A Hierarchical Physics-Informed Machine Learning Framework for Battery Pack Capacity Estimation With Limited Data

Accurate and robust estimation of battery pack capacity is essential for intelligent battery management. However, complex degradation mechanisms, cell-to-cell inconsistencies, and the scarcity of labeled data pose significant challenges. To address these issues, a hierarchical physics-informed machine learning (HPIML) framework is proposed to enhance model training and transferability with minimal labeled data. First, a simulation model of battery pack is constructed to emulate dynamic behavior across multiple aging scenarios, replacing costly and time-consuming experimental data acquisition. A dual-branch deep learning model is then developed, combining degradation mechanisms and data-driven information to improve estimation robustness and adaptability. Finally, the model is pretrained using simulated data and adapted to real-world scenarios through a transfer learning strategy to enhance adaptability. Experimental results demonstrate that using only four labeled samples, the HPIML framework can accurately estimate the capacity of battery packs under different materials and operating conditions, with the root mean square error and mean absolute percentage error maintained below 0.2 Ah and 0.5%, respectively. Compared to conventional methods, the proposed approach improves estimation accuracy by at least 27.89%, while saving 492 kWh of electricity and reducing CO2emissions by 383.76 kg.