Data-Driven Capacity Estimation Method for Retired Lithium-Ion Batteries Based on Electrochemical Impedance Spectroscopy

With the rapid expansion of global electric vehicles (EVs) deployment, the echelon utilization of retired lithium-ion batteries (LIBs) has emerged as a critical issue. Although these batteries typically retain over 70% of their initial capacity and remain suitable for stationary energy storage systems, the substantial variability in aging states poses safety risks. Conventional capacity estimation methods are often time-intensive and costly, while data-driven approaches face challenges from complex degradation mechanisms and limited historical usage data. This study uses the electrochemical impedance spectroscopy (EIS) method to create a model that estimates the capacity of retired batteries. EIS offers fast measurement, requires no historical cycling data, and provides rich state-of-health (SOH) information. An EIS dataset was acquired from 18650-type LFP and NCM cells aged under multiple cycling conditions. The real part and magnitude of the impedance spectra were extracted as input features for model training. A hybrid deep learning framework integrating the sparrow search algorithm (SSA), convolutional neural networks (CNN), gated recurrent units (GRU), and an attention mechanism was developed. SSA automatically optimize model hyperparameters, mitigating the overfitting risks, while the attention mechanism highlighted informative frequency-domain features, reducing manual feature engineering and enhancing prediction accuracy. Results show excellent performance: for LFP cells, the root-mean-square error (RMSE) and mean absolute error (MAE) are 0.24% and 0.19%, respectively, with a coefficient of determination (R²) of 98.96%; for NCM cells, the RMSE and MAE are 0.99% and 0.88%, with R² of 97.97%. On the mixed-material dataset, the RMSE, MAE, and R² reach 0.79%, 0.67%, and 97.84%. These results confirm that the proposed method maintains high accuracy across different cathode chemistries, while significantly reducing testing and modeling costs. The approach shows strong potential for large-scale, automated screening and classification of retired LIBs in practical second-life applications.