A method for predicting insulator flashover voltage in SF6 under nanosecond pulses

Abstract

The present invention proposes a method for predicting the flashover voltage of insulators in SF6 under nanosecond pulses, which solves the technical problem of the difficulty in balancing accuracy and generalization in existing flashover voltage prediction methods. The method uses empirical formulas to generate a wide range of input and output data and performs regularization on the data; then the above data is input into an MLP neural network for pre-training; then the input and output parameters are extracted from the results of artificial simulation experiments and regularized; the pre-trained MLP neural network is used for incremental learning, and finally the training set and test set errors obtained by the model are verified to obtain an MLP regression prediction model. By inputting any set of environmental parameters into the model, the corresponding insulator flashover voltage can be obtained. The present invention requires a small amount of experiments, has high prediction accuracy, and can also realize extrapolated prediction of flashover voltage outside the range of experimental conditions, and has strong versatility.

Description

Technical Field

[0001] The present invention relates to an insulation design method for a pulse power insulator, and in particular to a prediction method for an insulator flashover voltage in SF6 under nanosecond pulses. Background Art

[0002] Insulator surface flashover under pulse voltage poses a significant threat to the normal operation of high-voltage equipment. Therefore, insulator flashover voltage is a key consideration in the insulation design of pulse power equipment. However, due to the numerous factors that influence insulator surface flashover, accurately predicting insulator flashover voltage and ensuring a balanced insulation design that balances reliability and cost-effectiveness remains a major challenge.

[0003] Currently, methods for predicting the flashover voltage of insulators in SF6 gas under pulse voltages primarily rely on summarizing, analyzing, and fitting a large number of experimental results to derive empirical formulas. This method requires significant human and material resources for experimentation, and the empirical formulas often have low accuracy.

[0004] In recent years, artificial intelligence technologies represented by neural networks have gradually been applied in flashover voltage prediction. However, since the training data set of neural networks still comes from experimental data and is limited by their limited generalization ability, this method is difficult to achieve extrapolation prediction beyond the data range. Summary of the Invention

[0005] In order to solve the technical problem that it is difficult to strike a balance between accuracy and generalization in existing flashover voltage prediction methods, the present invention proposes a prediction method for the flashover voltage of insulators in SF6 under nanosecond pulses. The method is a "two-step" neural network flashover voltage prediction method based on pre-training of computational data and incremental learning based on experimental data.

[0006] To achieve the above object, the technical solution of the present invention is:

[0007] A method for predicting the flashover voltage of an insulator in SF6 under nanosecond pulses is characterized in that it comprises the following steps:

[0008] S1, using multiple sets of SF6 environmental parameters under nanosecond pulses as a first input set, and calculating a first output set using a flashover voltage calculation formula, wherein the first output set is the flashover voltage of the insulator under the multiple sets of environmental parameters; the multiple sets of environmental parameters are parameter sets consisting of different SF6 gas pressures, voltage polarities, voltage effective action times, and insulator creepage distances;

[0009] S2, performing regularization processing on the first input set and the first output set in step S1 to generate a first data set;

[0010] S3, inputting the first data set generated in step S2 into the MLP neural network for training to obtain a training model;

[0011] S4, extracting the input parameter set and the output parameter set from the artificial simulation experimental data, performing regularization processing on the data, generating a second data set, and dividing the second data set into a training set and a test set; the training set includes the second input set and the second output set; the test set includes the third input set and the third output set; the input parameter set includes the set of SF6 gas pressure, voltage polarity, voltage effective action time, and insulator surface distance parameters used in the experiment; the output parameter set is the set of insulator flashover voltages obtained through the experiment;

[0012] S5, inputting the training set in step S4 into the training model in step S3 for incremental learning to obtain a training output set and an optimized training model;

[0013] S6, verify the training model optimized in step S5:

[0014] S6.1 calculates the error between the training output set in step S5 and the second output set in step S4;

[0015] S6.2 Input the test set in step S4 into the training model optimized in step S5 for training, obtain a test output set, and calculate the error between the test output set and the third output set;

[0016] S6.3 If the errors calculated in S6.1 and S6.2 are both ≤5%, then the MLP regression prediction model is obtained and step S7 is executed; otherwise, the training is continued in step S3.

[0017] S7, by inputting any set of environmental parameters into the MLP regression prediction model, the insulator flashover voltage in SF6 under nanosecond pulses can be predicted.

[0018] Furthermore, in step S1, the flashover voltage calculation formula is: E br =k ± p 0.4 (dt eff ) -1 / 6

[0019] Among them, E br is the breakdown field strength, i.e. the flashover voltage of the insulator, in kV / cm; k is the polarity coefficient, when in SF6 gas, k is taken when the voltage polarity is positive + is a constant 44, and k is used when the voltage polarity is negative - is a constant 72; p is the SF6 gas pressure in atm; t effis the effective time, which refers to the time when the pulse voltage amplitude exceeds 89% of the peak voltage, in μs; d is the distance along the insulator surface, in cm.

[0020] Furthermore, in step S3, before training, the Optuna framework can be used to optimize the hyperparameters of the MLP neural network.

[0021] Furthermore, in step S4, 80% of the data in the second data set is used as a training set, and 20% of the data is used as a test set.

[0022] Furthermore, in step S4, 70% of the data in the second data set is used as a training set, and 30% of the data is used as a test set.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] 1. The present invention uses computational data to pre-train the neural network, and combines it with an incremental learning strategy to use experimental data under a small range of conditions to predict the flashover voltage under a large range of conditions. This overcomes the problem of inaccurate extrapolation prediction in existing prediction methods and achieves the accuracy and generalization of the flashover voltage prediction method.

[0025] 2. The present invention uses computational data to pre-train the neural network, which can generate a large amount of data within the required prediction conditions, thereby saving a lot of manpower and time.

[0026] 3. The present invention uses experimental data to perform incremental learning on the neural network, and can use the experimental data to correct the empirical formula to further enhance the accuracy of the prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 The present invention provides a flow chart for building a prediction model for an embodiment of a method for predicting the flashover voltage of an insulator in SF6 under nanosecond pulses. DETAILED DESCRIPTION

[0028] The technical solution of the present invention is described in detail below with reference to the accompanying drawings.

[0029] Combine Figure 1 As shown, the present invention provides a method for predicting the flashover voltage of an insulator in SF6 under nanosecond pulses, comprising the following steps:

[0030] S1 uses multiple sets of SF6 environmental parameters under nanosecond pulses as a first input set and calculates a first output set using a flashover voltage calculation formula. The first output set is the flashover voltage of an insulator in SF6 under the nanosecond pulse. The multiple sets of environmental parameters are parameter sets consisting of different SF6 gas pressures, voltage polarities, voltage effective durations, and average insulator creepage distances.

[0031] The present embodiment selects JCMartin's flashover voltage empirical formula, namely: E br =k ± p 0.4 (dt eff ) -1 / 6 , in the formula, E br is the breakdown field strength, i.e. the flashover voltage of the insulator, in kV / cm; k is the polarity coefficient, when in SF6 gas, k is taken when the voltage polarity is positive + is a constant 44, and k is used when the voltage polarity is negative - is a constant 72; p is the SF6 gas pressure in atm; t eff The effective time, expressed in μs, refers to the time during which the pulse voltage amplitude exceeds 89% of the peak voltage; d is the distance along the insulator surface, expressed in cm. Other empirical formulas for flashover voltage can also be used to calculate the flashover voltage. Specifically, different SF6 gas pressures, voltage polarities, effective voltage action times, and average distance along the insulator surface are used as inputs, and the insulator flashover voltage is used as the output. This method generates a large amount of input and output data to form the first input and output sets.

[0032] S2, performing regularization processing on the first input set and the first output set in step S1 to generate a first data set, where the first data set is a data set consisting of the second input set and the second output set after regularization.

[0033] The regularization formula is as follows: Among them, x is the actual value of the data; x' is the value after regularization; x min and x max are the minimum and maximum values of each input data respectively.

[0034] After regularization, the input data will be processed into the interval [-1,1], which helps to train the neural network when the input magnitude varies greatly.

[0035] S3, inputting the first data set generated in step S2 into the MLP neural network for training to obtain a training model.

[0036] Specifically, the Optuna framework based on Bayesian optimization is used to optimize the hyperparameters of the MLP (Multi-layer perception) neural network, and the first data set generated in the above step S2 is used to pre-train the neural network. Hyperparameters refer to the parameters used to configure the neural network training model. They cannot be estimated directly from data training and must be set before the model is trained. Adjusting the hyperparameters to the optimal configuration is conducive to establishing the best neural network model. Optuna is a hyperparameter optimization algorithm that uses orthogonal grid search and automatic loop attempts to find the best hyperparameter combination, thereby improving the accuracy of the training model.

[0037] In addition, applying the input used in the calculation of the empirical formula for flashover voltage and the output obtained by calculation to the MLP neural network for training can generate a large amount of data within the required prediction conditions, thereby saving a lot of manpower and time.

[0038] S4, extracting the experimental input parameter set and output parameter set from the artificial simulation experimental data, performing regularization processing on the input parameter set and the output parameter set respectively to generate a second data set, and dividing the second data set into a training set and a test set. The training set includes the second input set and the second output set; the test set includes the third input set and the third output set.

[0039] The input parameter set in this step includes the experimental SF6 gas pressure, voltage polarity, voltage effective duration, and insulator creepage distance. The output parameter set is the experimentally obtained insulator flashover voltage. In the second dataset, 80% of the data can be used as the training set, and 20% as the test set. Alternatively, 70% of the data can be used as the training set, and 30% as the test set.

[0040] In step S5, the training set from step S4 is input into the training model from step S3 for incremental learning, resulting in a training output set and an optimized training model. Incremental learning within the training model allows for full utilization of experimental data to correct empirical formulas, further enhancing the accuracy of the prediction model.

[0041] S6, verify the training model optimized in step S5:

[0042] S6.1 calculates the error between the training output set in step S5 and the second output set in step S4;

[0043] S6.2 Input the test set in step S4 into the training model optimized in step S5 for training, obtain a test output set, and calculate the error between the test output set and the third output set;

[0044] S6.3 If the errors calculated in S6.1 and S6.2 are both ≤5%, then the MLP regression prediction model is obtained and step S7 is executed; otherwise, the training is continued in step S3.

[0045] S7, by inputting any set of environmental parameters into the MLP regression prediction model, the insulator flashover voltage in SF6 under nanosecond pulses can be predicted.

[0046] The present invention combines a neural network hyperparameter optimization algorithm with training data generated by large-scale calculations using empirical formulas to pre-train the neural network, and then uses the pre-trained neural network to perform incremental learning on the experimental data. This not only ensures the accuracy of the predicted values, but also enables extrapolated prediction of flashover voltages within a wider range of conditions.

[0047] Although the embodiments of the present invention have been shown and described above, it will be apparent to those skilled in the art that any changes or modifications to the above embodiments should fall within the scope of protection of the present invention as long as they are within the spirit of the present invention.

Claims

1. A method for predicting the flashover voltage of an insulator in SF6 under nanosecond pulses, characterized in that: The following steps are involved: S1, using multiple sets of SF6 environmental parameters under nanosecond pulses as a first input set, and calculating a first output set using a flashover voltage calculation formula, wherein the first output set is the flashover voltage of the insulator under the multiple sets of environmental parameters; the multiple sets of environmental parameters are parameter sets consisting of different SF6 gas pressures, voltage polarities, voltage effective action times, and insulator creepage distances; S2, performing regularization processing on the first input set and the first output set in step S1 to generate a first data set; S3, inputting the first data set generated in step S2 into the MLP neural network for training to obtain a training model; S4, extracting the input parameter set and the output parameter set from the artificial simulation experimental data, performing regularization processing on the data, generating a second data set, and dividing the second data set into a training set and a test set; the training set includes the second input set and the second output set; the test set includes the third input set and the third output set; the input parameter set includes a set of SF6 gas pressure, voltage polarity, voltage effective action time, and insulator creepage distance parameters used in the experiment; The output parameter set is a set of insulator flashover voltages obtained through experiments; S5, inputting the training set in step S4 into the training model in step S3 for incremental learning to obtain a training output set and an optimized training model; S6, verify the training model optimized in step S5: S6.1 calculates the error between the training output set in step S5 and the second output set in step S4; S6.2 Input the test set in step S4 into the training model optimized in step S5 for training, obtain a test output set, and calculate the error between the test output set and the third output set; S6.3 If the errors calculated in S6.1 and S6.2 are both ≤5%, then the MLP regression prediction model is obtained and step S7 is executed; otherwise, the training is continued in step S3. S7, by inputting any set of environmental parameters into the MLP regression prediction model, the insulator flashover voltage in SF6 under nanosecond pulses can be predicted.

2. The method for predicting the flashover voltage of an insulator in SF6 under nanosecond pulse according to claim 1, characterized in that: In step S1, the flashover voltage calculation formula is: E br =k ± p 0.4 (dt eff ) -1 / 6 Among them, E br is the breakdown field strength, i.e. the flashover voltage of the insulator, in kV / cm; k is the polarity coefficient, when in SF6 gas, k is taken when the voltage polarity is positive + is a constant 44, and k is used when the voltage polarity is negative - is a constant 72; p is the SF6 gas pressure in atm; t eff is the effective time, which refers to the time when the pulse voltage amplitude exceeds 89% of the peak voltage, in μs; d is the distance along the insulator surface, in cm.

3. The method for predicting the flashover voltage of an insulator in SF6 under nanosecond pulses according to claim 2, wherein: In step S3, before training, the Optuna framework can be used to optimize the hyperparameters of the MLP neural network.

4. The method for predicting the flashover voltage of an insulator in SF6 under nanosecond pulses according to claim 3, characterized in that: In step S4, 80% of the data in the second data set is used as a training set, and 20% of the data is used as a test set.

5. The method for predicting the flashover voltage of an insulator in SF6 under nanosecond pulse according to claim 3, characterized in that: In step S4, 70% of the data in the second data set is used as a training set, and 30% of the data is used as a test set.

Images