A proximity feature method for efficient capacity estimation of fast-charging lithium-ion batteries

To address the challenge of insufficient capacity estimation accuracy in fast-charging lithium-ion batteries due to stage-dependent heterogeneous degradation, this study proposes a Long Short-Term Memory (LSTM) neural network model based on proximate feature inputs. This proposed method first employs the K-means clustering algorithm to quantitatively analyze capacity decay rates, dividing the battery degradation process into three consecutive stages: the slow decay stage, the transition stage, and the fast decay stage. To address the degradation characteristics across these stages, the model is trained using locally proximate data rather than all-historical data. This approach effectively mitigates the adverse impact of early-cycle data on capacity prediction during the middle and late degradation stages. The method is validated using aging data from 33 lithium iron phosphate cells subjected to six fast-charging protocols. Experimental results demonstrate that, during the transition and fast decay stages, the three-time proximity-based model achieves a 20%–57% improvement in estimation accuracy compared to a baseline model trained on all-historical data, while simultaneously reducing training time by 55%–61%. Furthermore, the proposed framework exhibits robust adaptability across diverse prediction windows, offering an efficient and accurate solution for capacity estimation in fast-charging lithium-ion batteries. This proposed method first divides the battery degradation process into three consecutive stages: the slow decay stage, the transition stage, and the fast decay stage employing the K-means clustering algorithm. The performance of A Long Short-Term Memory (LSTM) neural network model trained by locally proximate data excels the all-historical data model. Validation by 33 LFP batteries under 6 fast-charging schemes demonstrates that 20%-57% higher estimation accuracy in transition and fast decay stages while reducing training time by 55%-61%. • Battery degradation is divided into slow decay, transition, and fast decay stages. • Proximity Feature Method trains an LSTM model using adjacent historical data. • The proximity-based model achieves higher accuracy and shorter training time. • Provides an efficient, stage-specific solution for capacity estimation.