Research On Short-term Traffic Flow Prediction Method Based On Graph Neural Network

With the continuous development of China’s economic construction,the huge growth of private cars has brought tremendous pressure on the roads.The introduction of intelligent transportation system(ITS)has greatly alleviated this phenomenon,and traffic flow prediction is one of the most important things to realize ITS.Accurate and efficient traffic flow prediction can respond to the traffic road conditions in advance,giving managers sufficient decision scheduling time to cope with changes in traffic conditions,and can also provide some help for citizens to travel.Traffic flow data is not purely serial data,but contains both spatial and temporal characteristics.Some existing studies simply use serial feature extraction models to extract the temporal characteristics of traffic flow alone,ignoring the spatial correlation of road networks,which makes the accuracy of prediction greatly reduced.In this paper,we address this situation by introducing graph neural networks,combined with recurrent neural networks,in order to simultaneously uncover the spatio-temporal correlation features of spatio-temporal flow data.The specific work is as follows:A Graph Attention Network GAT is introduced to extract spatial information of the road network in traffic flow,and then combined with a fast recurrent neural network variant SRU to form an SR-GAT model for the traffic flow prediction task.The attention mechanism of GAT can fully fuse the feature information of road nodes as well as their neighbors to extract spatial features,which can improve the prediction accuracy.SRU has the same ability as other gated RNNs to ability of extracting temporal features,but due to its improved structure,it largely alleviates the serial processing in time and greatly improves the speed.On the basis of SR-GAT,an improved Bi SR-DGAT model was constructed by introducing dynamic attention mechanism and bidirectional mechanism.The static attention mechanism in GAT only pays attention to the features of the same neighbor in a fully bipartite graph-connected structure,while the dynamic attention mechanism can dynamically pay attention to the features of different neighbors in this structure,and the traffic road network contains many fully bipartite graph topologies.The introduction of dynamic attention mechanism can improve the representational capability of the model and thus improve the accuracy.The bidirectional SRU captures the temporal characteristics of traffic flow data from both forward and reverse directions,and the whole sequence information can be obtained at each time step,thus improving the prediction capability of the model.To verify the accuracy and efficiency of the model in this paper,validation is performed on a real dataset PEMSD7.The evaluation metrics of RMSE,MAE and MAPE are selected to evaluate the accuracy of this paper’s model and the baseline model,and the training time is used to evaluate the efficiency of the model.The experimental results demonstrate that the time efficiency of the SR-GAT model is better than other baseline models,and the 15-min prediction accuracy is better than other baseline models,and the 30-min and 45-min prediction accuracies are on par with other models.The accuracy of the improved Bi SR-DGAT model is better than that of the SR-GAT model before improvement and other baseline models.The experimental results show that the model in this paper can fully extract the spatio-temporal characteristics of the road,and the accuracy and efficiency are better.