On October 12, 2022, an interdisciplinary team led by academician Han Ding, professors Yunhui Huang and Ye Yuan published a cover article on Energy & Environmental Science entitled "Real-time personalized health status prediction of lithium-ion batteries using deep transfer learning".
With the national policy of carbon peaking and carbon neutrality, the electrification of power systems in a variety of fields is a mainstream trend. Battery health management has been seen as a high-value scientific topic with the hope of optimizing the energy output rate and effectively promoting the realization of the 'dual carbon' goal. In recent years, machine learning methods have been successfully applied to battery health status prediction applications, such as early cycle life prediction, state of health estimation and remaining useful life prediction. However, two significant challenges stand in the way of developing a reliable real-time personalized health status prediction strategy:
(1) Different end-users have different battery usage preferences, and the distribution of battery data varies significantly due to the inconsistency of discharging protocols. Directly evaluating new batteries by building models using battery data with other discharge protocols is hard to meet the actual needs. (2) Due to the severe dependence on historical data, it is impossible to evaluate battery health at any charge-discharge cycle.
Therefore, it is difficult for the existing technologies to establish a reliable and general model to assess the health status of different batteries for different end-users in real time. The industry urgently needs to customize personalized battery health management strategies to ensure end-user safety and provide scientific guidance for manufacturers to improve battery materials by providing scheduled battery usage strategies.
Academician Ding Han's team has devoted a long time to interdisciplinary research and has deeply explored the common scientific problems of multiple disciplines. Faced with the difficulties mentioned above, the team innovatively introduced the idea of transfer learning in artificial intelligence and customized a real-time personalized battery health assessment for end-users. This approach can realize the transfer of health assessment among batteries of the same material type and different discharge protocols and the transfer of health assessment among lithium-ion batteries of different chemistries. With only 30 historical cycles, the proposed approach can evaluate battery health status accurately in real time. Compared with the general framework developed by Ding Han's team [National Science Review, 2020, 7(2): 418-429], the proposed approach has improved the accuracy of battery health assessment by 3~5 times. Compared with the battery health assessment method jointly developed by MIT and Stanford University [Nature Energy, 2019, 4(5): 383-391], the proposed approach can not only customize real-time health assessment for different end-users but also reduce the amount of utilized historical data by 2/3. In addition, the study published the largest standard dataset of multi-stage discharge protocols of lithium-ion batteries so far in the world, which can be a benchmark for researchers in related fields. The "real-time personalized" transfer learning method proposed in this paper offers a new idea for battery health management, which is of great significance for battery manufacturing, testing and recycling fields. This study can be extended to the health assessment of solid-state batteries, quasi-solid-state batteries, lithium-sulfur batteries, sodium-ion batteries, etc.
This research was supported by the National Natural Science Foundation of China (92167201) and the HCP Program.
Related link: https://pubs.rsc.org/en/content/articlelanding/2022/ee/d2ee01676a
Source：School of Mechanical Science & Engineering of HUST
Edited by: Peng Yumeng