State of health estimation of lithium-ion battery packs based on global-divergent features and multi-layer stacking ensemble model

The state of health (SOH) of lithium-ion batteries is a critical indicator for assessing their operational status. However, accurately estimating the SOH of battery packs is challenging because of the inherent nonlinear degradation characteristics of batteries, performance fluctuations caused by dynamic working conditions, and inconsistencies among individual cells. To address this issue, a method for estimating the SOH of lithium-ion battery packs based on global-divergent features and a Multi-Layer Stacking ensemble model is proposed. First, the aging test data of the battery pack at various cycle stages is collected. Then, the area of the constant-current charging curve of the battery pack is extracted to characterize its global features. Additionally, the combined temperature-weighted voltage inconsistency feature is calculated as a divergent feature to reveal the degree of inconsistency among the battery cells. These two features work together to characterize both the global characteristics of the battery pack and the consistency of the cells. Next, a construction method for the multi-layer stacking ensemble model is proposed. In the first layer, the Long Short-Term Memory Network (LSTM), Convolutional Neural Network (CNN), and Multilayer Perceptron (MLP) are used as base learners to collaboratively model temporal dependencies, local pattern features, and global mapping capabilities. In the second layer, Ridge regression is used as the output result of the meta-learner, fusing each base learner to achieve weighted integration and error suppression of multi-feature information. The proposed method was validated using a laboratory lithium battery pack dataset. The experimental results demonstrate that the proposed method can accurately estimate the SOH of the battery pack, achieving a final Mean Absolute Error (MAE) of 0.0035 and a Root Mean Square Error (RMSE) of 0.0045. The coefficient of determination (R2) is 0.9917. Compared with traditional methods, the method proposed in this study demonstrates superior performance in estimating SOH.