Lithium-ion batteries remaining useful life prediction via Fourier-mixed window attention enhanced Informer with decomposition and adaptive error correction strategy

Abstract

Remaining useful life (RUL) prediction for lithium-ion batteries is crucial for safe and reliable operation in energy storage systems (ESS). However, the complex characteristics of capacity degradation make accurate and stable RUL prediction a critical problem. In response, this paper proposes a novel framework called FFWinformerAGA, which merges Fourier Decomposition Method (FDM), Fourier-mixed Window Attention-enhanced Informer (FWinformer), and adaptive error correction strategy via Gated Recurrent Unit-Attention Mechanism (AGA). The FDM initially decomposes original sequence into detail and trend components, effectively reducing the nonlinearity. Utilizing two distinct streams, the FWinformer then specially integrates local and global information in both time and frequency domains of the detail component, significantly enhancing the capture of abrupt changes and long-term dependencies under limited samples. Additionally, the AGA is incorporated to mine predictive relationships in the residual series. Experiments and analysis on four real-world datasets yielded the following conclusions: each component within the FFWinformerAGA is demonstrated to be necessity and superior, leading to enhanced results. The model outperforms state-of-the-art models, with improvements in Root Mean Square Error ranging from 30.39% to 85.30%, while maintaining reasonable performance times. Furthermore, the robustness of FFWinformerAGA is demonstrated by training it with fewer degradation samples and incorporating three different types of noise as additional inputs. Findings of this study hold potential application value in prognostics and health management of ESS.