Transfer Learning With Long Short-Term Memory Network for State-of-Health Prediction of Lithium-Ion Batteries

• Published in: IEEE transactions on industrial electronics (1982) Vol. 67; no. 10; pp. 8723 - 8731
• Main Authors: Tan, Yandan, Zhao, Guangcai
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.10.2020; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Batteries; Battery charge measurement; Current measurement; Degradation; Feature expression scoring (FES); Learning; Lithium; Lithium-ion batteries; lithium-ion battery (LIB); long short-term memory (LSTM); Noise prediction; Noise reduction; Noise sensitivity; Prediction models; Predictive models; Rechargeable batteries; state-of-health (SOH); Task analysis; Training; transfer learning; Voltage measurement
• ISSN: 0278-0046; 1557-9948
• DOI: 10.1109/TIE.2019.2946551

Existing state-of-health (SOH) data-driven prediction techniques for lithium-ion batteries are subject to mass training data, which leads to limited application. To face the challenge, in this article, we propose a novel SOH prediction method based on transfer learning. The long short-term memory (LSTM) combined with fully connected (FC) layers is designed as the base model. The LSTM can learn the long-term dependencies of battery aging to reduce the noise sensitivity of the prediction model, and the FC layers serve as the "firewall" during the transferring process. A feature expression scoring (FES) rule is developed to assess the relevance of multiple prediction tasks. Different from traditional transfer learning, we select the task with the highest FES score to obtain the base model with superior generalization performance. During transfer learning, the fine-tuning strategy is executed for the tasks with high scores, but rebuilding strategy for the low score one. Only using the first 25% of a dataset for transfer training, our technique can predict more phases compared to traditional data-driven methods, which will avoid more unreasonable operations from users. The experimental results verify that the proposed method can achieve accurate, fast, and steady SOH prediction. Compared to some existing data-driven methods, our method obtains optimal performance.