Learning-Based Cell Level Battery Digital Twin for Electric Vehicles

Electric vehicles rely on accurate estimation of battery states to operate safely and efficiently. Traditionally, the state estimation is pack level and based on empirical models developed to capture the dynamics of a representative battery pack and hence falls short in accounting for cell-to-cell variations. These variations become more pronounced as the cells age within a battery pack under non-homogeneous mechanical, thermal, manufacturing, and electrical conditions. It is challenging to adapt the traditional physics-based model to changing battery dynamics in real-time. To improve the state estimation at the cell level, a data-driven approach utilizing streamed data from vehicles enabled by connectivity has been shown in this paper. While traditional data-driven approaches result in large models and require large quantities of data for training, the proposed method relies on combining the underlying physics of the electrochemical model with novel data-driven modeling techniques. The developed physics-informed data-driven framework enables continuous adaptation to changing battery dynamics over its lifetime. The cell-level predictions are compared against the well-established empirical models and experimental data. Simulation results show that the proposed data-driven framework is capable of adapting to the battery dynamics based on real-world battery data with the prediction error less than 2%.