Capacity prediction method of lithium-ion battery in production process based on eXtreme Gradient Boosting

Abstract

Measuring capacity through the lithium-ion battery (LIB) formation and grading process takes tens of hours and accounts for about one-third of the cost at the production stage. To improve this problem, the paper proposes an eXtreme Gradient Boosting (XGBoost) approach to predict the capacity of LIB. Multiple electrochemical features are extracted from the cell voltage curves obtained during the formation and 20% grading processes, and these features are ranked using the grey relational analysis (GRA) method. The charging polarization voltage, fixed-voltage rise time, and static-stage voltage difference, which show a high degree of correlation, form the optimal feature set. The sparrow search algorithm (SSA) improves accuracy and efficiency by optimizing the hyperparameters of XGBoost model. The experimental results indicate that the root mean square error (RMSE) and mean absolute percentage error (MAPE) of this method are respectively 0.1543 Ah and 0.2456%, which are lower than those of other data-driven methods and predict low-capacity cells with equal accuracy. Economically, our method significantly reduces the energy consumption by approximately 278 Wh (56.5%) and shortens the time required for the grading stage by about 7 h 30 min (80.36%) for each cell, the generalization of the model is verified by other types of battery data.