Hybrid modelling of Lithium-ion batteries: Proof of concept for application of Physics-informed Neural Network in Electrochemical Battery Modelling

Abstract

Accurately forecasting lithium-ion batteries‘ lifetime and degradation mechanisms are crucial for their optimization and management. An accurate state of health estimation is essential for ensuring safety and preventing latent failures. However, complex and dynamic cell parameters and wide variations in usage conditions make the state of health prediction challenging. Physics-based models need a tradeoff between accuracy and complexity due to vast parameter requirements. On the other hand, machine-learning models require large training datasets and may fail when generalized to unseen scenarios. This indicates the need for a novel approach that combines the two approaches and can tap into their advantages.

This work aims to integrate the physics-based battery model and machine learning model to leverage their respective strengths to calculate the state of health of lithium-ion cells. This is achieved by applying the deep learning framework called Physics-Informed Neural Networks (PINN) in electrochemical battery modeling. Exploring PINN for battery modeling is essential for advancing the state-of-the-art in battery research and development and enabling the deployment of safer, more efficient, and sustainable energy storage systems. By integrating the partial differential equations of a single particle model into a neural network, this approach indicates how such a combination of models enhances one’s ability to predict the health of batteries.

The main contribution of this work is the training of a neural network on the basis of Fick’s law of diffusion, normally used for a single particle model. The results indicate that PINN can accurately estimate a battery’s state of charge and state of health even with limited training data. Compared to a purely physics-based model, PINN is less complex while still incorporating the laws of physics into the training process, resulting in adequate predictions, even for unseen situations. While the accuracy of this approach cannot be claimed to always outperform existing state estimation approaches, this study provides a promising starting point for the systematic integration of partial differential equations of complex P2D battery models into neural networks.