Robust state of charge prediction for lithium-ion batteries in diverse operating environments via machine learning

Abstract

Accurate estimation of the state of charge (SOC) in lithium-ion batteries is critical for optimizing performance and ensuring safety, particularly under dynamic and sub-zero conditions, where capacity utilization and internal resistance vary significantly. A key gap in existing literature was addressed by comprehensively investigating temperature profiles (25 degrees C to-10 degrees C) in conjunction with varying discharge rates (0.2C to 2C). Two machine learning (ML) techniques-Neural Networks (NN) and Gaussian Process Regression (GPR) - were applied to predict SOC using a feature set that includes C-rate, measured ambient temperature, battery surface temperatures from five different locations, and voltage. Real-world scenarios, including the New European Driving Cycle (NEDC), were replicated to capture simultaneous changes in temperature and current. Experimental results show that both ML models consistently achieve high coefficients of determination (R2)-ranging from 0.98 to nearly 1.00-across all tested conditions. In simpler scenarios, NN achieved slightly higher accuracy and reduced computational time by up to %30, making it suitable for real-time applications such as battery management systems (BMS). Conversely, GPR excelled in more complex conditions, accurately modeling nonlinear interactions among temperature, C-rate, and SOC. Furthermore, up to an %18.51 reduction in discharged energy capacity was observed under sub-zero temperatures combined with elevated C-rates, underscoring the severity of cold-temperature operation. Consequently, these results highlight the effectiveness of ML-based approaches for refining SOC estimation and guiding energy management decisions in demanding real-world environments.