基于注意力和线性层融合的动态图卷积交通量预测模型OA
Dynamic graph convolutional traffic flow prediction model based on attention and linear layer fusion
交通量的精准预测是优化路网运行效率、缓解城市交通拥堵的关键.针对传统模型依赖预定义静态图结构、难以捕捉动态时空相关性以及单一时间尺度建模难以全面提取多尺度特征的问题,提出了双重动态自适应时空建模框架.该框架在时间维度采用动态时间特征提取多头注意力机制,自适应调整时序权重以捕捉关键动态特征;在空间维度设计动态图卷积网络,通过自注意力机制实时生成邻接矩阵,以表征节点间动态空间依赖关系,从而实现时空双重动态协同建模.此外,该框架引入可学习的线性融合层,自适应整合多时间尺度预测结果,协同优化局部与全局特征表达.在真实道路数据集上的实验表明,该框架显著优于基线模型,验证了其优越的时空特征捕捉与预测性能.
Accurate traffic volume prediction is crucial for optimizing the operational efficiency of road networks and alleviating urban traffic congestion.Traditional models rely on predefined static graph structures,making it difficult to capture dynamic spatiotemporal correlations,and single-time-scale modeling struggles to comprehensively extract multi-scale features.To address these issues,a dual dynamic adaptive spatiotemporal modeling framework is proposed.In the temporal dimension,this framework employs a dynamic time feature extraction multi-head attention mechanism to adaptively adjust temporal weights to capture key dynamic features.In the spatial dimension,it designs a dynamic graph convolutional network that generates an adjacency matrix in real-time through a self-attention mechanism to represent dynamic spatial dependencies between nodes,thus achieving spatiotemporal dual dynamic collaborative modeling.Furthermore,this framework introduces a learnable linear fusion layer to adaptively integrate multi-time-scale prediction results and collaboratively optimize local and global feature representations.Experiments on real road datasets demonstrate that this framework significantly outperforms the baseline model,validating its superior spatiotemporal feature capture and prediction performance.
尉辉;肖洪波;邹北骥;奎晓燕;肖捡花;和佳聚;合尼古力
新疆交通职业技术学院人工智能工程学院,新疆 乌鲁木齐 831401中南大学计算机学院,湖南 长沙 410012怀化学院计算机与人工智能学院(软件学院),湖南 怀化 418000中南大学计算机学院,湖南 长沙 410012中南大学计算机学院,湖南 长沙 410012怀化学院计算机与人工智能学院(软件学院),湖南 怀化 418000怀化学院计算机与人工智能学院(软件学院),湖南 怀化 418000
信息技术与安全科学
交通量预测动态时间特征提取多头注意力机制动态图卷积线性层融合
traffic volume predictiondynamic time feature extraction with multi head attention mechanismdynamic graph convolutionlinear layer fusion
《大数据》 2026 (1)
126-145,20
国家自然科学基金项目(No.62202198、No.U22A2034、No.62177047)湘江实验室重点项目(No.23XJ02005)湖南省自然科学基金项目(No.2024JJ7372)湖南省教育厅科学研究重点项目(No.24A0550、No.24A0018)湖南省教育厅教学改革项目(No.202401001340、No.202502001316)湖南省科技厅重点研发计划(No.2024JK2135)湖南省普通高等学校科技创新团队支持项目"武陵山片区智慧农业信息处理与控制技术"(No.ZNKZD2024-3)新疆维吾尔自治区自然科学基金项目(No.2024D01A52)中南大学前沿交叉项目(No.2023QYJC020) The National Natural Science Foundation of China(No.62202198,No.U22A2034,No.62177047),Key Project of Xiangjiang Laboratory(No.23XJ02005),The Natural Science Foundation of Hunan Province(No.2024JJ7372),Key Scientific Research Projects of Hunan Provincial Department of Education(No.24A0550,No.24A0018),The Teaching Reform Project of Hunan Provincial Department of Education(No.202401001340,No.202502001316),The Key Research and Development Plan of Hunan Provincial Department of Science and Technology(No.2024JK2135),The Science and Technology Innovation Team Support Project for Ordinary Higher Education Institutions in Hunan Province"Wuling Mountain Area Smart Agriculture Intommation Processing and Control Technology"(No.ZNKZD2024-3),The Natural Science Foundation of Xinjiang Uygur Autonomous Region(No.2024D01A52),Frontier Cross Project of Central South University(No.2023QYJC020)
评论