State of health estimation for lithium-ion battery based on energy features

Abstract

There is a recognized need to forecast lithium-ion batteries' state of health (SOH) to guarantee their safety and reliability. However, the selected health indicators highly influence the prognostics accuracy of SOH. This paper's primary purpose is to assess the applicability and prediction accuracy of the proposed energy features-based SOH estimation model for different lithium-ion batteries under varied charging and discharging scenarios. These health indicators are energy in the constant current (CC) charging phase, constant voltage (CV) charging stage, and energy in the equal discharge voltage interval (EDVI). The proposed SOH estimation model employs a machine learning algorithm based on Gaussian process regression (GPR). The validation scheme utilizes two data training modes. In addition, data sets from MIT, CALCE, NASA, and Oxford containing different charge and discharge conditions and lithium-ion battery types are adopted. The experimental results reveal that the prediction errors are less than 0.5% for both training modes, while the coefficient of determination (R2) is more than 97%. In addition, 95% of tested cells had an R2 value of more than 98%. This research suggests that the proposed energy feature-based SOH estimation model has high prediction accuracy and excellent generalization ability.