Pressure-augmented physics-informed dynamic graph neural network and dual Kalman filter framework for robust battery state-of-charge estimation

Abstract

State-of-charge (SOC) estimation is a core function of battery management systems that directly impacts system safety and energy management efficiency. Under complex conditions involving multisource disturbances, parameter drift, and cross-regime operation, traditional models often fail to adapt because of simplified physical assumptions or rigid structural formulations. To address these challenges, this paper proposes an SOC estimation framework that integrates a physics-informed graph structure with a dual Kalman filtering mechanism. The framework constructs a structured graph that encodes electro–thermal–mechanical coupling relationships, explicitly modeling the physical dependencies among current, voltage, temperature, and internal pressure. A dynamic graph neural network is employed to extract spatiotemporal prior features from multisource signals. Furthermore, a decoupled dual-filter mechanism—comprising a cubature Kalman filter (CKF) for dynamic state estimation and an extended Kalman filter (EKF) for online circuit parameter adaptation—is introduced to increase model flexibility and accuracy. A pressure–temperature coupling compensation unit is additionally designed to improve robustness under extreme environmental perturbations. Extensive experiments conducted on real-world datasets across various operating conditions, temperatures, and battery chemistries demonstrate that the proposed method significantly outperforms conventional filtering algorithms and typical data-driven models in terms of estimation accuracy, stability, and generalizability. The results confirm the framework’s strong physical consistency and practical applicability, offering a novel and interpretable solution pathway for high-reliability SOC estimation under complex operating scenarios.