数据特征驱动的单变量预测建模研究

Data-Trait-Driven Univariate Predictive Modeling

预测是决策和规划的基础,包括单变量和多变量预测建模.由于单变量预测建模仅需利用时间序列本身的历史值,在农业、能源、环境、金融等领域得到了广泛应用.数据特征驱动模型是基于数据本身特征进行模型选择,以预测未来趋势.本文立足于数据特征驱动的预测建模研究范式,通过文献梳理和总结,提出了数据特征驱动的单变量预测建模的七种典型框架,即专家知识类、数据特征驱动类、专家知识驱动的分解-集成类、专家知识驱动的分解-聚类-重构-集成类、数据知识混合驱动的分解-集成类、数据知识混合驱动的分解-聚类-重构-集成类、知识数据混合驱动的分解-集成类.在此基础上,对数据特征分类与识别方法、分解-集成方法、聚类-重构方法和预测方法等进行了评述.最后,从混叠数据特征的识别检验、智能化预测建模、聚类-重构新方法、预测-集成新方法、时序预测大模型等方面讨论了未来的五大研究方向及其典型科学问题,以期为促进数据特征驱动的单变量预测理论与方法的研究提供参考.

Prediction is the foundation of decision-making and planning,including univariate and multivariate predictive modeling.Univariate predictive modeling,which only utilizes the historical values of time series,has been widely applied in fields such as agriculture,energy,environment,and finance.Data-trait-driven models are based on the traits of the data itself to select models and predict future trends.This article focuses on the research paradigm of data-trait-driven predictive modeling.Through literature review and summary,seven typical frameworks are proposed,including expert knowledge-based,data trait-driven,expert knowledge-driven decomposition-ensemble,expert knowledge-driven decomposition-clustering-reconstruction-ensemble,data-knowledge hybrid-driven decomposition-ensemble,data-knowledge hybrid-driven decomposition-clustering-reconstruction-ensemble,and know-ledge-data hybrid-driven decomposition-ensemble.Then,the methods of data trait classification and identification,decomposition-ensemble,clustering-reconstruction,and prediction methods are reviewed.Finally,future research directions and typical scientific problems are discussed,including the identification and verification of mixed data traits,intelligent predictive modeling,clustering-reconstruction new methods,prediction-ensemble new methods,and large-scale models for time series data,aiming to provide reference for the research of data-trait-driven univariate prediction theory and methods.