主管:中国科学院
主办:中国优选法统筹法与经济数学研究会
   中国科学院科技战略咨询研究院

中国管理科学 ›› 2026, Vol. 34 ›› Issue (11): 361-369.doi: 10.16381/j.cnki.issn1003-207x.2023.2035

• • 上一篇    

数据特征驱动的新能源汽车月度销量二次分解集成预测

王方1,3, 赵桉坤1, 余乐安2()   

  1. 1.西安电子科技大学经济与管理学院,陕西 西安 710126
    2.四川大学商学院,四川 成都 610065
    3.滨州魏桥国科高等技术研究院,山东 滨州 256600
  • 收稿日期:2023-12-04 修回日期:2024-06-11 出版日期:2026-11-25 发布日期:2026-11-10
  • 通讯作者: 余乐安 E-mail:yulean@amss.ac.cn
  • 基金资助:
    国家自然科学基金项目(72001165);国家自然科学基金项目(72331007);西安市科技计划项目软科学研究重点项目(23RKYJ0006);陕西省创新能力支撑计划项目(2022SR5016);滨州魏桥国科高等技术研究院“产学合作协同育人”项目(BINTECH-KJZX-20220831-67)

Quadratic Decomposition-ensemble Method for Multi-step Forecast of Monthly Sales Volume of New energy Vehicles

Fang Wang1,3, Ankun Zhao1, Lean Yu2()   

  1. 1.School of Economics & Management,Xidian University,Xi’an 710126,China
    2.Business School,Sichuan University,Chengdu 610065,China
    3.WQ-UCAS Graduate School of Business,Binzhou Institute of Technology,Binzhou 256600,China
  • Received:2023-12-04 Revised:2024-06-11 Online:2026-11-25 Published:2026-11-10
  • Contact: Lean Yu E-mail:yulean@amss.ac.cn

摘要:

针对新能源汽车月度销量时序数据具有的趋势性、季节性、复杂性等数据特征混叠的问题,基于“分解-集成”的建模思想,构建了新能源汽车月度销量预测的二次分解集成框架。首先,通过集合经验模态分解算法将原始序列分解,提取出趋势性数据特征。然后,采用STL(seasonal and trend decomposition using loess)算法从去趋势项序列中提取季节项及残差项。在此基础上,分别使用Logistic回归、差分自回归移动平均模型和深度极限学习机模型对趋势项、季节项以及残差项进行预测。最后,将各分项的预测值,通过线性加权集成的方法得到最终的预测结果。基于我国新能源汽车销量数据的实证分析结果表明,提出的二次分解集成模型在3步、6步和12步的预测中的平均绝对百分比误差分别为5.80%、6.10%和7.34%,较基准模型均下降了15%以上,预测建模效果提升显著。

关键词: 新能源汽车, 集合经验模态分解, 多步预测, 二次分解集成方法

Abstract:

Driven by policy support, technological progress and market demand, China's new energy vehicle market has shown a strong momentum of development, and the rapid development of the new energy vehicle market has attracted wide attention, and its sales forecast is crucial for enterprises and governments. At present, the sales forecast of new energy vehicles is mainly based on annual sales. Considering that the sales volume of new energy vehicles is susceptible to various factors, the monthly sales data has the phenomenon of aliasing data characteristics, which makes multi-step forecasting a challenge.In view of the complexity, seasonality, trend and other characteristics of monthly sales timing data of new energy vehicles, based on the modeling idea of “disintegration-integration”, a secondary disintegration-integration framework for monthly sales forecasting of new energy vehicles is constructed. First, the original sequence is decomposed by the ensemble empirical mode decomposition (EEMD) algorithm to extract the trend item data features. Secondly, the algorithm STL (Seasonal and Trend decomposition using Loess) is used to extract the seasonal terms and residual terms from the detrended item sequence. Then, the Logistic regression, autoregressive integrated moving average model (ARIMA) and deep Extreme Learning machine (DELM) models are respectively used to predict the trend term, seasonal term and residual term. Finally, the forecast results of each item are integrated to get the final forecast results.The empirical analysis based on China's new energy vehicle sales data shows that the MAPE of the proposed quadratic decomposition integrated model in the three-step, six-step and twelve-step forecasting is 5.80%, 6.10% and 7.34%, respectively, which is more than 15% lower than the benchmark model, and the improvement effect is significant. In the future, for the prediction of complex time series data, the corresponding characteristics of time series can be fully considered and extracted for modeling prediction.

Key words: new energy vehicle, ensemble empirical mode decomposition, multistep prediction, quadratic decomposition-ensemble method

中图分类号: