Lightweight Electrochemical Hybrid Modeling Approach for Li-Ion Batteries Using Gaussian Process Regression

The development of next-generation battery management systems needs models with enhanced performance to enable advanced control, diagnostic, and prognostic techniques for improving the safety and performance of lithium-ion battery systems. Specifically, battery models must deliver efficient and accurate predictions of physical internal states and output voltage, despite the inevitable presence of various system uncertainties. To facilitate this, we propose a lightweight hybrid modeling framework that couples a high-fidelity physics-based electrochemical battery model with a computationally-efficient Gaussian process regression (GPR) machine learning model to predict and compensate for errors in the electrochemical model output. This is the first time that GPR has been implemented to predict the output residual of an electrochemical battery model, which is significant for the following reasons. First, we demonstrate that GPR is capable of considerably improving output prediction accuracy, as evidenced by an observed average root-mean-square prediction error of 7.3 mV across six testing profiles, versus 119 mV for the standalone electrochemical model. Second, we employ a data sampling procedure to exhibit how GPR can use sparse training data to deliver accurate predictions at minimal computational expense. Our framework yielded a ratio of computation time to modeled time of 0.003, indicating ample suitability for online applications.