Federated Physics-Informed Neural Networks for Privacy-Preserving Lithium-Ion Battery Degradation Prognostics Under Sparse Data Conditions

Lithium-ion batteries (LIBs) are widely used in electric vehicles (EVs), renewable energy storage systems (ESSs), and smart grids, but accurate degradation and remaining useful life (RUL) prediction remains challenging under sparse data and distributed data-storage conditions. This study proposes a Federated Physics-Informed Neural Network (F-PINN) framework that combines federated learning with reduced-order physics-informed constraints for privacy-preserving battery degradation prognostics. The framework incorporates degradation monotonicity, a reduced-order Fickian lithium-diffusion constraint, and a battery thermal energy-balance constraint into the neural-network optimization. The effective lithium-concentration and temperature states are treated as latent physical states rather than independently supervised targets. Experiments using NASA lithium-ion battery datasets evaluate the framework under different training-data ratios and noise conditions against LSTM, GPR, SVR, and standalone PINN models. F-PINN substantially reduces the RMSE compared with standalone PINN. The results demonstrate improved sparse-data prognostics, physically constrained degradation trajectories, and privacy-preserving collaborative learning. The framework provides a potential scalable architecture for distributed battery-management applications.