Research On Traffic Prediction Model And Spatial-temporal Fusion Optimization Based On Graph Neural Network

As an important part of realizing intelligent transportation system,traffic prediction has received extensive attention of researchers.The changing trend of traffic state is not only affected by the temporal dimension,but also dominated by the spatial structure of the road.The spatial dependencies among sensor nodes around roads for capturing traffic states are different and change over time.However,the traffic state of the road network is affected by spatial and temporal factors while being susceptible to various potential traffic factors.The complex interaction of these latent factors not only leads to differences in the spatial dependencies of the road network,but also affects the collection of traffic data by sensors.Aiming at the above problems,this thesis optimizes the traffic prediction model of spatio-temporal fusion,proposes a spatial attention to mine the spatial dependence difference of roads,and verifies that the spatial dependence difference has an impact on the prediction results.On this basis,a traffic speed prediction model based on the disentangled representation of potential traffic factors is proposed to reveal the potential factors that cause this difference and affect traffic conditions.The main contributions are as follows:(1)A traffic speed prediction model that combines spatial attention and graph convolution is proposed.Aiming at how to improve the model’s ability to mine the spatial dependence difference of the road network,this thesis designs a spatial attention mechanism,which samples the nodes through the pooling operation to obtain the aggregated features of the nodes,so that the model pays more attention to the sampled nodes.At the same time,the spatial features of the road network are obtained by using diffusion convolution.Then the spatial features and aggregated features are integrated into the convolutional gated recurrent unit(Conv GRU)to obtain the spatio-temporal features of the road network.In addition,this thesis utilizes scheduled sampling to reduce the model’s reliance on real samples during training.The experimental results show that the difference of road network spatial dependence has an impact on the prediction results,and effectively capturing the difference of spatial dependence can further reduce the prediction error of the model.(2)A traffic speed prediction model based on the disentangled representation of latent traffic factors is proposed.Aiming at how to reveal the potential traffic factors that cause spatial dependence differences,this thesis introduces the idea of disentanglement to analyze the traffic data,and uses the neighborhood routing mechanism to divide neighbor nodes into different latent spaces according to the influence of different potential traffic factors,and performs graph convolution on the neighborhood nodes in different latent spaces to obtain the spatio-temporal features of the road network.In addition,to consider both local and long-term temporal dependencies,this thesis stacks multiple layers of temporal attention convolutional networks to enable the model to perceive longer temporal features.Experimental results show that disentangled representation learning not only reveals the underlying traffic factors behind complex interactions in spatiotemporal data,but also improves the prediction accuracy of the model.