Prediction of Battery Cycle Life Using Early-Cycle Data, Machine Learning and Data Management

Abstract

The prediction of the degradation of lithium-ion batteries is essential for various applications and optimized recycling schemes. In order to address this issue, this study aims to predict the cycle lives of lithium-ion batteries using only data from early cycles. To reach such an objective, experimental raw data for 121 commercial lithium iron phosphate/graphite cells are gathered from the literature. The data are analyzed, and suitable input features are generated for the use of different machine learning algorithms. A final accuracy of 99.81% for the cycle life is obtained with an extremely randomized trees model. This work shows that data-driven models are able to successfully predict the lifetimes of batteries using only early-cycle data. That aside, a considerable reduction in errors is seen by incorporating data management and physical and chemical understanding into the analysis.