Hybrid Extended Kalman Filtering and Feedforward Neural Correction for Lithium-Ion Battery State-of-Charge Estimation

Abstract Estimating the state of charge (SOC) of a lithium-ion battery from terminal measurements is difficult because voltage depends on load, temperature and internal polarization as well as remaining charge. This study evaluates a hybrid estimator coupling a calibrated, temperature-dependent first-order equivalent-circuit extended Kalman filter (EKF) with a feedforward neural network (FNN). Rather than predicting SOC independently, the FNN uses causal measurement history and EKF diagnostics to produce two bounded corrections: a voltage-model offset applied within the EKF observation update and an output-only SOC residual correction. The frozen network was developed using four Panasonic 18650PF recordings and two development-validation recordings. In a diagnostic replay of ten previously examined automotive load-profile recordings at 10°C and 25°C, equal-weighted mean recording-level RMSE against a charge-counter-derived SOC reference decreased from 0.5863 to 0.4516 SOC percentage points, a 22.98% relative reduction. Eight recordings improved, while both UDDS recordings deteriorated. Although the results demonstrate reproducible condition-dependent error correction, the reference relies on an assumed 2.79826 Ah capacity, benchmark recordings were previously inspected, and no independent multi-cell generalization or safety claim is made.