21 October 2026 in SSRN Electronic Journal
Cell-to-cell variations in degradation trajectories limit the cross-cell generalization of state of health (SOH) prediction models trained on one lithium-ion cell. Training with labeled data from multiple cells can improve generalization, but SOH labels can only be obtained through full-capacity tests that interrupt operation, making such data difficult to acquire. To solve this problem, this paper proposes a cross-cell SOH prediction framework based on the single particle model with electrolyte dynamics (SPMe). Automatic matching of baseline parameter sets and two-stage particle swarm optimization identify electrochemical parameters from the discharge voltage and current of each cycle, producing parameter sequences for each cell. For each dataset group, only one cell with SOH labels is used for feature selection and model training. Electrochemical features are selected by considering their correlations with SOH in the training cell, parameter redundancy, and physical meaning. A long short-term memory network learns the relationship between the selected feature sequences and SOH to predict future SOH values. The trained model is directly applied to target cells without adaptation. Target-cell SOH labels are used only for evaluation. The framework is validated on five dataset groups covering four cathode materials. With equal weighting across groups, the root mean square error and mean absolute error for 10-step-ahead prediction are 0.0269 and 0.0212, respectively. These values are 12.52% and 15.06% lower than those obtained using voltage statistical features. The results show that electrochemical parameters identified using the SPMe provide useful degradation information for cross-cell SOH prediction without target-cell SOH labels during model development.