摘要
为了解决液化天然气(LNG)日度价格因高波动性与复杂周期性导致预测精度不足的问题,构建了一种融合Informer架构与快速傅里叶变换(FFT)频率增强机制的短期动态预测方法。该方法利用Informer模型高效处理长序列依赖的优势,并引入FFT显式提取价格序列中的频域特征作为辅助输入,以同步捕捉价格波动中的时域动态与周期模式。基于2011—2025年中国LNG出厂价格日度数据开展了实证研究,研究结果表明:①该方法在测试集与验证集上的平均绝对百分比误差(MAPE)分别为2.37%和2.65%,显著优于LSTM及标准Transformer等基线模型;②频域分析发现,LNG价格序列中存在一个约955天的强主导周期,证明了引入FFT特征对于捕捉非传统长周期规律的有效性,能显著增强模型的预测能力;③模型在包含未知数据的验证集上依然表现出良好的泛化能力与鲁棒性,预测曲线与真实走势高度吻合。结论认为,该融合预测框架能够精准、稳定地预测LNG每日价格,可以为相关市场主体的风险管理与决策提供有效的技术支持。
There exists low accuracy in predicting daily price of liquefied natural gas(LNG)because of high price fluctuation and complex periodicity.Thus,integrated the Informer architecture to the Fast Fourier Transform(FFT)frequency enhancement mechanism(Informer-FFT),a short-term dynamic prediction method have been developed.It makes use of Informer's advantages in handling long-sequence dependency and introduces FFT to extract frequency-domain characters explicitly in price sequence as auxiliary input in order to synchronously capture both time-domain dynamics and periodic patterns in price fluctuation.Some empirical study was conducted by using daily data on China's LNG factory price from 2011 to 2025.Results show that(i)the mean absolute percentage er-ror(MAPE)from this method is 2.37%and 2.65%on both test and validation sets,respectively,much superior to that from the other two baseline models of long short-term memory(LSTM)and standard Transformer;(ii)frequency-domain analysis reveals a strong dom-inant period,approximately 955 days,in LNG price sequence,proving the effectiveness of introducing FFT characters in capturing non-traditional long-period patterns and obviously enhancing this model in predicting ability;and(iii)the model still takes on excel-lent generalization ability and robustness on the validation set containing unknown data,and the predicted curves are highly consistent with actual ones.It is concluded that this integrated method can precisely and stably predict LNG's daily price,and provide effective technical support for relevant market entities in risk management and decision-making.
作者
郑小强
袁燕如
郭玉博
孙逸林
李梦悦
ZHENG Xiaoqiang;YUAN Yanru;GUO Yubo;SUN Yilin;LI Mengyue(School of Economics and Management,Southwest Petroleum University,Chengdu,Sichuan 610500,China;Key Laboratory of Energy Security and Low-carbon Development,Chengdu,Sichuan 610500,China;China Energy Index Research Center,Southwest Petroleum University,Chengdu,Sichuan 610500,China)
出处
《天然气技术与经济》
2025年第4期39-45,共7页
Natural Gas Technology and Economy
基金
国家社会科学基金项目“能源转型背景下农村能源返贫风险测度及协同治理研究”(编号:23XGL029)
国家社会科学基金重大项目“统筹新能源发展与国家能源安全重要关系及实践路径研究”(编号:24&ZD106)。
作者简介
郑小强(1981-),教授,博士,博士研究生导师,主要从事能源经济与管理相关研究工作。E-mail:zhengxiao-qiang@163.com;通信作者:袁燕如(2001-),硕士,主要从事能源经济与管理相关研究工作。E-mai:yuanyanru_swpu@163.com。
