Multisource domain adaptation mutual information network for small sample state-of-health prediction of lithium-ion batteries

Abstract Lithium-ion batteries (LIBs) play an important role in energy systems with their functions of energy storage and regulation, and the state-of-health (SOH) prediction has received great attention. The data-driven approach based on deep neural networks has greatly promoted the advancement of SOH prediction technology. However, the diversity of battery electrochemistry and dynamic changes in working conditions can lead to differences in cross domain data distribution, which poses significant challenges for data-driven methods to accurately predict the SOH of LIBs. The small sample of battery aging data, compared with the number of batteries used in the application, makes accurate SOH prediction more difficult. To address these challenges, a novel transfer learning model, multisource domain adaptation mutual information network (MSDA-MIN) model is built to predict the SOH of LIBs on cross domain. This model integrates continuous time neural dynamics and mutual information to achieve simultaneous alignment of source joint distribution and target joint distribution. In three cross domain scenarios of constant current-constant voltage charging, staged charging, and cross chemical composition, the average root mean square error of MSDA-MIN is 1.17%, 0.80%, and 2.85%, respectively, which is superior to the state-of-the-art domain adaptation models.