Hybrid Recurrent Neural Network and Adaptive Linear Kalman Filter for Lithium‐Ion Battery State of Charge Estimation

ABSTRACT Determining the precise state of charge (SOC) is essential to maximizing the safety and efficiency of lithium‐ion batteries (LIBs) in electric vehicles (EVs), constituting a key function of battery management systems (BMS). However, conventional model‐based methods often depend on complex mathematical formulations and exact internal battery parameters, which can lead to estimation inaccuracies due to uncertainties and variations. Additionally, deep learning methods commonly demand a large amount of datasets for neural network parameter tuning and may struggle to generalize beyond training datasets. To address these challenges, this article presents a recurrent neural network combined with the adaptive linear Kalman filter (RNN–AKF) method. Initially, the recurrent neural network (RNN) is applied first to preestimate the battery state of charge using measured variables (current, voltage, and temperature). after that, the adaptive linear Kalman filter (AKF) is used to refine the preestimated battery state of charge to obtain accurate and stable estimates. The validation of the proposed method uses the dataset from Turnigy Graphene 5000 mAh 65C Li‐ion batteries, integrating various driving cycles under various temperature conditions (10°C, 25°C, and 40°C). The results reveal a remarkable improvement in SOC estimation accuracy compared to existing approaches, notably outperforming the second‐order equivalent circuit model with extended Kalman filter (ECM–EKF) and solo RNN methods. The RNN–AKF method demonstrates exceptional accuracy, robust generalization, and efficient convergence, providing a promising solution for SOC estimation in LIBs for EVs with minimal computational complexity.