Traffic Flow Prediction Based On Improved Dynamic Graph Spatiotemporal Convolutional Neural Network

At present,the modernization of Chinese-style is continuously advanced,the amount of automobile ownership has increased rapidly,and the problem of traffic congestion is becoming increasingly serious.How to effectively alleviate traffic congestion and improve the efficiency of road resources allocation is particularly important.As a core component of the Intelligent Transportation System(ITS),traffic flow forecast can effectively help the management department to master road resources in advance and formulate corresponding management measures to alleviate traffic congestion,so as to improve the travel efficiency of travelers.The traditional deep learning prediction model is mainly aimed at grid data,which lacks the concept of time and space in capturing the characteristics of traffic node attributes.Moreover,it cannot meet the prediction accuracy requirements of different time intervals under long time series dependence.Therefore,this paper proposes a traffic flow prediction model based on graph convolutional neural network to solve the above problems.(1)By analyzing the temporal characteristics between traffic flows,traffic flow data is periodically divided.On this basis,a local time disturbance term T_r is added to supplement recent,daily and weekly cycle data,enhance the data correlation with the predicted time period,and provide data support for the following model design.(2)A Diffusion Graph Spatiotemporal Convolutional Networks model(DCSTGCN)is proposed for traffic flow prediction.Considering the fixity of the road network topology,firstly,on the basis of the graph convolution neural network,a bidirectional random walk model is added to enhance the dynamics of the spatial structure;Secondly,by adding a spatial Transformer structure,multi-head self-attention mechanism is used to filter the correlation of multiple attributes to further improve the spatial dependence;In terms of time series,the temporal attention mechanism and two-dimensional convolutional network are used to capture temporal features,thus completing the overall spatiotemporal feature learning.The experimental results showed that the RMSE,MAE,and MAPE of the DCSTGCN model decreased by 3.24%,3.31%,and 4.91%compared with the ASTGCN model.(3)A Dynamic Adaptive Graph Diffusion Spatiotemporal Convolutional Network model(ADCSTGTN)is proposed for traffic flow prediction.Based on the DCSTGCN model,an improved adaptive mechanism is added to enhance the learning of potential attribute features of nodes,and eliminate the weak connections between nodes to reduce the computational complexity;Further,in order to address the issue of decreasing prediction accuracy in long time series,time transformer is proposed to learn the features of long time series,constructing an ADCSTGTN model,and perform ablation experiments on the ADCSTGTN model.The experimental results show that the overall performance of the ADCSTGTN model is superior to other similar models,with RMSE,MAE,and MAPE decreased by 2.78%,10.77%,and 11.18%,respectively,compared with the DGCN model.