《计算机应用》唯一官方网站 ›› 2025, Vol. 45 ›› Issue (12): 3855-3863.DOI: 10.11772/j.issn.1001-9081.2024121818

• 数据科学与技术 • 上一篇    下一篇

基于多层感知器的高频增强型时间序列预测模型

朱昶胜1, 杨琛1, 冯文芳2, 袁培文1   

  1. 1.兰州理工大学 计算机与通信学院,兰州 730000
    2.兰州理工大学 经济管理学院,兰州 730000
  • 收稿日期:2024-12-27 修回日期:2025-03-14 接受日期:2025-03-18 发布日期:2025-03-27 出版日期:2025-12-10
  • 通讯作者: 杨琛
  • 作者简介:朱昶胜(1972—),男,甘肃秦安人,教授,博士,主要研究方向:高性能计算与大数据、制造业信息化系统与工程
    杨琛(1999—),女,陕西咸阳人,硕士研究生,主要研究方向:时间序列预测、自然语言处理
    冯文芳(1973—),女,甘肃秦安人,副教授,硕士,主要研究方向:金融理论与资产定价、金融风险控制
    袁培文(1993—),男,安徽合肥人,硕士,主要研究方向:时间序列预测。
  • 基金资助:
    国家自然科学基金资助项目(72161026)

High-frequency enhanced time series prediction model based on multi-layer perceptron

Changsheng ZHU1, Chen YANG1, Wenfang FENG2, Peiwen YUAN1   

  1. 1.School of Computer Science and Communication,Lanzhou University of Technology,Lanzhou Gansu 730000,China
    2.School of Economics and Management,Lanzhou University of Technology,Lanzhou Gansu 730000,China
  • Received:2024-12-27 Revised:2025-03-14 Accepted:2025-03-18 Online:2025-03-27 Published:2025-12-10
  • Contact: Chen YANG
  • About author:ZHU Changsheng, born in 1972, Ph. D., professor. His research interests include high-performance computing and big data, information systems and engineering in manufacturing.
    YANG Chen, born in 1999, M. S. candidate. Her research interests include time series forecasting, natural language processing.
    FENG Wenfang, born in 1973, M. S., associate professor. Her research interests include financial theory and asset pricing, financial risk control.
    YUAN Peiwen, born in 1993, M. S. His research interests include time series forecasting.
  • Supported by:
    National Natural Science Foundation of China(72161026)

摘要:

简单线性模型的时间序列预测质量通常超过Transformer等深度模型;而在具有大量通道的数据集上,深度模型尤其是多层感知器(MLP)的性能反而可超过简单线性模型。针对简单线性模型和MLP在时间序列预测中的误差功率谱差异,提出一种基于MLP的高频增强型时间序列预测模型HiFNet(High-Frequency Network)。首先,利用MLP在低频段的拟合能力;其次,通过自适应序列分解(ASD)模块及分组线性层解决MLP高频段易过拟合以及通道独立策略不能有效应对通道冗余的问题,从而增强MLP在高频段的鲁棒性;最后,对HiFNet在气象、电力和交通等领域的标准数据集上进行实验。结果表明:HiFNet的均方误差(MSE)在最佳情况下相较于NLinear、RLinear、SegRNN(Segment Recurrent Neural Network)和PatchTST(Patch Time Series Transformer)分别降低了23.6%、10.0%、35.1%和6.5%,而分组线性层通过学习通道相关性的低秩表达减轻了通道冗余的影响。

关键词: 时间序列预测, 误差功率谱, 线性模型, 多层感知器, 序列分解

Abstract:

The prediction quality of simple linear models for time series forecasting often surpasses that of deep models such as Transformers. However, on datasets with a large number of channels, deep models, particularly Multi-Layer Perceptron (MLP), can outperform simple linear models. Aiming at the differences in error power spectrum between simple linear models and MLPs in time series forecasting, an High-frequency enhanced time series prediction model based on multi-layer perceptron — HiFNet (High-Frequency Network) was proposed. Firstly, the fitting capability of MLPs within low-frequency bands was utilized. Then, the Adaptive Series Decomposition (ASD) module and the grouped linear layer were adopted to address the overfitting issue of MLPs in high-frequency bands and the issue of channel independence strategy failing to handle the channel redundancy effectively, thereby enhancing the robustness of MLPs in high-frequency band. Finally, experiments were conducted to HiFNet on standard datasets in the fields of meteorology, power, and transportation. The results demonstrate that the Mean Squared Error (MSE) of HiFNet is reduced by up to 23.6%, 10.0%, 35.1%, and 6.5%, respectively, compared to those of NLinear, RLinear, SegRNN (Segment Recurrent Neural Network), and PatchTST (Patch Time Series Transformer). At the same time, the grouped linear layer alleviates the impact of the channel redundancy by learning low-rank representations related to channels.

Key words: time series forecasting, error power spectrum, linear model, Multi-Layer Perceptron (MLP), series decomposition

中图分类号: