Machine learning configurations for state of charge predictions of Li-ion batteries

Abstract This work investigates the development of a Nonlinear Autoregressive with Exogenous Input (NARX) Artificial Neural Network (ANN) model to predict the dynamic State-of-Charge (SOC) of a Li-ion battery (LIB) using voltage and Electrochemical Impedance Spectroscopy (EIS) data. To optimize the model’s performance, extensive ANN parameter adjustments were investigated and various data structuring techniques were explored. It was found that directly inputting all EIS data corresponding to different SOC levels led to a significant decrease in predictive accuracy. Specifically, the root mean square error (RMSE) for SOC prediction increased by approximately 40% when using frequency-separated EIS data. In contrast, utilizing EIS data from a specific SOC level (0%) significantly improved the model’s performance. By selectively excluding input features with lower input–output correlation, the RMSE was reduced by 62%, outlining the significant advantage of using EIS measured at 0% SOC. This result highlights the importance of careful data selection and preprocessing in enhancing the accuracy and efficiency of NN-based SOC estimation. The findings of this study provide valuable insights into the optimal data structuring and feature selection strategies for developing accurate and efficient NN models for battery SOC prediction. Graphical abstract