OmniTIEFormer: A tri-branch transformer with cross-scale transfer learning for multi-scale battery life-cycle forecasting

Accurately predicting the remaining useful life (RUL) of lithium-ion batteries is fundamentally challenging due to their multi-scale degradation dynamics, which range from approximately linear trends to complex nonlinear behaviors. Moreover, existing models often struggle to generalize across different battery types, limiting their practicality in real-world applications. To address these issues, we propose OmniTIEFormer, a novel multi-scale Transformer architecture designed for end-to-end (E2E) battery lifecycle management. OmniTIEFormer features a three-branch parallel architecture that simultaneously captures local fluctuations, regional trends, and global degradation patterns from historical capacity sequences. These multi-scale representations are fused by a Tri-branch Cross-Exchange Module (TCEM), which enables effective interaction and integration of information across scales. We demonstrate the effectiveness of OmniTIEFormer through extensive experiments on multiple public datasets exhibiting diverse degradation behaviors. OmniTIEFormer outperforms state-of-the-art baselines, reducing mean absolute error by up to 44% and end-of-life (EOL) prediction error by up to 74%. The model’s robustness and generalization capability are further demonstrated in a challenging cross-dataset, cross-scale transfer learning setting: when pre-trained on small-capacity cells, it adapts effectively to large-format industrial batteries. This indicates that OmniTIEFormer captures fundamental, transferable degradation patterns rather than overfitting to specific datasets. These findings position OmniTIEFormer as a robust, data-efficient solution for battery prognostics, particularly valuable in industrial scenarios where data for new battery types is often limited. The code is publicly available at: https://github.com/keepawakeyi/OmniTIEFormer . • Tri-branch Transformer captures local, regional, and global degradation cues. • TCEM enables cross-scale feature exchange and gated fusion for forecasting. • Outperforms SOTA: −44% MAE and −74% EOL error across multiple datasets. • Robust on CALCE, PANASONIC, and GOTION, handling non-monotonic patterns. • Transfers from small cells to large industrial batteries with strong accuracy.