Accurate State-of-Charge Estimation of Lithium-Ion Batteries Using Bayesian- and PSO-Optimized NARX Neural Networks

Accurate state-of-charge (SoC) estimation is essential for reliable energy management in electric vehicles, particularly under varying operating and temperature conditions. This study develops a recurrent neural-network-based SoC estimator that exploits battery cur rent, terminal voltage, ambient temperature, and delayed output information to capture the temporal dynamics of the battery system. To address the strong dependence of the nonlinear autoregressive network with exogenous inputs (NARX) architecture on its hy perparameters, two optimization strategies are considered, namely Bayesian Optimization (BO) and Particle Swarm Optimization (PSO). The resulting BO-NARX and PSO-NARX estimators are evaluated against a conventional Extended Kalman Filter (EKF) under dif ferent driving cycles and temperature conditions. The results show that both optimized NARX models provide accurate SoC tracking and generally outperform the EKF, while their relative performance varies depending on the operating conditions. These ndings highlight the importance of both temporal modeling and hyperparameter optimization for robust SoC estimation in battery management systems.