华中科技大学学报（自然科学版）2025，Vol.53Issue(5)：150-156,177,8.DOI:10.13245/j.hust.250812
一种频域注意力机制变量筛选的时序预测模型
A time series forecasting model with frequency-domain attention mechanism for variable selection
摘要
Abstract
In response to the issue that redundant time series variables in multivariate time series forecasting introduce noise and reduce forecasting accuracy,a multivariate time series forecasting model based on frequency-domain attention mechanism for important variable selection was proposed.First,the model decomposed the multivariate time series into seasonal and trend components through time series decomposition.A frequency-domain attention mechanism was designed for the seasonal component,which calculated the correlation functions between variables of the multivariate time series in the frequency domain and obtained the weights of each variable through an activation function.The weights reflected the importance of each variable,thereby reducing the impact of redundant variables on the model's forecasting performance.Then,a multi-scale feature extraction method for time series was designed,which divided the seasonal component into multiple segments with the same proportion in the time direction.The multi-layer perceptron was used to extract features from the segmented seasonal components and the trend component separately,in order to obtain the multi-scale features of the seasonal component and enhance the model's feature extraction capability for time series.Finally,the multi-scale features of the seasonal component and the features of the trend component were fused in the time direction and mapped through a fully connected layer to obtain the forecasting output of the multivariate time series.Experimental results on six public multivariate time series datasets show that the average forecasting accuracy of the proposed model is significantly improved compared with Autoformer,FEDformer,Crossformer,DLinear,iTransformer,and TimesNet models.关键词
多变量时间序列预测/多尺度特征/注意力机制/频域/特征融合Key words
multivariate time series forecasting/multi scale feature/attention mechanism/frequency domain/feature fusion分类
计算机与自动化引用本文复制引用
陈海燕,欧阳林威,任宝民..一种频域注意力机制变量筛选的时序预测模型[J].华中科技大学学报（自然科学版）,2025,53(5):150-156,177,8.基金项目
国家自然科学基金资助项目(62161019,62061024). （62161019,62061024）
