A short-circuit reactor magnetic field simulation method based on finite elements and deep learning

Abstract

The application discloses a short-circuit reactor magnetic field simulation method based on finite elements and deep learning, and relates to the field of short-circuit reactor magnetic field simulation. First, the obtained magnetic field data is divided into a training data set and a test data set; the magnetic field data short-circuit position is one-hot encoded and then one-dimensional normalized current is added as input; principal component analysis dimension reduction operation is performed on the training data set, and the obtained characteristic value is used as the output of training, and the mean value and principal component are saved; the mean value and principal component saved by the training data PCA dimension reduction are used for PCA dimension reduction of the test data, and the obtained characteristic value is used as the output of testing; the training data is used as the training of a deep learning model, and the test data is used for testing; prediction data is generated by using the trained deep learning model, and the mean value and principal component saved when the PCA transformation of the training data is performed are used for inverse transformation of the prediction data, so that the magnetic field simulation result of the reactor under normal or short-circuit conditions is obtained.

Description

Technical Field

[0001] This application relates to the technical field of reactor magnetic field simulation methods, and in particular to simulation methods for the magnetic fields of normal reactors and short-circuit reactors. Background Technology

[0002] Dry-type air-core reactors have the advantages of being dry and oil-free, having a simple structure, being easy to operate and maintain, having high insulation strength, high mechanical strength, and high reliability, and are therefore widely used in power systems.

[0003] Series reactors are widely used in power grid equipment, and the fire problems caused by their equipment failures cannot be ignored. Fires in series reactors are generally caused by short circuits. Therefore, it is particularly important to conduct magnetic field simulation on short-circuit reactors to study the magnetic field situation when a short circuit occurs. This is also one of the core aspects of digital power grids—equipment digitization. When equipment fails, digitization can provide its real-time status information.

[0004] Currently, the magnetic field simulation of reactors uses the finite element method (FEM). While FEM offers high accuracy, its computation time increases with model complexity and problem difficulty, resulting in slow magnetic field simulation results. Furthermore, different models are required to simulate reactors of the same model experiencing short circuits at different locations. Therefore, it is difficult to meet the requirements of high real-time production environments. Summary of the Invention

[0005] This application provides a method for simulating the magnetic field of a short-circuit reactor based on finite element method and deep learning, in order to solve the problems of excessive time consumption of finite element analysis and the need for different models to simulate different short-circuit points in the existing reactor magnetic field simulation.

[0006] The technical solution adopted in this application is as follows:

[0007] A method for simulating the magnetic field of a short-circuit reactor based on finite element method and deep learning mainly includes the following steps:

[0008] Step 1: Obtain magnetic field data of the same type of reactor under normal conditions and under short circuit conditions at different locations, as well as the basic parameters of the reactor;

[0009] Step 2: Establish a reference coordinate system and put the normal magnetic field data of the reactor into the established reference coordinate system to obtain the reference magnetic field data;

[0010] Step 3: Establish sample magnetic field data;

[0011] When a short circuit occurs at a certain location, the magnetic field data within a spherical range centered on the coordinates of that location is replaced with the reference magnetic field data to obtain a sample magnetic field data; this method is then used to process the magnetic field data obtained in step 1 when a short circuit occurs at all locations to obtain a set of sample magnetic field data.

[0012] Step 4: Divide the sample magnetic field data set into training and test sets.

[0013] Step 5: Perform one-hot encoding on the short-circuit locations in the sample magnetic field data set and normalize the basic parameters; perform PCA dimensionality reduction on the sample magnetic field data and extract the feature values ​​of the magnetic field data.

[0014] Step 7: Use the one-hot encoded and normalized basic parameters as the input to the deep learning network, and use the magnetic field data feature values ​​as the output of the deep learning network to train and test a deep learning network.

[0015] Step 8: In practice, the coordinates of the short-circuit point to be simulated are one-hot encoded, and the corresponding normalized basic parameters are input into the trained and tested deep learning network to obtain the corresponding magnetic field feature data, and then the magnetic field data is inverted.

[0016] Furthermore, the basic parameter in step 1 is current.

[0017] Furthermore, in step 5, PCA is used to reduce the dimensionality and extract the mean and principal components of the magnetic field data as the feature values ​​of the magnetic field data.

[0018] Furthermore, the deep model includes: an input layer, multiple hidden layers, and an output layer, wherein the number of nodes in the input layer is consistent with the dimension of the data after one-hot encoding of the short-circuit position and current normalization; the number of nodes in the output layer is consistent with the number of magnetic field data features extracted in step 5.

[0019] A method for determining the short-circuit location of a reactor based on finite element method and deep learning mainly includes the following steps:

[0020] Step 1: Obtain magnetic field data of the same type of reactor under normal conditions and under short circuit conditions at different locations, as well as the basic parameters of the reactor;

[0021] Step 2: Establish a reference coordinate system and put the normal magnetic field data of the reactor into the established reference coordinate system to obtain the reference magnetic field data;

[0022] Step 3: Establish sample magnetic field data;

[0023] When a short circuit occurs at a certain location, the magnetic field data within a spherical range centered on the coordinates of that location is replaced with the reference magnetic field data to obtain a sample magnetic field data; this method is then used to process the magnetic field data obtained in step 1 when a short circuit occurs at all locations to obtain a set of sample magnetic field data.

[0024] Step 4: Divide the sample magnetic field data set into training and test sets.

[0025] Step 5: Perform one-hot encoding on the short-circuit locations in the sample magnetic field data set and normalize the basic parameters; perform PCA dimensionality reduction on the sample magnetic field data and extract the feature values ​​of the magnetic field data.

[0026] Step 7: Use the normalized basic parameters and magnetic field data feature values ​​as input to the deep learning network, and use the one-hot encoding as the output of the deep learning network to train and test a deep learning network.

[0027] Step 8: In practice, magnetic field data around the faulty reactor is collected, and PCA dimensionality reduction is performed to extract the feature values ​​of the magnetic field data. These feature values ​​are then added to the normalized basic parameters of the reactor and input into the deep learning network trained in Step 7 to obtain the one-hot encoding of the short-circuit point location. Finally, the short-circuit point location is retrieved.

[0028] The beneficial effects of adopting the technical solution of this application are as follows:

[0029] This application discloses a method for analyzing the magnetic field of a short-circuit reactor based on finite element method and deep learning. The method includes: acquiring magnetic field data and basic reactor parameters generated by finite element analysis of the same type of reactor under normal conditions and at different short-circuit locations; processing the reference coordinates; interpolating all data onto the aforementioned coordinates to obtain magnetic field data with a consistent number of coordinate points; dividing the magnetic field data into training and testing datasets; performing One-Hot encoding on the short-circuit location of the magnetic field data and adding one-dimensional normalized current as input; performing PCA dimensionality reduction on the training dataset, using the resulting feature values ​​as the training output, while preserving the mean and principal components; using the mean and principal components preserved during PCA dimensionality reduction of the training data to perform PCA dimensionality reduction on the testing data, using the resulting feature values ​​as the test output; using the training data for training a deep learning model and the testing data for testing; establishing the deep learning model; generating prediction data using the trained deep learning model, and performing an inverse transformation on the prediction data using the mean and principal components preserved during the previous PCA transformation to obtain the magnetic field simulation results of the reactor under normal or short-circuit conditions. Compared to directly using finite element method (FEM) for reactor magnetic field simulation analysis, this application offers significantly faster simulation speed, enables real-time prediction of the reactor's magnetic field, and provides simulation data with very small errors compared to FEM simulation data. Furthermore, this application requires only one model to simulate the magnetic field of the same type of reactor under normal conditions and under short-circuit conditions at different locations, whereas FEM simulations require different models for such magnetic field simulations. Attached Figure Description

[0030] Figure 1 A flowchart illustrating a method for simulating the magnetic field of a short-circuit reactor based on finite element method and deep learning, provided for embodiments of this application;

[0031] Figure 2 A flowchart illustrating the coordinate processing provided in this application embodiment. Detailed Implementation

[0032] The embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described below do not represent all embodiments consistent with this application. They are merely examples of systems and methods consistent with some aspects of this application as detailed in the claims.

[0033] See Figure 1 The flowchart illustrates a method for simulating the magnetic field of a short-circuit reactor based on finite element method and deep learning, as provided in this application embodiment; see also... Figure 2 A flowchart illustrating the coordinate processing provided in this application embodiment.

[0034] Example 1

[0035] The magnetic field simulation method for short-circuit reactors based on finite element method and deep learning provided in this application mainly includes the following steps:

[0036] Step 1: Obtain the magnetic field data and basic parameters of the reactor generated by finite element analysis when the same type of reactor is short-circuited under normal conditions and at different locations.

[0037] The data includes magnetic field data of the same model of reactor under different currents during normal operation, generated by finite element analysis software simulation, and magnetic field data under different currents when a short circuit occurs at different locations. The basic parameters of the reactor include the number of coil turns, conductor cross-sectional area, and coil dimensions.

[0038] Step 2: Reference Coordinate Processing: Select a common coordinate system as the reference coordinate system. Then, crop the coordinates of the short-circuit magnetic field data corresponding to the short-circuit points, retaining only the coordinates of the short-circuit points. Next, crop a block of the same size from the reference coordinate system.

[0039] The coordinates of the cut-out short-circuit points of the same size are attached to the reference coordinates. This process is repeated for each short-circuit point to obtain a processed reference coordinate system.

[0040] Step 3: Interpolate all the data to the processed reference coordinates.

[0041] Step 4: Divide the processed data into training dataset and test dataset according to a certain ratio.

[0042] Step 5: Input data processing: Based on the characteristics of the reactor turns, the short-circuit location (including the number of turns, middle and end) and current of the original data are One-Hot encoded and normalized respectively.

[0043] Step 6: Perform PCA dimensionality reduction on the training dataset. The dimensionality-reduced data will be used as the data for training the model, while preserving the mean and principal components.

[0044] Step 7: Use the mean and principal components retained during PCA dimensionality reduction of the training dataset to perform dimensionality reduction on the test dataset, and use them as the output data when testing the model.

[0045] Step 8: Building the deep learning model. The number of nodes in the input layer is consistent with the dimension after One-Hot encoding of the short-circuit location and current normalization in Step 4. The number of nodes in the output layer is consistent with the dimension preserved by PCA dimensionality reduction in Step 5. The number of hidden layers and the number of nodes in each layer are adjusted based on experimental results.

[0046] Step 9: Determine the model training parameters, including the learning rate, activation function, optimization function, and loss function, and train the model using the processed training set data. These parameters should be adjusted as needed based on the model's training performance.

[0047] Step 10: Input the processed test dataset into the trained deep learning model to obtain the test results. Perform PCA inverse transformation on the results using the training data and the saved mean and principal components to obtain the predicted magnetic field square data. Calculate the parameters of the model's performance on the test set.

[0048] Step 11: Encode the input data as in step five, input it into the model, obtain the model output, and then perform inverse PCA transformation on the model output using the mean and principal components saved during PCA dimensionality reduction of the training data to obtain the corresponding magnetic field simulation data.

[0049] Example 2

[0050] The present application will now be described in detail with reference to specific embodiments.

[0051] This application uses a coil with 284.792 turns and a conductor cross-sectional area of ​​12.464269 mm². 2 Taking a reactor as an example, the explanation includes the following steps:

[0052] Step 1: Obtain the magnetic field data and basic parameters of the reactor generated by finite element analysis under normal conditions and when a short circuit occurs at different locations.

[0053] The reactor model to be analyzed was created in the finite element analysis software COMSOL, including models under normal conditions and models under short-circuit conditions. The reactor is a hollow cylinder with a total of 41 coil layers, exhibiting good symmetry; therefore, one cross-section can be selected. Short-circuit models were created using upper-end inter-turn short circuits and middle-end inter-turn short circuits, with a total of 41 turns. Therefore, there are 41 models each for upper-end and middle-end short circuits. Adding the normal condition model, a total of 83 models were created in COMSOL. Data from COMSOL was obtained for each model, with 51 sets of different currents ranging from 100A to 1100A at 20A intervals, resulting in 4233 data points across the 83 models.

[0054] Step 2: Processing the reference coordinates.

[0055] A total of 91,462 coordinate points were selected from the model under normal conditions. Then, the coordinates of the short-circuit magnetic field data corresponding to the short-circuit points were cropped, retaining the coordinates of the short-circuit points. Then, a block of the same size was cropped from the reference coordinates, and the coordinates of the cropped short-circuit points of the same size were attached to the reference coordinates. This process was repeated for each short-circuit point, resulting in a processed reference coordinate system with a total of 99,325 coordinate points.

[0056] Step 3: Interpolate all the data to the processed reference coordinates. This unifies the coordinates of all 4233 magnetic field data points, resulting in 99325 coordinate points.

[0057] Step 4: Divide the interpolated magnetic field data into two datasets with a ratio of 8:2, i.e., 3386 data points for the training dataset and 847 data points for the test dataset.

[0058] Step 5: Input data processing: Based on the characteristics of the reactor turns, the short-circuit location (including the number of turns, middle and end) and current of the original data are One-Hot encoded and normalized respectively.

[0059] After one-hot encoding the short-circuit location, a 22-dimensional dataset is obtained. Adding the one-dimensional data from the current normalization, the total input data is 23-dimensional. Both the training and test datasets are processed using this method. This ensures that the input dimension of the subsequent deep learning model is 23 dimensions.

[0060] Step 6: Dimensionality reduction of the training dataset using PCA:

[0061] Training dataset merging: There are a total of 3386 training data entries, each with 99325 coordinate points. Each data entry occupies one row. All data entries are concatenated into a matrix to obtain a 3386×99325 training dataset matrix.

[0062] The resulting matrix is ​​then subjected to column-wise PCA dimensionality reduction, retaining the first 82 eigenvectors. This transforms the dataset matrix into a 3386×82 dimensionality-reduced matrix. The mean and principal components from the PCA dimensionality reduction process are also preserved. Thus, the model's output is determined to be 82-dimensional.

[0063] Step 7: PCA dimensionality reduction of the test dataset:

[0064] Similar to step 6, the test dataset is also row-concatenated to obtain an 847×99325 matrix. The dimensionality is then reduced using the mean and principal components retained during PCA of the training dataset, resulting in an 847×82-dimensional matrix.

[0065] Step 8: Build the deep learning model. After repeated experiments, the final model structure used was 23×40×60×120×120×120×120×82. That is, the number of nodes in the input layer and the output layer are 23 and 82 respectively, which was determined by the dimension of the input data and the dimension retained after PCA dimensionality reduction of the dataset. The number of hidden layers and the number of neurons in each hidden layer can be freely adjusted. Finally, the experiment used 6 hidden layers, with 40, 60, 120, 120, 120, and 120 neurons in each layer, respectively.

[0066] Step 9: Determine the model's training parameters. After repeated training and testing, the learning rate used during model training is 1×10⁻⁶. -4 The activation function for each layer is the Rectified Linear Unit (ReLU), the optimization algorithm is the Adaptive Moment Estimation (Adam), and the loss function is the Mean-Square Error (MSE). Each training batch contains 3,500 data points, and the training is conducted a total of 800,000 times.

[0067] Step 10: Model Performance. The final model has an MSE of 1×10⁻⁶ during training. -5 The performance results on the test set are shown in Table 1:

[0068] Table 1. Performance of deep learning models on the test set

[0069]

[0070] Wherein, MSE is the calculation result when the model outputs directly, and the Mean Absolute Percentage Error (MAPE) is the calculation result after performing PCA inverse transformation on the result and comparing it with the true value.

[0071] Step 11: Encode the input data as in step five, input it into the model, obtain the model output, and then perform inverse PCA transformation on the model output using the mean and principal components saved during PCA dimensionality reduction of the training data to obtain the corresponding magnetic field simulation data.

[0072] Finally, the obtained model was used to predict 17 data points. The prediction time of the model was 0.21s, and the time from output to final data visualization was 11.84s.

[0073] In Table 1, MSE refers to the expected value of the squared difference between the true and predicted values. It equals 0 when the predicted values ​​perfectly match the true values, representing a perfect model; the larger the error, the larger this value. It's important to note that the true values ​​here refer to the test data after PCA dimensionality reduction, i.e., the 847×82 matrix mentioned above, while the predicted values ​​are predictions of this portion. MAPE refers to the percentage of the true value that the predicted value represents. A MAPE of 0% indicates a perfect model, while a MAPE greater than 100% indicates a poor model. It's also important to note that the true values ​​here are the original data before PCA dimensionality reduction, i.e., the 847×99325 matrix mentioned above, while the predicted values ​​refer to the data output from the model after inverse PCA transformation, resulting in a dimension of 847×99325. Both MSE and MAPE indicate that the difference between the predicted and true values ​​is very small.

[0074] Therefore, the magnetic field simulation method for short-circuit reactors based on finite element method and deep learning in this application has high accuracy in both simulation time and simulation accuracy. Compared with simple finite element analysis, the method in this application can predict the magnetic field of reactors under normal conditions and under short-circuit conditions at different locations in a short time using only a single model, which can meet the purpose of high real-time prediction of magnetic field data in actual production.

Claims

1. A method for simulating the magnetic field of a short-circuit reactor based on finite element method and deep learning, mainly including the following steps: Step 1: Obtain magnetic field data of the same model of reactor under normal conditions and under short circuit conditions at different locations, as well as the basic parameters of the reactor; the basic parameter is current; Step 2: Establish a reference coordinate system and put the normal magnetic field data of the reactor into the established reference coordinate system to obtain the reference magnetic field data; Step 3: Establish sample magnetic field data; When a short circuit occurs at a certain location, the magnetic field data within a spherical range centered on the coordinates of that location is replaced with the reference magnetic field data to obtain a sample magnetic field data; this method is then used to process the magnetic field data obtained in step 1 when a short circuit occurs at all locations to obtain a set of sample magnetic field data. Step 4: Divide the sample magnetic field data set into training and test sets. Step 5: Perform one-hot encoding on the short-circuit locations in the sample magnetic field data set and normalize the basic parameters; perform PCA dimensionality reduction on the sample magnetic field data and extract the feature values ​​of the magnetic field data. PCA was used to reduce the dimensionality of the magnetic field data and extract the mean and principal components as the feature values ​​of the magnetic field data. Step 7: Use the one-hot encoded and normalized basic parameters as the input to the deep learning network, and use the magnetic field data feature values ​​as the output of the deep learning network to train and test a deep learning network. The deep learning network includes: an input layer, multiple hidden layers, and an output layer. The number of nodes in the input layer is consistent with the dimension of the data after one-hot encoding of the short-circuit position and current normalization. The number of nodes in the output layer is consistent with the number of magnetic field data features extracted in step 5. Step 8: In practice, the coordinates of the short-circuit point to be simulated are encoded using one-hot encoding, and the corresponding normalized basic parameters are input into the trained and tested deep learning network to obtain the corresponding magnetic field feature data, and then the magnetic field data is inverted.

2. A method for determining the short-circuit location of a reactor based on finite element method and deep learning, mainly including the following steps: Step 1: Obtain magnetic field data of the same model of reactor under normal conditions and under short circuit conditions at different locations, as well as the basic parameters of the reactor; the basic parameter is current; Step 2: Establish a reference coordinate system and put the normal magnetic field data of the reactor into the established reference coordinate system to obtain the reference magnetic field data; Step 3: Establish sample magnetic field data; When a short circuit occurs at a certain location, the magnetic field data within a spherical range centered on the coordinates of that location is replaced with the reference magnetic field data to obtain a sample magnetic field data; this method is then used to process the magnetic field data obtained in step 1 when a short circuit occurs at all locations to obtain a set of sample magnetic field data. Step 4: Divide the sample magnetic field data set into training and test sets. Step 5: Perform one-hot encoding on the short-circuit locations in the sample magnetic field data set and normalize the basic parameters; perform PCA dimensionality reduction on the sample magnetic field data and extract the feature values ​​of the magnetic field data. PCA was used to reduce the dimensionality of the magnetic field data and extract the mean and principal components as the feature values ​​of the magnetic field data. Step 7: Use the normalized basic parameters and magnetic field data feature values ​​as input to the deep learning network, and use the one-hot encoding as the output of the deep learning network to train and test a deep learning network. The deep learning network includes: an input layer, multiple hidden layers, and an output layer. The number of nodes in the input layer is consistent with the dimension of the data after one-hot encoding of the short-circuit position and current normalization. The number of nodes in the output layer is consistent with the number of magnetic field data features extracted in step 5. Step 8: In practice, magnetic field data around the faulty reactor is collected, PCA dimensionality reduction is performed, magnetic field data feature values ​​are extracted, and the normalized basic parameters of the reactor are added to the deep learning network trained in step 7 to obtain the one-hot encoding of the short circuit point location. Finally, the short circuit point location is inverted.

Images