A machine-learning prediction method of lithium-ion battery life based on charge process for different applications. (15th June 2021)

• Record Type: Journal Article
• Authors: Yang, Yixin
• Is Part Of: Applied energy. Volume 292(2021)
• Publication Date: 2021-06-15
• Subjects: Cycle life early prediction -- Remaining useful life prediction -- Lithium-ion battery -- Hybrid convolutional neural network -- Feature and cycle attention; Power (Mechanics) -- Periodicals; Energy conservation -- Periodicals; Energy conversion -- Periodicals; 621.042
• DOI: 10.1016/j.apenergy.2021.116897
• Languages: English
• ISSNs: 0306-2619

Abstract

Highlights: Established a hybrid CNN for both battery life early prediction and RUL prediction. Described battery life by several charge V, I and T curves and their difference. Applied a feature attention algorithm able to reduce errors by up to 2.7 times. Applied a cycle attention algorithm able to reduce errors by up to 3.3 times. Achieved 1.1% test error for battery life early prediction and 3.6% for RUL. Abstract: For accelerating the technology development and facilitating the reliable operation of lithium-ion batteries, accurate prediction for battery cycle life and remaining useful life (RUL) are both critical. However, diverse aging mechanisms, significant device variability and random working conditions have remained challenges. A reasonable description and an effective prediction algorithm are indispensable for achieving accurate prediction results. In this paper, battery terminal voltage, current and temperature curves from several charge cycles and especially their difference between these cycles are first utilized for description of battery cycle life and RUL. Moreover, a hybrid convolutional neural network (CNN), which is based on a fusion of three-dimensional CNN and two-dimensional CNN, is designed for their predictions. The battery charge voltage, current and temperature and their curves are first fused for considering the strong relationships between them. And the features hidden in the curves are extracted and modelled automatically. Furthermore, a feature attention algorithm and a multi-scale cycle attention algorithm are proposed to estimate the relationships between different features and cycles respectively for further heightening the prediction performance. Experiments and comparisons are conducted. The results show that the proposed method is an accurate method for different applications. It achieved 1.1% test error for battery cycle life early prediction of different batteries under different charge policies, and 3.6% for RUL prediction.