Data-driven Capacity Estimation of Lithium-Ion Batteries Using Machine Learning Methods

On-line battery State of Health prediction is essential for the safe and reliable operation of electrically powered machines and electronics. In this work, we present a Machine Learning (ML) pipeline which uses pre-collected charging data in order predict the battery capacity on-line and without the need for a lengthy intrusive capacity measurement. The model is based on Gradient Boosting, namely XGBoost which is an efficient version of this algorithm. It achieves an average error rate of 1% (Mean Absolute Error) while being very quick in inference, in the order of a few seconds. Moreover, it can be easily expanded to work on other devices and scenarios by applying similar steps with new training data.