Physics-Informed Machine Learning Approach to Improve the Temperature Estimation Accuracy of Space-Resolved LPTNs

This paper focuses on physics-informed machine (PIML) learning modeling methods for space-resolved lumped parameter thermal networks (SLPTNs) of an interior permanent magnet synchronous machine (IPMSM) to improve the temperature estimation accuracy. It introduces two novel data-driven approaches to address this challenge while preserving as much physical knowledge as possible. The first approach uses neural networks (NNs) to approximate input data variations for SLPTNs, resulting in significant temperature estimation accuracy improvements. The second approach utilizes NNs for parameter identification of heat transfer coefficients (HTCs), offering more flexibility and adaptability compared to state-of-the-art methods. Both approaches show promising results, outperforming the purely physics-based model in most cases.