Research On Spatiotemporal Modeling And Prediction Of Traffic Flow Based On Deep Learning

With the continuous development of urbanization,the number of motor vehicles has steadily increased,and a series of problems such as traffic congestion and accidents have become very prominent.Utilizing massive urban traffic data such as road sensors,taxis,private cars,and public transportation to conduct big data analysis and use it for traffic control,travel mode selection,and road planning has become an indispensable part of the development of smart cities.However,due to the complex temporal and spatial dependencies of traffic data,accurate and reliable traffic prediction has always been a challenging task.Based on the theory of deep learning,this thesis proposes a traffic data restoration model,and two spatiotemporal combination traffic flow prediction models from the perspectives of road level and regional level to address some of the current problems in the field of traffic flow prediction.The main research findings are summarized as follows:(1)A data restoration model based on three traffic flow parameters has been proposed.Firstly,the process of preprocessing global positioning system(GPS)trajectory data was introduced,including data filtering,coordinate system conversion,map matching,etc;Then,based on the division of road sections at traffic intersections,a road network structure diagram is constructed,and the parameters of traffic flow,average speed,and traffic flow density are extracted by road sections;Secondly,a data restoration model is proposed to address the issues of incomplete feature extraction and neglect of data spatial features in current data restoration methods;Finally,the effectiveness of the proposed data repair method was verified through publicly available datasets.(2)A traffic flow prediction model based on the fusion of spatiotemporal features and attention mechanism has been proposed.Firstly,a weighted fusion method for traffic flow three parameters was proposed,which integrates the characteristics of traffic flow,average speed,and traffic flow density;Secondly,the fusion traffic flow sequence is divided into closeness sequences,period sequences,and trend sequences;Then,in response to the current feature extraction methods ignoring the issue of dynamic changes in traffic networks,a combination model of residual network(Res Net)and graph attention network(GAT)is used to deeply extract dynamic spatial features in traffic flow in spatial feature extraction.In temporal feature extraction,a bidirectional simple recurrent unit(SRU)is used,and a temporal attention mechanism is introduced to fully extract dynamic temporal features of traffic flow;Finally,the performance of the model was evaluated and validated through publicly available datasets.(3)A regional traffic flow prediction model based on deep combination modeling has been proposed.Firstly,in order to solve the problem of heavy workload and storage overhead in the regional large-scale traffic flow prediction based on graph theory,the regional grid is processed and the traffic flow is calculated in grid computing;Then,select closeness sequences,period sequences,and trend sequences,and design different shallow feature extraction methods for different sequences;Secondly,in the temporal feature extraction,the gate recurrent unit(GRU)is used to extract the temporal features of the traffic sequence.In the spatial feature extraction,aiming at the problem that the effective depth of the traditional Res Net is not enough,which leads to the small receptive field,the self-calibrated convolutions(Sc Net)is introduced to form the SCRes Net network,and the spatial features are extracted in depth through multi-layer SCRes Net;Finally,the effectiveness of the proposed model was verified through experiments.