Research On Traffic Flow Forecast Based On Data-driven

Traffic flow prediction,as one of the important research fields of intelligent transportation system,can reduce the travel cost of travelers and social operation cost.The existing traffic flow prediction model does not analyze the data quantitatively before the establishment,so it is impossible to adapt to different traffic flow.At the same time,most of the current research ignore the frequency domain characteristics of traffic flow or do not explore in-depth.In response to these problems,the paper analyzes the inherent characteristics of traffic flow,and considers whether the traffic flow characteristics change after modal decomposition,which lays the foundation for the subsequent model establishment and hyperparameter optimization.Aiming at the traffic flow prediction at the road network level,combined with the analysis results,the research proposes a combination of Variational Mode Decomposition and Attention based Spatial-Temporal Convolutional Network.The traffic flow prediction model VMD-ASTCN;For the traffic flow prediction of some nodes with high prediction accuracy requirements,a traffic flow prediction model Delay-VMD-ASTCN considering the delay characteristics is proposed.The specific work content and results are as follows:(1)Multi-dimensional characteristic analysis of traffic flow.Based on time-frequency analysis,the macroscopic characteristics and frequency domain characteristics of traffic flow are captured qualitatively and quantitatively,and the modal decomposition methods suitable for traffic flow are compared and optimized.Based on the correlation coefficient and considering the time delay,the traffic flow before and after the modal decomposition is analyzed in time and space.The results show that traffic flow data contains a lot of noise;traffic flow has temporal autocorrelation,specifically strong periodicity and can maintain autocorrelation over a long time span;traffic flow has spatial cross-correlation,and the cross-correlation is affected by distance The impact is time-delay and is affected by the road topology.(2)Based on the analysis results,a traffic flow forecasting framework was established.First,the data was decomposed modally to reduce data noise.Secondly,after considering the spatiotemporal characteristics of traffic flow and road topology,and combining the attention mechanism,temporal convolutional network and graph convolutional neural network,a spatiotemporal prediction model of traffic flow was established.The results were analyzed and the relevant parameters were determined based on the data.The modal components were predicted by the model.Finally,the predicted values were obtained by combining the modal components.At the same time,for the prediction of traffic flow nodes,the time delay characteristics was especially considered to ensure that the influence between spatial nodes does not have time lag.The paper evaluates the model proposed in the paper through real data sets with differences in two temporal and spatial dimensions,and compares it with a variety of representative benchmark models.The experimental results show that whether it is for the road network level traffic flow prediction model VMD-ASTCN,or Delay-VMD-ASTCN,a node-level traffic flow prediction model,has better performance than other traditional models or existing models.