Research On Urban Traffic Condition Prediction Methods Via Integrating Multi-Granular Information

In the field of intelligent transportation,the problem of urban traffic condition prediction is one of the important research topics,which has attracted extensive attention from academia and industry.The difficulty of urban traffic condition prediction is that there are many factors affecting urban road conditions,and its own change pattern is complex.It is difficult to effectively model the correlations in the data with traditional statistical methods and accurately capture the law of road condition changes.With the development of artificial intelligence technology,many deep learning methods for traffic condition prediction tasks have emerged in the academic world,all of which have achieved good results.However,due to the complex spatiotemporal relationship of urban traffic,the promotion space of the existing methods is relatively large.Therefore,this paper exploratively integrates two granularity-levels traffic characteristics of urban areas and roads,and combines various external influencing factors to model urban traffic conditions to predict short term traffic conditions.Firstly,in order to solve the problem that the spatial and temporal correlations of roads are difficult to model due to the complex structure of urban roads,a Multiple Granularities Convolution Recurrent Neural Networks(MGCRN)is proposed.The model contains two main modules,one is a road-based prediction module-Spatial-Temporal Dynamic Graph Neural Networks(STDGNN).In the time dimension,the recurrent neural networks are used to learn high-dimensional features with historical information,and the adjacency matrix of road nodes is adaptively learned from the historical features,through graph convolution layers,the road prediction is obtained.Another module is a regionalbased prediction module,Grid Convolutional Recurrent Neural Networks(GCRN).Convolutional operations is used to model the changing patterns between urban areas traffic,and recurrent neural networks are used to obtain the associations between traffic conditions at the predict time and the historical time.Finally,the two modules are fused to obtain the traffic condition prediction result.Secondly,in order to further improve the prediction effect of the model,a Multiple Granularities Dynamic Convolution Recurrent Neural Networks(MGDCRN)is proposed.MGDCRN has optimized and improved several modules of the MGCRN model.In the STDGNN module,a historical period graph is added to improve the road-level prediction ability,and in the GCRN module,historical period information and functional area feature embedding are added to obtain explicit macro-semantic information.In addition,the fusion mechanism is improved to a dynamic back-projection mechanism,so that the road can adaptively extract relevant information from regional traffic conditions.Finally,a large number of experiments were conducted on the Chengdu public dataset provided by Didi Chuxing.The experimental results show that the MGCRN model performs good on this dataset,and the MGDCRN model further improves the prediction performance.This shows that the model proposed in this paper can well model the complex urban traffic change patterns in the urban traffic condition prediction task,so as to obtain more accurate prediction results.