基于滑动时间窗和组合模型结合的中国LNG现货价格预测方法  

A prediction method of China's LNG spot price based on sliding time window and combination model

作　　者：孙仁金[1] 邓钰暄 李慧慧 刘子越 SUN Renjin;DENG Yuxuan;LI Huihui;LIU Ziyue(School of Economics and Management,China University of Petroleum,Beijing 102249,China)

机构地区：[1]中国石油大学(北京)经济管理学院

出　　处：《天然气工业》2025年第3期170-178,共9页Natural Gas Industry

基　　金：国家自然科学基金面上项目“能源绿色转型的路径与优化:基于生态足迹的视角”(编号:72273151)。

摘　　要：近年来,中国液化天然气(LNG)的生产量和进口量持续攀升,成为最重要的天然气供给来源之一。由于LNG供给灵活、市场参与主体众多、在产业链中市场化程度相对较高,科学准确预测LNG现货价格能够为市场参与者提供决策参考,降低市场风险。为此,建立了基于滑动时间窗以及二次分解思想的变分模态分解(VMD)—自适应噪声完全集合经验模态分解(CEEMDAN)—极限学习机模型(ELM)有机组合预测模型,并以内蒙古自治区的LNG价格数据为例进行实证分析。研究结果表明:(1)采用滑动时间窗可以有效提取LNG价格序列中用于分析建模的部分,将训练集数据进行分解、建模以及预测等环节,避免了待预测的LNG价格数据混入其中;同时可以在建模预测步骤完成后,舍弃最旧日期的LNG价格,将新一期的价格数据纳入其中,随时间推移有效把握数据规律,实现了模型的动态更新。(2)利用VMD对LNG价格进行初次分解,再通过CEEMDAN对VMD的残差序列展开二次分解,可以充分提取LNG价格的数据信息,以提高预测精度。(3)将二次模态分解LNG价格得到的分量模态序列分别带入ELM模型中预测,再将各价格分量预测结果加和集成得到LNG价格预测结果,可以显著提升价格预测的准确度。结论认为,该模型可以更好提取序列时频信息,有效规避了数据泄露问题,充分利用残差数据,显著提高了预测精度,是LNG现货价格预测的可行方法与有效手段。In recent years,the continuous rising production and import volumes of liquefied natural gas(LNG)in China makes LNG one of the most important sources of natural gas supply.In view that LNG has flexible supply,multiple market participants and a high degree of marketization in the industrial chain,the scientific and accurate prediction of LNG spot price can provide market participants with decision-making reference and reduce market risks.In this paper,an organically combined prediction model of variational modal decomposition(VMD),complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN)and extreme learning machine(ELM)based on the sliding time window and the idea of quadratic decomposition is established,and then analyzed with the LNG price data from Inner Mongolia as an example.The following results are obtained.First,the sliding time window can effectively extract some parts of LNG price series used for analysis and modeling,and perform the decomposition,modeling and prediction of training set data,so as to prevent the LNG price data to be predicted from mixing with them.After the modeling and prediction is completed,the earliest LNG price is discarded,and the latest price data is introduced to understand the data laws over time effectively,realizing dynamic model update.Second,VMD is adopted to the primary decomposition of LNG price,and CEEMDAN is employed for the quadratic decomposition of VMD residual sequence,which can sufficiently extract the data information of LNG price,so as to improve the prediction accuracy.Third,the component modal sequence obtained from the quadratic modal decomposition of LNG price is introduced into the ELM model for prediction,and the prediction results of each price component are summed and integrated to get the prediction result of LNG price,which can improve the accuracy of price prediction significantly.In conclusion,this model can better extract the time frequency information of the series,effectively avoid data leakage,and make full use of residual data,so

关 键 词：滑动时间窗 机器学习 二次分解 LNG 价格预测 

分 类 号：TE-9[石油与天然气工程]