Research on state of health prediction of lithium-ion batteries based on sparrow algorithm and dynamic physical constraint neural network

To address the limitations of insufficient model generalization under small-sample conditions, empirical dependence of physical priors, and inadequate characterization of nonlinear degradation features in the state of health (SOH) prediction for lithium-ion batteries, this paper proposes a hybrid prediction framework fusing the sparrow search algorithm (SSA) and dynamic physics-informed neural network (PINN), termed SSA-PINN-GRU-Attention. Firstly, SSA is employed to globally optimize the parameters of the double-exponential degradation model for constructing adaptive physical prior boundaries. Secondly, a gated recurrent unit combined with the self-attention mechanism (GRU-Attention) is introduced to strengthen the modeling capability for long-term temporal dependencies and local capacity regeneration phenomena. Finally, a dynamic physical constraint weight strategy based on the Huber loss is designed to achieve effective collaboration between physical laws and data-driven learning. Experimental results under leave-one-out validation on the NASA and CALCE CS2 datasets demonstrate the strong cross-battery and cross-dataset generalization capability of the proposed model. The mean e R 2 ,a root mean square error ( e RMSE ), and a mean absolute error ( e MAE ) are 0.9754, 1.3128%, and 0.8014% on the NASA dataset, and 0.9894, 1.2676%, and 0.9873% on the CALCE dataset, respectively. The proposed model also outperforms the representative Double-Exp, PI-GPR, and Dynamic PINN baselines.