Short-term Traffic Flow Prediction On Road Network Based On Graph Attention Network

Intelligent transportation system(ITS)plays an important role in urban traffic planning,management and control.In ITS,traffic flow prediction is an important part.At present,the mainstream traffic flow prediction research scheme is based on capture of the time series feature of a single road,and does not effectively use the road network spatial feature of the traffic flow.In recent years,with the rapid development of computer and computing methods,the research of deep learning is becoming more and more popular.In this paper,the deep learning algorithm is used to mine the road network spatial feature and time series feature of traffic flow and then accurately predict traffic flow to provide algorithm support for ITS.The main work of this paper is as follow:1.This paper introduces the related theory and main parameters of traffic flow in detail,and deeply analyzes the data features of traffic flow.The mathematical model and evaluation metric of traffic flow prediction are established,which lays the foundation for follow-up research and experimental work.2.Aiming at the shortcoming of traditional traffic flow prediction methods that do not effectively utilize road network spatial feature of traffic flow,this paper proposes to regard the road network as a topological graph structure,in which the traffic flow on the road is the feature of nodes.First of all,the Graph Attention Network is used to weighted sum the neighborhood node features of the observation node,extract the traffic flow features of the observation node and the influence of the neighborhood node on it.Secondly,the Fully Connected Network is used to integrate the collected traffic flow feature and adjust the output dimension to obtain the predicted value.Finally,by comparing the real highway dataset with the Autoregressive Integrated Moving Average Model,Support Vector Regression and Long Short-Term Memory neural network,the experimental results show that short-term traffic flow prediction model based on Graph Attention Network has significant improvement in accuracy,the root mean square error and mean absolute error are smaller,the value of the coefficient of determination and explained variance score is higher,which shows that the model has stronger ability to explain the dependent variable,and can effectively describe the change law of traffic flow.3.Aiming at the shortcoming that the Graph Attention Network ignores the time series feature of traffic flow,this paper proposes to construct the Spatiotemporal Graph Attention Network to obtain the road network spatial feature and time series feature of traffic flow at the same time.First of all,the road network is regarded as a topological graph structure,in which the traffic flow on the road is the feature of nodes.And the road network spatial feature and time series feature of the traffic flow are extracted through the Spatiotemporal Graph Attention Network.Secondly,the output of the Spatiotemporal Graph Attention Network is adjusted through the Global Time Attention Network to make the model focus on the features that play a key role.Finally,the Fully Connected Network is used to integrate the output of Global Time Attention Network and adjust the output dimension to obtain the predicted value.The experimental results on real highway dataset show that: compared with the Gated Recurrent Unit neural network and the Graph Attention Network,short-term traffic flow prediction model based on Spatiotemporal Graph Attention Network has stronger ability to extract time series feature and road network spatial feature.