Two-stage deep learning for online prediction of knee-point in Li-ion battery capacity degradation

Abstract

Accurate monitoring of capacity degradation of a lithium-ion battery is important as it enables the user to manage the battery usage for optimal performance/lifetime and to take preemptive actions against any potential explosion or fire. Battery capacity fades gradually through repetitive charging and discharging until it reaches the so called ‘knee-point’, after which it goes through rapid and irreversible deterioration to reach its end-of-life. It is crucial to forecast the knee-point early and accurately for safety and economic use of the battery. Machine learning based methods have been used to predict the knee-point with early cycles cell data. Despite some notable progress made, the existing methods make the unrealistic assumption of constant cycle-to-cycle charge/discharge operation. In this study, a novel two-stage deep learning method is proposed for online knee-point prediction under variable battery usage. A CNN-based model extracts temporal features across past and current cycles to sort out those cells should be monitored closely for near-term failures, and then predict the number of cycles left to reach the knee-point for them. The proposed method extracts features from time-series data and thus reflects dynamic changes in battery properties, resulting in improved prediction performance under realistic scenarios.