Abstract
Accurate estimation of the state of charge (SoC) is crucial for ensuring the reliability, efficiency, and safety of lithium-ion batteries in electric vehicles and renewable-energy systems. However, conventional model-based and data-driven techniques remain sensitive to noise, modeling uncertainties, and nonlinear dynamics. This paper proposes an adaptive hybrid Long Short-Term Memory Extended Kalman Filter (LSTM-EKF) framework that integrates the predictive capability of deep learning with the real-time correction of model-based estimation. The main novelty lies in an adaptive fusion factor (αk) that dynamically balances the contributions of the LSTM and EKF according to their instantaneous confidence levels, enhancing both accuracy and robustness under noisy and time-varying operating conditions. A comprehensive comparative study including BiLSTM, LSTM-Attention, and EKF methods demonstrates that the proposed adaptive LSTM-EKF achieves the lowest RMSE and MAE, with accuracy improvements of approximately 70 % compared with standalone approaches. These results highlight the framework's strong potential as a scalable and noise-resilient solution for advanced battery-man-agement systems, contributing to improved energy efficiency, extended battery lifespan, and safer operation in electric-mobility and renew-able-storage applications.