Research On Traffic Flow Prediction Method Based On Graph Convolutional Network

With the continuous economic development,the increasing surge of per capita vehicle ownership has brought a severe challenge to traffic control and management,making the urban traffic system under great pressure.The intelligent traffic system relieves the burden of artificial control of road network traffic patterns and is an effective tool to solve the problems of traffic congestion,traffic safety and environmental pollution,playing a key role in the construction and development of smart cities.Traffic flow prediction,as an important part of traffic guidance in intelligent traffic system,aims to predict the future traffic condition of the road network based on historical traffic data,and then improves the overall operation capacity as well as operational efficiency of the traffic system.However,there are still some weaknesses in the field of traffic flow prediction: on the one hand,the nonstationarity of traffic flow is obvious,and thus a single prediction model cannot specify the features of trend and detail terms of traffic flow;on the other hand,most current methods do not deeply explore the rich spatial semantic features and dynamically changing heterogeneous information among nodes in a road network,and have low performance when predicting long-term traffic flow.To address the above problems,two traffic flow prediction algorithm models based on graph convolutional network are proposed in this thesis,and the main research contents are as follows:(1)In order to deeply depict the spatio-temporal features of traffic flow,a DWT-GCNTAN model for traffic flow prediction is proposed.Firstly,the original traffic sequence can be decomposed into detailed and approximate components by Discrete Wavelet Transform(DWT)to reduce the impact of nonstationarity of traffic flow on prediction results.Secondly,distance factor term and adaptive matrix can be introduced to model the spatial association of road network,and the spatial contextual feature information can be captured by graph convolutional network.Meanwhile,a Temporal Channel Attention Network(TAN)is proposed to extract the temporal features of traffic flow in bilateral time dimension,and finely describe the features attention of different moments,and then optimize the model prediction performance.Finally,the output of each component model can be fused to obtain the final prediction results.(2)In order to fully model the spatio-temporal correlation of traffic flow in a road network,a spatial temporal graph neural network is proposed for traffic flow prediction.Firstly,the spatial dependency between nodes can be dynamically updated by the spatial attention mechanism,and the spatial correlation between upstream and downstream of road sections can be portrayed deeply by introducing diffusion graph convolutional network.Secondly,the longrange temporal features of traffic flow can be captured by adjusting the dilation factor of the one-dimension dilation convolutional network,and the gating mechanism can be used to control updating and forgetting of the features.Meanwhile,the spatio-temporal feature fusion mechanism can be constructed to further enhance the capability of the proposed model to capture the complex spatio-temporal features of traffic flow.Finally,the feature mapping of external factors can be accomplished through the fully connected layers,and then the influence of external factors on traffic flow prediction is integrated to improve the model prediction performance.(3)Experiments are designed on two publicly available datasets,PeMS04 and PeMS08,and the results show that the two models proposed in this thesis,with MAE,RMSE and MAPE,perform better than other baseline models,and can be effectively used in urban road network traffic flow prediction.