Comparing experimental designs for parameterizing semi-empirical and deep learning-based lithium-ion battery aging models

Design of Experiment (DOE) methods can be applied to optimize test plans of cycle life aging studies with the aim to efficiently parameterize lithium-ion battery aging models. Since different DOEs exist and their effect on the prediction performance of battery aging models has not yet been investigated, we conducted a cycle life aging study with six commonly used DOEs (One-factor-at-a-time, Taguchi, Box–Behnken, Central Composite, Full Factorial, and D-optimal) and compare their influence on the prediction performance of a semi-empirical and a deep learning-based battery aging model. The results show that the semi-empirical model benefits the most from statistically optimized test plans. Compared to randomly selecting test plans, applying DOE methods helps to consistently achieve one of the lowest possible prediction errors for a given number of test points. Furthermore, it is shown that a D-optimal test plan and the test plans obtained from response surface methods (Box–Behnken and Central Composite) require only half as many test points as a Full Factorial test design, but still result in semi-empirical models with a high prediction accuracy that is similar to the Full Factorial test design. In contrast, deep learning-based battery aging models benefit significantly less from statistically optimized test plans. The highest prediction accuracy is achieved by the Full Factorial test plan and all other DOEs result in higher prediction errors and are even outperformed by several randomly defined test plans. Instead of using static designs, deep-learning-based models profit from a dynamic test optimization, which reduces the number of tested batteries during cycle life testing based on their information gain. We demonstrate that with our proposed dynamic test reduction algorithm, which analyzes the information gain based on aging features extracted after 100 EFC of cycling, up to 50% of all tested batteries of a Full Factorial test plan can be excluded from the cycle life study without deteriorating the prediction accuracy of the resulting deep learning-based battery aging model. • A cycle life study with 42 li-ion batteries is presented and published. • Optimal test design differs between battery aging modeling approaches. • Semi-empirical aging models benefit from statistically optimized designs. • Deep learning aging models benefit from dynamically optimized test plans.