Transfer Learning Based on Transferability Measures for State of Health Prediction of Lithium-Ion Batteries

Abstract

Lithium-ion (Li-ion) batteries are considered to be one of the ideal energy sources for automotive and electronic products due to their size, high levels of charge, higher energy density, and low maintenance. When Li-ion batteries are used in harsh environments or subjected to poor charging habits, etc., their degradation will be accelerated. Thus, online state of health (SOH) estimation becomes a hot research topic. In this study, normalized capacity is considered as SOH for the estimation and calculation of remaining useful lifetime (RUL). A multi-step look-ahead forecast-based deep learning model is proposed to obtain SOH estimates. A total of six batteries, including three as source datasets and three as target datasets, are used to validate the deep learning model with a transfer learning approach. Transferability measures are used to identify source and target domains by accounting for cell-to-cell differences in datasets. With regard to the SOH estimation, the root mean square errors (RMSEs) of the three target batteries are 0.0070, 0.0085, and 0.0082, respectively. Concerning RUL prediction performance, the relative errors of the three target batteries are obtained as 2.82%, 1.70%, and 0.98%, respectively. In addition, all 95% prediction intervals of RUL on the three target batteries include the end-of-life (EOL) value (=0.8). These results indicate that our method can be applied to battery SOH estimation and RUL prediction.