Rethinking Activation Function: A Simple Path to Efficient and Accurate Time Series Forecasting
IEEE Transactions on Knowledge and Data Engineering · 2026-10 · 38 (10) · 6987-7004
DOI: 10.1109/tkde.2026.3716656 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Knowledge and Data Engineering, 2026-10. Verfasst von Fan Zhang und 2 weiteren Autor:innen. Die Arbeit wurde bisher 4-mal zitiert (Crossref) und nennt 94 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Special Funds for the Taishan Scholars Project, Youth Innovation Technology Project of Higher Education Institutions in Shandong Province.
Abstract (Originalsprache)
Long-term time series forecasting is a critical task in numerous fields, however, even state-of-the-art models often face a trade-off between prediction accuracy and computational efficiency. While a vast body of research has focused on architectural innovations, we have identified a fundamental yet overlooked bottleneck within a core component: the activation function. Widely-used functions such as the Gaussian Error Linear Unit (GELU), with their fixed functional form, lack the flexibility to adapt to diverse temporal patterns, thereby limiting the model’s expressive power and final performance. To address this issue, we propose the Piecewise Polynomial Activation (PPA) function, a novel function constructed via Hermite cubic spline interpolation that retains the beneficial properties of GELU while offering superior functional flexibility and training stability. Building on the foundation of PPA, we design a simple yet efficient time-frequency framework, TFPPA, which utilizes a multi-scale temporal learner to process the trend component and a frequency-domain learner for the seasonal component. Extensive experiments on multiple benchmark datasets demonstrate that TFPPA surpasses state-of-the-art baseline models, achieving an exceptional balance between prediction accuracy and efficiency. Furthermore, the proposed PPA can be easily integrated into existing models as a universal building block. Our work demonstrates that re-evaluating fundamental network components offers a powerful avenue for advancing the field of time series forecasting.
Förderung
- National Natural Science Foundation of China
- Special Funds for the Taishan Scholars Project
- Youth Innovation Technology Project of Higher Education Institutions in Shandong Province
Zitieren
APA
Zhang, F., Fan, S., & Wang, H. (2026). Rethinking Activation Function: A Simple Path to Efficient and Accurate Time Series Forecasting. IEEE Transactions on Knowledge and Data Engineering, 38(10), 6987-7004. https://doi.org/10.1109/tkde.2026.3716656
Vancouver
Zhang F, Fan S, Wang H. Rethinking Activation Function: A Simple Path to Efficient and Accurate Time Series Forecasting. IEEE Transactions on Knowledge and Data Engineering. 2026;38(10):6987-7004. doi:10.1109/tkde.2026.3716656
BibTeX
@article{Zhang2026,
author = {Zhang, Fan and Fan, Shiming and Wang, Hua},
title = {Rethinking Activation Function: A Simple Path to Efficient and Accurate Time Series Forecasting},
journal = {IEEE Transactions on Knowledge and Data Engineering},
year = {2026},
volume = {38},
number = {10},
pages = {6987--7004},
doi = {10.1109/tkde.2026.3716656}
}
EUVIMED