High-accuracy state-of-charge fusion estimation of lithium-ion batteries by integrating the Extended Kalman Filter with feature-enhanced Random Forest

Abstract

Fusing data-driven techniques with model-based methods is a key focus in lithium-ion battery state-of-charge (SOC) estimation research. Previous studies have often utilized data-driven techniques to compensate for errors inherent in model-based methods. However, challenges such as feature acquisition, interpretability, and overfitting limit their effectiveness. This paper proposes a novel method for high-accuracy SOC estimation. Parameters of the dual polarization (DP) model are identified and utilized as feature inputs for Random Forest (RF). The suitability of these features is evaluated using maximal information coefficient and RF feature importance scoring. An enhanced RF model with seven feature inputs (RF-7F) significantly improves estimation accuracy. An innovative Extract Segment Fusion method integrates the Extended Kalman Filter (EKF) and RF-7F, resulting in a high-accuracy and robust SOC estimation approach termed EKF-RF-ESF (ERFE) method. Validation across five driving cycle tests (DST, FUDS, US06, BJDST, and NEDC) shows that ERFE achieves mean absolute errors (MAE) and root mean squared errors (RMSE) below 0.080 % and 0.107 %, respectively. Compared to EKF and RF-7F, ERFE reduces MAE by an average of 89.762 % and 49.279 %, and RMSE by an average of 87.673 % and 69.426 %, respectively. This method shows significant potential for application in electric vehicles and large-scale energy storage systems.