Enhancing SOC estimation accuracy via incremental learning techniques for second-life batteries

Abstract

The available options for retired batteries include recycling or finding a secondary application. Repurposing the batteries for applications like stationary systems or low-speed vehicles may present a viable alternative. However, Second-Life Batteries (SLB) entail unique challenges since a battery pack comprises cells with diverse resistance and capacity characteristics. It makes the operation of BMS less efficient as the battery undergoes degradation. In light of these challenges, this study introduces an application of Incremental Learning in conjunction with a Random Forest model employing twenty estimators to predict a battery’s state of charge (SOC) at various levels of state of health. The proposed model was trained using the first three cycles and applied in real time during subsequent discharges. After each discharge cycle, the model underwent further training to refine its understanding of the underlying problem. Compared to other models, the implementation is more straightforward because the model learns with the data and does not require cell rest, actual capacity, or a battery model, such as known models in the literature. To validate the model, it has been applied for sixty discharge cycles for first-life batteries (Exp1), stationary (Exp2), and dynamic SLB systems (Exp3). The results demonstrated a consistently low RMSE SOC variation of less than 2% for Exp1 and Exp2 and less than 4% for Exp3, validating the use of it for first-life batteries and SLB with constant and dynamic loads.