Research On Urban Multi-Scenario Traffic Flow Prediction Methods Based On Deep Learning

The rapid growth of urban vehicles has made it necessary to control vehicle flow and conduct traffic diversion with advanced traffic management and control technology,to facilitate proper planning and allocation of road resources,so as to provide smooth transportation in urban areas,and maximize transportation efficiency of the road network.Accurate traffic flow prediction is an important foundation of the smart transportation system in a city.It helps to improve road capacity and transportation efficiency of the road network,alleviate traffic congestion,reduce traffic accidents,decrease energy consumption and mitigate environmental pollution as well.To predict traffic flow accurately,it is necessary to conduct in-depth research on the spatial and temporal scenes of various scales of intersections,segments,and areas,including crosssectional lanes,road-segments and full spatio-temporal road network.However,the traffic flow data in these scenes are highly dynamic,random and chaotic,making it rather difficult to predict the traffic flow.This dissertation,focusing on urban multiscene traffic flow prediction problems,is targeted at exploring the prediction methods applicable to multiple scenes with deep learning algorithms.The main research contents and results of this dissertation are listed as follows:(1)This dissertation introduces three scenes of urban traffic flow data collection,and analyzes the characteristics of missing,erroneous and noisy data.An improved wavelet threshold denoising method is proposed to process noisy data,which is compared with those collected via EMD denoising and wavelet threshold denoising.The results show that the improved wavelet threshold denoising effect is better.Apply the data before and after the improved wavelet threshold denoising method to train BPNN and RNN,then,input the test data into the two networks to make predictions.The results show that the prediction effect of the preprocessed data is better than that of the unprocessed data,which indicates the effectiveness of the improved wavelet threshold denoising method.(2)An improved fruit fly optimization algorithm(IFOA)is proposed and a prediction method(IFOA-DESN)which combines deep echo state network with IFOA is established for traffic flow prediction at the cross-sectional lane,where,in light of the number of vehicles in multiple sampling intervals,the number of vehicles passing the cross-sectional lane in the next sampling interval is predicted.Considering that the deep echo state network can deal with complex nonlinear characteristics,it is used to predict the cross-sectional lane traffic flow,and IFOA is adopted to optimize the hyperparameters of the deep echo state network.The single group of FOA is divided into multiple groups,and each subpopulation search in parallel to realize optimal information exchange in the subpopulation via crossover operators,thereby,to improve the global search capability of FOA.The IFOA-DESN model to predict the traffic flow of cross-sectional lane is compared with other prediction methods,and the simulation results indicate that the proposed model has better prediction effect in case of crosssectional lanes.(3)A prediction model(IFOA-WNN)which combines wavelet neural network with IFOA is proposed for traffic flow prediction of road-segments,where the number of vehicles passing through the segment at the next sampling moment is predicted in light of the number of vehicles at multiple sampling intervals.A mathematical model of the traffic flow of the road-segment is established,and the traffic flow data is calculated based on that collected from the cross-sectional lane.Considering that the time-frequency of wavelet neural network is more liable to be analyzed,wavelet neural network is used to predict the traffic flow of road-segments.In order to improve the prediction effect of the model,IFOA is used to optimize the structural parameters of wavelet neural network so as to be applicable to the data characteristics of traffic flow in the road-segments.The IFOA-WNN model to predict the traffic flow of roadsegment is compared with other prediction methods.The simulation results show that the proposed IFOA-WNN model has better prediction effect in case of road-segments.(4)A prediction model(Cheb Net-GRU)combining graph convolution network with gated recurrent unit network is proposed for traffic flow prediction of full spatiotemporal road network,where the number of vehicles passing through the monitoring points distributed in full spatio-temporal traffic network at the next sampling interval is predicted in light of the number of vehicles in multiple sampling intervals.Traffic flow in this situation changes dynamically which is superimposed on the static spatial structure of the road network,with complex spatial topology and strong temporal correlation,while the traditional spatio-temporal model cannot accurately extract the local characteristics of the road detection points.Cheb Net is used to extract spatial topology structure from traffic network,and the update and reset gates inside GRU are used to extract characteristics adaptively at various time series.The Cheb Net-GRU model to predict the traffic flow of full spatio-temporal road network is compared with other prediction methods.The simulation results indicate that the proposed Cheb NetGRU prediction model has better prediction effect in case of full spatio-temporal road network.This dissertation discusses the application of deep learning algorithms to the prediction of urban traffic flow in various scenes,and the traffic flow prediction methods are proposed for cross-sectional lane,road-segments and full spatial-temporal road network respectively.Relevant results can provide theoretical guidance and technical support for traffic management departments to predict traffic flow.