Bilevel-optimized continual learning for predicting capacity degradation of lithium-ion batteries

This study proposes a novel method to predict the capacity degradation trajectory of lithium-ion batteries in a real-world scenario in which datasets with different degradation patterns are continuously updated. Most traditional machine learning models, trained on fixed data and assuming static data distribution, are insufficient for capturing the varying capacity degradation patterns of batteries. To address this limitation, we propose a novel framework combining continual learning and bilevel optimization to predict the overall capacity degradation trajectory of batteries in a dynamic real-word environment. Elastic weight consolidation (EWC), a representative continual learning method, was applied to a convolutional neural network to predict knots in the capacity degradation curve. Next, the bilevel optimization technique was used to determine the optimal value of the importance rate of past data to learn the different degradation patterns of batteries. The proposed method was validated using 251 experimental cells grouped into six batches from two different datasets. The experimental results show that the average mean absolute error of the proposed method after training six batches was 22% lower than that of the existing methods without the bilevel-optimized EWC. We demonstrated that the proposed method is sufficient to learn the new degradation pattern while maintaining the information of past patterns by storing only 5% of all the data, which enables the application of the lithium-ion battery health diagnosis system in an actual environment.