基于动态时间自注意力机制的交通预测

Insulator defect recognition based on cross-level feature fusion and channel and spatial attention mechanism

针对智能交通系统(Intelligent transportation system,ITS)和出行导航中的实时交通预测问题,提出了一种动态时间自注意力图卷积网络(Dynamic Temporal Self-attention Graph convolutional Network,DT-SGN)模型。现有的交通预测方法大多使用简单的由0和1组成的邻接矩阵来捕捉空间依赖性,无法准确描述城市道路网络的拓扑结构和随时间变化的规律。为解决这个问题,将邻接矩阵视为可训练的注意力得分矩阵,并根据不同的输入调整网络参数。基于空间机制选择自注意力图卷积网络(Self-attention Graph convolutional Network,SGN)来捕捉空间依赖性,基于时间机制选择动态门控循环单元(Dynamic-GRU,DGRU)来捕捉时间依赖性并学习输入数据的动态变化。在SZ-taxi好Los-loop数据集上,实验证明所提方法在真实交通数据集上表现优于基于模型和数据驱动的对比方法。

This paper addresses the problem of real-time traffic prediction in Intelligent Transportation Systems(ITS)and travel navigation guidance.It proposes a Dynamic Temporal Self-Attention Graph Convolutional Network(DT-SGN)model.Existing traffic prediction methods often use a simple adjacency matrix consisting of 0 and 1 to capture spatial dependencies,which cannot accurately describe the topological structure of urban road networks and their temporal dynamics.To tackle this issue,this paper treats the adjacency matrix as a trainable attention score matrix and adapts network parameters to different inputs.Specifically,the Self-Attention Graph Convolutional Network(SGN)is employed to capture spatial dependencies,while the Dynamic Gated Recurrent Unit(DGRU)is utilized to capture temporal dependencies and learn the dynamic changes in input data.Experimental results on the SZ-taxi and Los-loop dataset demonstrate the superiority of our proposed method over model-driven and data-driven comparative approaches when applied to real-world traffic datasets.