Long-term Traffic Flow Forecasting Method Based On Spatial-temporal Graph Neural Network

Traffic prediction is an important part of intelligent transportation system and smart city management.Accurate traffic prediction can be deployed in urban applications such as road congestion alerts and route planning,thus providing appropriate and sustainable services to the public or relevant authorities.With the rapid development of modern intelligent transport systems and the increasing demand for various types of traffic data,the need for accurate traffic prediction has become an important issue that needs to be addressed.Traffic data is collected by sensors at fixed geographic locations at constant time intervals,which is typical of spatial-temporal data.Traditional time series models and existing deep learning methods mostly target temporal dependence and spatial dependence for modelling,neglecting key patterns in various types of spatial-temporal data,such as periodic patterns and spatial-temporal heterogeneity patterns,resulting in large long-term prediction errors that cannot meet practical application needs.In response to the above challenges,this thesis proposes three spatial-temporal graph neural network prediction models for effective prediction of road traffic conditions in traffic networks.The main works of the thesis include.(1)A multi-scale spatial-temporal fusion graph network MFSTGN is proposed in order to effectively learn the spatial-temporal correlation and spatial heterogeneity of traffic data.In particular,it has designed a spatial-temporal graph convolution module which dynamically models spatial-temporal correlation while preserving the inherent structure of the traffic network,and describes the trend variation of traffic flows through a kind of trend graph convolution,as well as modelling the spatial heterogeneity of the traffic network using spatial-temporal embedding.In addition,the time dependence is modelled explicitly periodically by introducing weekly and daily series,and a gated attention module is proposed to fuse the periodicity features.Experimental results on four traffic datasets show that the MFSTGN model improves the prediction performance over long term periods.(2)To effectively learn the long-term temporal dependence and spatial heterogeneity in traffic data and to reduce the impact of temporal heterogeneity,a trend graph attention network TGAN is proposed.In this context,a trend spatial attention module has been designed to construct a dynamic graph structure on the spatial through trend-trend pattern.Its main idea is to transfer information between nodes with similar attributes,thus solving the problem of spatial heterogeneity.For modelling the long-term temporal dependence,a trend construction module was introduced to construct local and global trend blocks,fully considering the local trends around consecutive value points at different scales,and then performing aggregation operations between time steps and trend blocks,which not only allows each time step to enjoy a local and global view,but also enables to reduce the impact of temporal heterogeneity effectively.Experimental results on five real-world traffic datasets show that the TGAN model outperforms most existing methods in terms of efficiency and performance for long-term prediction.(3)Aiming to learn both long-short term temporal dependence,and local and global spatial correlation in time series with heterogeneity,a parallel spatial-temporal Transformer architecture,ISTNet,is proposed.Within it a parallel temporal module is designed to explicitly transfer the advantages of convolution and maximum pooling for capturing local information and attention for capturing global information to the Transformer.Secondly,spatial dependence are handled at different levels of granularity by considering both local and global semantic information on the spatial level through parallel spatial modules.Lastly,the parallel temporal and parallel spatial modules achieve greater efficiency through a channel separation mechanism.Experimental results on six real-world traffic datasets show that the ISTNet model leads most existing methods in terms of efficiency and performance for long-term prediction.