Early prediction of Lithium-ion cell degradation trajectories using signatures of voltage curves up to 4-minute sub-sampling rates

Ibraheem, Rasheed and Wu, Yue and Lyons, Terry and dos Reis, Goncalo (2023) Early prediction of Lithium-ion cell degradation trajectories using signatures of voltage curves up to 4-minute sub-sampling rates. Applied Energy, 352. 121974. ISSN 0306-2619 (https://doi.org/10.1016/j.apenergy.2023.121974)

[thumbnail of Ibraheem-etal-AE-2023-Early-prediction-of-Lithium-ion-cell-degradation-trajectories]
Preview
Text. Filename: Ibraheem_etal_AE_2023_Early_prediction_of_Lithium_ion_cell_degradation_trajectories.pdf
Final Published Version
License: Creative Commons Attribution 4.0 logo

Download (6MB)| Preview

Abstract

Feature-based machine learning models for capacity and internal resistance (IR) curve prediction have been researched extensively in literature due to their high accuracy and generalization power. Most such models work within the high frequency of data availability regime, e.g., voltage response recorded every 1–4 s. Outside premium fee cloud monitoring solutions, data may be recorded once every 3, 5 or 10 min. In this low-data regime, there are little to no models available. This literature gap is addressed here via a novel methodology, underpinned by strong mathematical guarantees, called ‘path signature’. This work presents a feature-based predictive model for capacity fade and IR rise curves from only constant-current (CC) discharge voltage corresponding to the first 100 cycles. Included is a comprehensive feature analysis for the model via a relevance, redundancy, and complementarity feature trade-off mechanism. The ability to predict from subsampled ‘CC voltage at discharge’ data is investigated using different time steps ranging from 4 s to 4 min. It was discovered that voltage measurements taken at the end of every 4 min are enough to generate features for curve prediction with End of Life (EOL) and its corresponding IR values predicted with a mean absolute percentage error (MAPE) of approximately 13.2% and 2.1%, respectively. Our model under higher frequency (4 s) produces an improved accuracy with EOL predicted with an MAPE of 10%. Full implementation code publicly available.