Hybrid SOC Estimation for LiFePO4 Batteries Using Observability- and Innovation–Reliability-Regulated EKF with Reliability-Scaled Residual Learning

Abstract

Accurate state-of-charge (SOC) estimation of lithium iron phosphate (LiFePO4) batteries is challenging because the voltage feedback used for correction does not provide constant SOC-related information under different operating conditions. Conventional extended Kalman filters (EKFs) usually apply measurement correction based on predefined statistical assumptions, while overlooking variations in voltage-domain observability and innovation reliability. This paper proposes a hybrid estimation framework, termed observability- and innovation–reliability-regulated EKF with reliability-scaled residual learning (OIR-EKF-RSRL). The proposed method retains a first-order RC model and EKF as the physical estimation backbone, while regulating voltage correction according to local OCV-SOC sensitivity and normalized innovation reliability. A reliability-scaled residual learning module is further introduced after physical filtering to compensate for remaining SOC deviations rather than directly predicting SOC. The learned residual correction is modulated by a reliability-dependent scaling coefficient before fusion with the OIR-EKF estimate, after which the final SOC estimate is constrained to the physical interval [0, 1]. The framework is evaluated using the CALCE A123 LiFePO4 dataset under a frozen temperature-disjoint train–validation–holdout protocol. On the independent holdout set, OIR-EKF-RSRL reduces the RMSE from 3.331 percentage points for the conventional EKF to 2.942 percentage points. The results demonstrate that reliability-aware measurement utilization and reliability-scaled residual compensation provide an interpretable solution for LiFePO4 SOC estimation under varying voltage-information quality.