Abstract: Aiming at the problems of delayed traffic flow effects between upstream and downstream of road sections and insufficient excavation of spatio-temporal correlation characteristics among intersections in the existing research on urban road traffic flow prediction, a model based on dynamic graph neural delay differential equations was proposed to fine-grain the instantaneous changes of traffic flow and extract long-distance dynamic spatio-temporal features in order to improve the prediction accuracy. Firstly, considering that the traffic flow showed high similarity at different cycle scales, the spatio-temporal attention mechanism was used to model the weekly and daily scale traffic flow data to enhance the spatio-temporal correlation among intersection nodes. Secondly, the delay time between upstream and downstream roads was calculated and delay differential equations were introduced to extract the time lag characteristics among nodes of the road network, simulating the spatial information propagation process under the delay effect. Finally, the features of each time scale were integrated to obtain the predicted output values. Through the validation of real public traffic flow datasets such as PEMS04, PEMS07 and PEMS08, the results show that the average absolute error and the root mean square error of the proposed model are reduced by about 2.20% and 1.16% on average.

Key words: traffic engineering; traffic prediction; dynamic graph neural delay differential equations; delay; feature fusion

摘要： 基于动态图神经延迟微分方程的交通流预测

关键词: 交通工程；交通预测；动态图神经延迟微分方程；延迟；特征融合