System Identification for Lithium-Ion Batteries with Nonlinear Coupled Electro-Thermal Dynamics via Bayesian Optimization

TL;DR

The paper tackles the challenging problem of identifying parameters for a nonlinear, coupled electro-thermal LiB model (NDC-T) using maximum likelihood estimation. It develops a Bayesian-optimization framework with a Monte Carlo likelihood evaluation and a dynamic ellipsoid-based search-space reduction to efficiently locate the global optimum in a high-dimensional, nonconvex landscape. Simulation studies on a 3.3 Ah LiB demonstrate accurate recovery of ten parameters and accurate voltage/temperature predictions across multiple current profiles and ambient temperatures, highlighting improved convergence and computational efficiency. The approach provides a practical, continuous-time, data-driven parameter extraction workflow that can extend to a broad class of electrochemical models and ECMs.

Abstract

Essential to various practical applications of lithium-ion batteries is the availability of accurate equivalent circuit models. This paper presents a new coupled electro-thermal model for batteries and studies how to extract it from data. We consider the problem of maximum likelihood parameter estimation, which, however, is nontrivial to solve as the model is nonlinear in both its dynamics and measurement. We propose to leverage the Bayesian optimization approach, owing to its machine learning-driven capability in handling complex optimization problems and searching for global optima. To enhance the parameter search efficiency, we dynamically narrow and refine the search space in Bayesian optimization. The proposed system identification approach can efficiently determine the parameters of the coupled electro-thermal model. It is amenable to practical implementation, with few requirements on the experiment, data types, and optimization setups, and well applicable to many other battery models.

Figures (5)

• Figure 1: The NDC-T model, which couples the NDC submodel and the lumped thermal submodel.
• Figure 2: Upper and lower bounds on the search space for each parameter over four rounds of BayesOpt. "$\color{darkred}-$" for upper bound, "$\color{darkblue}-$" for lower bound, and "$-$" for the true parameter.
• Figure 3: Projection of the ellipsoid-shaped parameter search space on $C_s$, $R_\mathrm{surf}$ and $R_o$ in round one to three of BayesOpt.
• Figure 4: Average maximum log-likelihood over 20 independent runs. $"-"$ shows the log-likelihood evaluated at the true parameters.
• Figure 5: Comparison of the measured and predicted voltage and surface temperature under the UDDS discharging profile at $T_{\mathrm{amb}}=283$ kelvin.

Theorems & Definitions (1)

• Remark 1