Machine Learning Models For Internal Physical Fields Prediction Of Transformers

Transformers are important primary electrical equipment in the power system.However,the current finite element analysis(FEA)method used to obtain the distribution of internal physical fields in transformers has the issue of time-consuming simulations.Therefore,achieving high-precision and fast prediction of physical fields in transformers holds important research significance and practical value.The electromagnetic field data generated by FEA simulations were taken as examples in this thesis,and machine learning methods were used to conduct research on the high-precision and rapid prediction of internal physical fields in transformers.The main research content and results of this thesis are as follows:(1)In view of the problem of time-consuming nature and high computational resource requirements of FEA simulations,which hinder fast simulations,we conducted research on the prediction of internal physical fields in transformers under normal conditions.The electromagnetic field simulation data for real transformers and their components provided by the CSG Electric Power Research Institute,as well as the magnetic field simulation data for single-phase transformer and three-phase transformer generated by COMSOL were obtained.The simulation data were preprocessed by physical field mesh interpolation,the importance ranking of input features,input features normalization,and data dimensionality reduction based on principal components analysis(PCA).Electrical parameter datasets were constructed using voltage and current as inputs,the differences in various machine learning models for electromagnetic field prediction and the adaptability of the prediction model to the magnetic field simulation data of single-phase transformer were studied.Experimental results showed that:1)ridge regression(RR)and support vector regression(SVR)were more effective in predicting the electromagnetic fields of real transformers and their components,both achieving a R-squared(R~2)of 0.99 and prediction times within 2.00 s.2)deep neural network(DNN)was the most effective in predicting the magnetic field of the single-phase transformer with four-dimensional input parameters,with a mean absolute percentage error(MAPE)of 1.37%,R~2 of 0.99,and a prediction time of 0.21 s.3)convolutional neural network(CNN)was the most effective in predicting the magnetic field of the three-phase transformer with fifteen-dimensional input parameters,with a MAPE of 0.36%,R~2 of0.99,and a prediction time of 0.37 s.4)after modifying the three-dimensional geometric model and simulation parameters of the single-phase transformer,the optimal prediction model DNN,achieved a MAPE of 4.26%,R~2 of 0.99,and a prediction time of 0.32 s.5)compared with FEA,the average speed of yielding a piece of data by the magnetic field prediction model for transformers under normal condition was at least 14.62 times and at most 141.54 times.(2)In view of the problem that the magnetic fields of transformers change temporally when connected to a sinusoidal AC power source,we conducted research on the prediction of the temporal magnetic field for the single-phase transformer.The continuous time-series magnetic field data,reduced dimensionality with PCA,were utilized as inputs to construct a temporal magnetic field dataset.The temporal convolutional network(TCN)and long short-term memory(LSTM)network were employed to study the prediction of the magnetic field distribution at the next moment in continuous time.Experimental results showed that:1)TCN was the most efficient and effective in training and predicting the temporal magnetic field,with a MAPE of 5.83%,R~2 of 0.99,and a prediction time of 0.38 s.2)compared with FEA,the average speed of generating a piece of data with the time-series magnetic field prediction model of single-phase transformer based on TCN was 8.08 times.3)the effectiveness of the electrical parameter dataset used in this thesis was verified by comparing the differences in magnetic field prediction effects for single-phase transformer between the temporal magnetic field dataset and the electrical parameter dataset.(3)In view of the problem of transformers entering abnormal states due to different excitations,resulting in irregular variations in the magnetic fields,we conducted research on the prediction of internal physical fields for transformers under abnormal conditions.Four sets of simulation data representing the magnetic fields of transformers under abnormal conditions were obtained using COMSOL.The influences of state parameters on the electrical parameters in transformers were analyzed,and the input feature dimensions for magnetic field prediction were expanded.Four models were utilized to predict the internal magnetic fields of transformers under abnormal conditions.Experimental results showed that:1)CNN was the most effective in predicting the magnetic fields for overvoltage data in three-phase transformer and direct current bias data in single-phase transformer,with MAPEs of 2.57%and 2.03%,R~2 values of 0.99,and prediction times of 0.56 s and 0.33 s,respectively.2)DNN was the most effective in predicting the magnetic fields for magnetizing inrush current data and winding deformation data in single-phase transformer,with MAPEs of 4.02%and 2.46%,R~2values of 0.99,and prediction times of 0.42 s and 0.87 s,respectively.3)compared with FEA,the average speed of yielding a piece of data by the magnetic field prediction model of transformers under four abnormal conditions was at least 5.59 times and at most 96.95times.(4)In view of the current lack of a standardized process for machine learning-based transformer electromagnetic field prediction,a preliminary online prediction platform was developed for predicting the internal physical fields of transformers in this thesis.The process involved in the transformer electromagnetic fields prediction was analyzed,and corresponding functional modules were designed for different requirements.A user interface(UI)was developed using Py Qt5.Users can import FEA simulation data into the platform,customize model parameters,train prediction models,and visualize the predicted results.