Novel Data-Efficient Mechanism-Agnostic Capacity Fade Model for Li-Ion Batteries

Abstract

Accurate capacity fade prediction of Li-ion batteries is essential to reduce the time spent by manufacturers in performing quality assurance tests and to ensure the safety and durability of these batteries for end users. Various complicated aging mechanisms and the resulting capacity fade phenomena of Li-ion batteries make such predictions challenging; thus, mechanism-agnostic approaches using empirical and data-driven models are considered to be promising. This article proposes a mechanism-agnostic capacity fade empirical model called aging density function model (ADFM) for Li-ion batteries. Developed by innovating existing empirical models, the proposed ADFM predicts capacity fades for arbitrary battery input current trajectories, requires no additional experiments at the prediction phase, and reflects real batteries phenomena such as the varying amount of capacity fade for each cycle. As the proposed ADFM could generate a large amount of synthetic data, it was augmented with Bayesian neural networks (BNNs) to enhance its data efficiency. As a result, it can completely utilize the experimental data and achieve reasonable prediction accuracy regardless of the amount of experimental data. This BNN-augmented ADFM can also provide the reliability of the capacity fade prediction to ensure safety. Through charge/discharge cycle tests with an NCM/graphite Li-ion battery, the proposed BNN-augmented ADFM was shown to provide good performance in terms of the capacity fade prediction accuracy, with a mean absolute error of approximately 0.5% and maximum absolute error of approximately 2.5%.