Early-Charging Features and Heterogeneous Ensemble Learning for High-Fidelity Lithium-Ion Battery Capacity Estimation

Abstract

Accuracy capacity estimation in lithium-ion batteries serves as a cornerstone for battery management systems. Although recent data-driven methods have advanced this field, practical applications still face the challenge of reconciling high-quality feature extraction with minimal charging time. To resolve this conflict, we propose a capacity estimation framework that synergistically integrates early-stage charging segment features with a heterogeneous ensemble model. Specifically, four fixed local windows representing different early fast-charging stages are extracted from the first 5 min of charging data and concatenated into a compact 2.5 min sequence comprising SOC, current, voltage, and charged capacity as four-channel inputs. This sequence then feeds into a heterogeneous ensemble learning model composed of a convolutional neural network, a long short-term memory network, and a support vector machine, which collectively exploit spatial, temporal, and nonlinear characteristics of the data. Experiments conducted under different charging conditions confirm that the method achieves a root mean square error of 0.0170 Ah using early charging-feature data, outperforming traditional models in both accuracy and stability. Thus, the proposed framework offers a practical and efficient solution for online capacity estimation, striking a favourable balance between data economy and model generalizability.