Magnetic Field Prediction And Relationship Analysis Between Data Set And Prediction Accuracy

In view of the traditional finite element-based methods generate high-precision data,they are difficult to meet the practical requirements for fast prediction of transformer magnetic fields due to their long computation time and large data volume.Therefore,it is of great research significance and practical value to carry out research on fast and highprecision prediction of transformer magnetic fields and the relationship between prediction accuracy and data sets.we toke a leakage magnetic single-phase transformer as an example to study magnetic field prediction and the relationship between prediction accuracy and data sets.The main research content and results are as follows:(1)In view of the problem that the magnetic field of single-phase leakage magnetic transformer changes nonlinear with time,resulting in difficulty to predict,the research on magnetic field prediction of single-phase leakage magnetic transformer based on electrical parameters was carried out.Firstly,the time and voltage characteristics were extracted.Then,the FEA simulation data is subjected to data cleaning,data partitioning preprocessing to obtain two different sets of sparse data and dense data based on different grid divisions.On this basis,data set partitioning and Principal Components Analysis(PCA)are conducted according to the time intervals corresponding to the magnitude of magnetic field strength,resulting in valley and non-valley data sets for the two different grid divisions.Finally,different machine learning models are constructed to predict the magnetic field of the single-phase transformer with leakage flux.Experimental results show that the deep neural network(DNN)model yields the best results for predicting the magnetic field of the single-phase transformer with leakage flux.The average mean absolute percentage errors(MAPE)for the sparse data set and dense data set are 1.57%and 1.56%,respectively.The prediction time is 0.46 s for the sparse data set and 0.61 s for the dense data set,with prediction speeds 43.48 times and 83.61 times faster than the corresponding FEA simulation speeds.(2)In view of the problem that the collected magnetic field data cannot predict the full picture of the internal magnetic field of single-phase leakage magnetic transformer due to the small number of magnetic field sensors deployed at present,the research on magnetic field prediction of single-phase leakage magnetic transformer based on sampling points based on the grid data points obtained by FEA was conducted.Firstly,sparse and dense data sets were obtained based on FEA simulation data with different grid divisions,and outlier data were cleaned for each set.Then,different sampling regions were set according to the variation of magnetic flux density at the grid points on the iron core and coils.The grid points in these regions were sampled and subjected to PCA and data normalization preprocessing.Finally,different machine learning models were constructed to predict the magnetic field of the single-phase transformer with leakage flux.Experimental results showed the following: 1)The DNN model yielded the best prediction results.In sampling region 1 on the iron core and coils,the MAPE for the prediction results of the sparse data set with only 4 points was 8.26%,and the MAPE for the prediction results of the dense data set with only 17 points was 26.83%.2)On the sparse data set,with 4 sampled points distributed in sampling region 2 on the edges of the iron core and in sampling region 3 on the surface and edges of the iron core,the MAPE for the prediction results in sampling region 2 and sampling region 3 were 5.07%and 6.95%,respectively.On the dense data set,with 17 sampled points,the MAPE for the prediction results in sampling region 2 and sampling region 3 were 10.36% and15.89%,respectively.Therefore,when the number of sampled points was the same,the spatial distribution of the sampled points had a certain impact on the accuracy of magnetic field prediction,and the sampling region 2 with sampled points distributed on the edges of the iron core had the highest prediction accuracy.(3)In view of the issue of redundancy in the data generated by FEA simulations,the study on the relationship between the data set size and prediction accuracy based on predicted data was conducted.Firstly,based on the previous prediction experimental results,data set A(1,078 samples)and data set C(2,500 samples)with 1,305 grid points,as well as data set B(30,934 grid points)and data set D(1,559 grid points)with 420 samples,were obtained.By reducing the number of data samples,the relationship between the number of data samples and the prediction accuracy under a fixed number of grid points was investigated.Finally,based on the established relationship between the number of data samples and the prediction accuracy,the influence of the number of grid points and data samples on magnetic field prediction accuracy was studied.Experimental results show the following: 1)With a fixed number of grid points(30,934),when data set B has a maximum of 420 data samples,the highest prediction accuracy was 98.34%.With a fixed number of data samples(420),when the number of grid points is maximized to1,559,the highest prediction accuracy was 96.47%.2)Using the prediction accuracy function established with data set B,the prediction of the number of data samples in data set D yielded a MAPE of 8.76%,while using the prediction accuracy function established with data set C,the prediction of the number of data samples in data set A yielded a MAPE of 20.72%.Therefore,the DNN model can be used to quickly and accurately predict the magnetic field of single-phase transformers with leakage flux.When computing the model data to be exported in FEA simulations,the number of data samples should be considered first,followed by the number of grid points,and finally the distribution of grid points.Given a specific magnetic field prediction accuracy,the number of data samples,the number of coordinate grid points,and the spatial distribution of coordinate grid points can be determined accordingly.