A Data-Driven Approach to Rapidly Identify the Critical Current of Fast-Charging Lithium-Ion Batteries

Abstract

Traditional evaluation of battery charging protocols typically requires hundreds of electrochemical cycles and months of experimentation to select charging schemes that maximize the battery performance without compromising the cycle life. In this work, by nesting clustering and classification algorithms, a data-driven method using only data within a few tens of cycles is proposed to accurately classify constant-current charging protocols and rapidly identify the critical current, beyond which rapid degradation tends to occur within a specified lifetime. Specifically, by utilizing unsupervised clustering to process early-stage features and generate prediction labels, a model for early-stage prediction of the rapid degradation is established with an accuracy higher than 92.75%. Subsequently, the critical current is determined by intersecting the classification boundary with the physical distribution domain of the features. The reliability and generalizability of the proposed method is also discussed, which suggests that only ∼30 cycles and ∼40 samples are required to accomplish acceptable identification. The method is also proven to suitable for different battery systems. Therefore, the data-driven method proposed in this work provides a novel pathway to rapidly evaluate fast-charging batteries and charging protocols.