Modeling of Li-ion batteries for real-time analysis and control: A data-driven approach

This paper presents a data-driven model (DDM) of Li-ion batteries (LIBs). Accurate real-time modeling of LIBs allows for faster and more aggressive inputs and operations, improving their performance and safety. The DDM was developed using the enhanced single-particle model with electrolyte as the plant dynamics. A sparse model of the system was developed based on a library of potential terms. A sparsity-promoting optimization algorithm was used to balance the trade-off between the model accuracy and complexity. We compared the Sequentially Thresholded Ridge regression (STRidge) and a LASSO optimization and showed the advantages of using the STRidge. First, we developed a model using only voltage and current signals (Model I). Model I’s performance and robustness were assessed via validation and generalization tests, where the model achieved normalized root mean square error (NRMSE) values of less than 1.6%. Additionally, we evaluated Model I’s robustness to noise, achieving NRSME< 2%. We showed the trend of Model I parameters with the charge/discharge curves. However, this model required switching parameters as state-of-charge (SOC) changes. Therefore, we augmented the model by including terms related to the SOC in the library (Model II) to avoid switching. We showed the performance of Model II using the US06 highway driving cycle as the cell’s charge varied from 40%SOC to 20%SOC with small errors (NRMSE=4.78%). The results showed that the models accurately predict the dynamic response of the cells in the simulated scenarios. The models are interpretable as they have explicit input and output terms and allow for efficient control design.