A Data-Centric Framework for OOD-Aware RUL Prediction of Lithium-Ion Batteries with Risk-Aware Evaluation
Resumo
Predicting the Remaining Useful Life (RUL) of lithium-ion batteries is a safety-critical machine learning task. A central obstacle is domain shift: models trained in one experimental context frequently fail to generalise to others, producing overoptimistic predictions that translate into direct operational risk, a failure mode that aggregate metrics such as mean absolute error (MAE) and root mean square error (RMSE) systematically conceal. We propose a data-centric framework that addresses this challenge through three contributions. First, we construct a multi-domain OOD evaluation benchmark integrating six heterogeneous public datasets spanning calendar ageing, industrial, and aeronautical degradation profiles under a unified schema and standardised evaluation protocols. Second, we introduce a physics-informed inductive bias via the Shadow Impedance proxy, an estimate of internal resistance derived directly from raw voltage and current signals, which enriches the feature space and acts as a routing signal in a Regime-Aware Ensemble of specialised XGBoost regressors following the Mixture-of-Experts principle. Third, we propose a risk-aware evaluation framework centred on the Danger Bias metric, which quantifies the fraction of operationally dangerous overestimates of remaining life. Evaluation under LOCO and LODO protocols shows strong intra-distribution performance (R2=0.9662, MAE < 5%), but reveals critical cross-domain degradation, with Danger Bias reaching 88.8% in EVERLASTING, a two-cell calendar-aging dataset, the most severe OOD scenario. These findings demonstrate that reliable AI for safety-critical physical domains requires not only expressive models, but also OOD-aware benchmarks and risk-sensitive evaluation metrics.
Referências
Bills, A. et al. (2021). eVTOL battery dataset. Kilthub – Carnegie Mellon University. Available at: [link].
Celik, B. et al. (2022). Prediction of battery cycle life using early-cycle data, machine learning and data management. Batteries, 8(12):266.
Chen, T. and Guestrin, C. (2016). XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 785–794, New York. ACM.
Fricke, K. A. et al. (2023). Randomized and recommissioned battery dataset. Technical report, NASA Prognostics Center of Excellence. Available at: [link].
Howey, D. A. and Birkl, C. (2017). Oxford battery degradation dataset 1. University of Oxford. DOI: 10.5287/bodleian:KO2kdmYGg.
Jacobs, R. A., Jordan, M. I., Nowlan, S. J., and Hinton, G. E. (1991). Adaptive mixtures of local experts. Neural Computation, 3(1):79–87.
Jiang, Y. and Song, W. (2023). Predicting the cycle life of lithium-ion batteries using data-driven machine learning based on discharge voltage curves. Batteries, 9(8):413.
Kouw, W. M. and Loog, M. (2021). A review of domain adaptation without target labels. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(3):766–785.
Li, K. and Chen, X. (2025). Machine learning-based lithium battery state of health prediction research. Applied Sciences, 15(2):516.
Lu, J. et al. (2023). Deep learning to estimate lithium-ion battery state of health without additional degradation experiments. Nature Communications, 14(1):2760.
Mayemba, Q. et al. (2024). General machine learning approaches for lithium-ion battery capacity fade compared to empirical models. Batteries, 10(10):367.
Ni, Y. et al. (2023). Hybrid estimation of residual capacity for retired LFP batteries. In Proceedings of ICEIV 2022, volume 1018 of Lecture Notes in Electrical Engineering, pages 44–51. Springer, Heidelberg.
Rahm, E. and Bernstein, P. A. (2001). A survey of approaches to automatic schema matching. The VLDB Journal, 10(4):334–350.
Sambasivan, N. et al. (2021). Everyone wants to do the model work, not the data work: data cascades in high-stakes AI. In Proceedings of the 2021 ACM CHI Conference on Human Factors in Computing Systems, pages 1–15, New York. ACM.
Severson, K. A. et al. (2019). Data-driven prediction of battery cycle life before capacity degradation. Nature Energy, 4(5):383–391.
Stroebl, F. et al. (2024). Multi-stage lithium-ion battery aging study. Figshare. DOI: 10.6084/m9.figshare.25975315.
Takahashi, A., Allam, A., and Onori, S. (2023). Evaluating the feasibility of batteries for second-life applications using machine learning. iScience, 26(4):106547.
Tan, Y. et al. (2022). Early prediction method for remaining useful life of retired batteries in second-life applications. In Proceedings of 2022 IEEE International Conference on Energy Internet and Energy System Integration (EI2), pages 522–530.
Trad, K. (2021). Calendar ageing test results on commercial 18650 Li-ion cell @ 25°C and 45°C. 4TU.ResearchData. DOI: 10.4121/13804304.v1.
Valizadeh, A. and Amirhosseini, M. H. (2024). Machine learning in lithium-ion battery: applications, challenges, and future trends. SN Computer Science, 5(6):697.
Vilsen, S. B. and Stroe, D.-I. (2023). Lithium-ion battery degradation dataset based on a realistic forklift operation profile.
Waag, W., Fleischer, C., and Sauer, D. U. (2014). Critical review of the methods for monitoring of lithium-ion batteries in electric and hybrid vehicles. Journal of Power Sources, 258:321–339.
Zhai, Q. et al. (2024). Machine learning for full lifecycle management of lithium-ion batteries. Renewable and Sustainable Energy Reviews, 202:114647.
Zhang, H. et al. (2025). Battery lifetime prediction across diverse ageing conditions with inter-cell deep learning. Nature Machine Intelligence, 7(2):270–277.
