Physics-informed battery degradation prediction: Forecasting charging curves using one-cycle data

Accurately predicting battery degradation is crucial for battery system management. However, due to the complexities of aging mechanisms and limitations of historical data, comprehensively indicating battery degradation solely through maximum capacity loss assessment is challenging. While machine learning offers promising solutions, it often overlooks domain knowledge, resulting in reduced accuracy, increased computational burden and decreased interpretability. Here, this study proposes a method to predict the voltage-capacity (V-Q) curve during battery degradation with limited historical data. This process is achieved through two physically interpretable components: a lightweight interpretable physical model and a physics-informed neural network. These components incorporate domain knowledge into machine learning to improve V-Q curve prediction performance and enhance interpretability. Extensive validation was conducted on 52 batteries of different types under different testing conditions. The proposed method can accurately predict future V-Q curves for hundreds of cycles using only one-present-cycle V-Q curve, with root mean square error and mean absolute error basically less than 0.035 Ah and R² basically less than 98.5%. This means that incremental capacity curves can be extracted from the predicted results for a more comprehensive and accurate battery degradation analysis. Furthermore, the method can flexibly adjust prediction length and density to cater to the practical needs of long-cycle prediction and data generation. This study provides a viable method for rapid degradation prediction and is expected to be generalized to in-vehicle implementations.