Joseph Sujoy Pulivarthi, Suman Jana, Sai Sudheer Tadi, Lokeshwar Sai Gummidi, Veeranjaneyulu Rajamahendravarapu, P. Abdul Subhahanalla
Abstract
Battery State-of-Health (SOH) prediction is critical for ensuring reliability and safety in energy storage systems. In this work, battery degradation is analyzed using Remaining Useful Life (RUL) trends and physics-informed features. A group-aware evaluation framework is proposed to analyze model performance under different data splits. Three models, LSTM with attention, Linear Regression, and Random Forest are evaluated under both random and time-based splits. Experimental results show that while all models perform strongly under random split conditions (R² > 0.99), deep learning models exhibit severe performance degradation under distribution shift (R² = -3.06), whereas Linear Regression remains robust (R² = 0.9975). A robustness score is introduced to quantify performance degradation across splits. The findings demonstrate that simpler models outperform deep learning when the underlying data exhibits strong monotonic degradation behavior. This highlights the importance of selecting models based on data characteristics rather than model complexity.
Keywords
Battery State-of-Health, Remaining Useful Life, Physics-Informed Features, LSTM, Random Forest, Distribution Shift, Robustness Analysis
Conclusion
This study demonstrates that model performance is strongly influenced by data characteristics rather than model complexity. While deep learning models perform well under random splits, they fail under temporal distribution shift. Linear models exhibit strong robustness, making them more suitable for real-world deployment. These results highlight that model robustness is a critical factor for real-world deployment in battery health monitoring systems. Future work includes extending the approach to larger and more diverse datasets, and exploring hybrid modeling techniques for improved robustness under distribution shift. This work demonstrates that model selection should prioritize robustness over complexity in real-world battery health prediction systems
References
[1]I. Goodfellow, Y. Bengio, and A. Courville,
*Deep Learning*, MIT Press, 2016. [2]S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,” Neural Computation, 1997. [3]X. Zhang et al., “Battery RUL Prediction Using Machine Learning,” IEEE, 2020. [4]F. Pedregosa et al., “Scikit-learn: Machine Learning in Python,” JMLR, 2011. [5]Y. Zhang et al., “Data-driven battery degradation modeling,” IEEE Transactions on Industrial Electronics, 2021. [6]J. Liu et al., “Battery remaining useful life prediction using machine learning,” Applied Energy, 2020.
How to Cite This Paper
Joseph Sujoy Pulivarthi, Suman Jana, Sai Sudheer Tadi, Lokeshwar Sai Gummidi, Veeranjaneyulu Rajamahendravarapu, P. Abdul Subhahanalla (2026). Battery State-of-Health Prediction using Physics-Informed Feature Engineering and Group-Aware Machine Learning. International Journal of Computer Techniques, 13(2). ISSN: 2394-2231.
