Physics Informed Neural Network for thermal modeling of an Electric Motor

Abstract

Abstract Artificial intelligence and machine learning are becoming increasingly significant, and the need to investigate the potential for different areas arises. This project investigated the potential of utilizing data-driven techniques for the thermal model of the motor drive system at Volvo GTT. The aim was to incorporate the already-known physics of the system into the data-driven models through different constraints to achieve higher performance. The physics-informed neural network was built using the PyTorch framework, and the project tested multiple types of networks and hyperparameters. Another model was created using the Nerve framework developed by Volvo. The Nerve model only took one week to develop, which is significantly shorter than the four months it took to develop the PINN model. The Nerve model underwent training on a large amount of data, but its ability to accurately predict the output of the thermal model was inconsistent. It was shown that the self-developed data-driven gated recurrent unit model can model the system effectively. Further, the data-driven model with physical constraints performed better than the models without incorporating previously known physics. The best performance acquired in this investigation showed an 80 % reduction in MAE loss for power loss estimations and a 53 % reduction in winding temperature estimations when incorporating physics compared to a purely data driven model. Based on the results, it is clear that PINNs have great potential for applications where there is a shortage of available data. In such cases, traditional data-driven models may not be able to accurately capture the dynamics of the thermal model as effectively as the PINNs