Research On Traffic Flow Prediction Methods Based On Adaptive Graph Convolutional Networks

Traffic prediction is an important issue in urban traffic management and planning.Traffic congestion,accidents,and other issues not only have an impact on the economic development of cities and the quality of life of citizens,but also lead to wasteful energy consumption and environmental contamination.Therefore,accurately predicting traffic flow and congestion can provide an essential reference and decision-making basis for urban traffic management and planning.However,due to the complex spatiotemporal correlation of traffic flow data,and the susceptibility to various unexpected traffic conditions,traffic flow prediction tasks have always been considered a challenging research topic.Existing traffic flow prediction models struggle to adequately capture the hidden spatiotemporal characteristics and nonlinear relationships present in traffic data,with their modeling of the spatiotemporal dependence relationship of data being inadequate.To address this issue,this paper constructs two traffic flow prediction models based on spatiotemporal adaptive graph convolutional neural networks.Specifically,the main research contents of this paper are as follows:(1)A recurrent neural network model based on adaptive graph convolution is proposed.Existing traffic flow prediction models based on graph convolutional networks typically use predefined graph structures to extract spatial correlations between nodes.However,the topology of real road networks is often very complex and it is difficult to accurately quantify the correlations between nodes in advance.In addition,existing models have poor performance in long-term traffic flow prediction.To address these issues,this model introduces two adaptive modules to enhance the functionality of traditional graph convolutional networks for better capturing spatial relationships.Firstly,the adaptive adjacency matrix generation module automatically infers the graph structure based on the input data.Then,the node adaptive parameter learning module captures node-specific patterns for each traffic sequence.Finally,a long short-term memory network is used to capture the temporal dependencies of the data.(2)An adaptive graph learning model based on spatiotemporal neural networks is proposed.Firstly,a novel graph learning module is designed,which can adaptively capture more possible correlations between nodes during the training process,providing more precise spatial dependencies for the prediction module.Then,the parameters of the graph structure and prediction network are optimized separately using an alternating training method.At the same time,a carefully designed loss function is employed to ensure the sparsity of the generated relationship matrix.Finally,the graph structure is updated by a weighted sum method.The prediction module of this model can be used with most traffic flow prediction models based on graph convolutional networks.(3)Finally,the performance of the two proposed traffic flow prediction models was evaluated on publicly available traffic datasets.Compared with some classic models in traffic flow prediction research,the traffic flow prediction models based on spatiotemporal adaptive graph convolutional neural networks proposed in this paper achieved better prediction accuracy.