Data-driven Battery Modeling based on Koopman Operator Approximation using Neural Network

Physics-based battery dynamical models are typically described by a set of highly nonlinear differential-algebraic equations or partial differential equations. These models provide ground-truth information of highly-detailed and accurate battery dynamics for various studies. However, complexity of the battery models makes it difficult to adopt the models in existing analytical framework for battery control and optimization. Hence, reducing model complexity is essential to achieve computational efficiency and optimality of battery control problems. However, most reduced-order models lack details of the battery dynamics which degrades accuracy of the solutions and also lose important dynamical behaviors. In this work, we propose a data-driven modeling framework to construct a linear surrogate model of battery dynamics based on Koopman operator. The model introduced in the paper is linear described in infinite function space
¹
based on Koopman operator theory. Furthermore, we leverage neural network to learn high-dimensional space where the Koopman operator is approximated using time-series data collected from simulations of the original nonlinear battery models. As a result, the Koopman operator model estimated by the proposed method reduces the model complexity while maintaining prediction accuracy compared to the original battery models. The validity of the proposed framework is demonstrated by constructing a Koopman operator model for the single particle model, one of the full-order physics-based Li-ion battery models, considering dynamic operating conditions from PJM regulation market signals as well as periodic charging/discharging cycles.
¹
Function space is a set of functions, e.g., polynomials, and forms a topological vector space whose points are functions.