The heterogeneous stacking ensemble learning model for cross-cell capacity estimation by using discharging segments

Abstract

Accurate and robust capacity estimation is crucial for battery management systems to avoid failures and ensure reliable operation for lithium-ion batteries. This paper proposes a cross-cell capacity estimation method based on feature extraction of discharging voltage segments and a heterogeneous Stacking ensemble learning model. First, a voltage partition strategy is adopted to extract features from discharging segments, and the most important voltage range (4.0 V–3.3 V) is identified through correlation analysis. Next, a novel Stacking ensemble learning model is constructed, integrating the complementary advantages of eXtreme Gradient Boosting (XGBoost), Temporal Convolutional Network (TCN), and Long Short-Term Memory Network (LSTM). Moreover, a battery group-fold cross-validation strategy and one-hot encoding of cell ID are proposed to enhance the cross-cell generalization ability of the model. Finally, a case study is implemented to verify the effectiveness, and the results show that the proposed capacity estimation method achieves a root mean square error within 0.0085 Ah and a mean absolute percentage error ≤0.67 % for 1.1 Ah cells. Compared with single models, the proposed method demonstrates the lowest error. The results highlight its superior accuracy, robustness, and adaptability to different degradation patterns, enabling cross-cell capacity estimation.