State-of-Charge Estimation of Lithium-ion Battery Based on a Combined Method of Neural Network and Unscented Kalman filter

An accurate estimation of the battery State of Charge (SoC) is essential for reliable and energy-efficient operation of electric vehicles (EVs). Model-based algorithms have been ubiquitously accepted for SoC estimation due to their promising features. However, challenges remain concerning the elevated modeling precision requirement and appropriate filter parameter selection. This paper presents a novel combined model-based algorithm that instead of the prevailing implementation of the equivalent circuit model (ECM), an offline trained neural network (NN) is configured with an unscented Kalman filter (UKF) considering its capability of highly nonlinear battery modeling. Distinct profiles are employed to compare the modeling performances between NN and ECM. Subsequently, the proposed method is further explored with the residual-based adaptive covariance matching algorithm aiming to tune filter parameters dynamically. For comparison, ECM based EKF and UKF, along with the adaptive algorithm are also constructed. Ultimately, the presented filters are assessed and discussed under situations of initial offset, capacity error, and current sensor drift considering shunt thermal effects with real data obtained from WLTP lab results.