Adaptive Spectral Enhancement Improves Multivariate Time Series Forecasting Accuracy Across Eight Datasets

A novel multivariate time series forecasting framework called ASE-TS achieves superior forecasting accuracy and empirically reduces the under-utilization of low-energy mid-to-high-frequency components across eight real-world datasets, according to a study published in Data Mining and Knowledge Discovery by researchers Jiayi Zhu, Zhihua Ding, and Weiqiang Sun.

Time series forecasting represents a critical task across diverse operational domains, where mid-to-high-frequency components capture essential short-term variations required for improved predictive performance. However, existing representative models frequently under-utilize low-amplitude mid-to-high-frequency signals whenever a spectrum is dominated by low-frequency energy, resulting in suboptimal forecasting outcomes.

To address these spectral imbalances, the proposed Adaptive Spectral Enhancement (ASE) framework incorporates two core components: a Spectral Enhancement Module and a Spectral Restoration Module. The Spectral Enhancement Module adaptively amplifies informative low-energy mid-to-high-frequency components to mitigate under-utilization, while the Spectral Restoration Module restores the original spectral energy distribution to prevent distortion during modeling.

Due to its modular design, the adaptive spectral enhancement approach functions as a general frequency-domain strategy capable of integration into various deep learning forecasting architectures. To facilitate reproducibility and further investigation, the researchers have made the underlying implementation code openly accessible via an online GitHub repository.

Subject of Research: Adaptive spectral enhancement for deep learning time series forecasting

Article Title: ASE-TS: adaptive spectral enhancement for time series forecasting with deep learning models

Article References: Zhu, J., Ding, Z., & Sun, W. (2026). ASE-TS: adaptive spectral enhancement for time series forecasting with deep learning models. Data Mining and Knowledge Discovery, 40(6), Article 121. https://doi.org/10.1007/s10618-026-01287-8

Image Credits: AI Generated

DOI: 10.1007/s10618-026-01287-8

Keywords: time series forecasting, adaptive spectral enhancement, ASE-TS, deep learning, multivariate time series, frequency-domain enhancement, spectral enhancement module, spectral restoration module, mid-to-high-frequency components, short-term variations, low-frequency energy, spectral energy distribution, energy distortion prevention, suboptimal forecasting, real-world benchmark datasets, open-source forecasting code, Data Mining and Knowledge Discovery