|
Title:
|
TFBFORMER: DECOMPOSITION TRANSFORMERS WITH TIME-FREQUENCY FOR LONG-TERM SERIES FORECASTING |
|
Author(s):
|
Hao Cang, Xuejie Zhang and Shufang Xu |
|
ISBN:
|
978-989-8704-62 |
|
Editors:
|
Paula Miranda and Pedro Isaías |
|
Year:
|
2024 |
|
Edition:
|
Single |
|
Keywords:
|
Long-Time Series Forecasting, Transformer, Time Series Decomposition |
|
Type:
|
Full |
|
First Page:
|
117 |
|
Last Page:
|
126 |
|
Language:
|
English |
|
Cover:
|
|
|
Full Contents:
|
click to dowload 
|
|
Paper Abstract:
|
Although Transformer-based models demonstrate strong capabilities in long-term sequence forecasting, they require
substantial computational resources during training and inference. These models typically focus on processing time series
patterns from a single domain (such as temporal or frequency), which limits their ability to capture the overall distribution
characteristics of the time series. To address these issues, we introduce a seasonal-trend decomposition method during
preprocessing, which decomposes the original time series into seasonal and trend components for separate modeling and
prediction. The trend component describes the overall development trend of the series, while the seasonal component
provides more detailed information. To further enhance the model's performance in long-term forecasting, we adaptively
select an appropriate kernel for decomposition based on different input sequences and model the complex seasonal
components from both temporal and frequency domains. Extensive experiments on four real-world datasets demonstrate
that the proposed TFBformer model significantly outperforms in terms of prediction accuracy. TFBformer reduced MSE
by 17% in multivariate forecasting compared to Autoformer. |
|
|
|
|
|
|
