Early Prediction of Lithium Battery Cycle Life With Variable Sequence Feature Extraction Based on Recursive Kernel Canonical Correlation Analysis

Different machine learning (ML) models are widely applied to predict battery cycle life using early cycle data. However, they often only utilize a few key features extracted from the data as the model inputs, which reduces the efficiency of information utilization and subsequently lowers the accuracy and stability of the estimation results. This article proposes a novel feature extraction approach. First, horizontal and vertical sequence sets are derived from different battery variables. Then, these sequences are fit to the battery cycle life using kernel canonical correlation analysis (CCA), and the resulting fits are used as the input features to the ML models. Subsequently, a validation group is employed to optimize the hyperparameters. To address the issue of long training time caused by extensive feature extraction and to accommodate larger datasets, the recursive kernel CCA (RKCCA) method is introduced, which reduces the solving time to 1/267 of the original at a minor cost of precision. Finally, a specified threshold for correlation coefficients is applied to select the high-quality features. The retained features are tested using four different ML models, among which random forest (RF) achieves the highest accuracy of 8.2% mean absolute percentage error (MAPE).