Physics-informed SOC estimation without explicit parameter identification via a dual-system error modeling framework

Accurate state of charge (SOC) estimation is pivotal for battery management systems. However, electrochemical model-based estimation methods are constrained by chemistry-specific parameter identification, while data-driven methods avoid this requirement but lack physical interpretability, resulting in poor generalization under dynamic conditions. To address these challenges, this study establishes a dual-system error modeling framework. The framework theoretically derives the error dynamics under parameter mismatch, proving that concentration errors follow an integral solution form dependent on the history of current and concentration under mismatched parameters. Based on this mathematical structure, a neural network with temporal memory capabilities is selected to implicitly solve this integral process, and a dual-task dual-stream fusion network (DTDSFN) is proposed to achieve physical reconstruction of terminal voltage. Through architectural hard constraints, DTDSFN decouples the terminal voltage into open-circuit voltage, concentration polarization overpotential, and coupled overpotential to enhance model interpretability. On this basis, by establishing a quasi-static mapping between solid-phase surface concentration and SOC, the DTDSFN is integrated into an unscented Kalman filter for closed-loop SOC estimation. Validation across diverse battery chemistries, temperatures, and aging states using only literature-reported lithium cobalt oxide (LCO) parameters achieves an average RMSE of 3.5 mV for terminal voltage and 0.4% for SOC estimation. Long-term validation over 9000 WLTP cycles confirms that the SOC estimation error remains within ±1% (average RMSE: 0.23%) without systematic drift. This study establishes a new paradigm based on physics-informed error compensation, providing a new avenue for enhancing the accuracy and interpretability of battery management systems.