基于改进长短期记忆网络的时间序列预测研究  

Research on Time Series Forecasting Based on Improved LSTM Model

作　　者：陈孝文 苏攀 吴彬溶 成承 王林[2] CHEN Xiaowen;SU Pan;WU Binrong;CHENG Cheng;WANG Lin(China Tobacco Hubei Industrial Co.LTD,Wuhan 430040,China;不详)

机构地区：[1]湖北中烟工业有限责任公司,湖北武汉430040 [2]华中科技大学管理学院,湖北武汉430074

出　　处：《武汉理工大学学报（信息与管理工程版）》2022年第3期487-494,499,共9页Journal of Wuhan University of Technology：Information & Management Engineering

基　　金：国家社科基金重大招标项目阶段性成果(20&ZD126)。

摘　　要：时间序列预测是研究时间数据行为和预测未来值的一项重要技术,为进一步扩展时间序列预测方法论,提出了一种新颖的时间序列预测框架来处理时间序列预测问题,即VMD-JADE-基于注意力机制的双向长短期记忆网络。变分模态分解用来分解历史时间序列数据,具有降噪的功能;改进的差分进化算法JADE用来优化LSTM的超参数;最后采用基于注意力机制的双向LSTM进行预测,双向机制可以从顺序和逆序两个方向挖掘输入变量的重要信息,注意机制通过对输入的特征赋予不同的权重来捕获重要的因素,有助于提升LSTM的预测性能。在两个时间序列数据集上的实验结果表明,与其它常用的预测方法相比,改进的LSTM模型具有更好的预测性能。Time series forecasting is an important technology to study the behavior of temporal data and predict future values. To further expand the time series forecasting methodology, this study proposes a novel time series forecasting framework to deal with the time series forecasting problem, namely VMD-JADE-Bidirectional LSTM Model based on attention mechanism. Variational mode decomposition is used to decompose historical time series data and has the function of noise reduction;JADE algorithm is used to optimize the hyperparameters of the LSTM model;finally, the bidirectional LSTM model based on the attention mechanism is used for prediction, and the bidirectional mechanism can mine from two directions of order and reverse order of the important information of input variables. The attention mechanism captures important factors by assigning different weights to the input features, which helps to improve the prediction performance of LSTM. The experimental results on two time series datasets show that the improved LSTM model obtains better forecasting performance than other commonly used forecasting methods.

关 键 词：时间序列预测 深度学习 长短期记忆网络 变分模态分解 玉米期货价格 

分 类 号：F272.1[经济管理—企业管理]