Mixing High-Frequency Bands Based on Wavelet Decomposition for Long-Term State-of-Charge Forecasting of Lithium-Ion Batteries

Although state-of-charge (SoC) forecasting has received considerable attention, long-term prediction remains a challenging task due to disrupted temporal dependencies and the neglect of battery signal characteristics. In this study, we propose a novel deep learning-based long-term SoC forecasting method that effectively captures temporal dynamics while preserving temporal order, thereby improving long-term predictive performance. Our approach first decomposes battery signals into low- and high-frequency bands using the discrete wavelet transform, enabling separate analyses of steady-state trends and localized events. Then, we introduce a feed-forward attention mechanism that selectively emphasizes informative high-frequency bands while suppressing irrelevant noise. Finally, we integrate the low- and high-frequency features generated solely by linear transformations that help maintain temporal structure and improve long-term forecasting accuracy. A series of experiments on a lithium-ion battery dataset demonstrate the superiority of the proposed method by achieving outstanding performance in long-term SoC forecasting.