Robust and data-efficient battery state of charge estimation via transfer learning-enhanced physics-informed neural networks
Seonri Hong, Hyejin Kim, Jonghoon Kim, Jongbok Baek
Abstract
Accurate state of charge (SOC) estimation is essential to ensure the reliability and safety of battery management systems. Conventional data-driven methods rely on large labeled datasets and exhibit poor generalization across battery chemistries. By contrast, physics-based models offer interpretability but require complex parameter identification and incur high computational costs. To address these challenges, this study proposed a transfer learning-enhanced physics-informed neural network (TL-PINN) framework that combined the interpretability of physics-based models with the adaptability of deep learning. The framework integrated three key components: a physics-informed neural network for SOC estimation, a TL mechanism for domain adaptation, and a hybrid loss function that balances physical consistency and data efficiency. Experimental evaluations on the nickel manganese cobalt, lithium polymer, and nickel cobalt aluminum cells demonstrated that the TL-PINN outperformed conventional methods by achieving a lower root mean square error (RMSE), faster inference (approximately 2.4 ms), and enhanced robustness to sensor noise. Notably, under data-scarce conditions, the TL-PINN reduced the RMSE by more than 63% compared with the extended Kalman filter and more than 50% compared with recurrent neural network models. Furthermore, it consistently maintained a prediction error within 2%, even with noisy measurements. These results validate TL-PINN as a robust and scalable solution for SOC estimation under practical constraints and suggest its potential extension to broader battery health diagnostics, such as SOH and remaining useful life estimation.