Parameter identification using transfer learning and influence functions : The case of modeling lithium-ion battery

Parameter identification through transfer learning utilizes pre-identified source models to improve the identification performance of the target system, especially under challenging conditions such as large noise or the presence of outliers. This paper focuses on the problem of source selection for the multi-source transfer scenario without repeated identification procedures. Employing influence functions can reliably quantify the effect of different sources on the model accuracy of the target system. This enables the selection of the optimal source model from all candidates, thereby maximizing the target model's performance. The proposed approach is validated through a numerical case study and an application involving a equivalent circuit model of lithium-ion batteries in electric vehicles, demonstrating its effectiveness and robustness.