Research On Urban Short-term Traffic Congestion Prediction Based On Deep Learning

Many large cities around the world have been plagued by traffic congestion,and with the expansion of the city scale,the siphon effect of large cities is very obvious.The increase in urban population,the increase in car ownership,and the backwardness of road infrastructure have led to congestion problems.The urban short-term traffic congestion prediction problem models the congestion evolution process,provides path decision guidance for urban residents’ travel,and can also provide effective traffic force deployment decision information for the traffic management department.Therefore,the short-term prediction problem of congestion can provide a certain auxiliary role to solve the urban congestion problem to some.Although the short-term forecasting problem of urban traffic congestion has always been a hot spot in the transportation field,urban congestion is affected by multiple factors,and the previous congestion prediction methods have certain limitations.Therefore,short-term predictions of traffic congestion have always been a difficult problem.In the past,when solving the problem of short-term forecasting of traffic congestion,the commonly used methods are mainly divided into two categories: one is based on statistical theory models,such as AR/MA/ARMA/ARIMA and Kalman filtering,but such methods have poor ability to capture features such as nonlinearity and uncertainty of traffic flow.The other is a kind of model based on knowledge discovery based on machine learning and deep learning.This kind of model performs well in mining unknown patterns from data and can continuously learn the characteristics in the data.When using deep learning for congestion prediction,the traditional CNN method has problems such as Coarse grain and multiple possibilities for mapping grid to road segment,so it is limited in practical applications.Models such as LSTM,GRU,and RNN can only capture temporal dimension features,ignoring the spatial correlation of traffic congestion.Based on the above problems,this paper uses the road segment as the basic unit,and designs a multi-scale segmentation method for road segmentation based on the characteristics of traffic congestion.Then,using the topology property of the road network itself,the urban road network is represented as a map.The structural form uses a short-term prediction of congestion using a graph convolution model.The main research contents and results of this paper include:(1)Analysis of spatio-temporal distribution characteristics of traffic congestion.Firstly,for the time distribution characteristics,the congestion has three kinds of congestion modes: working day,weekend,and holiday.The congestion characteristics of working days from Monday to Friday are similar,showing a narrow-width doublepeak mode;weekend shows a pattern with a long duration and a slightly smaller peak;holidays are characterized by severe congestion in the early holidays,weakened mid-term congestion,and late peaks in the later stages.For the spatial distribution characteristics,it shows that the formation of congestion hotspots is related to traffic events,and there is a certain correlation between traffic congestion and road grade.(2)Short-term forecast of traffic congestion.A multi-scale road segmentation method is proposed to divide the road,and the segmented road is represented to be graph according to the topological relationship of the road network.The spatiotemporal graph convolution neural network model is used to predict the short-term urban traffic congestion.The results show that the STGCN model has the best prediction performance,followed by the T-GCN model,which has great advantages over traditional models such as HA and ARIMA.It verifies the effectiveness of the spatio-temporal graph convolution neural network used in traffic prediction problems.In addition,experiments show that GCN has a high adaptive ability to the network structure,and perform well w.r.t.different road network structures.