摘要
时空联合分析可反映研究对象在时空维的变化规律,对揭示区域过程的时空交互关系和机制具有重要意义。聚焦时空联合特征的学习与交通流物理特性的建模问题,提出一种层次化的动态网络模型JST-DHNet,以融合不同尺度下的时空联合学习与内嵌领域知识学习。利用基于图乘积运算替代以往矩阵拼接方式构建多种时空图结构。结合时空小波变换与时空傅里叶变换,设计2种不同层次的时空同步学习模块,分别学习交通流的全域与局域时空特征。针对交通流的宏观流体动力学性质,通过基于图的广义偏微分方程设计一种新的时空扩散卷积,以学习真实场景下的交通波传播机制。在此基础上,采用注意力机制将不同尺度的时空联合特征进行融合。在4种不同路网规模的真实交通流数据集上进行测试,结果表明,JST-DHNet的预测性能优于采用时空分离式学习模块的预测模型,相比STSGCN时空联合学习模型,JST-DHNet预测精度的平均绝对百分比误差、平均绝对误差和均方根误差分别降低4.46%、6.65%、10.11%,且训练时间缩短近80%。
A joint spatio-temporal analysis can reflect the changing pattern of a studied object in the spatio-temporal dimension,which is significant for revealing the spatio-temporal interactions and mechanisms of regional processes.With a focus on joint spatio-temporal feature learning and modeling of traffic flow physical characteristics,this paper proposes a dynamic hierarchical network model called the JST-DHNet,which involves multi-scale joint spatio-temporal learning and physics-informed learning.First,instead of simply connecting adjacent graph snapshots like in previous studies,we developed multiple spatio-temporal graph structures using the graph product.Next,based on the joint time-vertex wavelet transform and Fourier transform,two spatial-temporal synchronous learning modules with different scales are designed to learn the global and local spatio-temporal characteristics of traffic flow,respectively.Based on the macroscopic fluid-dynamical properties of the traffic flow,we developed a novel spatio-temporal diffusion convolution with a graph-based partial differential equation,which enables the learning of the propagation mechanism of traffic waves in actual physical scenarios.Finally,the fusion of joint spatio-temporal features at different scales is performed by adopting an attention mechanism.After testing on four datasets of actual road network traffic flow with different sizes,the experimental results show that JST-DHNet outperforms other separated learning models.Compared with the existing joint spatio-temporal learning model called Spatial-Temporal Synchronous Graph Convolutional Network(STSGCN),JST-DHNet not only improves the prediction accuracy of the Mean Absolute Percentage Error(MAPE),Mean Absolute Error(MAE),and Root-Mean-Squared Error(RMSE)by 4.46%,6.65%,and 10.11%,respectively,but also shortens the training time by nearly 80%.
作者
葛宇然
付强
GE Yuran;FU Qiang(Key Laboratory of Road and Traffic Engineering of the Ministry of Education,Tongji University,Shanghai 201804,China)
出处
《计算机工程》
CAS
CSCD
北大核心
2023年第1期270-278,共9页
Computer Engineering
基金
国家自然科学基金重点项目(71734004)
上海市科技攻关项目(19DZ1208900)。
关键词
智能交通系统
时空域联合
交通流预测
图信号处理
交通流理论
Intelligent Traffic System(ITS)
spatio-temporal union
traffic flow prediction
Graph Signal Processing(GSP)
traffic flow theory
