System Identification of physical chemical model of Li-ion cell using deep learing

Lithium-ion batteries (LIBs) are essential components of electromobility, where safe, long-lasting, and fast-charging performance is crucial. However, during fast charging and extended use, internal degradation phenomena such as lithium plating and loss of active material which lead to a decline in capacity and increased safety risks. Accurate state estimation for battery management relies on high-fidelity physical chemical models with well-calibrated parameters. As the battery ages, deviations between nominal and true parameter values increase, necessitating continuous model adaptation through parameter identification. This thesis presents a deep learning-based framework for identifying internal electrochemical parameters of a detailed battery model, using only external measurements specifically, terminal voltage and current time series data. These quantities are commonly available from vehicle sensor systems and can serve as non-invasive inputs for tracking internal cell behaviour in real life. A global sensitivity analysis was performed using synthetic data generated from a Doyle–Fuller–Newman (DFN) model to determine which parameters significantly influence the terminal voltage and are theoretically identifiable. Based on this analysis and R-square fit of the deep learning model, a subset of sensitive parameters was selected for training a Transformer-based deep learning model. The model learns the complex, nonlinear mapping from a very short time-series data (300 s) of voltage and current time series data. The results show that deep learning provides a viable and flexible method for dynamic parameter identification using only these external measurements. This approach enables improved model adaptability and accurate state estimation over the battery’s lifecycle, paving the way for more robust and data-driven battery management systems.