Full-Period One-Shot Prediction of Li-Ion Battery Aging Trajectories via Multiview Encoding

Data-driven methods for battery aging trajectory prediction are limited in performance due to feature redundancy. To conquer this defect, this paper proposes a prediction framework via multi-view encoding. The method introduces three complementary views: capacity decay, material activity, and polarization shift, and uses independent encoders to extract aging information from each view. It then employs a multi-head attention mechanism to adaptively aggregate features, and finally generates the aging trajectory using a BiLSTM decoder. Experimental results show that, compared with early-fusion baselines, the proposed framework reduces MAPE by 47.5%, achieving a value of 1.05%. In addition, an R2 of 0.87 confirms that the model can accurately reconstruct the aging trajectory. Moreover, the model demonstrates excellent adaptability across aging stages, with prediction error decreasing from 1.16% in the early stage to 0.65% in the later stage. This work establishes a principled multi-view encoding paradigm and provides novel insights for achieving high-reliability state estimation in battery management systems.