一种基于相互聚合图注意网络的交通流量预测算法  

A Mutually Aggregated Graph Attention Network Based Approach for Traffic Flow Prediction

作　　者：刘雪霞 Liu Xuexia(Zhejiang Business College,Hangzhou,China)

机构地区：[1]浙江商业职业技术学院,浙江杭州

出　　处：《科学技术创新》2025年第12期84-87,共4页Scientific and Technological Innovation

基　　金：2024年浙江省教育厅一般科研项目资助(编号:Y202456024);2022年浙江省教育厅访问工程师校企合作项目(FG2022102)。

摘　　要：交通流预测是智能交通系统的关键部分,传统预测方法采用时间序列法作为解决方案,其性能无法满足。基于图神经网络(GNN)的方法缺乏考虑连续时间阶段相邻图节点间传递的信息,限制潜在的图异质性建模;基于卷积神经网络CNN的方法在长序列预测任务中存在梯度消失问题。本文提出了一种相互聚合图注意力网络(MAGAN),关注相邻图节点在连续时间阶段的时空相关性,且以自注意为基础,具备长时间范围的预测能力,并通过大量实验验证其有效性。Being as a key part of Intelligent transportation system,Traditional traffic flow prediction approaches utilize time series method as the solution,whose performance cannot be satisfied.The existing GNN based approaches rarely consider the mutually information transmitted between neighboring graph nodes during consecutive time phases,which restrict the potential graph heterogeneity modelling.The Recurrent Neural Network(RNN)suffers the gradient vanishing problem on long sequence prediction tasks.We build a novel model,called as Mutually Aggregated Graph Convolutional Network(MAGAN),which essentially focusing on the spatial temporal correlations of neighboring graph nodes during consecutive time phases,and utilize self-attention as the base to enable the prediction ability of long horizon.Extensive experiments are conducted to validate the efficiency of MAGAN.

关 键 词：图神经网络 自注意机制 时空相关性 

分 类 号：U491.112[交通运输工程—交通运输规划与管理]