State-of-charge estimation for a pouch-type lithium cobalt oxide battery using an adaptive generalized maximum correntropy strong-tracking cubature kalman filter

Accurate state-of-charge estimation is essential for the safe and efficient operation of lithium-ion battery management systems. However, measurement outliers, time-varying noise statistics, model mismatch, and abrupt operating-condition changes can degrade the accuracy and stability of conventional Kalman-filter-based estimators. To address these problems, this study develops an adaptive generalized maximum correntropy strong-tracking cubature Kalman filter (AGMCC-STCKF) for a pouch-type lithium cobalt oxide battery. The principal methodological contribution is a residual-statistics-based adaptive kernel-width strategy. At each sampling instant, the kernel width is updated using a weighted combination of the standard deviation of a sliding innovation-residual window and the instantaneous innovation magnitude, followed by bounded projection to prevent kernel-width degeneration. The generalized maximum correntropy criterion is used to attenuate abnormal measurements, while the strong-tracking fading factor and Sage–Husa estimator improve responsiveness to model mismatch and time-varying noise statistics, respectively. A second-order RC equivalent-circuit model is established, and its parameters are identified using forgetting-factor recursive least squares. Experimental validation is conducted at 25 °C under UDDS and DST operating conditions. Across 30 paired Gaussian-noise realizations, the proposed method achieves mean MAE, RMSE, and MaxAE values of 0.4040%, 0.4571%, and 0.8653%, respectively, corresponding to reductions of 19.15%, 28.51%, and 72.50% relative to CKF. Under weak, medium, and strong non-Gaussian disturbances, the proposed method also achieves the lowest mean errors among the compared CKF-based estimators. External validation using the CALCE CS2_35 dataset yields an RMSE of 0.7183%. The mean single-step runtime is 0.871 ms, which is substantially lower than the 1 s sampling interval. These results demonstrate improved estimation accuracy and robustness within the investigated cell type, temperature, and operating conditions.