Combining electrochemistry and data-sparse Gaussian process regression for lithium-ion battery hybrid modeling

Abstract

The widespread adoption of lithium-ion batteries is driving the concurrent development of advanced battery management systems, which seek to maximize safety and performance through state-of-the-art control, diagnostic, and prognostic techniques. To enable these capabilities, battery models must provide accurate predictions of output voltage and physical internal states, which is challenging due to the inevitable presence of system uncertainties and limited online computational resources. In response, a computationally-efficient hybrid modeling framework is proposed that integrates a physics-based electrochemical battery model with a Gaussian process regression (GPR) machine learning model to compensate for output prediction errors due to system uncertainties. A key feature of the framework is a proposed data sampling procedure that mitigates computational expense by leveraging the prediction capability of GPR under sparse data. The hybrid model was experimentally validated, yielding an average prediction root-mean-square error (RMSE) of 7.3 mV across six testing profiles, versus 119 mV for the standalone electrochemical model. The observed ratio of computation time to modeled time was 0.003, which is amply sufficient for online BMS applications. Finally, in a simulated BMS demonstration, the hybrid model was observed to reduce parameter estimation errors by one order of magnitude, the voltage prediction RMSE by 63 %, and the state estimation RMSE by 52 % when compared against the standalone electrochemical model.