Online Tuning of Extended Kalman Filter Using Reinforcement Learning for Improved Battery State-of-Charge Estimation

Accurate state-of-charge (SoC) prediction is important to determine the achievable service time of lithiumion batteries. Kalman filter (KF) is one of the most widely used methods for battery state prediction yielding promising SoC estimation results. However, since KF is a model-based approach, its performance degrades in the presence of modeling nonlinearities resulting in poor estimation accuracy, e.g., in low-SoC operating conditions. To address this issue, this paper puts forward an unorthodox approach for the online calibration of KF in battery SoC estimation. The proposed method is based on the classic extended KF (EKF) and battery Thevenin model, which are improved with reinforcement learning (RL). RL is used for online tuning of the EKF's noise covariance matrices to handle varying modeling inaccuracies during battery operation, which is hard to balance in EKF using fixed filtering settings. The results show that the proposed method reduces the estimation error by about 0.5% compared to the EKF tuned based on the well-established genetic optimization.