Dynamic Spatio-temporal traffic flow prediction based on multi fusion graph attention network

Abstract

Spatio-temporal prediction is very important in the application of traffic management and traffic planning. Urban traffic network is a complex network system, which is affected by various uncertain factors, such as major traffic events, weather and so on. Traffic flow prediction is a typical spatiotemporal prediction task. The traffic flow data in the road network has strong nonlinear dynamic characteristics; In addition to the two typical characteristics of temporal correlation and spatial correlation, such as the characteristics of road structure is also an important factor to be considered in traffic flow prediction, which still faces great challenges. This paper proposes a multi fusion graph network dynamic spatiotemporal traffic flow prediction model (DGAT) based on graph attention. The traffic graph network is constructed by comprehensively considering a variety of road structure information, and combined with graph attention to capture the dynamic transformation of traffic flow data, so as to complete the task of traffic flow prediction. The results show that the prediction accuracy of the DGAT proposed in this paper is significantly improved compared with the existing benchmark model.