Feature Selection for Cycle Life Prediction of Fast-Charged Lithium-ion Batteries

Abstract

Machine learning algorithms are widely used in data-driven predictive maintenance to address prognostics of the condition of lithium-ion batteries over their cycle life. However, selecting relevant features remains a critical issue when predicting the remaining useful life (RUL) of these batteries using data-driven approaches. This issue can significantly affect the performance of machine learning algorithms and lead to time loss. In this paper, we investigate the effectiveness of two feature selection techniques that use the Recursive Feature Elimination (RFE) method for predicting the RUL of fast-charged lithium-ion batteries. We use the RFE-LASSO and RFE-XGB methods for feature selection and the Elastic Net and Relevance Vector Regression models for RUL prediction. Experimental results using Nature Energy’s battery dataset show that the RFEXGB feature selection method can provide stable prediction performance using 33 or more features. Furthermore, when integrated with the Elastic Net model, RFE-XGB achieves the lowest prediction error at a train-test split of 80%-20%.