功率谱密度引导下的时间序列预测模型  

Time series forecasting model guided by power spectral density

作　　者：梁立河 崔锦莹 张雪松 高妮玲 赵涓涓[2,4] 强彦 LIANG Li-he;CUI Jin-ying;ZHANG Xue-song;GAO Ni-ling;ZHAO Juan-juan;QIANG Yan(College of Computer Science and Technology(College of Data Science),Taiyuan University of Technology,Taiyuan 030000,China;School of Software,Taiyuan University of Technology,Taiyuan 030000,China;School of Software,North University of China,Taiyuan 030000,China;School of Big Data,Jinzhong College of Information,Jinzhong 030600,China)

机构地区：[1]太原理工大学计算机科学与技术学院(大数据学院),山西太原030000 [2]太原理工大学软件学院,山西太原030000 [3]中北大学软件学院,山西太原030000 [4]晋中信息学院大数据学院,山西晋中030600

出　　处：《计算机工程与设计》2025年第4期1087-1095,共9页Computer Engineering and Design

基　　金：国家自然科学基金重点基金项目(U21A20469);中央引导地方科技发展基金项目(YDZJSX2022C004)。

摘　　要：为增强时间序列预测模型的高解释性、高稳定性、高准确性,从能量的角度分析,提出一种基于功率谱密度的时间序列预测编解码器模型(PSDformer)。通过引入多粒度能量选择模块、注意力知识引导模块和序列去噪分解模块,能够有效提取并融合序列的长短期特征、实现未来“先验”信息的有效传递和降低异常数据对序列预测的负面影响,提高模型的预测准确性。在3个数据集上进行的实验验证了PSDformer模型的可行性和有效性。To enhance the high interpretability,the stability,and the accuracy of time series prediction models,a time series prediction encoder-decoder model based on power spectral density(PSDformer)from an energy perspective was proposed.By introducing the multi-granularity energy selection module,attention knowledge guidance module,and sequence denoising decomposition module,the long-and short-term features of the sequences were effectively extracted and fused,the effective transmission of future“a priori”information was realized,the negative impact of abnormal data on sequence prediction was reduced,and the prediction accuracy of the model was ultimately improved.The feasibility and effectiveness of the PSDformer model are demonstrated through experiments conducted on three datasets.

关 键 词：时间序列预测 功率谱密度 编解码器模型 多粒度能量选择 注意力知识引导 序列去噪分解 长短期特征 “先验”信息 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]