrix

To address the issues of insufficient extraction of spatio-temporal intrinsic correlation and inadequate characterization of spatial dynamics in existing traffic flow prediction research, a traffic flow prediction model is proposed based on an adaptive dynamic spatial correlation matrix.The model constructs a spatio-temporal feature consistency extraction module by embedding a spatial feature extraction module and a multi-feature hierarchical fusion module inside the Transformerencoder model, combined with its multi-head self-attention mechanism, to achieve the full extraction of the spatio-temporal intrinsic correlation of traffic flow.The spatial feature extraction module constructs the adaptive dynamic spatial correlation matrix to fully characterize the dynamic spatial features of the road network’s neighborhood, while the spatial feature filtering network eliminates redundant information and highlights key influential features.Additionally, the multi-feature hierarchical fusion module conducts detailed and comprehensive mining of spatial and temporal correlations by hierarchically fusing the original temporal and spatial features of traffic flow.The experiments are tested on the PeMS04 and PeMS08 public datasets of the high-speed road network, and the results show that the proposed model has better prediction results compared to other baseline models.