Mechanistically guided residual learning for battery state monitoring throughout life

Battery monitoring requires high accuracy and robustness throughout the entire lifespan to ensure safe and optimal operations. Here we introduce mechanistic leading residual learners to enhance the monitoring of battery charge and health states, as well as guide safety warnings, targeting large-scale applications. Leveraging prior knowledge from real-time filtering as primary guidance, complemented by mechanistic and statistical features, our approach significantly improves accuracy and robustness. We propose and validate two general residual learning pipelines, namely the correction model and compensation model, across various scenarios, encompassing different battery types, loading profiles, aging conditions, and environmental conditions, using three aging datasets under dynamic cycling. Our method achieves a relative root mean square error reduction of over 50% from the results observed in prior estimations. The extrapolation capability under unseen conditions, along with interpretability, enhances both accuracy and trustworthiness. The models remain effective even with reduced training data and sampling frequency, maintaining the potential for practical electric vehicle applications. Application demonstrations confirm the efficacy in providing continuous monitoring across lifespan without the need for offline testing and model calibration during operation.