一种改进的CEEMDAN-LSTM-Neural Prophet Net 模型:用于 COVID-19背景下我国月度消费预测  

An Improved CEEMDAN-LSTM-Neural Prophet Net Model for Monthly China's Total Retail Sales of Consumer Goods Prediction under the Background of COVID-19

作　　者：郇松桦 刘秀丽[1,3,4] Huan Songhua;Liu Xiuli

机构地区：[1]中国科学院数学与系统科学研究院,北京100190 [2]中国科学院大学 [3]中国科学院预测科学研究中心,北京100190 [4]中国科学院大学,北京100049

出　　处：《复印报刊资料（统计与精算）》2022年第6期77-95,共19页STATISTICS AND ACTUARIAL SCIENCE

基　　金：国家自然科学基金(71874184)。

摘　　要：针对社会消费品零售总额(total retail sales ofconsumergoods,TRSCG)月度预测中存在突变点情况下的样本外预测精度低等问题,本文构建了CEEMDAN-LSTM-NeuralProphetNet组合模型,开展了不同情景下的月度TRSCG样本外预测,并比较了该模型与NeuralProphet、BPNetwork等7种常用模型的预测精度。结果表明:深度学习模型表现出对非线性一般数据的良好适应能力,其对数据的分类识别及组合预测效果俱佳;CEEMDAN-LSTM-NeuralProphetNet组合模型在有突变点的样本外预测中表现更佳,且该预测模型在不同国家的TRSCG、不同的预测指标及学习比例上均具有一定的稳健性。本文为存在突变点的时间序列分析和样本外预测提供了新思路。Facing with the problems existing in the current total retail sales of consumer goods(TRSCG)monthly prediction,such as low accuracy of the out-sample prediction under the condition with outliers,this paper proposes a CEEMDAN-LSTM-Neural Prophet Net combination model to make the out-sample monthly prediction on TRSCG in different scenarios,it also compares the predicting accuracy with seven other frequently-used models such as Neural Prophet and BP Network.The results show that the deep learning model shows good adaptability to nonlinear data,and its ability of data classification,recognition and combination prediction is fantastic;CEEMDANLSTM-Neural Prophet Net combination model has better out-sample predicting accuracy to the data with outliers.It also shows robustness in predicting different indicators in different countries and learning rates,etc.This paper provides a novel idea for time series analysis and out-sample prediction with outliers.

关 键 词：社会消费品零售总额 样本外月度预测 CEEMDAN-LSTM-NeuralProphetNet组合模型 COVID-19 突变点 

分 类 号：F724[经济管理—产业经济]