基于注意力机制的多图神经网络交通预测模型  

Attention-based multi-graph neural network for traffic prediction

作　　者：李博 崔高峰 冯泽涛 LI Bo;CUI Gaofeng;FENG Zetao(College of Computer Science and Technology,Shandong Technology and Business University,Yantai 264005,Shandong,China;College of Statistics,Shandong Technology and Business University,Yantai 264005,Shandong,China)

机构地区：[1]山东工商学院计算机科学与技术学院,山东烟台264005 [2]山东工商学院统计学院,山东烟台264005

出　　处：《智能计算机与应用》2025年第3期56-63,共8页Intelligent Computer and Applications

摘　　要：由于复杂的时空相关性和非线性的交通模式,实现精确的预测仍然是一个挑战。针对于此,本文提出了一种新颖的多变量时间序列预测框架(MSTGCN+AL),尝试使用多图神经网络进行预测。引入2种新的图类型,一是通过自适应邻接矩阵得到自适应邻接图,能更好地获取交通节点之间的位置关系;另一个是潜在图,通过使用全局变量拟合三角函数,可以更好地从交通数据中提取周期性和上下文信息。为了对齐图节点及其时间戳,采用了一个基于注意力机制的多图融合模块,包括多图空间嵌入、空间注意力和图注意力。为了验证本文方法的有效性,在METR-LA数据集上进行了广泛的实验。实验结果表明,与基线方法相比,所提MSTGCN+AL模型在预测性能上表现更佳。Accurate traffic flow prediction is essential for intelligent transportation systems;however,due to intricate spatiotemporal correlations and nonlinear traffic patterns,attaining precise predictions remains challenging.To address these issues,this paper presents a novel framework for multivariate time series prediction(MSTGCN+AL),attempting to use multi-graph neural networks for prediction.The paper has introduced two new types of graphs.One is the adaptive adjacency graph obtained through an adaptive adjacency matrix,which can better capture the positional relationships between transportation nodes.The other is the latent graph,which utilizes global variables to fit trigonometric functions,enabling it to extract periodicity and contextual information from traffic data more effectively.To align graph nodes and their timestamps,the paper devises a multi-graph fusion module based on attention mechanisms,including multi-graph spatial embedding,spatial attention,and graph attention.To validate the effectiveness of the propsed approach,extensive experiments are conducted on the METR-LA dataset.The experimental results show the superior predictive performance of the proposed MSTGCN+AL model when compared to baseline methods.

关 键 词：图神经网络 自适应邻接矩阵 注意力机制 多图融合 交通预测 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]