AGLFormer: Compact feature selection and parameter-budget-aware global-local modeling for lithium-ion battery state-of-health estimation

Accurate state-of-health (SOH) estimation is essential for lithium-ion battery management, while practical deployment requires an appropriate balance among input compactness, prediction accuracy, model capacity, and performance stability. This study proposes a deployment-aware SOH estimation framework based on an Adaptive Global-Local Former (AGLFormer). In the feature-engineering stage, statistical health features are extracted from charging profiles and ranked using an equally weighted combination of the Pearson and Spearman correlation coefficients. Four nested feature subsets, namely FS-1, FS-25, FS-50, and FS-100, are subsequently constructed to investigate the effect of feature quantity on SOH prediction performance. In the prediction stage, AGLFormer combines convolution-enhanced linear attention with an adaptive multi-scale feed-forward network to model global degradation dependencies and local degradation variations. Experiments on the CALCE and Oxford datasets show that increasing the number of input features does not consistently improve prediction accuracy. FS-25 achieves the best validation performance on CALCE, whereas the more compact FS-1 performs best on Oxford, demonstrating that an appropriately compact feature subset can maintain or improve prediction accuracy, enhance the reachability of high-accuracy results across different hyperparameter configurations, and reduce feature-extraction and data-preprocessing overhead. Compared with long short-term memory (LSTM), Transformer, convolutional neural network–Transformer (CNN-Transformer), iTransformer, and ModernTCN, AGLFormer achieves the best average performance on both datasets. Its average R 2 , mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) are 0.9886, 0.0143, 0.0206, and 0.0250 on CALCE and 0.9943, 0.0028, 0.0035, and 0.0034 on Oxford, respectively. Parameter-budget analysis further shows that AGLFormer achieves the best overall budget-matched performance on both datasets, with consistent advantages on Oxford and capacity-dependent gains on CALCE. These results demonstrate that the proposed framework provides a favorable balance among feature compactness, prediction accuracy, performance stability, and model capacity for lithium-ion battery SOH estimation.