Adaptive SOC Estimation for Lithium-Ion Batteries Using Cluster-Based Deep Learning Models Across Diverse Temperatures

Accurate State of Charge (SOC) estimation for lithium-ion batteries is crucial but challenging due to their complex nonlinear behaviour and sensitivity to ambient temperature. This paper assesses a novel Cluster-Based Learning Model (CBLM) integrating K-Means clustering with machine and deep learning algorithms like LSTM, BiLSTM, Random Forest, and XGBoost for SOC estimation. The key innovation is developing a framework that allows tailored learning of distinctive operational behaviour of the battery using the proposed CBLM. Additionally, the application of centroid proximity mechanism that dynamically assigns test data to the specialised models, real-time dynamic SOC estimation that is adapted to current charging conditions is the novelty of this paper, paired with the effective assessment of the CBLM framework under varied thermal conditions. Across temperatures from -20°C to 40°C, CBLM demonstrates superior accuracy over current state-of-art standalone model, with over 73% RMSE and 50% MAE reduction at 10°C and 40°C. Statistical validation confirms significant difference in performance, favouring the proposed framework.