Abstract
Time series forecasting in the frequency domain often encounters "spectral entanglement," where trends, seasonality, and noise overlap, complicating accurate prediction. We present the Frequency Decomposition Network (FreDN), which utilizes a learnable Frequency Disentangler and an efficient ReIm Block to separate these components without the high computational costs of complex-valued operations. FreDN achieves up to 10% improvement in Mean Squared Error (MSE) over state-of-the-art methods while reducing parameter counts by over 50% and training time by up to 75%.
Introduction
Accurate time series forecasting is essential for various domains, including energy management and traffic control. Recent advancements have leveraged Fourier transforms to provide neural networks with a global perspective on periodic patterns. However, real-world data is often non-stationary, containing trends and noise that lead to spectral entanglement. This phenomenon, caused by spectral leakage, smears the energy of different signal components across the frequency spectrum, making clean separation difficult for traditional methods like moving averages or Top-K frequency selection.
Methodology
FreDN addresses these issues through two primary innovations:
- Learnable Frequency Disentangler: Instead of using rigid mathematical filters, FreDN applies a learnable mask directly to the frequency spectrum. It adaptively assigns a "trend" and "seasonal" weight to every frequency component. The disentangled trend is processed in the time domain to capture gradual variations, while the seasonal part remains in the frequency domain.
- ReIm Block: To avoid the massive overhead of complex-valued matrix multiplications, the authors introduce the ReIm (Real-Imaginary) Block. It splits complex representations into real and imaginary parts and processes them through a shared real-valued MLP. Mathematical proofs show this lightweight setup retains the rotational and phase-shifting expressiveness of heavy complex-valued layers.
Additionally, the model utilizes a Frequency-Domain MAE Loss. Analysis reveals that this loss provides globally coordinated gradient updates, helping the model learn long-range patterns faster than time-domain losses that focus on isolated time steps.
Experimental Results
FreDN was evaluated on seven standard long-term forecasting benchmarks, including ETT, Weather, Electricity (ECL), and Traffic datasets.
- State-of-the-Art Accuracy: FreDN ranked first in 30 out of 35 MSE tasks, outperforming recent SOTA models (like FreDF, SOFTS, and TimeMixer) by up to 10%.
- Efficiency Gains: The ReIm Block enables a 50% reduction in parameters and cuts average training time by up to 75% on certain datasets without sacrificing performance.
- Scalability: Unlike many forecasting models that degrade with longer historical lookback windows, FreDN consistently improves as the input sequence length increases (tested up to 720 steps).
Conclusion and Future Work
FreDN offers a robust and efficient solution for frequency-domain time series forecasting by effectively disentangling spectral components. While highly adaptable, the current model relies entirely on data-driven learning. Future work may explore incorporating explicit signal processing priors (like enforced smoothness) and more expressive complex modeling techniques for highly irregular systems.
Summary updated