(435f) Residual Learning-Based Model for Lithium-Ion Battery End-of-Life Prediction

Predicting the lifetime of lithium-ion batteries (LIBs) is critical for ensuring safety and optimizing operational strategies. The end-of-life (EOL) of LIB is commonly defined as the point where the state-of-health (SOH), the current cycle's capacity over nominal capacity, is 80%. C-rate, depth-of-discharge (DOD), and ambient temperature (T_amb) are the major factors that determine the EOL of LIBs [1]. Predicting EOL becomes particularly challenging when LIBs are subjected to harsh conditions, e.g., rapid charge and discharge, too high or low , as Li plating can cause "knee", where capacity starts to decrease rapidly at a certain point under these conditions.

While various public datasets describe LIB’s capacity fade [2-3], to the best of our knowledge, no dataset exists that would reflect the variation of all three factors and cycled for EOL due to high experimental costs. Furthermore, due to cell-to-cell variability [4] (i.e. the deviation between cells due to uncontrollable changes during manufacturing stages), cells’ EOL would differ in some regions even if they are cycled under the same operating conditions.

In this study, we propose a systematic approach for lifetime prediction of LIBs, where two models are sequentially applied to effectively capture the effects of cell-to-cell variability. To do this, we first generate datasets by simulating the operation of LIBs up to EOL with the cell-to-cell variability considerations. For this purpose, we modify an electrochemical model (EM) simulator called LIONSIMBA [6] by adding Li plating effect [7] (overpotential η) to consider some cells where knees exist. Then, we conduct a large number of cycling tests under various major factors and internal cell parameters. With these datasets, the gaussian process regression (GPR) model is first used to predict the nominal EOL (i.e. EOL without considering cell-to-cell variability). Then, the concept of residual learning [5] is adopted to add a bias to our prediction so that the cell-to-cell variability or minor factors’ changes can be captured. We calibrate the nominal EOL by calculating the change in EOL when the cell-to-cell variability exists or minor factors change. Specifically, after calculating EOL bias with refined health indexes (RHIs), e.g., -related factors directly, we present additional neural networks that would calculate RHIs from direct health indexes (DHIs) from early cycles, e.g., voltage (V), current (I), and temperature (T) directly and predict final EOL with only operating conditions and DHIs.

[1] Panchal S, Pierre V, Cancian M, Gross O, Estefanous F and Badawy T. (2023). Development and validation of cycle and calendar aging model for 144Ah NMC/graphite battery at multi temperatures, DODs, and C-rates, SAE Technical Paper.

[2] Severson KA, Attia PM, Jin N, Perkins N, Jiang B, Yang Z, et al. Data-driven prediction of battery cycle life before capacity degradation. Nat. Energy 2019;4:383-91.

[3] Attia PM, Grover A, Jin N, Severson KA, Markov TM, Liao Y-H, et al. Closed-loop optimization of fast-charging protocols for batteries with machine learning. Nature 2020;578:397-402.

[4] Beck D, Dechent P, Junker M, Sauer DU and Dubarry M. Inhomogeneities and cell-to-cell variations in lithium-ion batteries, a review. Energies 2021;14:3276.

[5] He K, Zhang X, Ren S and Sun J. Deep residual learning for image recognition. CVPR;2016.

[6] Torchio M, Magni L, Gopaluni RB, Braatz RD and Raimondo DM. Lionsimba: a matlab framework based on a finite volume model suitable for li-ion battery design, simulation, and control. J. Electrochem. Soc. 2016;163:A1192.

[7] Yang X-G, Leng Y, Zhang G, Ge S and Wang C-Y. Modeling of lithium plating induced aging of lithium-ion batteries: Transition from linear to nonlinear aging. J. Power Sources 2017;360:28-40.