Anode Potential Estimation in Lithium-Ion Batteries Using Data-Driven Models for Online Applications

Three anode estimation methods are presented and evaluated for their accuracy and storage requirements. After generating training data using a Pseudo-2D Physiochemical model, these models are fit and trained to estimate the anode potential during fast charge events. A simplified linear and non-linear model show an estimationerror of ca. 13 mV and the lowest memory demand, however, a novel random forest model reduces the error to 2.6 mV. The empirical methods are suitable for a lithium plating warning detection system during fast charging and are further evaluated for over-fitting and robustness using an out-of-sample dataset.