智能系统学报2025，Vol.20Issue(4)：776-786,11.DOI:10.11992/tis.202402012
基于时空动态图的交通流量预测方法研究
Research on traffic flow prediction method based on spatial-temporal dynamic graph convolutional network
摘要
Abstract
To address the limitations of existing traffic flow prediction methods in modeling spatio-temporal data and capturing dynamic spatial correlations,a spatio-temporal dynamic graph network(STDGNet)is proposed.This model adopts an encoder-decoder architecture with an embedding layer and utilizes a dynamic graph generation module to un-cover potential spatio-temporal relationships from a data-driven perspective,reconstructing the dynamic correlation graph of nodes at each time step.The embedding layer employs a spatio-temporal adaptive embedding method to model the intrinsic spatio-temporal relationships and temporal information of traffic data.The encoder part uses a spatio-tem-poral memory attention mechanism to model spatio-temporal features from a global perspective.The decoder part incor-porates a graph convolution module into a recurrent neural network to simultaneously capture temporal and spatial de-pendencies and output future traffic conditions.Experimental results show that,compared with the optimal baseline model decoupled dynamic spatial-temporal graph neural network(D2STGNN),the proposed model reduces mean abso-lute error by an average of 1.63％and reduces model training time by nearly 2.5 times.This study effectively improves the accuracy and efficiency of traffic flow prediction,providing strong support for the development of intelligent trans-portation systems.关键词
交通流量/时空数据/混合模型/注意力机制/时空动态图/图卷积神经网络/循环神经网络/深度学习Key words
traffic flow/spatio-temporal data/hybrid model/attention mechanism/spatio-temporal dynamic graph/graph convolutional neural network/recurrent neural network/deep learning分类
信息技术与安全科学引用本文复制引用
孟祥福,谢伟鹏,崔江燕..基于时空动态图的交通流量预测方法研究[J].智能系统学报,2025,20(4):776-786,11.基金项目
国家自然科学基金面上项目(61772249). （61772249）
