Deep autoregressive models with spectral attention

Highlights：

• Two spectral attention models, global and local to the time series, integrate global trends and seasonality patterns within the forecast and perform spectral filtering to remove time series’s noise.

• The proposed Spectral Attention module, responsible of all frequency domain operations, can be easily incorporated into well-known forecast frameworks.

• Experiments unveil how Spectral Attention stands out, consistently out-performing the base models it is integrated into.

摘要

•We propose a forecasting architecture that combines deep autoregressive models with a Spectral Attention (SA) module, merging global and local frequency domain information in the model’s embedded space.•Two spectral attention models, global and local to the time series, integrate global trends and seasonality patterns within the forecast and perform spectral filtering to remove time series’s noise.•The proposed Spectral Attention module, responsible of all frequency domain operations, can be easily incorporated into well-known forecast frameworks.•Experiments unveil how Spectral Attention stands out, consistently out-performing the base models it is integrated into.

论文关键词：Attention models,Deep learning,Filtering,Global-local contexts,Signal processing,Spectral domain attention,Time series forecasting

论文评审过程：Received 28 December 2021, Revised 12 July 2022, Accepted 28 August 2022, Available online 31 August 2022, Version of Record 20 September 2022.