A LightGBM-Regulated and Residual-Compensated Adaptive Extended Kalman Filter for State of Charge Estimation Under Unseen Driving Cycles

Data-driven methods have been extensively used for state of charge (SOC) estimation because of their nonlinear regression capability. However, when the driving cycle differs from the training conditions, their estimation accuracy may deteriorate because of insufficient generalization ability. This problem becomes more pronounced under complex unseen dynamic conditions, while some data-driven methods also suffer from high computational complexity, which further limits their practical application. To overcome these issues, this study presents a LightGBM-regulated and residual-compensated adaptive extended Kalman filter method for SOC estimation. In the proposed method, battery dynamics are represented by a second-order equivalent circuit model, while SOC is recursively estimated using an adaptive extended Kalman filter (AEKF). LightGBM is used to predict covariance adjustment coefficients for adaptively regulating Q and R, and to compensate for the remaining SOC estimation error after AEKF processing. The proposed method is validated under multiple dynamic driving cycles and different temperature conditions. The experimental results of the main unseen-cycle tests show that the proposed method achieves good estimation accuracy and generalization performance under unseen driving cycles, with the maximum error (ME) and root mean square error (RMSE) not exceeding 2.26% and 1.28%, respectively.