Accurate Battery Life Prediction Using Early-Cycle Features

Data-driven machine learning (ML) applications on battery prognosis are thriving with compatible accuracy owing to advancements in computing and data storage technologies. Unfortunately, several experimental obstacles including sparse operating conditions and time cost affect the data quality which restricts the ML model generalizability and applicability given the nonlinear degradation pattern for lithium-ion batteries (LIBs). From a broader perspective, multiple electrochemical attributes have not been investigated thoroughly and therefore, we proposed to build an early-cycle lifetime prediction model by exploring the capability of features extracted from the battery’s extrinsic properties (ex: temperature, dis/-charge time) using cycling data taken from only up to the first 10 cycles. A simple linear model is employed and validated on datasets amounting about 200 cells in total with various environmental temperature settings, cycling protocols, and lithium-based cathode chemistries such as lithium - iron phosphate (LFP), nickel cobalt aluminium oxide (LNCA), and nickel manganese cobalt oxide (LNMC) with mean absolute percentage error (MAPE) cycle life prediction less than 17.3%. Integrating the short-listed explored features together with state-of-the-art features on a fast-charging dataset of 124 LFP cells, the model achieved a satisfying 11.8% overall MAPE using data from the first 10 cycles. This work provides an impactful insight into simple feature engineering on early-cycle prediction applications. Details on the methodology, data-standard challenges, and wider applications will be discussed to explore further derivation of potent early-cycle features.