FreqMLNet: Non-transformer network with frequency domain reconstruction and multi-scale representation for time...

Area of Science:

• Artificial Intelligence
• Data Science
• Signal Processing

Background:

• Traditional time-domain forecasting methods struggle with complex patterns.
• Existing frequency-domain methods often discard high-frequency information crucial for accuracy.
• Multilevel analyses inadequately integrate local and global time series features.

Purpose of the Study:

• Introduce FreqMLNet, a novel non-transformer architecture for enhanced time series forecasting.
• Combine frequency-domain reconstruction and multilevel feature representation for comprehensive analysis.
• Improve prediction accuracy by capturing both periodic patterns and multi-scale features.

Main Methods:

• Developed FreqMLNet, a novel non-transformer architecture.
• Implemented a frequency-domain reconstruction module to extract periodic patterns.
• Utilized multilevel feature representation to integrate information across multiple scales.

Main Results:

• Achieved an average 14.39% improvement in mean squared error on long-term datasets.
• Demonstrated an 11.03% improvement in symmetric mean absolute percentage error on short-term datasets.
• FreqMLNet showed superior robustness and prediction accuracy on complex forecasting tasks.

Conclusions:

• FreqMLNet offers a significant advancement in time series forecasting.
• The model effectively captures complex patterns by integrating frequency and multilevel features.
• FreqMLNet provides a more robust and accurate solution for diverse forecasting applications.