Convolution Neural Operator-Based Fast Simulation Proxy Model for Physical Fields

To meet the real-time simulation needs of physical fields in the digital twin of electrical equipment, this paper proposes a fast simulation proxy model based on a convolutional neural operator (CNO). First, physical field data under different operating conditions are generated using numerical simulation software. Next, KD-tree space partitioning and adaptive point density grid interpolation are employed to homogenize the data. Finally, the homogenized data are used for feature learning via a CNO, establishing the physical field simulation proxy model. Simulation experiments on a three-phase transformer show that the proposed model achieves a Test Loss of 0.0703 with a single prediction time of 0.328 seconds. Compared to other models, the proposed model exhibits the lowest Test Loss on the test set and the best overall performance. Although its prediction time is slightly longer, it is still significantly faster than the 2.094 seconds required by numerical simulation software, greatly improving simulation efficiency while accurately capturing complex physical field variations.