Research On Short-term Traffic Flow Prediction Method Based On An Orthogonal Spatial-Temporal Graph Convolution Neural Network

With the increase of urbanization scale and population size,traffic flow prediction has become increasingly important in urban management.Accurate traffic flow prediction information can not only assist traffic managers to make reasonable traffic decisions,but also provide a strong basis for drivers to choose road travel,thus effectively alleviating traffic congestion.However,existing prediction models have some drawbacks,which hinders the improvement of traffic flow prediction accuracy.It is difficult for the existing methods to model the dynamic traffic data in the spatio-temporal dimension.However,the existing models rarely model the temporal and spatial correlation at the same time.Moreover,it is difficult for existing models to achieve accurate long-term prediction.Neural network models suffer from gradient disappearance problems in long sequence modeling tasks.As a result,the prediction performance usually drops sharply when prediction intervals increase.Therefore,extracting spatio-temporal features of traffic flow and improving prediction accuracy in long prediction intervals is the key problem faced by current traffic flow prediction methods.In response to the problem,this paper proposes a traffic flow prediction method based on an orthogonal graph convolution neural network,which can simultaneously capture the implicit temporal and spacial dependence of traffic data.Moreover,the proposed method can alleviate the problem of gradient disappearance,so as to realize accurate traffic flow prediction in a long prediction interval.The main contributions of this paper are as follows:1.Based on the liquid neural network and self-attention mechanism,this pa-per proposes multi-dimensional modeling of time series traffic data.This paper models the time correlation both in the local and global level.In this pa-per,the liquid neural network is used to encode the traffic input at each moment,and the self-attention mechanism is used to capture the global time dependence.They are combined to learn the time correlation of historical traffic data more ac-curately.2.Based on the graph convolution,this paper proposes a traffic flow pre-diction algorithm that takes the temporal and spatial correlation into ac-count.This paper designs the adjacency matrix between the road network.Based on the adjacency matrix,this paper applies the graph convolution operation on the temporal characteristics of the traffic flow.For any road,its own time series char-acteristics integrate the time series characteristics of adjacent roads.Therefore,the proposed algorithm can learn the temporal and spatial correlation of traffic flow at the same time.3.Based on the parameter orthogonalization,this paper realizes the traffic flow prediction task in a long prediction interval.This paper applies or-thogonal constraints to the neural network parameters in the algorithm.By us-ing the norm-preserving property of the orthogonal matrix,the proposed method can alleviate gradient disappearance problems and improves the prediction ac-curacy of the model in a long prediction interval.Moreover,this paper realizes parameter orthogonalization based on the Lie group and Lie algebra theory.Take self-attention mechanism as an example,this paper discusses the universality of parameter orthogonalization.Comprehensive experiments on several classical traffic data sets verified the effectiveness of the proposed orthogonal graph convolution network prediction method.Compared with other traffic flow prediction methods,the proposed method can significantly improve the accuracy of traffic flow prediction and guarantee an acceptable accuracy with a long prediction interval.