EMD-LSTM-LB分频时序预测算法  

EMD-LSTM-LB frequency division time series prediction algorithm

作　　者：孔繁苗 高鹭[1] 李鹏程 张飞[1,2] 张晓琳 秦岭[1] KONG Fan-miao;GAO Lu+;LI Peng-cheng;ZHANG Fei;ZHANG Xiao-lin;QIN Ling(School of Information Engineering,Inner Mongolia University of Science and Technology,Baotou 014010,China;Renewable Energy College,North China Electric Power University,Beijing 102206,China)

机构地区：[1]内蒙古科技大学信息工程学院,内蒙古包头014010 [2]华北电力大学可再生能源学院,北京102206

出　　处：《计算机工程与设计》2023年第10期3021-3030,共10页Computer Engineering and Design

基　　金：国家自然科学基金地区科学基金项目(62161041);内蒙古自治区科技计划基金项目(2021GG0046、2021GG0048);政府间国际科技创新合作重点专项子基金项目(2017YFE0109000)。

摘　　要：针对现有时间序列预测模型存在的误差大、不稳定等问题,提出一种基于EMD的时间序列预测模型。该模型适应于预测数据量大、波动性大的时间序列,提出一种分配函数,解决分解后子序列的分配问题。EMD将原始序列分解为一系列子序列,分配函数根据子序列的波动状态划分出极端子序列和非极端子序列。极端子序列由LSTM进行训练预测,非极端子序列子序列由LB进行训练预测,将所有子序列整合为最终预测结果。与已有的一些模型方法相比,各数据集的预测精度均有所提高。Aiming at the problems of the large error and instability of the existing time series prediction model,a hybrid model based on EMD was proposed.This model was suitable for predicting time series with large data and strong fluctuation.A distribution function was proposed to solve the problem of subsequence allocation after decomposition.EMD was used to decompose the original time series into a series of sub-series,and the distribution function was used to put the extremely fluctuant sub-series into LSTM for predicting according to the fluctuation state of the sub-series,and the nonextreme sub-sequence was put into LB model for predicting.All sub-sequences were integrated into the final prediction result.Compared with some existing model methods,the prediction accuracy of experiments is improved.

关 键 词：深度学习 经验模态分解 长短期记忆神经网路 LB模型 时间序列预测 分配函数 神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]