摘要
现有多变量时间序列(multivariate time series,MTS)预测方法模型主要采用循环神经网络和注意力机制提取MTS的复杂时空特征,这些方法对MTS变量之间的空间依赖关系的捕获能力不足。图卷积网络对复杂数据的空间特征提取能力较强。为此提出一种融入图卷积网络、注意力机制和深度学习中的卷积神经网络的三通道网络框架模型,将该框架模型用于多变量时间序列预测任务。实验结果表明,该模型在国际汇率这一多变量时间序列数据集上的性能表现要优于目前较先进的几个基线模型。
The existing models of multivariate time series(MTS)prediction methods mainly use recurrent neural network(RNN)and attention mechanisms to extract complex spatiotemporal features of MTS.These methods have insufficient ability to capture spatial dependencies between MTS variables.The graph convolution network has strong ability to extract spatial features of complex data.A three-channel network framework incorporating graph convolutional networks,attention mechanisms and convolutional neural networks in deep learning was proposed for multivariable time series prediction tasks.Experimental results show that the performance of the model on the multivariable time series data set of international exchange rate is better than that of several advanced baseline models.
作者
李继龙
霍纬纲
李勤
LI Ji-long;HUO Wei-gang;LI Qin(School of Computer Science and Technology,Civil Aviation University of China,Tianjin 300300,China;Beijing Platform IT Service Delivery Centre,Northking Information Technology Limited Company,Beijing 100032,China)
出处
《计算机工程与设计》
北大核心
2022年第3期895-900,F0003,共7页
Computer Engineering and Design
关键词
注意力机制
图卷积网络
时间序列预测
卷积神经网络
深度学习
attention mechanism
graph convolution network
time series forecasting
convolution neural network
deep learning