Early prediction of battery life using an interpretable health indicator with evolutionary computing

Abstract

Accurate prediction of battery lifespan is crucial for optimizing energy management, enhancing safety, and ensuring system reliability, particularly when only early-stage battery data is available. Health indicators (HIs) play a pivotal role in monitoring battery degradation by providing a link between the current state and the battery's end of life (EOL). However, existing methods for HI extraction often depend on extensive expert knowledge, large volumes of lifecycle data, and complex models to map HIs to battery lifespan. This study introduces an intelligent and interpretable methodology for generating HIs using improved genetic programming (GP) to enable rapid and precise battery lifespan prediction based solely on data from two early discharge cycles. Four HI candidates are derived from statistical features of the differences between discharge voltage curves. Unlike conventional methods that employ root mean square error (RMSE) as a fitness function, we introduce a novel correlation-based fitness function using cosine similarity within GP. This approach generates a transparent composite mathematical formula for extracting interpretable HIs. It automatically filters irrelevant HI candidates and combines relevant ones through specific mathematical operations. The resulting composite mathematical expression, universally applicable for constructing interpretable HIs across various cycle selections, enables rapid and early battery lifespan prediction through regression models. Validation on 124 battery cells shows that the proposed composite HI, expressed as an explicit mathematical function, achieves a mean absolute percentage error of approximately 15 % when predicting battery lifespan using data from just two cycles within the first 20 cycles across diverse operating conditions. Moreover, the proposed approach surpasses benchmark HIs in both prediction accuracy and stability across different regression models.