A physics-informed ensemble learning method for battery aging prediction across diverse charging conditions

Regulating battery charging is one of the most effective approaches to extending battery lifetime, which relies on accurate prediction of battery aging under diverse charging conditions. Data-driven methods have been widely adopted. To tackle the problem of limited data, physics-informed neural networks (PINNs) and data augmentation show strong potential for guiding data-driven models by leveraging aging-related physical knowledge. However, this often requires accurate physical aging models or laws. To overcome such a challenge, this study proposes a new method based on ensemble learning. A batch of multi-layer perceptron (MLP) based aging prediction networks are first trained using real data. Each network can predict average degradation rate over the next 200 cycles using one cycle charging data. By using charging data under multiple charging conditions generated by an electro-thermal model, the physical consistency of each network was further evaluated and ranked with respect to well established rules that can reflect the influence of various charging factors on aging. Based on the results, the most well-performing MLPs were selected. By using ensemble learning, the aging prediction can be obtained through combining predicted results from each selected MLP. An aging dataset covering 45 charging conditions was constructed for validation. Cross-validation demonstrated that the proposed method can reduce prediction root mean squared error (RMSE) by 7–17% on degradation rate across different charging conditions, simultaneously guaranteeing physical consistency.