摘要
为了解决高速公路流量预测空间建模受限于仅考虑静态空间依赖性或动态空间相关性的不足,提出一种特征融合的时空图混合网络车流量预测模型(FF-STGM),引入多头注意力机制模块,结构上使用双重并行的优化时空卷积网络提取交通流的时空特征和高速公路网络结构属性,并捕捉数据中非连续时间动态相关性。同时,采用蚁群优化算法动态调整模型结构参数,以优化时空图混合网络的隐层结构。试验结果表明,与STSGCN、ASTGCN和CNN-LSTM相比,FF-STGM的MAE值分别降低了6.45、7.69、16.80,改进的时空图混合网络模型预测精度优于其他对比模型。
Mining the spatio-temporal correlations between different nodes is the key to improving the accuracy of highway traffic prediction.However,spatial modeling is limited by methods that only consider static spatial dependencies or dynamic spatial correlations.To solve the above problems,a feature fusion spatio-temporal graph mixed network traffic flow prediction model(FF-STGM)is proposed.A multi-attention mechanism module is introduced,the spatio-temporal features of traffic flow and the structural attributes of the freeway network are structurally extracted using a doubly-parallel optimized spatio-temporal graph convolutional network,and the non-continuous time dynamic correlations in the data are captured.Meanwhile,an optimized adaptive ant colony optimization algorithm based on the improved Lévy flight strategy is used to adjust dynamically the structural parameters of the model to optimize the hidden layer structure of the spatio-temporal graph mixed network.The experiment results show that the MAE values are reduced by 6.45,7.69,and 16.8 compared with STSGCN,ASTGCN and CNN-LSTM.The improved spatio-temporal graph mixed network model prediction accuracy outperforms other comparative models.
作者
张阳
周晨峰
陈燕玲
ZHANG Yang;ZHOU Chenfeng;CHEN Yanlin(School of Transportation,Fujian University of Technology,Fuzhou 350118,China;Fuzhou Software Technology Vocational College,Fuzhou 350211,China)
出处
《大连交通大学学报》
2025年第4期15-25,共11页
Journal of Dalian Jiaotong University
基金
福建省自然科学基金项目(2023J01946)。
关键词
高速公路
交通流预测
注意力机制
深度学习
highway
traffic flow prediction
attention mechanism
deep learning
作者简介
第一作者:张阳(1983-),男,教授。E-mailzhang_yang1983@163.com;通信作者:周晨峰(2000-),男,硕士研究生。E-mail:17858419245@163.com。
