(472d) A Physics-Informed Machine Learning Model for Battery Capacity Fading Prediction with Early Cycling Data

High-performance lithium-ion battery (LIB) technology is a key component to achieving high penetrations of renewable energy. However, there are numerous challenges that still need to be solved, such as increasing the energy density from 250 Wh/kg to 500 Wh/kg, achieving fast charging, and extending the cycle life to above 10,000 cycles. Much research has been carried out on addressing material problems, developing advanced battery management systems, and designing modern cells [1–4]. However, relatively little efforts have been made on developing battery models, especially that can predict capacity fading. Given the current capacity fading models, we can classify them into two categories: electrochemical models and regression model [5–7]. Despite the incorporation of electrochemical principles and newfound understanding of capacity fading mechanisms, the well-known electrochemical model, the pseudo-2D model, still cannot provide accurate prediction of capacity fading. For the regression model, thanks to the recent advancements in machine learning techniques, many models have been proposed. However, many of these models are physics-blind (i.e., pure black box), data-driven models which are unable to adapt to the changing internal environment.

To address these challenges, in this study, we proposed a physics-motivated, data-driven model to predict capacity fading during battery cycling. The pipeline is shown in Figure 1. For Li-ion batteries, there are two major contributors to the loss of active Li-ion inventory: the growth of a solid-electrolyte interface (SEI) and Li plating. Depending on which fading mechanism is dominant, the battery will experience a different fading rate. Compared with the growth of SEI, Li plating will result in rapid capacity fading [8]. To distinguish the dominant fading mechanism within the battery, based on the electrochemical principles of the Li-ion battery, we propose using, ∆V, as the primary feature for fading classification. The K-means clustering technique is employed and draws a boundary at 15.6 mV. With their average capacity fading at the 50^th cycle, batteries with ∆V<15.6 mV are denoted as the fast-fading group, others are denoted as the slow-fading group.

For the slow-fading group, the loss of capacity is caused by the growth of SEI. Therefore, the resistance of the SEI is a good feature for capacity fading prediction. However, there is no way to attain the resistance of the SEI from the cycling data. Instead of obtaining resistance of the SEI from the cycling data, which is not feasible, in this study, we propose using the internal resistance (R) as a substitute feature obtained from current-voltage curve. In contrast, Li plating is the dominant contributor to the capacity loss for batteries in the fast-fading group. Therefore, ∆V is chosen as the regression feature. To validate the model performance, a benchmark with log variance of the discharging curve difference ∆Q_N-1(V) is selected [9]. Given data from the 5^th cycle, we achieve 19% and 26% test mean percentage error (MPE) for batteries in the slow-fading group and in the fast-fading group, while the benchmark has 334% test MPE. Furthermore, by adding the prospective cumulative efficiency and past average fading rate as the second and third feature, we achieve 7%, 11%, and 40% test MPE for slow fading, fast fading, and the benchmark, respectively.

Reference

[1] Sheha MN, Powell KM. Dynamic Real-Time Optimization of Air-Conditioning Systems in Residential Houses with a Battery Energy Storage under Different Electricity Pricing Structures. vol. 44. Elsevier Masson SAS; 2018. https://doi.org/10.1016/B978-0-444-64241-7.50416-X.

[2] Sheha MN, Powell KM. An economic and policy case for proactive home energy management systems with photovoltaics and batteries. Electr J 2019;32:6–12. https://doi.org/10.1016/j.tej.2019.01.009.

[3] Hesse HC, Martins R, Musilek P, Naumann M, Truong CN, Jossen A. Economic optimization of component sizing for residential battery storage systems. Energies 2017;10. https://doi.org/10.3390/en10070835.

[4] Suo L, Borodin O, Sun W, Fan X, Yang C, Wang F, et al. Advanced high-voltage aqueous lithium-ion battery enabled by “water-in-bisalt” electrolyte. Angew Chemie 2016;128:7252–7.

[5] Torchio M, Magni L, Gopaluni RB, Braatz RD, 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:A1192–205. https://doi.org/10.1149/2.0291607jes.

[6] Yao J, You F. Simulation-based optimization framework for economic operations of autonomous electric taxicab considering battery aging. Appl Energy 2020;279:115721. https://doi.org/10.1016/j.apenergy.2020.115721.

[7] Yang XG, Leng Y, Zhang G, Ge S, Wang CY. Modeling of lithium plating induced aging of lithium-ion batteries: Transition from linear to nonlinear aging. J Power Sources 2017;360:28–40. https://doi.org/10.1016/j.jpowsour.2017.05.110.

[8] Gao T, Han Y, Fraggedakis D, Das S, Zhou T, Yeh CN, et al. Interplay of Lithium Intercalation and Plating on a Single Graphite Particle. Joule 2021;5:393–414. https://doi.org/10.1016/j.joule.2020.12.020.

[9] Severson KA, Attia PM, Jin N, Perkins N, Jiang B, Yang Z, et al. Data-driven prediction of battery cycle life before capacity degradation. vol. 4. 2019. https://doi.org/10.1038/s41560-019-0356-8.