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 perceptrons (MLP), can outperform simple linear models. Based on the differences in error power spectra between simple linear models and multi-layer perceptrons in time series forecasting, an MLP-based high-frequency enhanced model (HiFNet), was proposed. First, this model utilized the fitting capability within the low-frequency band and addresses the overfitting issue of MLPs and the weakness of channel independence in handling the channel redundancy issue by introducing the adaptive series decomposition module and the grouped linear module. Thus, it enhanced the robustness on high-frequency band. Finally, experiments on standard datasets in the fields of meteorology, power, and transportation demonstrate that the mean squared error of HiFNet is reduced by up to 23.6%, 10.0%, 35.1%, and 6.5% compared to NLinear, RLinear, SegRNN, and PatchTST, respectively. At the same time, the grouped linear module provides a low-rank representation of channel dependency and thus alleviates the impact of the channel redundancy issue.

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