Adaptive Energy Amplification for Robust Time Series Forecasting

Runchang Liang, Haihua XU, Yuanbo Xu, Pengyang Wang

Keywords: Time Series Forecasting; Frequency-Domain Learning; Adaptive Energy Amplification; Spectral Bias

Abstract:

Deep learning models for time series forecasting often exhibit a spectral bias, prioritizing high-energy, low-frequency components while underfitting predictive but low-energy, high-frequency signals.
Existing efforts attempt to correct this by amplifying high-frequency components but suffer from indiscriminate amplification, enhancing both meaningful signals and task-irrelevant noise, which destabilizes training and impairs generalization.
To address this, we propose AEA (Adaptive Energy Amplification), a novel framework that reframes the problem as one of adaptive signal enhancement.
AEA introduces two synergistic innovations: (1) a Spectral Mirroring mechanism that constructs a phase-preserving, low-frequency surrogate to guide targeted, distortion-free amplification of high-frequency signals;
and (2) a lightweight Differential Embedding module that operates in a latent space to adaptively suppress common-mode noise.
By decoupling signal amplification from noise suppression, AEA selectively enhances only informative features.
Extensive experiments show that our model-agnostic framework consistently improves the forecasting performance of state-of-the-art backbones in both long-term and short-term forecasting tasks, while significantly enhancing training stability and generalization.
The code repository is available at https://anonymous.4open.science/r/AEA-685E/.

Primary Area: learning on time series and dynamical systems

Submission Number: 15734