Abstract:

In the use of lithium-ion batteries, real-world operating conditions often restrict the availability of extensive, fully labeled data. This limitation presents a significant challenge in accurately predicting battery capacity degradation. To address this issue, this study introduces an approach that combines a capacity degradation curve augmentation algorithm with traditional neural network techniques to predict the trajectory of battery capacity degradation. First, a small set of fully labeled battery capacity degradation data, polynomial functions, and the Monte Carlo method are used to generate virtual capacity degradation curves. These augmented curves are subsequently filtered using KL divergence and Euclidean distance metrics. Next, four widely used neural network models—Multilayer Perceptron, Convolutional Neural Network, Gated Recurrent Unit, and Long Short-Term Memory—are developed to map the virtual degradation curve data to the actual battery capacity. Finally, the models are pretrained with a small amount of fully labeled battery data using the virtual capacity degradation curve data as input and the actual capacity as output. They are then fine-tuned with early degradation data from the target battery to predict its capacity degradation trajectory. The effectiveness of the proposed method is validated using data from 77 lithium-ion batteries subjected to various discharge schemes. The results demonstrate that, even with just three fully labeled battery capacity degradation datasets, the prediction performance of the proposed method remains robust regardless of the type of neural network used. All four neural networks effectively predict the capacity degradation trajectories of the remaining batteries, achieving a mean MAPE and RMSE of less than 2.3% and 31 mAh, respectively.

Key words: lithium-ion battery, capacity degradation trajectory, data-scarce condition, neural network