Design of a machine learning pipeline for data-driven li-ion degradation modeling

Abstract

The growing demand for Lithium ion (Li-ions) storage solutions requires continuous improvements in the battery system, especially with regard to degradation issues, which is the main challenge to be addressed in storage technologies. Determining battery state parameters, such as State of Health (SoH), is a crucial task for the Battery Monitoring System (BMS), as it is used as a common health indicator of the battery system. Data-driven methods can be very advantageous for complex problems such as degradation, where it is too difficult to write detection algorithms. The programs are shorter and very accurate, but they need a metric to evaluate performance and quantify uncertainty.

In this paper we design and evaluate a Machine Learning (ML) pipeline to estimate SoH, for 8 Kokham 740 mAh Lithium-ion cells cycled under CC conditions an opensource Battery Dataset that is not part of the work. Using segments of charging voltage and time curves, the pipeline engineers 10 features and uses Random Forest (RF) as an algorithm for the estimation that achieves a R squared (R2) of 0.9441, Mean Absolute Percentage Error (MAPE) of 0.0196 and Root Mean Squared Percentage Error (RMSPE) of 0.1405. The pipeline methodology combines experimental data with ML modeling and could be applied to estimate other battery states needed by the BMS.