State of charge prediction for new and second-life lithium-ion batteries based on the random forest machine learning technique

Abstract

Accurate state of charge (SOC) estimation is a critical requirement for the safe and efficient operation of lithium-ion batteries (LIBs), particularly in second-life battery (SLB) applications where battery ageing, nonlinear degradation, and measurement noise introduce uncertainty. Although numerous SOC estimation techniques have been proposed, reliable prediction for new and second-life batteries under varied operating conditions remains challenging. In this study, a comparative investigation of the conventional coulomb counting (CC) method and a data-driven random forest (RF) model is conducted for SOC prediction in new and second-life LIBs. Experimental data are obtained from Murata US18650VTC5D cells under pulse discharge tests (PDT), constant discharge tests (CDT), and dynamic stress tests (DST) across a wide range of C-rates. PDT is conducted at 0.24 C, CDT at 0.2 C, 0.5 C, 1 C, and 2 C, while DST is performed at C-rates ranging from 0.5 C to 4 C at a controlled ambient temperature of 25 °C. The RF model is trained using voltage, current, and time features and evaluated against CC using MAE, MSE, RMSE, and R² metrics. Results show that RF consistently outperforms CC under all conditions, particularly for SLBs, achieving significantly lower errors and R² values approaching 0.998. These findings confirm the effectiveness of RF-based SOC estimation for intelligent battery management systems (BMS).