Research On Traffic Flow Forecasting Based On Spatial-Temporal Attention Networks

With the rapid development of urbanization,traffic problems are becoming more and more serious,especially traffic congestion and low utilization rate.Traffic flow forecasting aims to forecast the future state of road networks according to the historical data in the traffic network,which can help users to take effective countermeasures in advance.So it has become one of important methods to solve related traffic problems.Due to the complex spatial-temporal dependence of traffic flow data and the complex interaction between spatial-temporal features,dealing with massive nonlinear and dynamic traffic data and forecasting has always been a challenging problem in the field of intelligent transportation.At present,the focus of traffic flow forecasting is to reduce the error.This paper reduces the error from three aspects:improving the ability of capturing the spatial-temporal dependence,establishing a model that can learn the relevance between spatial and temporal features,using more kinds of historical data.This paper constructs spatial-temporal attention networks to capture the spatial-temporal dependence better by multi-head attention models and two kinds of graph neural networks.In the model,a gated fusion mechanism is designed to fuse temporal modules and spatial modules for capturing the correlation between spatial and temporal features.Finally,using long-term and short-term historical data to achieve more accurate traffic flow forecasting by modifying transformer models.The main work of this paper is as follows:(1)A kind of spatial-temporal attention networks is constructed to capture the spatial-temporal dependence of traffic flow data.The method uses multi-head attention models to construct temporal modules.It can capture the temporal dependence well and avoid using recurrent neural networks commonly used in most related studies.The proposed model’s temporal modules in this paper are more convenient for parallel computing than works using recurrent neural networks.According to Non-European data structure of traffic network data,this paper uses multi-head attention models,graph convolution networks and graph attention networks to construct spatial modules for capturing spatial dependence.(2)Two gated fusion networks are designed,one is used to fuse spatial-temporal modules for learning the correlation between spatial-temporal features,and the other is used to fuse two graph neural network modules to further improve the ability of processing Non-Euclidean data.This paper considers that the spatial-temporal features of traffic data are not independent but have certain relevance.Therefore,a gated fusion network is used to fuse spatial modules and temporal modules for learning the interaction between spatial-temporal features.In spatial modules,a gated fusion network is used to fuse graph attention networks and graph convolution networks to enhance the ability of processing Non-Euclidean data.Finally,experiments on two real-world datasets demonstrates that spatial-temporal attention networks based on gated fusion achieve good results.(3)On the basis of gated fusion based spatial-temporal attention networks,realizing traffic flow forecasting method using long-term and short-term historical data by modifying transformer model,which can further improve the accuracy of forecasting.General traffic flow forecasting researches only use short-term historical data,but this paper uses long-term and short-term historical data.Introducing transformer in the field of natural language processing into the field of traffic flow forecasting,and extracting temporal and spatial features of long-term and short-term historical data by multi-head attention models of transformer.Finally,experimental results demonstrate that the forecasting strategy using long-term and short-term historical data can further improve the accuracy on the basis of gated fusion based spatial-temporal attention networks.