Bootstrap-Based Sparse Modeling for Temperature-Dependent State-of-Charge Prediction of Batteries

Abstract

Abstract. Data-driven modeling of complex systems, such as Li-ion batteries with explicit terms, has shown promising results. However, these methods typically work on a single dataset and are sensitive to noise. Here, we present a physics-informed and temperature-dependent data-driven model of Li-ion batteries by aggregating experimental data collected at various operating conditions using statistical bootstrapping techniques. The modeling process starts by creating a set of bootstrapped samples and including terms based on first principles and temperature-related terms in the model library. Then, multiple models were created using the bootstrapped data and aggregated into an ensemble with improved accuracy and robustness. The final model takes the median of the coefficients of the bootstrapped models and has temperature as a model input. We conducted several experiments to create the dataset for developing and validating the novel temperature-dependent battery model. The experiments included charging a Li-ion cell with a constant-current-constant-voltage algorithm and discharging it with custom and standard driving profiles at various temperatures from −10 °C to 40 °C. The baseline parsimonious model was developed at 25 °C and augmented using temperature-dependent terms to determine the sparse ensemble model, which achieved an error <2.5% on unseen test data.