Dual-phase optimized deep learning framework for accurate, efficient, and robust battery SoC estimation

The growing adoption of electric vehicles (EVs) requires accurate and robust State of Charge (SoC) estimation to ensure optimal battery performance, reliable driving range, and operational safety. This paper introduces KANBiLSTMAtt, a novel hybrid deep learning model that integrates the Kolmogorov-Arnold Network (KAN), Bi-directional Long Short-Term Memory (BiLSTM), and attention mechanisms to capture nonlinear interactions and long-term temporal dependencies in lithium-ion battery data. The framework incorporates Optuna for efficient hyperparameter tuning and NSGA-II for multi-objective optimization, achieving high predictive accuracy with minimal computational overhead. Validation on two distinct battery chemistries under varying temperatures, using the LG dataset and driving cycles from the CALCE dataset, demonstrates strong generalization and robustness. KANBiLSTMAtt achieves an RMSE of 0.02%, a MAE of 0.01%, and an R² of 99% for both datasets, utilizing a lightweight architecture and converging within 90 s, making it highly suitable for real-time and embedded battery management systems. By combining hybrid deep learning and evolutionary optimization, the proposed model addresses limitations of traditional SoC estimation methods, offering a scalable solution for next-generation EV energy management.