Wavelet Transformer: An Effective Method on Multiple Periodic Decomposition for Time Series Forecasting.

Area of Science:

• Artificial Intelligence
• Data Science
• Signal Processing

Background:

• Transformers show promise for long-term time series forecasting (LTSF) but struggle with point-wise self-attention for local/global characteristics.
• Real-world time series often exhibit multiple seasonal components and outliers, challenging existing forecasting models.

Purpose of the Study:

• To develop an advanced attention mechanism for time series forecasting that accurately captures local and global characteristics.
• To introduce a novel prediction technique, Waveformer, that leverages wavelet analysis for improved LTSF performance.

Main Methods:

• Utilized Maximal Overlap Discrete Wavelet Transform (MODWT) to create a Wavelet Attention (WA) mechanism.
• Proposed the Waveformer prediction technique integrating WA for enhanced time series analysis.
• Applied seasonal-trend decomposition methods to mitigate anomaly influence and extract periodic features.

Main Results:

• Waveformer effectively extracts multiple periodic features from time series data.
• The proposed method demonstrates improved precision in time series prediction, especially under seasonal-trend decomposition.
• Experimental evaluations on six real-world datasets show superior forecasting performance compared to state-of-the-art methods.

Conclusions:

• Waveformer successfully captures complex time series seasonal patterns through its multiple periodic decomposition strategy.
• The novel wavelet attention mechanism addresses limitations of point-wise self-attention in Transformers for LTSF.
• Waveformer offers a significant advancement in accurate and robust long-term time series forecasting.