关键词: 多变量长时预测, 频域特征提取, Gabor变换, transformer, 时间序列分解, 时间序列, 深度学习

Abstract: In response to the limitations of existing Transformer-based multivariate long-term prediction models, which primarily extract features from the time domain and struggle to identify reliable dependencies from dispersed time points in long time series, this paper proposes a novel model based on decomposition and frequency domain feature extraction. Firstly, a period-trend decomposition method based on the frequency domain is proposed, reducing the time complexity of the decomposition process. Secondly, based on the extraction of trend features using period-trend decomposition, a Transformer network that captures periodic dependencies through frequency domain feature extraction using the Gabor transform is utilized, enhancing stability and robustness of forecasting. Experiments were conducted on five benchmark datasets, and the results show that compared with the current state-of-the-art methods, the mean square error (MSE) of the proposed model in multivariate long-term time series prediction is reduced by an average of 7.4%, with a maximum reduction of 18.9%, which demonstrates that the proposed model effectively improves prediction accuracy.

Key words: multivariate long-term series forecasting, frequency domain feature extraction, Gabor transform, transformer, time series decomposition, time series, deep learning