摘要
时间序列是一种广泛存在于现实各领域之中的海量高维数据,时间序列预测是该领域的一个研究重点.传统的时间序列预测方法仅仅从时间的维度对时间序列进行分析,忽略了外界影响因素对时间序列可能产生的影响.针对传统时间序列预测方法存在的问题,提出一种基于深度学习的时间序列预测模型DAFDCRNN (dual-stage attention and full dimension convolution based recurrent neural network).该模型引入目标注意力机制来学习输入特征与被预测特征之间的相关性,引入全维度卷积机制来学习输入特征之间的相关性,并引入时间注意力(temporal attention)机制来学习时间序列的长期时间依赖性.在实验部分首先确定模型的超参数,然后对模型部件的有效性进行验证,最后通过对比实验验证了所提出的DAFDC-RNN模型在大特征量数据集上具有最佳的预测效果.
Time series is a kind of high-dimensional data that exists widely in various fields in reality, and time series prediction is a research focus in research activities related to time series. The traditional time series prediction methods only analyze the time series from the time dimension, ignoring the influence of external influence factors on the time series. Aiming at the problems existing in traditional methods, this paper proposes a time series prediction model based on deep learning named dual-stage attention and full dimension convolution based recurrent neural network(DAFDC-RNN).The model introduces a target attenion mechanism to learn the correlation between the input features and the predicted features, introduces a full dimension convolution mechanism to learn the correlation among input features, and introduces a temporal attention mechanism to learn the long-term temporal dependencies of time series. In the experimental part,we firstly determine the hyperparameters of the model, and then verify the validity of the model’s components. Finally,the comparative experiments show that the proposed DAFDC-RNN model has the best prediction effect on the dataset with large features.
作者
鹿天柱
钱晓超
何舒
谭振宁
刘飞
LU Tian-zhu;QIAN Xiao-chao;HE Shu;TAN Zhen-ning;LIU Fei(College of Software Engineering,South China University of Technology,Guangzhou 510006,China;Shanghai Institute of Mechanical and Electrical Engineering,Shanghai 201108,China)
出处
《控制与决策》
EI
CSCD
北大核心
2021年第3期645-652,共8页
Control and Decision
基金
国家重点研发计划项目(2018YFC0830900)
装备预研领域基金项目(61400010205)
上海航天科技创新基金项目。
关键词
时间序列
预测分析
深度学习
目标注意力
全维度卷积
时间注意力
time series
predictive analysis
deep learning
target attention
full dimension convolution
temporal attention