Early prediction of cycle life for lithium-ion batteries based on evolutionary computation and machine learning

Accurate early cycle life prediction of lithium-ion batteries is critical for efficient and rational battery energy distribution and saving the technology development period. However, relatively little research has been carried out on the early prediction based on evolutionary computation approaches. The principal purpose of this study is to explore the accurate early prediction of cycle life for lithium-ion batteries based on evolutionary computation techniques and machine learning approaches. Firstly, twenty features related to the capacity degradation are extracted using data from only the first 100 cycles. Next, four filter methods, four wrapper feature selection methods based on evolutionary computational strategies, and fusion feature selection methods are used to determine their feature subsets combined with the elastic net, respectively. Afterward, seven machine learning methods are selected to conduct comparative performance studies combined with the optimal feature set chosen. Finally, the research results show that the fusion feature selection method that combines Pearson correlation coefficient and differential evolution-based wrapper approach performs the best cycle life early prediction performance. Among, the prediction and average percentage errors are as small as 43.38 cycles and 5.21%, respectively; the coefficient of determination is as high as 0.98.

Automated summary

This study aims to accurately predict the early cycle life of lithium-ion batteries using evolutionary computation techniques and machine learning approaches. The research results show that the fusion feature selection method performs the best in terms of cycle life early prediction performance.