Enhancing Battery Monitoring under Real-World Driving Profiles through Physics-Guided Artificial Intelligence

Reference performance tests (RPTs) are crucial for calibrating the battery capacity and power capability; yet conducting time-consuming RPTs for large-scale electric vehicles is infeasible. Accurately predicting battery behaviors under RPTs via the physics-based model (PBM) promises to overcome this challenge, and additionally offers physical insights for more sophisticated battery management. However, using real-time driving profiles to identify PBM parameters can lead to unreliable prediction of RPTs due to the parameter variations under various conditions and potentially less informative data. To bridge the gap, this paper proposes the physics-guided differential evolution (PGDE) that combines the differential evolution (DE) with parameter sensitivity analysis. PGDE evaluates the similarity between the driving profile and RPT based on sensitivity results, and then remove non-similar data from driving profiles so that the remaining data are universally and densely informative. To validate PGDE, experiments under real-world driving profiles and RPTs are conducted using the commercial lithium-ion pouch cell. Our results show that PGDE achieves 0.89% and 3.41% test errors for estimating capacity and direct-current resistance, which outperforms vanilla DE. Additionally, the parameter sensitivity results highlight that electrolyte diffusion dominates the 1C-discharge capability.