SOC estimation for a lithium-ion pouch cell using machine learning under different load profiles

Estimating the state of charge of lithium-ion battery systems is important for efficient battery management systems. This work conducts a thorough evaluation of multiple SOC estimate methods, including both classic approaches Coulomb Counting and extended Kalman filter and machine learning techniques under different load profile on lithium-ion pouch cell. The assessment included a variety of experimental data collected from entire cycles, shallow cycles, and dynamic operations utilizing the Worldwide Harmonized Light Vehicles Test Procedure and Hybrid Pulse Power Characterization tests done from 100% to 10% SOC. While traditional approaches performed well under ordinary settings, they had severe limits during shallow cycling. In contrast, machine learning technologies, notably the random forest method, performed better across all testing conditions. The random forest approach showed outstanding accuracy while minimizing error metrics (RMSE: 0.0229, MSE: 0.0005, MAE: 0.0139) and effectively handled typical issues such as SOC drift and ageing effects. These findings validate random forest as a dependable and robust approach for real-time SOC estimation in battery management systems.