A generalizable, data-driven online approach to forecast capacity degradation trajectory of lithium batteries

作     者：Xinyan Liu Xue-Qiang Zhang Xiang Chen Gao-Long Zhu Chong Yan Jia-Qi Huang Hong-Jie Peng Xinyan Liu;Xue-Qiang Zhang;Xiang Chen;Gao-Long Zhu;Chong Yan;Jia-Qi Huang;Hong-Jie Peng

作者机构：Institute of Fundamental and Frontier SciencesUniversity of Electronic Science and Technology of ChinaChengdu 611731SichuanChina Beijing Key Laboratory of Green Chemical Reaction Engineering and TechnologyDepartment of Chemical EngineeringTsinghua UniversityBeijing 100084China Advanced Research Institute of Multidisciplinary ScienceBeijing Institute of TechnologyBeijing 100081China 

出 版 物：《Journal of Energy Chemistry》 (能源化学（英文版）)

年 卷 期：2022年第31卷第5期

页      面：548-555页

核心收录：

学科分类：0808[工学-电气工程] 08[工学] 

基　　金：supported by the Beijing Municipal Natural Science Foundation (Z20J00043) the National Natural Science Foundation of China (21825501, 22109020, 22109082, and U1801257) the National Key Research and Development Program(2016YFA0202500) the Tsinghua University Initiative Scientific Research Program the University of Electronic Science and Technology of China for its financial support through the Start-Up Fund for Outstanding Talent with grant number A1098531023601307 

主　　题：Battery prognosis Machine learning Time series forecasting Online prediction Lithium metal batteries 

摘      要：Estimating battery degradation is vital not only to monitor battery’s state-of-health but also to accelerate research on new battery chemistries. Herein, we present a data-driven approach to forecast the capacity fading trajectory of lab-assembled lithium batteries. Features with physical meanings in addition to predictive abilities are extracted from discharge voltage curves, enabling online prediction for a single cell with only its historical data. The robustness and generalizability allow for the demonstration on a compromised quality dataset consisting of batteries varying in battery architectures and cycling conditions,with superior accuracy for end of life and degradation trajectory prediction with average errors of 8.2%and 2.8%, respectively. Apart from the impressive prediction accuracy, the as-extracted features also provide physical insights, the incorporation of which into material design or battery operation conditions further enlightens the development of better batteries. We highlight the effectiveness of time-seriesbased techniques in forecasting battery cycling performance, as well as the huge potential of datadriven methods in unveiling hidden correlations in complicated energy chemistries such as lithium metal batteries.