Fusion State-of-Health Estimation of Lithium-Ion Batteries Based on Improved XGBoost Algorithm and Adaptive Kalman Filter

Abstract

Abstract Accurately predicting the state-of-health of lithium-ion batteries (LIBs) is of paramount significance for safety and stability of battery systems. This paper introduces a fusion model, which integrates the characteristic of data-driven model and equivalent circuit model to enhance precision. The first step is to preprocess data, including extracting health features, correlation screening, and compressing data. Subsequently, the hyperparameters of XGBoost algorithm are optimized using a weighted artificial bee colony algorithm, resulting in an improved XGBoost (IXGB) data-driven model. Finally, the observed values from the data-driven model and the prior values based on the equivalent circuit model are combined through adaptive Kalman filter (AKF), developing an improved XGBoost and adaptive Kalman filter (IXGB-AKF) fusion model, which makes the most of historical experience and the current state of LIBs. Validation is conducted using publicly available NASA Li-ion Battery Aging Datasets, with different datasets under various operating conditions, including different battery cells, different discharge depths and rates of LIBs. The resulting root mean square error values of the former three operating conditions are 1.834%, 2.570%, and 3.456%, respectively. The results indicate that the IXGB-AKF fusion model exhibits good accuracy and robustness under different operating conditions.