Early Prediction of Knee Point and Knee Capacity for Fast-Charging Lithium-Ion Battery With Uncertainty Quantification and Calibration

Knee point has been observed in the capacity degradation of lithium-ion (Li-ion) batteries under fast charging, such as electric vehicle applications, which divides the degradation into slow aging and fast aging. Early prediction of knee point and knee capacity can help to guarantee the safe and optimal operation of batteries. Existing works fail to predict knee capacity and also lack uncertainty quantification of the early prediction results. Moreover, how to construct suitable features for early knee point and knee capacity prediction remains open. To this end, this article proposes a Bayesian neural network (BNN)-based method for early prediction of knee point and knee capacity of fast-charging batteries with uncertainty quantification and calibration. A general feature extraction method based on sliding on capacity matrix is introduced for identifying battery characteristics at early aging cycles. BNN equipped with Monte Carlo dropout method is developed to establish the early prediction model for knee point and knee capacity, simultaneously providing uncertainty quantification. To obtain a reasonable prediction interval, a standard deviation scaling method is proposed for uncertainty calibration. The proposed method has been demonstrated on three publicly available fast-charging Li-ion battery datasets, generating earlier predicting point, improved prediction accuracy, and more advanced uncertainty quantification and calibration capability.