A Fast Calculation Method of 3D Temperature Field of Oil-Immersed Transformer Based on Point Cloud U-Net++ Neural Network

To address the challenges in real-time 3D temperature field analysis for intelligent power systems, we propose a fast calculation method based on point cloud U-net++ neural network. Taking a 35 kV oil-immersed transformer as an example, initially, we input key temperature-influencing factors into our algorithm. These input features are randomly combined in a limited range according to a specific step. The sets of 3D temperature are computed by Fluent on the Jinan Shanhe supercomputing platform. And the three-dimensional mathematical model is then converted into point clouds. Finally, we determined the optimal hyperparameters and proceeded with parameter training, evaluation and debugging. The results demonstrate that the method proposed can reduce single calculation time to 0.04 s with the vast majority of the error in the region of 0K or so, significantly improving the efficiency of the calculation. Meanwhile, the U-net++ neural network also achieves significantly higher accuracy than the U-net network. To validate the algorithm's effectiveness, we establish a platform for assessing the temperature increase. The experimental results indicate that the temperature rise trend from U-net++ neural network calculations aligns closely with the experimental data, and the temperature difference is within only 4K.