华中科技大学学报（自然科学版）2026，Vol.54Issue(5)：68-75,8.DOI:10.13245/j.hust.240904
一种基于时-频特征融合的多变量时间序列预测模型
A multivariate time series prediction model based on time-frequency feature fusion
摘要
Abstract
To address the problem in multi-variate time series prediction that existing methods were difficult to effectively identify and capture long-short-term features and periodic characteristics simultaneously,a multi-variate time series prediction model based on temporal-frequency feature fusion was proposed.The model was composed of an input module,a long-short-term feature extraction module,a periodic feature extraction module,and a fusion output module.The input module was used to preprocess data through reversible instance normalization(RevIN);the long-short-term feature extraction module was combined with Mamba and TCN networks in the temporal domain to capture the long-term and short-term dependencies of the sequence,respectively;the periodic feature extraction module was used to capture the periodic characteristics of the sequence through a frequency-domain attention mechanism;the fusion output module fused the temporal and frequency-domain features,and the prediction result was output through a fully connected layer.Experiments on seven public datasets show that the average prediction accuracy of this model is increased by 6.63％,8.55％,18.64％,48.07％,20.71％,9.09％,16.44％,10.66％,and 1.99％compared with that of the iTransformer,PatchTST,FEDformer,Crossformer,TiDE,RLinear,DLinear,TimesNet,and MICN models,respectively.The model is able to maintain high prediction accuracy while also having high computational efficiency.关键词
多变量时间序列预测/时-频特征融合/可逆实例归一化(RevIN)/Mamba/TCN网络/频域注意力机制Key words
multivariable time series prediction/time-frequency feature fusion/reversible instance normalization(RevIN)/Mamba/TCN networks/frequency domain attention mechanism分类
信息技术与安全科学引用本文复制引用
陈海燕,任宝民..一种基于时-频特征融合的多变量时间序列预测模型[J].华中科技大学学报（自然科学版）,2026,54(5):68-75,8.基金项目
国家自然科学基金资助项目(62161019,62061024). （62161019,62061024）
