西安电子科技大学学报 ›› 2024, Vol. 51 ›› Issue (3): 103-112.doi: 10.19665/j.issn1001-2400.20231205

• 信息与通信工程 • 上一篇    下一篇

双向长短期记忆网络的时间序列预测方法

管业鹏1(), 苏光耀1(), 盛怡2()   

  1. 1.上海大学 通信与信息工程学院,上海 200444
    2.上海体育学院 竞技运动学院,上海 200438
  • 收稿日期:2023-07-26 出版日期:2024-06-20 发布日期:2023-12-27
  • 通讯作者: 盛 怡(1981—),女,副教授,E-mail:549316264@qq.com
  • 作者简介:管业鹏(1967—),男,教授,E-mail:shugyp@yeah.net
    苏光耀(2000—),男,上海大学硕士研究生,E-mail:13235373474@163.com
  • 基金资助:
    国家重点研发计划(2019YFC1520500)

Time series prediction method based on the bidirectional long short-term memory network

GUAN Yepeng1(), SU Guangyao1(), SHENG Yi2()   

  1. 1. School of Communication and Information Engineering,Shanghai University,Shanghai 200444,China
    2. School of Competitive Sports,Shanghai University of Sport,Shanghai 200438,China
  • Received:2023-07-26 Online:2024-06-20 Published:2023-12-27

摘要:

时间序列预测即利用历史时间序列数据,预测未来一段时间内的数据信息,以便提前制定相应策略。目前,时间序列的类别复杂繁多,而现有的时间序列预测模型面对多种类型数据时无法取得稳定预测的结果,进而难以同时满足对现实中多种复杂的时序数据预测的应用需求。针对上述问题,提出了一种基于时间注意力机制双向长短期记忆网络的时间序列预测方法。笔者提出的网络模型采用改进的正向和反向传播机制提取时序信息并通过自适应权重分配策略推理未来的时序信息。具体来说,设计了一个改进的双向长短期记忆网络,通过结合双向长短期记忆和长短期记忆网络提取深度时间序列特征,挖掘上下文的时序依赖关系。在此基础上,融合所提出的时间注意力机制,实现对深度时间序列特征进行自适应加权,提升深度时序特征的显著性表达能力。通过与同类代表性方法在多个不同类别数据集上的客观定量对比,实验结果表明,该方法能够在多种类别的复杂时间序列数据上更优的预测性能。

关键词: 时间序列, 双向长短期记忆网络, 长短期记忆网络, 注意力机制, 深度学习

Abstract:

Time series prediction means the use of historical time series to predict a period of time in the future,so as to formulate corresponding strategies in advance.At present,the categories of time series are complex and diverse.However,existing time series prediction models cannot achieve stable prediction results when faced with multiple types of time series data.The application requirements of complex time series data prediction in reality are difficult to simultaneously meet.To address the problem,a time series prediction method is proposed based on the Bidirectional Long and Short-term Memory(BLSTM) with the attention mechanism.The improved forward and backward propagation mechanisms are used to extract temporal information.The future temporal information is inferred through an adaptive weight allocation strategy.Specifically,an improved BLSTM is proposed to extract deep time series features and explore temporal dependencies of context by combining BLSTM and Long Short-term Memory(LSTM) networks,on the basis of which the proposed temporal attention mechanism is fused to achieve adaptive weighting of deep time series features,which improves the saliency expression ability of deep time series features.Experimental results demonstrate that the proposed method has a superior prediction performance in comparison with some representative methods in multiple time series datasets of different categories.

Key words: time series, Bidirectional Long Short-Term Memory, Long Short-Term Memory, attention mechanism, deep learning

中图分类号: 

  • TP391.41