Research On Spatial-Temporal Prediction And Analysis Method Of Traffic Flow Based On Deep Learning

The rapid increase in the number of motor vehicles has led to increasingly serious problems such as traffic congestion and exhaust emissions,and urban traffic management is facing arduous challenges.Intelligent Traffic System(ITS)can realize traffic control and traffic guidance,which is an effective way to solve traffic congestion and improve road capacity.Accurate prediction of traffic flow-related factors is the key task to assist ITS in effectively carrying out traffic control and guidance activities.Therefore,this paper focuses on the task of traffic flow forecasting.Firstly,the research background and significance of the traffic flow forecasting task are introduced,and the domestic and international status quo of single-point traffic flow forecasting and road network traffic flow forecasting are analyzed respectively.Then the literature is summarized,and the research blank in the traffic flow forecasting task is obtained as the theoretical basis of this paper,and the main research contents and technical routes of each chapter of this paper are determined accordingly.Secondly,taking expressway and urban road networks as the research objects,the collection methods of traffic flow are introduced,and data preprocessing such as missing value filling and normalization are carried out,which provides data support for the research of the traffic flow prediction model.In order to analyze the temporal and spatial characteristics of traffic flow parameters,the traffic flow parameters in different time periods,different sampling points and different types are compared,and their correlation coefficients are calculated.The results show that the traffic flow of expressway and urban road network has daily correlation and weekly correlation in time characteristics.In terms of spatial characteristics,the correlation of traffic flow parameters between sections is related to the accessibility relationship and the distance between sections;In addition,many factors of traffic flow show correlation,which provides theoretical support for the construction of traffic flow forecasting model.Thirdly,according to the deficiency of current research and the analysis of the spatiotemporal characteristics of traffic flow,appropriate deep learning algorithms are selected,including convolutional neural network,cyclic neural network,graph-volume product neural network,etc.,and the basic theory of the algorithms is discussed,which provides technical support for the construction of short-term traffic flow forecasting models of expressways and road networks.In addition,aiming at the multi-step forecasting task of single-point traffic flow,a combined forecasting model of convolutional neural network and gated circulation unit is proposed.By applying the encoder-decoder framework to alleviate the step-by-step accumulation of errors,the multi-step forecasting of traffic flow series is realized,and the feature extraction of multiple traffic flow time series is used to improve the forecasting effect of the model.In terms of model evaluation,the effect of the model and the availability of each module are verified on the data set to ensure the prediction accuracy of the model and provide model support for the multi-step prediction task of single-point traffic flow for traffic management departments.Finally,aiming at the task of road network traffic flow forecasting,this paper proposes a spatio-temporal multi-head graph attention network,which considers the interaction between traffic flow factors in the input dimension,optimizes the model in the time and space dimensions respectively,and in the time dimension,uses the full convolution structure instead of the circular neural network to capture the characteristics of the time dimension,thus solving the problem of error accumulation step by step.At the same time,it uses the gated linear unit to realize parallel calculation.In the spatial dimension,the traffic network is modelled as a space-time directed graph,and the accessibility relationship and distance relationship between road nodes are taken into account.The multi-head graph attention network is used to capture the characteristics of the spatial dimension,so that each attention mechanism can deal with a subspace separately,which improves the expression ability of the model for the spatial dimension.In the aspect of model evaluation,the model effect and the availability of each module are verified by expressway data set and urban road network data set,which ensures the prediction accuracy of the model,broadens the application scenarios of the model,and provides model support for the traffic management department’s expressway and urban road network traffic flow prediction task.