Physics-Informed NN for Improving Electric Vehicles Lithium-Ion Battery State-of-Charge Estimation Robustness

• Published in: IEEE International Conference on Smart Energy Grid Engineering (Online) pp. 245 - 250
• Main Authors: Baraean, Abdullah, Kassas, Mahmoud, Abido, Mohammed
• Format: Conference Proceeding
• Language: English
• Published: IEEE 18-08-2024
• Subjects: Accuracy; Adaptation models; battery SoC estimation; Computer architecture; Estimation; Lithium-ion batteries; Measurement errors; neural networks prediction; Noise; physics-informed neural networks; Predictive models; Uncertainty; Vehicle dynamics
• ISSN: 2575-2693
• DOI: 10.1109/SEGE62220.2024.10739432

Electric vehicles (EVs) are often powered by lithium-ion batteries. To ensure the reliability of EVs, it is essential to model and predict the remaining useful life of these batteries. Accurate estimation of the state of charge (SoC) is a critical aspect for effectively managing and ensuring operational reliability of battery systems. Building principled accurate models is challenging due to the complex electrochemistry that governs battery operation. This research paper presents a novel approach that utilizes Physics-Informed Neural Networks (PINNs) to enhance the accuracy of SoC estimation. PINNs combine the data-driven learning capabilities of neural networks with the fundamental physical laws that govern battery dynamics. By incorporating both aspects, PINNs offer a novel way to improve the accuracy of SoC estimation in lithium-ion batteries and enable better management of battery systems in the context of electric vehicles. We demonstrate how incorporating physical constraints into the learning process enhances model prediction performance and ensures physically plausible solutions. The approach is validated using data publicly available through the Mendeley Data website by McMaster University in Hamilton, Ontario, Canada. Results showed that our proposed robust approach can successfully overcome the measurement's errors and noise. Moreover, the model can obtain an SoC estimation accuracy of less than 2.85% root mean squared error (RMSE).