A method for estimating lithium-ion battery state of health based on physics-informed machine learning

Abstract

Data-driven approaches are widely applied in estimating the State of Health (SOH) of Lithium-Ion Batteries (LIB). However, these methods often suffer from a lack of interpretability. To address this issue, this article proposes a method called Battery Physics-Informed Neural Network (BPINN) to enhance the interpretability of the Feedforward Neural Network (FNN) in SOH prediction. This article is based on the concept of Physics-Informed Neural Networks (PINN). Features are initially extracted from Incremental Capacity (IC) curves to characterize the battery aging process. Notably, IC curve peaks (P-IC) reflect electrochemical reactions during charge and discharge cycles. The degradation of these peaks is directly related to the loss of active materials, causing SOH reduction. This article transforms the monotonic relationship between P-IC and SOH into a physical constraint, which is embedded into model training. Furthermore, during prediction, a secondary “training” based on physical constraints is applied to the FNN prediction results, enhancing the model's interpretability and accuracy. The proposed method is validated using publicly available battery datasets from NASA and the University of Oxford. Results show BPINN effectively improves SOH prediction accuracy, reducing the Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) to below 0.4 %.