Frequency-Guided Transformer Modeling with Spectral Enhancement for Accurate Time Series Forecasting

Abstract

This paper proposes a forecasting framework that integrates frequency-aware spectral enhancement and frequency-guided attention mechanisms to address the challenges of insufficient modeling of periodic structures and interference from redundant information in time series prediction. The method first applies a frequency-aware spectral enhancement module to perform Fourier Transform on the input sequence. It selects and amplifies key frequency components to highlight dominant periodic information and suppress the influence of redundant disturbances. Then, during the encoding phase, a frequency-guided attention mechanism is introduced. It uses spectral response to guide attention distribution, allowing the model to focus more on high-weight structural segments when constructing contextual representations. This improves the ability to model cross-period dependencies and long-term relationships. Both modules are modular and structurally independent. They can be flexibly integrated into various Transformer architectures to enhance structural representation in feature extraction and modeling. Experimental results on multiple real-world time series datasets demonstrate that the proposed method improves prediction accuracy across several evaluation metrics while maintaining computational efficiency, showing strong adaptability to complex periodic structures and dynamic fluctuations.