Enhanced Bayesian physics-informed neural network for lithium-ion battery state-of-health estimations

To achieve accurate state-of-health (SOH) predictions for lithium-ion batteries this paper proposes an enhanced Bayesian physics-informed neural network (EB-PINN) that is integrated with split-conformal prediction. By the incorporation of physical constraints that are derived from battery-degradation dynamics into a Bayesian neural-network framework, the EB-PINN enables probabilistic modeling of the SOH. It addresses the limitations of traditional deterministic approaches, including the lack of uncertainty quantification, unreliable coverage, and insufficient physical consistency. A shared encoder is used to combine split-conformal prediction with Monte Carlo (MC) dropout, thereby enabling robust uncertainty quantification during battery SOH predictions. Additionally, physics-informed loss terms are included to ensure that the predictions adhere to the intrinsic physical laws of battery degradation. To validate the effectiveness of EB-PINN, experiments were conducted using publicly available datasets Health features, including the voltage and current statistics, constant-current and constant-voltage charging capacities, and entropy indicators, were extracted after outlier removal and dual normalization were conducted. After 300 epochs of training were performed, the model achieved a mean absolute percentage error (MAPE) of 0.403% and a 95% coverage rate of 95.272% for the uncertainty calibration. Furthermore, the EB-PINN enhances the interpretability and generalization capability across diverse battery chemistries and operating conditions.