Research On Traffic Flow Prediction Method Based On Graph Convolutional Network

Accurately predicting the traffic flow in different geographic areas of a city plays a key role in intelligent transformation systems,such as intelligent transportation and public risk assessment.Due to the high nonlinearity and complexity of traffic flow,traditional forecasting methods cannot meet the requirements of mid-and long-term forecasting tasks,and they often ignore spatio-temporal correlation.Recently,some deep-learning research work suggests using convolutional neural networks to model the correlation between adjacent regions,and at the same time using recurrent neural layers in the time dimension.Although the spatial and temporal correlations have been considered in the existing methods,they cannot fully capture the unique characteristics of the traffic map.In addition,some models use multiple graph convolutional layers to capture the spatial features of traffic maps,however,as the number of layers increases,it becomes more difficult to distinguish between distant nodes and nearby nodes.When the number of layers reaches a certain level,the entire network presents a stable fixed point.At the same time,most of the current forecasting methods only focus on modeling nearby geographic correlations,while ignoring cross-regional correlations in the global context.In order to overcome these problems,in this thesis,the research of traffic flow prediction method based on graph convolutional neural network is carried out.The advantages of graph convolution and recurrent neural network and other related technologies in traffic flow prediction are analyzed in this thesis.Combined with the principle of attention mechanism,and fully considered the spatiotemporal characteristics and complexity of traffic flow,a new traffic flow prediction framework is proposed,which provides theoretical basis and feasible methods for the following specific traffic flow forecasting model research.On this basis,in this thesis,a traffic flow forecasting model based on Graph Convolutional Dynamic Recurrent Network,GCDRN)is proposed,which adopts a new graph convolution block to learn the complete spatial features,and saves the spatial features to the greatest extent through a combination of local and global.In addition,a new Dynamic Gated Recurrent Unit(DGRU)is used,where a new fusion recursive unit and adjust the recursive unit are added to the original GRU module to capture the long-range time correlation more accurately.Experiments have proved the effectiveness of this method in improving the accuracy of traffic prediction.Furthermore,based on GCDRN,an Attention-enhanced Graph Convolutional Dynamic Recurrent Network(AGCDRN)for traffic forecasting is proposed,which adds a spatial attention mechanism to local spatial features module,dynamically adjusts the importance of nearby nodes to center node,and greatly reduces the parameters and complexity of the model.In addition,a node attention unit is added to the DGRU to adaptively measure the importance of each node from the road network information at different time points,and enhance the node information.The experimental results show that this method can further improve the accuracy of traffic flow prediction.The research results of this thesis can provide new ideas for the research of traffic flow forecasting,and can also be used in practical applications,thus,has good theoretical value and broad application prospects.