Network-scale Traffic Forecasting Model Based Spatial-temporal Feature Fusion

With the acceleration of urbanization,traffic congestion and a series of derivative problems have become increasingly severe.Traffic control and guidance are important means to alleviate traffic congestion,and accurate traffic flow prediction is an important reference for realizing traffic control and guidance.In recent years,with the enrichment of information collection methods,traffic flow data is no longer limited to a single road section or a certain area,but is often presented in the form of a large-scale road network.The development of deep learning theory,especially graph model theory,provides an important research approach for traffic flow prediction at the road network level.Road network traffic flow data is essentially a spatio-temporal sequence.This paper combines graph theory and time series prediction theory to propose a road network traffic flow prediction method based on the fusion of time and space features,aiming to provide data support for traffic guidance and route planning.Firstly,by combing the research background and relevant research status,the modeling goal of this article is clarified: integrating the temporal and spatial information of traffic flow,realizing large-scale road network-level traffic flow prediction,and exploring the possibility of long-term prediction.At the same time,related theories are elaborated according to modeling needs.Secondly,from the data point of view,explain the source,quality and scale of the data,and repair the data according to the characteristics of the data and subsequent modeling needs.At the same time,the basic characteristics and temporal and spatial correlation of road network traffic flow data are analyzed to provide data support and theoretical basis for subsequent modeling.Thirdly,from the perspective of spatiotemporal feature fusion,a road network traffic flow prediction model(STAtt)based on the spatiotemporal attention mechanism is proposed.The model uses the graph attention network(GAT)to describe the directionality,difference and variability of the interaction of the road sections in the road network,and embeds it in the cyclic unit instead of the gate structure to achieve spatiotemporal information fusion.At the same time,the Seq2 Seq architecture and timing attention mechanism are used to expand the model’s receptive field and reduce the cumulative error of multi-step prediction.Experimental results show that the model is better than other models in short-term prediction tasks within 1 hour.Finally,in view of the short-term prediction of the STAtt model,a road network traffic flow prediction model(STGAN)based on the spatio-temporal generation confrontation network is proposed.Based on the periodic characteristics of long-term traffic flow,the model constructs a sequence of time conditions to fuse historical traffic flow information.At the same time,in view of the difficulty of STGAN model training,the model training process and convergence judgment are introduced in detail.The experimental results show that the model is better than other models in the long-term prediction task within 24 hours,and the model is more interpretable,and it has realized the long-term prediction of traffic flow at the road network level.