Active-learning-driven error control for data-driven state of charge estimation across the lithium battery lifecycle

Abstract

Accurate estimation of lithium-ion battery state of charge (SOC) is crucial for the safe and efficient operation of electric vehicles (EVs). However, both data-driven and model-driven SOC estimation methods face significant challenges under battery aging, which alters internal resistance and electrochemical properties, especially across complex aging trajectories. Most existing deep learning and model-based approaches operate in an open-loop manner, lacking mechanisms for uncertainty quantification, accuracy prediction, or adaptive correction—leading to uncontrolled estimation errors during aging. To address this, we propose an innovative closed-loop SOC estimation framework that integrates active learning with uncertainty-aware correction into deep learning networks, enabling real-time feedback on SOC prediction confidence levels without the need for additional sensors or reference data. Specifically, we quantify the performance degradation of mainstream data-driven methods, including long short-term memory (LSTM) networks and Gaussian process regression (GPR), under complex aging paths. We demonstrate that our model-disagreement-based active learning correction strategy maintains robustness throughout the battery lifecycle. Experimental results show that with only four active retraining sessions over the full aging process, our method reduces average SOC estimation error to below 1.5 %, and maximum cycle-based average error to below 2 %. This work establishes a path toward uncertainty-informed, lifecycle-resilient, and data-efficient SOC estimation, marking a significant advancement in battery management systems for real-world EV applications. Graphical abstract: [Figure not available: see fulltext.]