An entropy-based, self-adaptive predictive algorithm for battery degradation

Journal article, 2025

In order to avoid excess waste generation and provide much needed energy storage capacity, lithium ion (Li-ion) batteries, when retired from their 1st-life, can be repurposed or given a 2nd-life in lower-stress storage roles. To do so, and to determine for what purpose, accurately predicting the degradation rate of 2nd-life Li-ion batteries’ state of health is highly important, yet difficult, owing to the lack of available data from cells of sufficient aging variety. Additionally, as there are no formal standards on what information may come with potential 2nd-life batteries, it is hard to predict their subsequent behavior. While certain models do exist for predicting degradation of certain cell types/chemistries, such models typically rely on extensive data from the battery's 1st-life and do not generalize well over different types of cell. This work aims to establish a novel entropy-based theoretical approach, and a novel entropy-based algorithm, for predicting 2nd-life batteries’ behavior. The proposed model hybridizes simple machine learning methods with a light weight model based on physics, centered around approximating the amounts of generated irreversible thermodynamic entropy and Shannon entropy. Tests of this model on three different Li-ion battery types (LFP, LCO, NMC) show that the model is able to make accurate predictions on 2nd-life battery lifetime while only requiring data from one single cycle. Subsequent sampling is shown to further improve model accuracy, placing this novel algorithm on par with state of the art ML-estimates, but without the need for extensive training or reliance on extensive data from 1st-life.