Insulator pollution flashover prediction system and method based on generative adversarial network

Abstract

The invention discloses an insulator pollution flashover prediction system based on a generative adversarial network, and the system comprises a data processing module which is used for dividing a test data set and a training data set according to historical insulator equipment data, generating an optimal generation model weight file based on sample data, and then constructing corresponding generation sample data according to the optimal generation model weight file; generated sample data is expanded to a training data set after being evaluated and screened; the salt deposit density prediction model construction module is used for training an insulator equivalent salt deposit density prediction model by using the expanded training data set, evaluating model prediction accuracy through a test data set, and obtaining an optimal weight file of the prediction model; and the pollution flashover voltage reasoning module is used for accessing meteorological environment parameter monitoring data into the prediction model to obtain an insulator surface equivalent salt deposit density change increment, and inputting the insulator surface equivalent salt deposit density change increment into the pollution flashover voltage reasoning model to obtain an insulator pollution flashover critical voltage prediction value. The problem of training process overfitting caused by insufficient data quantity can be solved, and the model prediction performance is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power system monitoring technology, and specifically refers to an insulator flashover prediction system and method based on a generative adversarial network. Background Art

[0002] The dirt layer on the surface of the insulator is one of the main factors causing flashover phenomena, and the equivalent salt deposit density reflects the degree of dirt on the insulator surface.

[0003] Research shows that the change of the insulator equivalent salt density has complex non-linear characteristics. Therefore, the excellent non-linear modeling ability of the neural network is proposed to improve the prediction accuracy and stability. The training and prediction of the neural network rely on a large amount of sufficient and balanced historical data related to insulators, which can effectively improve the generalization ability and prediction accuracy of the salt density prediction model. However, in actual situations, the collected historical data often has problems of insufficiency and imbalance, which may lead to deviations and instabilities in the prediction results, limiting the reliability and accuracy of the neural network in the application of salt density prediction. In addition, a large number of actual accident cases show that lightning strikes on insulators will also trigger flashover accidents, and the lightning impulse flashover voltage of the insulator after contamination will decrease to varying degrees. When inferring the critical flashover voltage based on the insulator contamination situation and meteorological conditions, the existing technical methods related to flashover early warning fail to consider the influence of external thunderstorm activities on flashover.

[0004] Therefore, inventing an insulator flashover prediction system based on a generative adversarial network, which can solve the problems of instability and overfitting in the training process of the neural network model caused by insufficient training data set, and comprehensively consider the influence of various meteorological environment parameters and thunderstorm activities, and infer the insulator flashover voltage, thus ensuring the effectiveness of flashover early warning has become an urgent technical problem to be solved. Summary of the Invention

[0005] The purpose of the present invention is to provide an insulator flashover prediction system based on a generative adversarial network, which can solve the problems of instability and overfitting in the training process of the neural network model caused by insufficient training data set, and comprehensively consider the influence of various meteorological environment parameters and thunderstorm activities to ensure the effectiveness of flashover early warning.

[0006] To achieve this purpose, an insulator flashover prediction system based on a generative adversarial network designed by the present invention includes:

[0007] The data processing module is used to establish an equivalent salt deposit density time series dataset of the insulator surface based on the multi-variable data of the insulator equipment collected. The equivalent salt deposit density time series dataset of the insulator surface and the historical insulator equipment data in the multi-variable data are divided into a test dataset and a training dataset according to a setting. An optimal generation model weight file is generated based on the existing sample data. The optimal generation model weight file and the generative adversarial network model are used to construct corresponding generated sample data. The generated sample data is expanded into the training dataset after being evaluated and screened to obtain an expanded training dataset.

[0008] The salt deposit density prediction model construction module is used to train the insulator equivalent salt deposit density prediction model using the expanded training dataset to obtain a trained insulator equivalent salt deposit density prediction model. The prediction accuracy of the insulator equivalent salt deposit density prediction model is evaluated through the test dataset, and an optimal weight file of the trained insulator equivalent salt deposit density prediction model is obtained according to the model prediction accuracy. The optimal weight file of the trained insulator equivalent salt deposit density prediction model is loaded into the trained insulator equivalent salt deposit density prediction model to obtain a final insulator equivalent salt deposit density prediction model.

[0009] The flashover voltage inference module is used to construct a flashover voltage inference model according to a hybrid model of a multi-layer perceptron and a support vector machine. The meteorological environment parameter monitoring data in the multi-variable data collected during a set time period is connected to the final insulator equivalent salt deposit density prediction model to obtain the change increment of the equivalent salt deposit density on the insulator surface during the set time period. The change increment of the equivalent salt deposit density on the insulator surface during the set time period is input into the flashover voltage inference model to obtain a predicted value of the insulator flashover critical voltage.

[0010] Preferably, the multi-variable data includes insulator material model data, distribution location data, historical equivalent salt deposit density data, historical meteorological environment parameter data around the insulator distribution location, operating voltage of the line where the insulator distribution location is located, and meteorological environment data.

[0011] Preferably, after preprocessing the multi-variable data of the insulator equipment collected, an equivalent salt deposit density time series dataset of the insulator surface is established. The data preprocessing of the two types of information of the predicted areas collected specifically includes: data cleaning, data reduction, and dataset division. Among them, data cleaning includes: outlier identification and outlier processing; data reduction includes: data transformation and normalization processing; dataset division includes: dividing the historical dataset processed by the above method into a training dataset and a test dataset according to a ratio of 8:2.

[0012] Preferably, the specific method for the optimal generation model weight file is as follows: On the basis of GAN, the Wasserstein distance is introduced as a method to measure the similarity of two distributions, replacing the objective function of the GAN network to construct a WGAN model; the Lipschitz continuous condition is introduced to limit the maximum local variation amplitude of the function, and it is added to the discriminator loss function in the form of a gradient penalty term, replacing the weight pruning strategy in WGAN to construct an insulator time-series data generation model; using the existing sample data as input to train the insulator time-series data generation model, after completing one round of the training process, a weight file will be generated and saved. Iteratively train the insulator time-series data generation model, and select a model weight file with the lowest value of the generator loss function and the discriminator loss function within the set range as the optimal generation model weight file.

[0013] Preferably, the specific evaluation and screening method for expanding the sample data into the original training data set after evaluation and screening is as follows: Use the t-distributed stochastic neighbor embedding algorithm to perform dimensionality reduction analysis on the collected multivariate data and sample data, and select the spatial clustering algorithm DBSCAN to mark and delete the generated samples that deviate from the distribution of the collected multivariate data. Expand the remaining sample data after marking and deletion into the original training data set.

[0014] Preferably, the specific method for constructing the insulator equivalent salt deposit density prediction model and evaluating the prediction accuracy of the model is as follows: Use CNN as a feature extraction module to extract short-term dependence features between time series data; apply a convolutional layer to capture short-term time dependence by sliding the convolutional kernel to obtain local features and extract feature maps; use a pooling layer to reduce the dimension while retaining important features; use an LSTM layer for prediction: Pass the features extracted by CNN through the LSTM layer to capture global dependence relationships and generate the final prediction; then perform model training. The model training is to iteratively construct the salt density prediction model using the training data set, and adjust the hyperparameters of the model according to the output results to optimize the performance. Use the evaluation metrics root mean square error and mean absolute error to evaluate the error between the model output prediction results and the test data set.

[0015] Preferably, the specific method for inferring the critical flashover voltage of insulators is as follows: the pollution characteristics and insulator characteristics are directly used as input features; the environmental characteristics are the most representative features selected after performing a correlation analysis between meteorological parameters and the flashover voltage; the thunderstorm characteristics generate the thunderstorm intensity index as a feature; all input features are subjected to Z-score standardization, preprocessing, and evaluation and screening; a flashover voltage inference model based on a multi-layer perceptron and a support vector machine is constructed, and the MLP and SVM models are trained using the features after standardization, preprocessing, and evaluation and screening, and the best model parameters are selected using cross-validation, and the hyperparameters are further optimized through Bayesian optimization; the prediction results of the MLP and SVM are used as new features and input into a regression model, or weighted averaging is performed, or multiple MLP and SVM models are integrated through Bagging or Boosting methods to further improve the prediction stability, and the model performance is evaluated in the final test dataset, and by combining the advantages of the two models, the final predicted value of the critical flashover voltage is obtained.

[0016] Preferably, it further includes a flashover warning and risk map generation module, which determines different warning levels according to the predicted value of the critical flashover voltage of the insulator and the operating voltage of the line where the insulator is located in the multi-variable data, and generates a flashover risk map based on the warning level in combination with the distribution position data of the target insulator; the specific method for generating the flashover risk map according to the warning level is as follows: using a geographic information system tool to generate a risk map, stratifying the area according to the risk value, using the set color coding to display the risk level, and associating the risk map with the obtained warning signals of different levels to ensure that the map can reflect the changes in real time.

[0017] Advantages of the present invention: The present invention proposes an insulator flashover prediction system based on a generative adversarial network. In the data preprocessing stage, a training dataset expansion method combining the WGAN-GP generative model with t-sne dimensionality reduction analysis and the DBSCAN algorithm is adopted. Compared with the traditional neural network training process, it makes up for the negative effects of insufficient and unbalanced historical data sample quantities on the model training performance, and improves the final recognition accuracy; Based on a convolutional neural network as a feature extraction module, and using an LSTM model for inference and prediction, it combines the advantages of CNN in local feature extraction and LSTM in capturing long-term time series data dependency relationships, and improves the reliability of insulator equivalent salt deposit density prediction; For the inference of the critical flashover voltage of insulators, on the basis of extracting various meteorological parameter information around the insulators, considering the influence of thunderstorm activity characteristics on the flashover voltage, a hybrid model based on MLP and SVM is established, which has a warning effect on flashover accidents caused by lightning strikes on insulators; Different warning signals are issued according to the operating voltage and the critical flashover voltage respectively, and a risk map is generated using a geographic information system tool, etc., and the risk map is connected to the real-time data stream to ensure that the map can reflect changes in real time. It can assist operators to visually identify high-risk areas through the map, adjust the maintenance plan or take emergency measures when necessary, and ensure the effectiveness of flashover warning. Brief Description of the Drawings

[0018] Figure 1 is a schematic structural diagram of the present invention;

[0019] Figure 2 is a schematic flow diagram of the present invention. Detailed Embodiments

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Therefore, the detailed description of the embodiments of the present invention provided below in the drawings is not intended to limit the scope of the claimed present invention, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0021] The following further describes the present invention in detail in conjunction with the drawings and specific embodiments:

[0022] Embodiment 1

[0023] An insulator flashover prediction system based on a generative adversarial network, as Figure 1 shown, it includes:

[0024] The data processing module is used to establish an equivalent salt deposit density time series dataset of the insulator surface based on the multi-variable data of the insulator equipment collected. The equivalent salt deposit density time series dataset of the insulator surface and the historical insulator equipment data in the multi-variable data are divided into a test dataset and a training dataset according to a set rule. An optimal generation model weight file is generated based on the existing sample data. The optimal generation model weight file and the generative adversarial network model are used to construct corresponding generated sample data. The generated sample data is expanded into the training dataset after being evaluated and screened to obtain an expanded training dataset.

[0025] The salt deposit density prediction model construction module is used to train the insulator equivalent salt deposit density prediction model using the expanded training dataset to obtain a trained insulator equivalent salt deposit density prediction model. The prediction accuracy of the insulator equivalent salt deposit density prediction model is evaluated through the test dataset, and an optimal weight file of the trained insulator equivalent salt deposit density prediction model is obtained according to the model prediction accuracy. The optimal weight file of the trained insulator equivalent salt deposit density prediction model is loaded into the trained insulator equivalent salt deposit density prediction model to obtain a final insulator equivalent salt deposit density prediction model.

[0026] The flashover voltage inference module is used to construct a flashover voltage inference model according to a hybrid model of a multi-layer perceptron and a support vector machine. The meteorological environment parameter monitoring data in the multi-variable data collected within a set time period is connected to the final insulator equivalent salt deposit density prediction model to obtain the change increment of the equivalent salt deposit density on the insulator surface within the set time period. The change increment of the equivalent salt deposit density on the insulator surface within the set time period is input into the flashover voltage inference model to obtain a predicted value of the insulator flashover critical voltage.

[0027] In the above technical solution, the multi-variable data of the collected insulator equipment is put into a multi-variable database, and historical data is also stored in the multi-variable database. The existing sample data is included in the historical data.

[0028] In the above technical solution, the equivalent salt deposit density time series dataset of the insulator surface includes salt deposit density data and climate environment data that affect the change of salt deposit density, specifically the following data types: insulator attachment pollution parameters (equivalent salt deposit density increment), meteorological data (air relative humidity (RH), rainfall, wind speed, temperature, air pressure, lightning strike times, lightning current peak value); environmental data (total suspended particulate concentration, inhalable particulate concentration, types of pollution sources in the region).

[0029] In the above technical solution, the historical insulator equipment data includes: specific distribution location data of the insulator (latitude and longitude coordinates); material of the insulator (glass, ceramic, silicone rubber), model and structural parameters (diameter, structural height, surface area); voltage level of the operating line where the insulator is located.

[0030] In the above technical solution, the optimal weight file stores all the parameters in the insulator equivalent salt deposit density prediction model. These parameters are continuously optimized during the training process and represent the learning results of the model for the training data distribution. The model after loading the weights can perform the salt density prediction task, which helps to output the salt density prediction value, and this value is input into the flashover voltage inference model.

[0031] In the above technical solution, the multi-variable data includes insulator material model data, distribution location data, historical equivalent salt deposit density data, historical meteorological environment parameter data around the insulator distribution location, operating voltage of the line where the insulator distribution location is located, and meteorological environment data.

[0032] In the above technical solution, after preprocessing the collected multi-variable data of insulator equipment, an insulator surface equivalent salt density time series data set is established. Among them, data preprocessing is performed on the two types of information of the areas to be predicted collected, specifically including: data cleaning, data reduction, and data set division. Among them, data cleaning includes: outlier identification and outlier processing; data reduction includes: data transformation and normalization processing; data set division includes: dividing the historical data set processed by the above method into a training data set and a test data set according to a ratio of 8:2.

[0033] In the above technical solution, the specific method of data cleaning is as follows: Duplicate item processing: Duplicate items in the database are considered as incorrect data caused by duplicate records or input errors, and only the first item is retained for multiple duplicates; Outlier processing: Process values that deviate from the normal data distribution, and mark values that exceed three times the standard deviation of the mean as outliers and delete them; Missing value processing: Adopt a filling method to process missing values. If the number of missing values is too large so that the distribution law loses meaning, then delete this section of data; If the number of missing values is small, use the polynomial interpolation filling method to maintain the smoothness of the time series data.

[0034] In the above technical solution, the specific method of data reduction is as follows: Data transformation: It includes normalization processing, mapping the time series data to the interval [0,1], so that the data in each dimension is at the same order of magnitude, thereby eliminating the influence of the dimension; Feature selection: Data that can improve the model effect is regarded as relevant features and retained, and the remaining features are regarded as irrelevant or redundant features and deleted. The judgment method uses the mutual information method, and the formula of the mutual information method is as follows:

[0035]

[0036] Among them, P(x,y) is the joint probability distribution of X and Y taking values x and y at the same time, P(x) and P(y) are the marginal probability distributions of X and Y respectively, and I(X;Y) is the calculation result of the mutual information method.

[0037] In the above technical solution, the two types of information of the regions to be predicted are the data information selected for the two predictions of salt density prediction and pollution flashover voltage prediction.

[0038] In the above technical solution, the specific division method of dividing the test data set and the training data set according to the setting is as follows: extraction is carried out by adopting the common random extraction method, and it is divided according to a fixed ratio of 8:2; the time in the time series data is used as the data label, and the other collected multiple parameters with the same time label are attributed to the same piece of data, and then divided according to the previous random extraction method.

[0039] In the above technical solution, through data cleaning, data reduction and data set division, the data quality is improved, the model training efficiency is increased, and the generalization ability of the model is ensured. Among them, data cleaning ensures the accuracy of data by removing outliers; data reduction standardizes features through transformation and normalization processing so that the model can treat all features equally; data set division divides the data into a training set and a test set to ensure that the model can not only accurately predict the training data, but also perform well on unseen data.

[0040] In the above technical solution, the preprocessing step can improve the reliability and consistency of the data, reduce the interference of noise on the model, improve the training efficiency, accelerate the model convergence speed, and ensure that the model has a balanced sensitivity to different features. The reasonable division of the data set can effectively avoid overfitting and ensure that the model has good generalization ability, thereby improving the prediction accuracy and reliability.

[0041] In the above technical solution, the specific method for the optimal generated model weight file is as follows: on the basis of the GAN network, the Wasserstein distance is introduced as a method for measuring the similarity of two distributions, the objective function of the GAN network is replaced, and the WGAN model is constructed; the Lipschitz continuous condition restriction function is introduced to limit the maximum local change amplitude, and it is added to the discriminator loss function in the form of a gradient penalty term to replace the weight pruning strategy in the WGAN, and the insulator time series data generation model is constructed;

[0042] Using the existing sample data as the input to train the insulator time series data generation model, after completing one round of training process, a weight file will be generated and saved. Iteratively train the insulator time series data generation model, and select a model weight file with the lowest and stable value of the generator loss function and the value of the discriminator loss function within the set range as the optimal generated model weight file.

[0043] In the above technical solution, the generator and discriminator loss functions of the GAN network are as follows:

[0044]

[0045] Among them, G(z) is the output of the generator, representing the sample generated from the noise vector z, and L G is the loss function of the generator in the generative adversarial network, which is the expectation of the random noise z, and z comes from the noise distribution P z , D(·) is the discriminator output, and P z is the distribution of the noise vector z;

[0046]

[0047] Among them, u is the collected multivariate data, Pdata is the distribution of the multivariate data, λ is the gradient penalty coefficient, ▽ is the function for calculating the gradient, is the linear difference between the collected multivariate data and the sample data, is the L2 norm of the original input gradient, and L D represents the loss function of the discriminator.

[0048] In the above technical solution, the specific generation process of the sample data is as follows: read the optimal generator weight file saved during the previous training process and load it into the generator network of the generation model; the generator network performs a non-linear transformation on the noise and outputs the generated sample data; the generated sample data is saved.

[0049] In the above technical solution, the sample data is used to expand the scale of the dataset, increase the diversity of the training data, and help reduce the overfitting risk of the salt density prediction model and the pollution flashover voltage inference model; by adding samples with different feature distributions, the model can better learn the potential characteristics of the data, thereby improving the generalization ability of the model to unknown data; in the case where the number of data in some categories is relatively lacking, the generated samples are used to supplement such insufficient data and improve the balance of the data distribution.

[0050] In the above technical solution, by iteratively training the insulator time series data generation model to optimize the performance of the generator and the discriminator, finally select the model weight file with the minimum generator loss and the discriminator loss within the set range as the optimal generation model weight; improve the accuracy and stability of the model to generate data, ensure that the generated results are close to the collected multivariate data, avoid overfitting at the same time, and improve the prediction accuracy and generalization ability of the model.

[0051] In the above technical solution, since good and sufficient training set data can effectively improve the performance of the subsequent prediction model, obtaining the generation model and the optimal generation model weight for generating data and supplementing the training dataset can effectively improve the accuracy of subsequent salt density prediction and pollution flashover voltage prediction.

[0052] In the above technical solution, after the sample data is evaluated and screened, it is expanded into the original training dataset. The specific evaluation and screening method is as follows: The t-distributed stochastic neighbor embedding algorithm is used to perform dimensionality reduction analysis on the collected multivariate data and the sample data. The spatial clustering algorithm DBSCAN is selected to mark and delete the generated samples that deviate from the distribution of the collected multivariate data. The remaining sample data after marking and deletion is expanded into the original training dataset.

[0053] In the above technical solution, the specific method of the evaluation and screening method is as follows: The t-distributed stochastic neighbor embedding algorithm (t-SNE) is used to perform dimensionality reduction analysis on the collected multivariate data and the sample data. During the dimensionality reduction process, 2D fusion features are generated, without physical dimensions. The distribution interval of the sample data basically coincides with the distribution of the collected multivariate data, proving that the sample data at this point has high similarity with the collected multivariate data. A small part of the sample data deviates from the distribution of the collected multivariate data, indicating that the quality of the sample data at this point is poor, and the sample data at this point needs to be deleted to ensure the quality of the overall dataset.

[0054] The spatial clustering algorithm (DBSCAN) is selected to mark and delete the low-quality sample data that deviates from the distribution of the collected multivariate data; based on the data point density as the standard for type division, it is specifically divided into core points, boundary points, and noise points, where the noise points are the target data to be deleted; during the implementation process, all the collected multivariate data is used as core points, and the core points are traversed in a loop and their ε-neighborhoods are checked to determine whether the sample data belongs to the neighborhood of the collected multivariate data. The sample data outside the neighborhood range is regarded as the sample data to be deleted.

[0055] In the above technical solution, the generated sample data is expanded through t-SNE dimensionality reduction analysis and DBSCAN clustering screening to optimize the quality of the training dataset and improve the model performance of the insulator pollution flashover prediction system.

[0056] In the above technical solution, the screened sample data is closer to the actual distribution, avoiding the negative impact of deviated data on model training, and effectively improving the prediction accuracy and robustness of the model.

[0057] In the above technical solution, the specific method for constructing the insulator equivalent salt deposit density prediction model and evaluating the prediction accuracy of the model is as follows: CNN is used as a feature extraction module to extract the short-term dependence features between time series data; the convolutional layer is applied to capture the short-term time dependence through a sliding convolutional kernel for obtaining local features and extracting feature maps; the pooling layer is used to reduce the dimension while retaining important features; the LSTM layer is used for prediction: the features extracted by CNN are passed through the LSTM layer to capture the global dependence relationship and generate the final prediction;

[0058] Then, model training is carried out. Model training is to iteratively build the salt density prediction model using the training dataset, and adjust the hyperparameters of the model according to the output results to optimize the performance. The root mean square error and mean absolute error of the evaluation metrics are used to evaluate the error between the model output prediction results and the test dataset.

[0059] In the above technical solution, the specific method for constructing the salt density prediction model is as follows:

[0060] Model construction: Input stage: The shape of the input tensor is rearranged by permuting dimensions to conform to the expected input format of the model; Convolution and pooling layer stage: The first convolutional layer receives the rearranged input tensor, performs a convolution operation, processes and convolves with the ReLU activation function, and then performs a max pooling operation; The second convolutional layer receives the output of the first pooling layer, performs convolution, ReLU activation function processing, and max pooling operations; LSTM layer stage: Receives the output of the second pooling layer, performs sequence modeling through the LSTM layer, and outputs; Fully connected layer stage: The fully connected layer maps and changes the shape of the output tensor of the LSTM layer, and selects the output of the last time step as the final prediction result.

[0061] Model training: Use historical data to train the model, adjust the hyperparameters of the model to optimize the performance, use cross-validation technology to evaluate the generalization ability of the model, and adjust the model to prevent overfitting. Use appropriate evaluation metrics (root mean square error RMSE, mean absolute error MAE) to evaluate the performance of the model on the training set and the test set. Adjust the parameters or algorithms of the model according to the evaluation results to improve the prediction accuracy and robustness. The specific calculation formulas for the root mean square error RMSE and mean absolute error MAE are as follows:

[0062]

[0063] Where: y i is the time series of the equivalent salt deposit density output by the model, x i is the measured time series of the equivalent salt deposit density, and n is the total number of data in this time series.

[0064] In the above technical solution, the salt density prediction model is iteratively trained with the training dataset and the hyperparameters are adjusted to optimize the model performance and make its output results more accurate.

[0065] In the above technical solution, using the root mean square error and mean absolute error to evaluate the model prediction results can effectively reduce the error and improve the prediction accuracy and generalization ability of the model.

[0066] In the above technical solution, the specific method for inferring the critical voltage of insulator pollution flashover is as follows: the pollution characteristics and insulator characteristics are directly used as input features; for the environmental characteristics, the most representative features are selected after performing a correlation analysis between meteorological parameters and the pollution flashover voltage; for the thunderstorm characteristics, a thunderstorm intensity index is generated as a feature;

[0067] Perform Z-score standardization, preprocessing, and evaluation and screening on all input features;

[0068] Construct a pollution flashover voltage inference model based on a multi-layer perceptron and a support vector machine. Use the standardized, preprocessed, and evaluated and screened features to train the MLP and SVM models, and use cross-validation to select the best model parameters. Further optimize the hyperparameters through Bayesian optimization;

[0069] Use the prediction results of the MLP and SVM as new features and input them into the regression model, either perform weighted averaging, or integrate multiple MLP and SVM models through the Bagging or Boosting method to further improve the prediction stability. Evaluate the model performance in the final test dataset, and combine the advantages of the two models to obtain the final predicted value of the critical voltage of pollution flashover.

[0070] In the above technical solution, the specific method for constructing a multi-layer perceptron model is as follows:

[0071] Network structure design: The number of nodes in the input layer is the same as the number of input features. Select 2 - 3 hidden layers for the hidden layer, and the number of nodes in the hidden layer decreases layer by layer. For example, the first hidden layer has 100 nodes, the second layer has 50 nodes, and the third layer has 20 nodes. The output layer selects 1 node to output the predicted pollution flashover voltage value;

[0072] Activation function selection: Select the ReLU activation function for the hidden layer to avoid the problem of gradient disappearance; select a linear activation for the output layer;

[0073] Regularization: Use L2 regularization to prevent the model from overfitting; use Dropout to randomly discard a certain proportion of neurons during training to further prevent overfitting;

[0074] Optimizer: Select the Adam optimizer, which is an adaptive learning rate optimizer;

[0075] Loss function: Select the mean squared error MSE to measure the squared error between the predicted value and the actual value.

[0076] In the above technical solution, the specific method for constructing a support vector machine SVM model is as follows:

[0077] Kernel function selection: Select the radial basis function kernel RBF to capture non-linear relationships. The parameter γ of this kernel function controls the influence range of the support vectors;

[0078] Hyperparameter Optimization: The regularization parameter C is used to control the error tolerance of the model. The values of C and the kernel function parameter γ are adjusted through grid search or random search to find the best combination.

[0079] In the above technical solution, the specific method for constructing the hybrid model is as follows:

[0080] Weighted Average Method: Directly perform weighted averaging on the prediction results of MLP and SVM according to certain weights to generate the final predicted value. The specific calculation formula is:

[0081]

[0082] Among them, y i is the true value of the pollution flashover voltage in the dataset, y i,MLP and y i,SVM are the predicted values of the two models respectively. MSE MLP represents the mean square error between the prediction result of the MLP model and the true value of the test set, and MSE SVM represents the mean square error between the prediction result of the SVM model and the true value of the test set;

[0083]

[0084] y i,final = ω SVM ·y i,MLP + ω SVM ·y i,MLP

[0085] The weights ω MLP and ω SVM will be adjusted according to the errors of the models. The model with smaller errors will be assigned a larger weight, and finally a more accurate predicted value y i,final is obtained through weighted averaging. y i,final represents the final predicted value output after being corrected by this weighted average method;

[0086] Finally, hyperparameter tuning is performed. The hyperparameters are further optimized through Bayesian optimization, such as the learning rate and the number of hidden layer nodes of MLP, and the regularization parameter C and the kernel function parameter γ of SVM.

[0087] In the above technical solution, in practical applications, it is also necessary to dynamically adjust the weight coefficients of MLP and SVM according to real-time data to adapt to environmental changes and maintain the model; an adaptive learning mechanism is introduced, and the model performs online learning and updates according to new data during operation to ensure long-term effectiveness; an anomaly detection module is added to identify and process abnormal data in extreme situations such as thunderstorms to prevent misjudgment of the model.

[0088] In the above technical solution, by taking the prediction results of MLP and SVM as new features and inputting them into the regression model, or performing weighted average fusion, or combining the advantages of multiple MLP and SVM models through ensemble methods such as Bagging / Boosting, the prediction performance and stability of the model for the critical voltage of pollution flashover are improved.

[0089] In the above technical solution, the fusion method can make full use of the non-linear fitting ability of MLP and the processing advantages of SVM for high-dimensional small sample data, reducing the bias or variance problems that may be brought by a single model; evaluating the model performance in the final test dataset can improve the stability and accuracy of the prediction results, so as to obtain the predicted value of the critical voltage of pollution flashover more reliably.

[0090] In the above technical solution, it further includes a pollution flashover warning and risk map generation module, which determines different warning levels according to the predicted value of the critical voltage of insulator pollution flashover and the operating voltage of the line where the insulator is located in the multi-variable data, and generates a pollution flashover risk map based on the warning level in combination with the distribution position data of the target insulator;

[0091] The specific way of generating a pollution flashover risk map according to the warning level is as follows: use a geographic information system tool to generate a risk map, stratify the area according to the risk value, display the risk level using the set color coding, and link the risk map with the obtained warning signals of different levels to ensure that the map can reflect the changes in real time.

[0092] In the above technical solution, the warning levels are set as:

[0093] Level 1 warning (high risk): At this time, the operating voltage is close to or exceeds the critical voltage of pollution flashover, and the pollution flashover risk is extremely high. It is recommended to take immediate emergency measures, such as power outage or increasing the frequency of insulator cleaning. The condition is 0.85U ≤ U 50％ <0.95U;

[0094] Level 2 warning (medium risk): The operating voltage is between 70% - 90% of the critical voltage, and there is a certain pollution flashover risk. It is recommended to carry out condition monitoring and preventive maintenance. The condition is 0.95U ≤ U 50％ <U;

[0095] Level 3 warning (low risk): The operating voltage is relatively low and the risk is small, but regular monitoring and maintenance are still required. The condition is U ≤ U 50％ <1.05U;

[0096] Normal operation: The operating voltage is much lower than the critical voltage of pollution flashover, and the pollution flashover risk is extremely low. Regular monitoring can be maintained. The condition is 1.05U ≤ U 50％ .

[0097] In the above technical solution, the map drawing tool is: a Geographic Information System (GIS) tool (such as ArcGIS, QGIS) or a Python library (such as Matplotlib, Basemap, Folium) to generate a risk map.

[0098] In the above technical solution, risk stratification and color coding are used to stratify regions according to risk values, and color coding (such as red for high risk, yellow for medium risk, and green for low risk) is used to visually display the risk level.

[0099] In the above technical solution, the risk map includes a real-time data update function, connecting the risk map to real-time data streams (such as weather stations, radar data) to ensure that the map can reflect changes in real time.

[0100] In the above technical solution, the risk map includes an interactive function, providing an interactive function for the map. Operators can click on any area to view detailed risk assessment data and obtain corresponding early warning information and recommended measures.

[0101] In the above technical solution, the risk map is integrated with an early warning system. When the risk value in a certain area reaches the above warning level, an early warning is automatically triggered and a notification is sent; historical risk map data is saved for trend analysis and long-term planning decisions.

[0102] Embodiment 2

[0103] An insulator pollution flashover prediction method based on a generative adversarial network, as Figure 2 shown, input insulator characteristic information, and collect multi-variable information such as meteorological and environmental parameters; preprocess the multi-variables and construct a training set database for model training; use a generative model to generate training data samples, perform dimensionality reduction analysis on the collected data and the generated samples, and mark and delete low-quality generated samples; predict the change in the equivalent salt deposit density of the insulator, and output the increment of the change in the equivalent salt deposit density of the insulator; combine the meteorological environment parameter information and the prediction result of the equivalent salt deposit density, and infer the critical voltage U 50％ of insulator pollution flashover; combine the inference result of the data pollution flashover critical voltage U 50％ and the operating voltage of the line, and issue insulator pollution flashover early warning signals of different levels according to the calculation results, and generate a pollution flashover risk map.

[0104] The insulator pollution flashover prediction method includes the following steps:

[0105] The data processing module is used to establish an equivalent salt deposit density time series data set of the insulator surface based on the multi-variable data of the insulator equipment collected. The equivalent salt deposit density time series data set of the insulator surface and the historical insulator equipment data in the multi-variable data are divided into a test data set and a training data set according to a setting. An optimal generation model weight file is generated based on the existing sample data. The optimal generation model weight file and a generative adversarial network model are used to construct corresponding generated sample data. The generated sample data is expanded into the training data set after being evaluated and screened to obtain an expanded training data set;

[0106] The salt deposit density prediction model construction module is used to train an insulator equivalent salt deposit density prediction model using the expanded training data set to obtain a trained insulator equivalent salt deposit density prediction model. The prediction accuracy of the insulator equivalent salt deposit density prediction model is evaluated through the test data set, and an optimal weight file of the trained insulator equivalent salt deposit density prediction model is obtained according to the model prediction accuracy. The optimal weight file of the trained insulator equivalent salt deposit density prediction model is loaded into the trained insulator equivalent salt deposit density prediction model to obtain a final insulator equivalent salt deposit density prediction model;

[0107] The flashover voltage inference module is used to construct a flashover voltage inference model according to a hybrid model of a multi-layer perceptron and a support vector machine. The meteorological environment parameter monitoring data in the multi-variable data collected within a set time period is connected to the final insulator equivalent salt deposit density prediction model to obtain the change increment of the equivalent salt deposit density on the insulator surface within the set time period. The change increment of the equivalent salt deposit density on the insulator surface within the set time period is input into the flashover voltage inference model to obtain a predicted value of the insulator flashover critical voltage.

[0108] Embodiment 3

[0109] A computer program product includes a computer program, and when the computer program is executed by a processor, it implements the steps of the method described in Embodiment 2.

[0110] The content not detailedly described in this specification belongs to the known prior art of those skilled in the art.

Claims

1. An insulator flashover prediction system based on generative adversarial network, characterized in that: It includes: The data processing module is used to establish an insulator surface equivalent salt density time series data set according to the collected insulator equipment multivariate data, divide the insulator surface equivalent salt density time series data set and the historical insulator equipment data in the multivariate data into a test data set and a training data set according to the settings, generate an optimal generation model weight file based on the existing sample data, use the optimal generation model weight file and the generation adversarial network model to construct corresponding generation sample data, expand the generation sample data into the training data set after evaluation and screening, and obtain the expanded training data set; The salt density prediction model construction module is used to use the expanded training data set to train the insulator equivalent salt density prediction model to obtain the trained insulator equivalent salt density prediction model, evaluate the prediction accuracy of the insulator equivalent salt density prediction model through the test data set, and obtain the optimal weight file of the trained insulator equivalent salt density prediction model according to the prediction accuracy of the model, load the optimal weight file of the trained insulator equivalent salt density prediction model into the trained insulator equivalent salt density prediction model, and obtain the final insulator equivalent salt density prediction model; The pollution flashover voltage inference module is used to construct a pollution flashover voltage inference model based on a hybrid model of a multi-layer perceptron and a support vector machine. The meteorological environmental parameter monitoring data of a set time period in the collected multivariate data is connected to the final insulator equivalent salt density prediction model to obtain the incremental change of the insulator surface equivalent salt density in the set time period. The incremental change of the insulator surface equivalent salt density in the set time period is input into the pollution flashover voltage inference model to obtain the predicted value of the insulator pollution flashover critical voltage.

2. The insulator flashover prediction system based on generative adversarial network according to claim 1 is characterized by: The multivariate data includes insulator material model data, distribution location data, equivalent salt density historical data, surrounding meteorological environment parameter historical data of the insulator distribution location, line operating voltage of the insulator distribution location, and meteorological environment data.

3. The insulator flashover prediction system based on generative adversarial network according to claim 1 is characterized by: After preprocessing the collected multivariate data of insulator equipment, an insulator surface equivalent salt density time series data set is established, in which the two types of collected regional information to be predicted are preprocessed, specifically including: data cleaning, data reduction, and data set division, in which data cleaning includes: outlier identification and outlier processing; data reduction includes: data transformation and normalization processing; data set division includes: dividing the historical data set processed by the above method into a training data set and a test data set according to a set ratio.

4. The insulator flashover prediction system based on generative adversarial network according to claim 1 is characterized by: The specific method of the optimal generation model weight file is as follows: based on the GAN network, the Wasserstein distance is introduced as a method for measuring the similarity of two distributions, the objective function of the GAN network is replaced, and the WGAN model is constructed; the Lipschitz continuous condition is introduced to limit the maximum local variation of the function, and it is added to the discriminator loss function in the form of a gradient penalty term, replacing the weight pruning strategy in the WGAN, and constructing the insulator time series data generation model; The insulator time series data generation model is trained based on the existing sample data as input. After one round of training, a weight file will be generated and saved. The insulator time series data generation model is iteratively trained, and a model weight file with the lowest generator loss function value and a discriminator loss function value within the set range is selected as the optimal generation model weight file.

5. The insulator flashover prediction system based on generative adversarial network according to claim 1, characterized in that: The sample data is expanded to the original training data set after evaluation and screening. The specific evaluation and screening method is: use the t-distribution random neighbor embedding algorithm to perform dimensionality reduction analysis on the collected multivariate data and sample data, select the spatial clustering algorithm DBSCAN to mark and delete the generated samples that deviate from the distribution of the collected multivariate data, and expand the remaining sample data after the mark deletion to the original training data set.

6. The insulator flashover prediction system based on generative adversarial network according to claim 1, characterized in that: The specific method of constructing the insulator equivalent salt density prediction model and evaluating the prediction accuracy of the model is as follows: using CNN as a feature extraction module to extract short-term dependency features between time series data; applying a convolution layer to capture short-term time dependencies by sliding convolution kernels to obtain local features and extract feature maps; using a pooling layer to reduce dimensions while retaining important features; using an LSTM layer for prediction: passing the features extracted by CNN through the LSTM layer to capture global dependencies and generate a final prediction; Then the model is trained. The model training is to use the training data set to iteratively construct the salt density prediction model, and adjust the model's hyperparameters according to the output results to optimize the performance. The evaluation indicators root mean square error and mean absolute error are used to evaluate the error between the model output prediction results and the test data set.

7. The insulator flashover prediction system based on generative adversarial network according to claims 2 and 6 is characterized in that: The specific method for inferring the critical voltage of insulator pollution flashover is: pollution characteristics and insulator characteristics are directly used as input characteristics; Environmental characteristics are the most representative features selected after correlation analysis between meteorological parameters and pollution flashover voltage; thunderstorm characteristics are the thunderstorm intensity index generated as a feature; Perform Z-score standardization, preprocessing, and evaluation screening on all input features; Construct a pollution flashover voltage inference model based on multi-layer perceptron and support vector machine, train MLP and SVM models using standardized, preprocessed and evaluated features, select the best model parameters using cross-validation, and further optimize hyperparameters using Bayesian optimization; The prediction results of MLP and SVM are input into the regression model as new features or weighted averaged, or multiple MLP and SVM models are integrated through Bagging or Boosting methods to further improve the prediction stability. The model performance is evaluated in the final test data set, and the advantages of the two models are combined to obtain the final pollution flashover critical voltage prediction value.

8. The insulator flashover prediction system based on generative adversarial network according to claim 1, characterized in that: It also includes a pollution flashover warning and risk map generation module, which determines different warning levels according to the predicted value of the pollution flashover critical voltage of the insulator and the operating voltage of the line where the insulator is located in the multivariate data, and generates a pollution flashover risk map based on the warning level combined with the target insulator distribution position data; The specific method of generating a pollution flashover risk map based on the warning level is: using geographic information system tools to generate risk maps, stratifying regions according to risk values, using set color codes to display risk levels, and linking risk maps with different levels of warning signals obtained to ensure that the map can reflect changes in real time.

9. An insulator flashover prediction method based on generative adversarial network, characterized in that: It includes the following steps: The data processing module is used to establish an insulator surface equivalent salt density time series data set according to the collected insulator equipment multivariate data, divide the insulator surface equivalent salt density time series data set and the historical insulator equipment data in the multivariate data into a test data set and a training data set according to the settings, generate an optimal generation model weight file based on the existing sample data, use the optimal generation model weight file and the generation adversarial network model to construct corresponding generation sample data, expand the generation sample data into the training data set after evaluation and screening, and obtain the expanded training data set; The salt density prediction model construction module is used to use the expanded training data set to train the insulator equivalent salt density prediction model to obtain the trained insulator equivalent salt density prediction model, evaluate the prediction accuracy of the insulator equivalent salt density prediction model through the test data set, and obtain the optimal weight file of the trained insulator equivalent salt density prediction model according to the prediction accuracy of the model, load the optimal weight file of the trained insulator equivalent salt density prediction model into the trained insulator equivalent salt density prediction model, and obtain the final insulator equivalent salt density prediction model; The pollution flashover voltage inference module is used to construct a pollution flashover voltage inference model based on a hybrid model of a multi-layer perceptron and a support vector machine. The meteorological environmental parameter monitoring data of a set time period in the collected multivariate data is connected to the final insulator equivalent salt density prediction model to obtain the incremental change of the insulator surface equivalent salt density in the set time period. The incremental change of the insulator surface equivalent salt density in the set time period is input into the pollution flashover voltage inference model to obtain the predicted value of the insulator pollution flashover critical voltage.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method described in claim 9 are implemented.

Images