Estimate state of charge in lithium-ion batteries with unknown data

Lithium-ion batteries (LIBs) are vital for sustainable energy solutions, with the state of charge (SOC) serving as a critical indicator of their energy levels, directly influencing safety and efficiency. However, accurately estimating SOC remains challenging due to complex internal chemical processes and factors such as unknown data distribution, error accumulation, delayed feedback, and sensitivity to input data. To tackle these challenges, we propose a physics-guided meta-learning framework for cross-task adaptation in SOC estimation. During the model adaptation phase, the challenge of acquiring the maximum capacity prevented the calculation of the SOC at the beginning of the discharge process, which is critical for adaptation or fine-tuning. This framework adapts quickly to new data distributions with the most similarly distributed data and learns adaptive strategies, enabling rapid model updates across various attributes, such as LIB types, temperatures, and operating conditions. Moreover, to further enhance the accuracy and generalization performance of the model, the Coulomb counting method is integrated into the network training process. The physical parameters utilized in Coulomb counting are provided and refined by the neural network, which also generates a separate estimate. Additionally, physical constraints are added to the loss function to guide the update of network parameters. This approach combines physical guidance with the estimation of neural network, thereby partially mitigating the errors inherent in the training process with similar data. We evaluated the estimation model through five-fold cross-validation on 56 datasets. The framework demonstrates strong generalization, achieving robust SOC estimation across varying capacities and conditions. Our work highlights the potential of meta-learning for fast adaptation in SOC estimation under unknown distributions and demonstrates the importance of physical information guidance in improving robustness and performance. This framework can be applied to a wide range of batteries, providing robust support for battery management systems (BMS).