Physics-Informed Recurrent Neural Network Framework for Parameter Identification of Li-ion Batteries

Accurate parameter identification is fundamental to the reliable operation of Battery Management Systems (BMS). However, traditional identification methods often struggle to balance computational efficiency with the physical interpretability of model parameters. To address this challenge, this paper proposes a novel Physics-Informed Recurrent Neural Network (PIRNN) framework for the online parameter identification of the first-order Thevenin Equivalent Circuit Model (ECM). Unlike conventional black-box approaches (e.g., least square method), the proposed architecture explicitly maps the discrete state-space equations of the ECM onto the RNN topology. A physics constraint layer is integrated to enforce electrochemical principles, ensuring that identified parameters remain strictly within feasible domains. Furthermore, the Adagrad optimization algorithm is employed to resolve convergence difficulties arising from the multi-scale nature of battery parameters. The framework is validated using experimental data from Panasonic NCR18650B (NCA) battery under dynamic discharge profiles at 25 °C. Experimental results demonstrate that the PIRNN achieves high-fidelity voltage tracking with a global Root Mean Square Error (RMSE) of 38.84 mV. Crucially, the algorithm successfully extracts physically meaningful parameter trajectories, accurately capturing nonlinear behaviors such as the surge in polarization resistance at low State of Charge (SOC).