Ultra-Early Prediction of Lithium-Ion Battery Cycle Life Based on Visualized Single-Cycle Data

This article proposes a battery cycle life prediction framework based on the visualized data of a single charging-discharging cycle during the ultra-early stage of the battery operation. To develop the framework, a sliding window-based image construction method is proposed that divides the raw sequential data extracted from a single cycle into multiple sub-sequences and uses the Euclidean distance between any two sub-sequences to construct the images. The framework employs three AlexNet blocks to build a sophisticated convolutional neural network model to capture the features from the images. Comprehensive evaluations of the framework are conducted using the Severson dataset (Severson et al., 2019) with 124 batteries. All the four models trained using the three measurements (i.e., voltage, current, and capacity) and the combination of them result in acceptably low cycle life prediction errors for the 29 batteries in the test set. Among the four models, the “Full” model based on the combination of the three measurements performs the best with an average root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination ($\mathrm{R}^{2}$) of 76.81 cycles, 7.05%, and 0.9178, respectively. As a feature-free method, the “Full” model outperforms three feature-based methods and another three feature-free methods, demonstrating its effectiveness in predicting the battery cycle life during the ultra-early stage with only the data of a single cycle.