Research On Traffic State Prediction Algorithm Based On Graph Neural Network

The spatiotemporal prediction has become one of the research hotspots in recent years,and it has been widely used in various fields.Traffic state prediction,as a typical spatiotemporal prediction task,has also gained a lot of attention.Traffic state prediction refers to inputting road network and traffic historical state data into the model,and outputting the prediction result of future traffic state after calculation,where the state can be vehicle speed,traffic flow,etc.A good traffic state forecast can not only help the government build a "smart city",but also facilitate peoples lives.Artificial intelligence technology is changing with each passing day,providing new ideas and methods for spatiotemporal prediction tasks,and more and more traffic state prediction research is based on deep learning.Based on a variety of classic methods in deep learning,the thesis builds an end-to-end multi-step prediction model for the difficulties and challenges of traffic state prediction,and achieves effective and accurate traffic state prediction.The main contributions of the thesis are as follows:(1)Adaptive Dynamic Graph Convolutional Networks.To deal with the complex spatiotemporal features in traffic state data,the thesis constructs an end-to-end model: Using graph convolutional networks to capture spatial correlations in traffic topology graph,and using recurrent neural networks to extract temporal dependencies in time series data.Innovative improvements are made to the graph convolutional network to better handle the latent information between node pairs and the difference information between consecutive time points.A convolutional neural network is then used to generate multi-step prediction results.Extensive experiments on two real-world public datasets validate the effectiveness of the proposed neural network.(2)Attribute Augmentation Adaptive Dynamic Graph Convolutional Networks for and Independent Attribute Augmentation Modules.This thesis proposes a time module to address the time periodicity in traffic data,and introduces the time module in the previous work to improve the feature extraction ability and prediction accuracy of the model.Subsequently,an independent attribute augment cell is further proposed,which can act on many deep learning methods to strengthen these methods through attribute augment to improve accuracy at a cost.Theoretical analysis and experimental results prove the effectiveness of attribute augment for deep learning methods.