Recursive multilayer perceptron-based data-driven identification for a parameterized polarization model of rechargeable Li-ion battery

Abstract

A battery model is intended to reflect the dynamics of a battery in terms of terminal voltage under different stress factors. The terminal voltage further depends on the open-circuit voltage (OCV) and polarizations of the battery. Therefore, an important concern of battery modeling is how to accurately map the impact of stress factors on the OCV and polarizations of the battery. Commonly, the OCV curve obtained experimentally by the DC pulse technique is fitted empirically to model the OCV as a function of the operating conditions by using different regression techniques, which may be suitable for offline modeling. However, for online implementation, intelligent fitting techniques are necessary to adjust the model according to changing conditions. Similarly, different combinations of resistors and capacitors are selected to model the polarizations, and these components are parameterized by curve fitting of the terminal voltage. Although the cumulative effect of the components may result in good curve fitting, the first-principle justification of the selection of components is lacking. To address this problem, data-driven identification of the parametric model of a battery and its subsequent parameterization by a multilayer perceptron (MLP) with a gradient-descent back-propagation algorithm are introduced. The proposed algorithm is intelligent, and it provides justification for the selection of components to model polarizations. Furthermore, this algorithm can be easily extended to develop a model adaptable to changes in battery life and operating conditions.