Comparing experimental designs for parameterizing semi-empirical and deep learning-based lithium-ion battery aging models
Year: 2024
Type: article
Abstract: Conducting cycle life aging studies for lithium-ion batteries to parameterize battery aging models is a time-consuming and expensive process. Design of Experiment (DoE) methods provide a solution to reduce testing costs by minimizing the number of required test cases. However, the impact of different DoEs on modeling performance has not been thoroughly analyzed. Therefore, we conducted a lithium-ion battery aging study combining the test plans of six different DoE approaches. Based on this aging study, which is published alongside this publication, it is shown that the choice of the applied DoE method affects the prediction performance and that the optimal choice of DoE varies between semi-empirical and deep learning-based aging models. For the semi-empirical aging model, all investigated static DoE approaches yield higher prediction accuracy scores compared to randomly selecting test points. Notably, the D-optimal test design provides the highest prediction accuracy, even outperforming the same model fitted with a full factorial test plan that involved twice as many batteries. In contrast, the deep learning-based aging model, which generally achieved lower prediction errors, did not benefit from statistically optimized test plans. It achieved the highest prediction accuracy ($MAPE_{avg} = 1.77\%$) with a full factorial test design. All static DoEs, which are defined prior to the test execution, result in lower prediction scores and are outperformed by many random selections of test points. To mitigate this problem of static designs, we demonstrate the potential of a dynamic test reduction approach, which removes batteries with low information gain for model training based on aging features during cycling. Using this dynamic approach, 45\% of the test points from the full factorial test plan could be eliminated without compromising the prediction accuracy of the deep learning-based aging predictor.
Source: Journal of Energy Storage
Institutions: Technical University of Munich, TÜV SÜD (Germany)
Language: English
Subfield: Automotive Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $3,030
OpenAlex ID: W4404855970