GLF-Former: Dual-Attention Global–Local Fusion for Battery Remaining Useful Life Prediction

Abstract

Accurate prediction of the remaining useful life (RUL) of lithium-ion batteries is essential for operational safety and predictive maintenance. However, accurate RUL estimation from incomplete degradation trajectories remains challenging because of limited observable history, nonlinear capacity degradation, local fluctuations, and error accumulation during long-term forecasting. This study proposes GLF-Former, a Transformer-based framework with a global–local dual-branch architecture for battery capacity and RUL prediction. The global branch represents long-term degradation trends, whereas the local branch focuses on short-term capacity variations. A structure-aware attention (SAA) module and an aligned channel-fusion (ACF) module are further introduced to enhance global–local feature interaction and adaptively recalibrate the fused degradation representations. The proposed model was evaluated on representative cells from the CALCE, NASA, and TJU public datasets under multiple prediction starting points. Across the 12 evaluated cell starting point settings, GLF-Former achieved capacity-prediction MAE values of 0.0019–0.0068 Ah and RMSE values of 0.0026–0.0152 Ah, with R2 values ranging from 0.9644 to 0.9997. The corresponding mean AE values ranged from 0.4 to 1.1 cycles across the evaluated settings. The results demonstrate that GLF-Former achieves competitive capacity-prediction and RUL-estimation performance under the evaluated settings, particularly when the available degradation history is limited.