基于深度学习的时间序列预测方法综述  

A survey of time series forecasting methods based on deep learning

作　　者：毛远宏 孙琛琛 徐鲁豫 刘曦[1] 柴波[1] 贺鹏超 MAO Yuanhong;SUN Chenchen;XU Luyu;LIU Xi;CHAI Bo;HE Pengchao(Xi'an Microelectronics Technology Institute,Xi'an 710054,Shaanxi,China;School of Computer Science and Engineering,TianJin University of Technology,TianJin 300384,China;China Academy of Aerospace Electronics Technology,Beijing 100094,China)

机构地区：[1]西安微电子技术研究所,陕西西安710054 [2]天津理工大学计算机科学与工程学院,天津300384 [3]中国航天电子技术研究院,北京100094

出　　处：《微电子学与计算机》2023年第4期8-17,共10页Microelectronics & Computer

基　　金：西安微电子技术研究所2022年创新支持项目(YL-220009)。

摘　　要：时间序列预测通过分析时间序列找到其内在规律性对未来发展进行预测,其研究有着重要的学术意义和应用价值.特别随着传感器和网络技术的发展,如何基于大量历史时序数据进行更加精准和高效的预测分析成为了需要解决的迫切问题.时间序列预测任务充分借鉴了深度学习的技术研究成果,在近些年取得了快速发展.本文分析了时间序列预测技术的研究现状,论述了时间序列预测所涉及到深度学习方法的相关理论和方法,包括卷积神经网络、循环神经网络、注意力机制和图神经网络等方法在时间预测领域的应用,归纳总结近年来基于深度学习的时间序列研究成果,比较了基于各种深度学习时间序列方法的优缺点,在此基础上对基于深度学习时间序列预测方法的发展进行了展望.Time series forecasting finds its internal regularity by analyzing time series to forecasts its future.Its research has important academic and application.Especially with the development of sensor and network technology,how to make more accurate prediction and analysis based on a large number of historical time series data has become an urgent problem to be solved.At present,time series forecasting methods fully use the research results of deep learning,and have made rapid development in recent years.This paper analyzes the research status of time series forecasting technology,discusses the relevant theories and methods of deep learning methods involved in time series forecasting of time overview,including the application of convolutional neural network,recurrent neural network,attention mechanism,graph neural network and other methods in the field of time forecasting,and summarizes the research achievements of time series based on deep learning in recent years,The advantages and disadvantages of various time series methods based on deep learning are compared.Finally,this paper forecasts the development trend of time series prediction methods based on deep learning.

关 键 词：时间序列预测 深度学习 卷积神经网络 循环神经网络 注意力机制 图神经网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]