Hybridization of an electro-thermal model with neural networks for real-time prediction of voltage and the volume-average temperature of lithium-ions batteries

In the context of the energy transition, lithium-ion batteries have become indispensable, from consumer electronics to electric vehicles; however, their safe and optimal operation relies on reliable models for state estimation and control. This study proposes two hybrid architectures that couple the single particle model with a lumped thermal module (SPMT) and multilayer perceptrons (MLP), designed to predict simultaneously the terminal voltage and the internal temperature of cells. Simulations performed across various current profiles show an average improvement in accuracy of more than 85% for voltage and 90% for temperature compared with the SPM-T alone, while retaining moderate computational complexity. Furthermore, additional observations indicate that, under aging, the hybrid architectures maintain their advantage over the SPM-T. A calibration method based on learning the aging phenomenon separately and creating a new variable that is injected online to adjust the cell dynamics according to aging enables the hybrids to preserve robustness up to end of life. These results underscore the strong potential of hybrid models for onboard battery management and pave the way for future developments toward even more adaptive architectures.