Omidreza presented at the IMECE 2023, International Mechanical Engineering Congress & Exposition, ASME.

Title: Interpretable Machine Learning Modeling of Li-ion Batteries

Abstract: Lithium-ion batteries (LIBs) are present in many modern applications due to several desirable properties such as high energy and power densities. Accurate real-time modeling of LIBs improves their operation and safety. Traditionally, equivalent circuit models (ECM) have been used to model LIBs due to their simplicity. These models utilize passive electrical components such as resistors and capacitors to model the battery’s responses. However, their lack of connection to physics results in poor extrapolation performance, and they require limiting the operating range and life of LIBs. These drawbacks have led to an increase in the popularity of physics-based models of LIBs in real-time applications over the past few years. However, developing these models is very cumbersome and requires several measurements and information that often are unavailable and change with the operational conditions and life of the batteries. We propose a novel solution for battery management systems through interpretable machine learning (ML) modeling. We identify the governing equations of LIBs without requiring the in-situ measurements and proprietary information needed by physics-based models. To address the common issue of overfitting and finding a wrong fit with many ML techniques, which results in poor performance in unseen scenarios, we propose a novel physics-informed reduced-order nonlinear model of LIBs. The model’s input is the electrical current and the measurable output is the voltage. We seek an input/output based formulation that predict the dynamics of LiBs., and SOC as the output of the ML model. We used Sequentially Thresholded Ridge regression (STRidge) to promote the model’s sparsity. The technique includes physics-based functions and employs sparse regression to balance the accuracy and complexity of the model using measured data. These terms were associated with the solution of the Doyle-Fuller-Newman (DFN) model, which is a mechanistic model. The added observables (functions of the input/output data) are solid and electrolyte concentration, the Butler-Volmer equation, and solid and electrolyte electric potentials. Furthermore, we improve our ML models by i) augmenting the method to determine the battery’s governing equations with noisy measurements accurately, and ii) restructuring the optimization problem and introducing the sparsifying parameters as hyperparameters and tuning them using a training and a validation dataset, hence resulting in a more generalizable model. Sparsifying parameters are thresholds and regularization parameters that nullify the less important terms and adjust the coefficient values, respectively to balance model accuracy and complexity. We use training data with known input/output to find sparse models for different ranges of hyperparameters. The identified sparse model for each hyperparameter is assessed with the validation data set whose input is only known. The hyperparameters that provide the lowest cost function for the validation data are selected. In addition, we present a robust modeling technique for noisy measurements. This model uses a Kalman filter approach, where the sparse terms are updated based on the Kalman filter to remove the noise from the voltage data and enhance the state of charge (SOC) prediction. We have tested the method using data from advanced chemistry that is used to develop new electric vehicle applications (21700 cylindrical cells, NMC811). We used uniformly distributed electrical current signals up to 2C/4C charge/discharge rates for training the model and the US highway profile (US06) for the validation set. We showed the model’s accuracy using the Urban Dynamometer Driving Schedule (UDDS) as the unseen test data. The model predicted the response with less than 8.3×10^−5 normalized root mean square error (NRMSE) for SOC and Voltage predictions.

Related Publications:

1. O. Ahmadzadeh, R. Rodriguez, Y. Wang, D. Soudbakhsh, A physics-inspired machine learning nonlinear model of li-ion batteries, in: 2023 American Control Conference (ACC), IEEE, 2023, pp. 3087–3092. 10.23919/ACC55779.2023.10156368
2. O. Ahmadzadeh, Y. Wang, D. Soudbakhsh, Sparse modeling of energy storage systems in presence of noise, IFAC-PapersOnLine 56 (2) (2023) 3764–3769. https://doi.org/10.1016/j.ifacol.2023.10.1546