计算机工程与应用2026，Vol.62Issue(17)：353-362,10.DOI:10.3778/j.issn.1002-8331.2507-0135
DFSTNet:动态图结构与周期性挖掘的交通流预测
DFSTNet:Traffic Flow Prediction with Dynamic Graph Structure and Periodicity Mining
摘要
Abstract
Addressing the key limitations of existing deep learning models in traffic flow prediction,such as difficulties in capturing dynamic changes in road network structures,insufficient utilization of temporal information,limited prediction accuracy,and delayed response to unexpected events,this paper proposes a novel traffic flow prediction model named dynamic fourier spatio-temporal network(DFSTNet).The model first employs a gated TCN to extract short-term depen-dency features from traffic flow data.Subsequently,a spatial module is designed to construct a dynamic adjacency matrix using cosine similarity,which preserves the static topological structure of the road network while dynamically perceiving real-time changes in road structures,thereby overcoming the constraints of traditional models that rely on fixed adjacent nodes.Additionally,a periodicity module is introduced to extract periodic features of traffic flow data,and a temporal Transformer is utilized to effectively capture long-term temporal dependencies.Finally,experiments on two real-world traffic datasets,PEMS04 and PEMS08,demonstrate that the DFSTNet model significantly outperforms existing baseline models,with mean absolute errors reduced by 50.69%,46.91%,41.16%,34.89%,19.71%,16.33%,and 10.92%,respectively.关键词
交通流预测/门控TCN/动态邻接矩阵/TransformerKey words
traffic flow prediction/gated TCN/dynamic adjacency matrix/Transformer分类
信息技术与安全科学引用本文复制引用
申博彦,王志文,王浩旭,令国壁,姚子珩,王梦泽,成晓龙..DFSTNet:动态图结构与周期性挖掘的交通流预测[J].计算机工程与应用,2026,62(17):353-362,10.基金项目
国家自然科学基金(62263019). （62263019）
