KEDformer: Knowledge extraction seasonal trend decomposition for long-term sequence prediction

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Time series forecasting is crucial for energy, finance, and meteorology.
• Existing Transformer models struggle with computational inefficiency and long-term sequence generalization.

Purpose of the Study:

• Introduce KEDformer, a novel framework to address limitations in Transformer-based time series forecasting.
• Improve computational efficiency and generalization for long-term sequences.

Main Methods:

• KEDformer integrates knowledge extraction and seasonal-trend decomposition.
• Utilizes sparse attention and autocorrelation mechanisms.
• Reduces computational complexity from O(L^2) to O(L log L).

Main Results:

• KEDformer demonstrates superior performance across five public datasets (energy, transportation, weather).
• Achieved an average improvement of 10.4% in Mean Squared Error (MSE) prediction accuracy.
• Achieved an average improvement of 2.9% in Mean Absolute Error (MAE) prediction accuracy.

Conclusions:

• KEDformer effectively captures short-term fluctuations and long-term patterns in time series data.
• The proposed framework offers a more efficient and accurate solution for complex forecasting tasks.
• KEDformer outperforms traditional models in various real-world applications.