Simulation Of Physical Field For Power Equipment Based On Deep Learning

Although the traditional physical field simulation of power equipment based on finite element analysis（FEA）has achieved high simulation accuracy,the simulating is very time-consuming.Thus,it is impractical to solve the real-time physical field simulation of power equipment.In this thesis,the deep neural networks（DNN）approach is developed for the physical field simulation of power equipment.The magnetic field simulation about short-circuit reactor,leakage magnetic field simulation about transformer,magnetic field simulation about non-short-circuit reactor,temperature field simulation about transformer,magnetic field and heating power density field simulation about 60 Hz high-voltage bushing were carried out,respectively.The main research contents and experimental results about the thesis are listed as below:（1）Aiming at the simulation of magnetic field about short-circuit reactor,DNN was introduced into the simulation of magnetic field for short-circuit reactor,and two deep-learning-based methods about the magnetic field simulation for short-circuit reactor were carried out.Based on coding short-circuit points,reducing data dimensionality,and constructing DNN model,the data points of middle and end for reactor were used to simulate the magnetic field of short-circuit position and the other non-short-circuit positions.Based on extracting the magnetic field coordinates of the short-circuit points in the middle and end of the reactor,clipping the magnetic field of the short-circuit points and the background,reducing data dimensionality,and constructing DNN model,only the data of the short-circuit position at the middle and end of reactor was used as the training data to simulate the magnetic field of any short-circuit positions.Compared with the data obtained by FEA,the experimental results showed that the mean absolute percentage error（MAPE）of the two models were up to 0.18 % and 0.07 %,respectively.But the simulating time with the experimental platform were up to 1.27 s and 1.65 s,respectively.（2）In view of the fact that the magnetic field of leakage transformer presents the characteristics of periodic decay similar to the sinusoidal curve,the common DNN model is difficult to simulate its magnetic field with high precision.The magnetic field data of leakage transformer was divided into two parts including wave trough data and the other data,and their magnetic field simulation based on DNN was carried out respectively.By processing wave trough data and the other data,reducing data dimensionality,and constructing DNN model,the simulation experiments with ten-fold cross-validation were carried out.The experimental results showed that the MAPE was1.68 %,but the simulating time of the approach with the experimental platform was up to 0.02 s.（3）Aiming at the practicability of DNN in the physical field simulation of other normal power equipment,experiments including the magnetic field simulation about non-short-circuit reactor,temperature field simulation about transformer,magnetic field and heating power density field simulation about 60 Hz high-voltage bushing were carried out,respectively.According to the characteristics of different physical field data of different power equipment,based on interpolating,reducing data dimensionality,and constructing DNN model,a lot of experiments were performed,respectively.The experimental results showed that the MAPE for the magnetic field simulation about non-short-circuit reactor,temperature field simulation about transformer,magnetic field and heating power density field simulation about 60 Hz high-voltage bushing were0.01%,0.01%,0.11%,0.86%,respectively,and the simulating time were 0.16 s,0.16 s,0.16 s,and 0.17 s,respectively.In summary,the experimental results showed that the simulation accuracies of the DNN model was very close to the counterparts of the finite element model,but the simulating time based on DNN model was decreased from hours to seconds compared with the simulating time based on the finite element model.