FreAPNet: Learnable Amplitude–Phase Parameterization and Dual‐Path Frequency Learning for Multivariate Time Series Forecasting

View at publisher Summarise with VIRAT
Authors: Nannan Xu, Kecheng Xu, Yuanyuan Cui, Guo Yang, Meng Huang
Cited by: 0
References: 30

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Multivariate time series forecasting remains challenging. Key hurdles include long‐range temporal dependencies, temporal misalignment, and complex cross‐variable interactions. This paper presents FreAPNet, a frequency‐domain forecasting framework built upon two complementary components. The first component is a learnable amplitude–phase parameterization. It separately models spectral magnitude and phase to refine periodic strength and temporal alignment. The second is a dual‐path frequency learning architecture, which captures both temporal‐frequency evolution and channel‐frequency coupling. Together, these designs provide a structured and physically meaningful representation. This representation overcomes the limitations of conventional time‐domain approaches. Extensive experiments on multiple benchmark datasets show that FreAPNet achieves state‐of‐the‐art or competitive accuracy in long‐horizon forecasting. Notably, the model maintains a lightweight size and favorable computational efficiency. Ablation results further reveal a clear contribution hierarchy: amplitude modeling provides the greatest performance gain, followed by phase refinement and dual‐path frequency modeling. These findings demonstrate the effectiveness of structured, learnable frequency‐domain modeling for multivariate time series forecasting.

Fields of Study & Subjects

Time Series Analysis and Forecasting Forecasting Techniques and Applications Traffic Prediction and Management Techniques

Publication details

Journal / source
Concurrency and Computation Practice and Experience
Volume
38
Issue
7
Pages
e70675
Publication date
31 March 2026
Language
EN