Analysis And Prediction Of Urban Traffic Flow Based On Deep Learning

With the continuous development of society and the acceleration of urbanization,the complexity and scale of the urban transportation network structure increase dramatically,and the residents’ travel demand continues to grow.The number of motor vehicles in China has reached 417 million,and traffic problems such as road congestion inevitably appear.In order to deal with these traffic problems,many cities began to build Intelligent Transportation System(ITS).A key part of ITS is the traffic flow prediction system,which can use the data generated by traffic vehicles or sensors(such as taxi track data and mobile signaling data)to predict the traffic flow in the future period of time by using the deep learning model.According to the forecast results,traffic management departments can focus on areas that may be congested in the future,implement traffic control in time,and reduce the incidence of road congestion and traffic accidents.Currently,traffic flow prediction models based on deep learning have achieved significant success,but there are still some limitations.First,as traffic flow is a spatiotemporal sequence data,how to fully exploit its spatiotemporal correlations and periodicity has been a challenge in traffic flow prediction.Second,traffic flow is easily affected by various natural and social factors,such as weather,holidays,Points of Interest(POI),etc.How to accurately model external factors is another challenge in traffic flow prediction.Although existing models have used external information to assist in predicting traffic flow,they have not fully utilized these data.In order to deal with the problems existing in the field of traffic flow prediction,this paper proposes a Self-Attention residual network STA-RNIMEI which integrates various external information.For the modeling part of external information,STARNIMEI model uses different embedding methods for different heterogeneous data.Specifically,for POI data,time-sensing and channel attention mechanisms are used to capture the dynamics of urban regional functions in time dimension,while for meteorological data,multi-layer perceptron and convolution are used to model.In the main structure of STA-RNIMEI model,in order to consider the proximity and periodicity of traffic flow,we screened out three historical data components that are highly correlated with the forecast time: short-term data,daily cycle data and weekly cycle data,and extracted the spatio-temporal correlation of corresponding components respectively from three submodules with the same structure.In each sub-module,Conv LSTM and stacked 3D residual convolution unit(3D-Res Unit)were used to extract the spatio-temporal characteristics of the traffic data.Conv LSTM could capture the long-term trends in the traffic data and learn the spatial correlation of the adjacent regions at the same time.3D-Res Unit consists of 3D convolutional layers and residual network that captures traffic correlations across time and senses historical traffic characteristics in neighboring areas.Finally,Vision Transformer(Vi T)was used to capture potential remote spatial dependencies between urban areas.The experimental results on Beijing taxi track data set and Changchun mobile signaling data set show that the STA-RNIMEI model proposed in this paper has better prediction performance than other benchmark models.