Generative adversarial neural network-based power grid fault detection system

Abstract

The invention relates to a power grid fault detection system based on a generative adversarial neural network, and the system comprises a data collection and preprocessing module which is used for collecting real-time operation data and historical storage data from a bus of a power grid and each database, and carrying out the denoising and standardization processing of the real-time operation data and historical storage data; the data generation module is used for generating virtual fault data similar to the actual fault data according to the GAN model; the fault detection and analysis module is used for fusing physical calculation characteristics through a fault detection system constructed by a deep learning model; the alarm and response module is used for triggering an alarm mechanism after the fault type is determined; and the data storage and management module is used for storing the fault data, the generated data and the detection result thereof, and providing historical data support for the system for future system optimization and update. By adopting the system provided by the invention, the training of the fault detection model is enhanced by generating data similar to faults, so that various power grid faults can be identified and distinguished more sensitively and more accurately.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, and in particular to the field of power grid fault detection, and specifically refers to a power grid fault detection system based on a generative adversarial neural network. Background Art

[0002] The power grid is an important infrastructure in modern society. With the advancement of industrialization and urbanization, the demand for electricity continues to grow, and the structure of the power system is becoming increasingly complex. This complexity is not only reflected in the dynamic changes in power load and the expansion of the power grid scale, but also involves the diversity of power grid topology and the access of distributed energy. These changes have put forward higher requirements for the management and control of the power grid, especially in fault detection and processing. The power grid fault detection and diagnosis system is mainly used to monitor the operating status of the power grid, detect and locate faults in a timely manner, and ensure the safe operation of the power grid. Power grid failures not only affect the reliability of power supply, but may also cause serious economic losses and social impacts. Therefore, improving the accuracy and response speed of power grid fault detection technology has become an important research direction in the operation of power systems.

[0003] At present, power grid fault detection technology has made great progress. Traditional methods mainly include rule-based methods, signal processing methods, machine learning methods, etc. These methods can meet the needs of power grid fault detection to a certain extent, but with the complexity and intelligence of power grid systems, the requirements for fault detection are getting higher and higher.

[0004] The rule-based grid fault detection method is the earliest technology applied to grid fault detection. It determines whether the grid has a fault based on preset operating rules and thresholds. By monitoring key parameters such as voltage and current, once the detection value exceeds the preset threshold, the system will trigger an alarm mechanism. Its advantages are simple implementation and good real-time performance, but it also has significant limitations. Especially in the case of complex and changeable grid environment, it is difficult to adapt to different working conditions and is prone to false alarms or missed alarms.

[0005] Signal processing technology extracts possible fault information by analyzing the signal characteristics in the power grid operation data. Common methods include Fourier transform and wavelet transform. Through these transformation technologies, abnormal signals such as harmonics and transients in the power grid operation can be identified to determine the fault. Although signal processing methods can process complex time series data and improve the accuracy of fault detection to a certain extent, this method relies on expert knowledge for feature extraction and has high computational complexity in high-frequency data processing.

[0006] With the development of data-driven technology, machine learning has gradually become an important means of power grid fault detection. By learning historical fault data, machine learning models can automatically identify and classify fault types. Common algorithms include support vector machines, decision trees, and random forests. Machine learning methods can effectively process large-scale data and improve the degree of automation of detection. However, these methods are highly dependent on the amount of data, especially when power grid fault data is scarce, the model training effect is not good, and the generalization ability of the model is limited when dealing with new or complex faults.

[0007] Although existing technologies have made some progress in power grid fault detection, facing the increasingly complex power grid environment, these technologies still have many shortcomings and cannot meet the high requirements of modern power systems for fault detection. There are the following technical defects: (1) Insufficient sensitivity: When facing rare or complex fault types, existing detection methods often show the problem of insufficient sensitivity. Due to its fixed threshold setting, rule-based detection methods cannot dynamically adapt to changes in power grid operating conditions, resulting in high false alarm and missed alarm rates. Although signal processing methods can capture the signal characteristics of faults, the accuracy of feature extraction will be affected in a noisy environment, which in turn affects the sensitivity of fault detection. (2) Lack of data: The scarcity of power grid fault data is one of the main factors affecting the effectiveness of existing technologies. In particular, rare fault types are usually difficult to be fully reflected in the training data set, resulting in insufficient model training and difficulty in covering all possible fault scenarios. Machine learning methods rely on a large amount of historical data to train models. In the case of insufficient data, the model often cannot fully learn the fault characteristics and shows poor detection performance. (3) Limited differentiation ability: Existing technologies show obvious deficiencies in distinguishing similar types of faults. Different fault types may show similar signal patterns in the feature space, and traditional methods have difficulty effectively distinguishing these faults. For example, short circuit and ground faults may show similar current waveforms in some cases, which makes existing detection methods prone to misjudgment when classifying faults, thus affecting the accuracy of fault handling. Summary of the invention

[0008] The purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a power grid fault detection system based on generative adversarial neural network.

[0009] In order to achieve the above object, the power grid fault detection system based on generative adversarial neural network of the present invention is as follows:

[0010] The main features of the power grid fault detection system based on generative adversarial neural network are as follows:

[0011] The data acquisition and preprocessing module is used to collect real-time operation data and historical storage data from the power grid bus and various databases, obtain data sets of different fault types, and perform denoising and standardization processing on them;

[0012] A data generation module, connected to the data acquisition and preprocessing module, for generating virtual fault data similar to actual fault data based on a generative adversarial neural network (GAN);

[0013] A fault detection and analysis module, connected to the data generation module and the data acquisition and preprocessing module, is used to integrate physical computing features into the fault detection system built by the deep learning model, and to efficiently detect and distinguish different types of faults through model training;

[0014] An alarm and response module, connected to the fault detection and analysis module, is used to trigger an alarm mechanism after a fault is detected and the fault type is determined, and to provide corresponding processing suggestions or automatically perform corresponding fault processing operations according to the fault type; and

[0015] The data storage and management module is connected to the data acquisition and preprocessing module and is used to store all fault data, generated data and test results, and provide historical data support for the system for future system optimization and update.

[0016] Preferably, the generative adversarial neural network GAN consists of a generator and a discriminator, wherein the generator is used to input random noise or normal operation data to generate fault-like data; the discriminator is used to evaluate the generated data and compare it with the real fault data, so as to guide the generator to generate more realistic data.

[0017] Preferably, the generator is composed of multiple layers of fully connected layers and deconvolution layers, and specifically performs the following processing:

[0018] (1.1) Input random noise z or normal operation data X normal Feature extraction is performed through the fully connected layer, and the output is a high-dimensional vector;

[0019] The high-dimensional vector described in (1.2) is gradually upsampled through the deconvolution layer to generate simulated data X with the same dimension as the actual fault data fake ;

[0020] (1.3) Batch Normalization technology is used and Leaky ReLU activation function is added to each layer to make the generated simulation data as close as possible to the actual fault data in the feature space.

[0021] Preferably, the discriminator is composed of multiple layers of fully connected layers and deconvolution layers, and specifically performs the following processing:

[0022] The discriminator input described in (2.1) includes the real fault data X real and the simulated fault data X generated by the generator fake ;

[0023] (2.2) The convolution layer extracts local features from the data through filters and downsamples through the Max Pooling layer to reduce the amount of computation while retaining the main features;

[0024] The output of the discriminator described in (2.3) is a multidimensional vector, each data represents the probability that the input corresponding fault category is a real fault data, and the Sigmoid activation function is used to make the output value between 0 and 1, and the Dropout technology is used to prevent overfitting;

[0025] (2.4) By distinguishing the authenticity of the input data, the classification ability of real fault data and simulated fault data is optimized, and the error loss between categories is maximized, so as to enhance the discrimination of simulated samples and the difference relationship between physical characteristics.

[0026] Preferably, the generative adversarial neural network GAN performs model training on the data generation module in the following manner:

[0027] (a) using the original real fault data to train the discriminator and maximize the output of the discriminator for the real data;

[0028] (b) training the discriminator using the simulated fault data generated by the generator, minimizing the output error of the discriminator to the generated data, and maximizing the classification loss of samples of different categories;

[0029] (c) fixing the discriminator and back-propagating the parameters required to update the generator so that the data generated by the generator can deceive the discriminator;

[0030] (d) Repeat the above steps until the GAN network model converges, that is, the discriminator cannot distinguish between real data and generated data, so that the generator can generate simulated data that is close to the distribution of real fault data.

[0031] Preferably, the GAN network model expresses the objective function in the following manner:

[0032]

[0033] Where x represents real fault data, G(z) represents simulated fault data generated by the generator, and D(x) and D(G(z)) represent the output probabilities of the discriminator for real fault data and generated data, respectively.

[0034] Preferably, the fault detection and analysis module specifically performs the following processing:

[0035] (e) Extracting the temporal features of the input power grid data through the convolutional layer;

[0036] (f) The physical calculation features of the power grid are fused with the high-level features extracted by the convolutional layer through the physical feature fusion layer;

[0037] (g) The fusion features are further extracted through the fully connected layer for the input of the classifier, thereby outputting the probability distribution of the fault type and distinguishing the fault type.

[0038] Preferably, the fault detection and analysis module minimizes the difference between the actual fault type and the predicted fault type through a customized multi-objective loss function, specifically: using the following method to weight the cross entropy loss L CE and mean square error L MSE Jointly optimize the multi-objective loss function L:

[0039] L=α·L CE +β·L MSE

[0040] Among them, α and β are weight coefficients, and the cross entropy loss calculation formula is:

[0041]

[0042] Among them, Y true To minimize the actual fault type, Y pred To predict the fault type, i represents the i-th category label, represents the predicted probability of the i-th category.

[0043] The power grid fault detection system based on the generative adversarial neural network of the present invention has the following technical advantages: (1) Improving detection sensitivity: GAN is used to generate a large amount of high-quality simulated fault data to make up for the deficiency of actual fault data, so that the detection model can still maintain high sensitivity and accuracy when facing rare or complex fault types. (2) Expanding the coverage of the data set: By generating diversified simulated data, the training data set of the fault detection model is expanded so that it can cover more potential fault scenarios and improve the generalization ability of the model. (3) Enhancing the ability to distinguish faults: By combining the generated simulated data with the actual fault data and incorporating the loss calculation of physical characteristics to optimize the feature extraction and classification capabilities of the fault detection model, it can accurately distinguish similar types of faults and improve the accuracy of fault handling. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a schematic diagram of the overall framework of the power grid fault detection system based on generative adversarial neural network of the present invention.

[0045] Figure 2 The figure is a schematic diagram of the GAN network structure used to generate simulated power grid fault data in the present invention.

[0046] Figure 3 It is a schematic diagram of the calculation of the custom function of the fault detection and analysis module of the present invention. DETAILED DESCRIPTION

[0047] In order to more clearly describe the technical content of the present invention, further description is given below in conjunction with specific embodiments.

[0048] Before describing in detail embodiments according to the present invention, it should be noted that, hereinafter, the terms "comprises", "includes" or any other variations are intended to cover non-exclusive inclusion, whereby a process, method, article or apparatus comprising a series of elements includes not only these elements, but also other elements not explicitly listed or inherent to such process, method, article or apparatus.

[0049] See also Figure 1 As shown, the power grid fault detection system based on generative adversarial neural network, wherein the system includes:

[0050] The data acquisition and preprocessing module is used to collect real-time operation data and historical storage data from the power grid bus and various databases, obtain data sets of different fault types, and perform denoising and standardization processing on them;

[0051] A data generation module, connected to the data acquisition and preprocessing module, for generating virtual fault data similar to actual fault data based on a generative adversarial neural network (GAN);

[0052] A fault detection and analysis module, connected to the data generation module and the data acquisition and preprocessing module, is used to integrate physical computing features into the fault detection system built by the deep learning model, and to efficiently detect and distinguish different types of faults through model training;

[0053] An alarm and response module, connected to the fault detection and analysis module, is used to trigger an alarm mechanism after a fault is detected and the fault type is determined, and to provide corresponding processing suggestions or automatically perform corresponding fault processing operations according to the fault type; and

[0054] The data storage and management module is connected to the data acquisition and preprocessing module and is used to store all fault data, generated data and test results, and provide historical data support for the system for future system optimization and update.

[0055] As a preferred embodiment of the present invention, the generative adversarial neural network GAN is composed of a generator and a discriminator, wherein the generator is used to input random noise or normal operation data to generate fault-like data; the discriminator is used to evaluate the generated data and compare it with the real fault data, so as to guide the generator to generate more realistic data.

[0056] See also Figure 2 As shown, as a preferred embodiment of the present invention, the generator is composed of multiple layers of fully connected layers and deconvolution layers, and specifically performs the following processing:

[0057] (1.1) Input random noise z or normal operation data X normal Feature extraction is performed through the fully connected layer, and the output is a high-dimensional vector;

[0058] The high-dimensional vector described in (1.2) is gradually upsampled through the deconvolution layer to generate simulated data X with the same dimension as the actual fault data fake ;

[0059] (1.3) Batch Normalization technology is used and Leaky ReLU activation function is added to each layer to make the generated simulation data as close as possible to the actual fault data in the feature space.

[0060] As a preferred embodiment of the present invention, the discriminator is composed of multiple layers of fully connected layers and deconvolution layers, and specifically performs the following processing:

[0061] The discriminator input described in (2.1) includes the real fault data X real and the simulated fault data X generated by the generator fake ;

[0062] (2.2) The convolution layer extracts local features from the data through filters and downsamples through the Max Pooling layer to reduce the amount of computation while retaining the main features;

[0063] The output of the discriminator described in (2.3) is a multidimensional vector, each data represents the probability that the input corresponding fault category is a real fault data, and the Sigmoid activation function is used to make the output value between 0 and 1, and the Dropout technology is used to prevent overfitting;

[0064] (2.4) By distinguishing the authenticity of the input data, the classification ability of real fault data and simulated fault data is optimized, and the error loss between categories is maximized, so as to enhance the discrimination of simulated samples and the difference relationship between physical characteristics.

[0065] As a preferred embodiment of the present invention, the generative adversarial neural network GAN performs model training on the data generation module in the following manner:

[0066] (a) using the original real fault data to train the discriminator and maximize the output of the discriminator for the real data;

[0067] (b) training the discriminator using the simulated fault data generated by the generator, minimizing the output error of the discriminator to the generated data, and maximizing the classification loss of samples of different categories;

[0068] (c) fixing the discriminator and back-propagating the parameters required to update the generator so that the data generated by the generator can deceive the discriminator;

[0069] (d) Repeat the above steps until the GAN network model converges, that is, the discriminator cannot distinguish between real data and generated data, so that the generator can generate simulated data that is close to the distribution of real fault data.

[0070] As a preferred embodiment of the present invention, the GAN network model represents the objective function in the following manner:

[0071]

[0072] Where x represents real fault data, G(z) represents simulated fault data generated by the generator, and D(x) and D(G(z)) represent the output probabilities of the discriminator for real fault data and generated data, respectively.

[0073] As a preferred embodiment of the present invention, the fault detection and analysis module specifically performs the following processing:

[0074] (e) Extracting the temporal features of the input power grid data through the convolutional layer;

[0075] (f) The physical calculation features of the power grid are fused with the high-level features extracted by the convolutional layer through the physical feature fusion layer;

[0076] (g) The fusion features are further extracted through the fully connected layer for the input of the classifier, thereby outputting the probability distribution of the fault type and distinguishing the fault type.

[0077] See also Figure 3 As shown, as a preferred embodiment of the present invention, the fault detection and analysis module minimizes the difference between the actual fault type and the predicted fault type through a customized multi-objective loss function, specifically: the following method is used to weight the cross entropy loss L CE and mean square error L MSE Jointly optimize the multi-objective loss function L:

[0078] L=α·L CE +β·L MSE

[0079] Among them, α and β are weight coefficients, and the cross entropy loss calculation formula is:

[0080]

[0081] Among them, Y true To minimize the actual fault type, Y pred To predict the fault type, i represents the i-th category label, represents the predicted probability of the i-th category.

[0082] The technical solution will be further described in detail below:

[0083] The data generation module generates data of simulated power grid faults based on the GAN network model, which is used to supplement the training set of the fault detection module to improve the robustness and accuracy of the fault detection system. The specific structure of GAN is as follows:

[0084] Generator (G): The generator consists of multiple layers of fully connected layers and deconvolution layers. First, the input random noise z or normal operation data X normal After the fully connected layer performs feature extraction, the output is a high-dimensional vector. The high-dimensional vector is gradually upsampled through the deconvolution layer to generate simulated data X with the same dimension as the actual fault data. fake In order to improve the generation ability of the generator, batch normalization technology is used, and the Leaky ReLU activation function is added to each layer. The goal of the generator is to make the generated simulated data as close as possible to the actual fault data in the feature space through learning.

[0085] Discriminator (D): The discriminator consists of multiple convolutional layers and fully connected layers. The input of the discriminator includes the real fault data X real and the simulated fault data X generated by the generator fake . The convolution layer extracts local features from the data through filters, and downsamples through the Max Pooling layer to reduce the amount of calculation while retaining the main features. The output of the discriminator is a multidimensional vector, and each data represents the probability that the input corresponding fault category is real fault data. The discriminator uses the Sigmoid activation function to make the output value between 0 and 1, and uses the Dropout technology to prevent overfitting. The task of the discriminator is to distinguish the true and false of the input data, optimize its classification ability for real fault data and simulated fault data, and maximize the error loss between categories, and enhance the distinction of simulated samples and the relationship between physical feature differences.

[0086] GAN training process: Step 1: Train the discriminator with real fault data to maximize the output of the discriminator for real data. Step 2: Train the discriminator with simulated fault data generated by the generator to minimize the output error of the discriminator for generated data and maximize the classification loss of samples of different categories. Step 3: Fix the discriminator and back-propagate to update the parameters of the generator so that the data generated by the generator can deceive the discriminator. Step 4: Repeat the above steps until the GAN model converges, that is, the discriminator cannot distinguish between real data and generated data. Finally, the generator can generate simulated data with a distribution close to that of real fault data.

[0087] The objective function of the GAN network model is expressed as:

[0088]

[0089] Where x represents real fault data, G(z) represents simulated fault data generated by the generator, and D(x) and D(G(z)) represent the output probabilities of the discriminator for real fault data and generated data, respectively.

[0090] This part specifically represents the target optimization process of the generator and the discriminator in the GAN process. V(D,G) refers to the objective function of the generative adversarial network, which includes two parts: 1. The logarithmic probability expectation for real data, which reflects the discriminator's ability to identify real data; 2. The logarithmic probability expectation for the data generated by the generator, which reflects the discriminator's ability to identify generated data. The optimized min-max problem describes the different goals of the generator and the discriminator. maxD maximizes the discriminator's ability to distinguish between real data x and generated data G(z) as much as possible. It expects the probability of the output of real data D(x) to be as close to 1 as possible, and the output of generated data D(G(z)) to be as close to 0 as possible. Therefore, the discriminator maximizes V(D,G) to improve its ability to distinguish between real and simulated data. minG minimizes the loss of the generator, so that it can generate realistic simulated data as much as possible, so that the discriminator cannot distinguish these data from real data. It expects the probability of the output of D(G(z)) to be as close to 1 as possible. Therefore, the generator minimizes V(D,G) so that the discriminator cannot distinguish whether the generated data is simulated.

[0091] x~pdata(x) means sampling x from the real data distribution pdata, that is, taking the real data point; the formula is the expected value of the real data, which means that the discriminator hopes to make the probability D(x) of the real data as close to 1 as possible, maximizing this part to indicate that the data x is real data.

[0092] z~pz(z) means sampling noise z from the noise distribution pz(z); the data generated by the generator is required to convince the discriminator, thus maximizing the expectation of log(1-D(G(z))).

[0093] The generator G obtained through training generates a large amount of simulated fault data, including data of rare fault types. These data can supplement the deficiencies of actual fault data and expand the diversity of the data set. The generated data is compared and analyzed with the actual historical fault data, and the validity of the generated data is verified through indicators such as feature distribution and time series graph similarity. The generated data is evaluated using an independent validation set to ensure that it is consistent with the statistical distribution of the real fault data.

[0094] The main purpose of comparing and analyzing the generated data with the actual historical fault data is to further extract the simulated data generated by the trained generator through numerical analysis, and retain the data that maintains a consistent statistical distribution with the real fault data, so as to obtain higher quality simulated data. The specific processing process is supplemented as follows:

[0095] Key features are extracted from the generated fault data and actual historical fault data, including voltage and current peak values, mean and variance of voltage and current, harmonic content, fault duration and phase angle, etc. Each feature is extracted as a feature vector, and the features of the generated data and historical data are compared using the probability density method to check whether the distribution of the two sets of data on each feature is consistent, and the difference index of the two distributions is calculated at the same time. To ensure the validity of the generated data, this module will also extract an independent validation set from the actual historical fault data that does not participate in the generator training. The validation set should contain various types of fault data to ensure its representativeness and diversity. The module uses the non-parametric test KS method to verify whether the cumulative distribution functions of the generated data and the validation set data are similar, and eliminate the simulated data with poor statistical distribution consistency. The final data generation module can obtain high-quality simulated generated data, which can better ensure the validity of the data and the recognition accuracy of the subsequent fault detection module.

[0096] In the power grid, common fault types include single-phase grounding fault, multi-phase short circuit fault, open circuit fault, etc. GAN generates various types of fault data with high discrimination by maximizing the classification loss, and trains the detection model so that it can accurately distinguish different types of faults. Several common types of power grid faults and their characteristic data are as follows: (1) Single-phase grounding fault: Single-phase grounding fault is usually manifested as a sudden decrease in voltage and a sharp increase in current between a single power line and the ground. The generator generates data with similar voltage drop and current peak characteristics based on the input noise or normal data; (2) Multi-phase short circuit fault: Multi-phase short circuit fault is manifested as a sharp increase in current between two-phase or three-phase power lines and a significant decrease in phase voltage. The generator generates voltage and current waveform data simulating short circuit faults based on the characteristics of multi-phase short circuit; (3) Open circuit fault: A open circuit fault causes the current of one or more power lines to suddenly become zero. The open circuit fault data generated by the generator is characterized by the sudden return of current to zero. The generated fault data flows to the fault detection module, and is combined with real fault data for deep learning training, so that the model can accurately distinguish various types of faults.

[0097] The fault detection and analysis module is based on the deep learning model and integrates physical computing features, which can detect and identify various types of faults in real time. Its input is the grid operation data (including generated virtual data and actual data). The model detects the occurrence of faults and distinguishes the fault types by learning the timing characteristics of parameters such as voltage and current.

[0098] The fault detection model integrates physical features, which integrates the physical calculation features of the power grid with the time series features of the simulated data generated by GAN. Specifically, the calculated physical features are used as part of the loss function, together with the high-level features output by the convolutional layer, and connected to the fully connected layer for further feature extraction and fusion. The final fused features are used as input to the classifier to help distinguish the fault type.

[0099] The fault detection model structure mainly includes: (1) convolutional layer: used to extract the time series features of the input power grid data; (2) physical feature fusion layer: fuses the physical calculation features of the power grid with the high-level features extracted by the convolutional layer; (3) fully connected layer: further extracts the fused features and outputs the probability distribution of the fault type.

[0100] Through this design, the network is not only able to perform fault detection based on data-driven, but also combines the precise calculation of the physical model to improve the overall recognition ability of the model.

[0101] Assume that the input data is X, and the output is the probability distribution of fault classification P(Y|X), where Y represents the fault type. The model minimizes the actual fault type Y through a custom multi-objective loss function. true and predicted fault type Y pred The difference between them can improve the accuracy of fault detection. The loss function is a joint optimization of weighted cross entropy loss (Weighted Cross Entropy Loss) and mean squared error (Mean Squared Error, MSE):

[0102] L=α·L CE +β·L MSE

[0103] Among them, α and β are weight coefficients, which are adjusted through experiments to obtain the optimal model. The cross entropy loss calculation formula is:

[0104]

[0105] Among them, Y true To minimize the actual fault type, Y pred is to predict the fault type, i is the i-th category label, Represents the predicted probability of the i-th category. The cross entropy loss calculation is to traverse the logarithmic probability product of all categories and compare the difference between the true category label and the predicted probability.

[0106] Physical characteristics include characteristic calculations for power system fault analysis:

[0107] (1) Single-phase grounding fault:

[0108] The voltage at the fault point V f It can be expressed as:

[0109]

[0110] Where Z f is the fault point impedance, Z line is the line impedance, V source is the source voltage.

[0111] Fault current I f It can be expressed as:

[0112]

[0113] (2) Multi-phase short circuit fault:

[0114] Assume that the fault occurs when the phase-to-phase voltage is V ab , the fault current is I ab :

[0115]

[0116] Among them, Z ab is the impedance between the two phases at the fault point.

[0117] These power system calculation logics are combined with the virtual data generated by GAN to enhance the model's detection and fault differentiation capabilities through feature analysis.

[0118] In actual application, the alarm and response module can automatically trigger the alarm mechanism and provide corresponding fault handling suggestions after detecting a fault. When the fault detection and classification model detects an abnormality and determines the type of fault, the alarm module will be automatically triggered. Different fault types will trigger different alarm levels and notify the operation and maintenance personnel in real time through a visual user interface. At the same time, the module provides a detailed fault report, including the type of fault, the time of occurrence, the power lines or equipment involved, and the recommended handling measures. The system supports sending alarm notifications to relevant personnel through SMS, email, etc.

[0119] In actual application, the data storage and management module stores all fault detection data, generated data, calculated physical characteristics and fault handling results in the system database. The database adopts a hierarchical storage structure and stores data according to time, equipment, fault type, etc., which is convenient for retrieval and subsequent analysis. The data storage module supports the tracing and export of historical data, providing data support for the continuous optimization and improvement of the system.

[0120] This technical solution has the following technical advantages in practical application:

[0121] 1. Improve the sensitivity of fault detection: Generate diversified fault data by maximizing classification loss through GAN, and enhance the detection system's ability to identify various faults.

[0122] 2. Enhanced differentiation capability: Compared with traditional methods, this system integrates physical features to more accurately distinguish different types of faults and reduce the probability of false alarms and missed alarms.

[0123] 3. Data expansion and model optimization: By generating virtual fault data, the problem of scarce power grid fault data is solved, and the detection model can be continuously optimized.

[0124] Through the implementation of the above technical solutions, the system can effectively improve the accuracy and sensitivity of power grid fault detection, while greatly improving the automation level of fault handling, making it easier for operation and maintenance personnel to carry out daily management and fault handling.

[0125] For example, in an actual operation, the system successfully identified a multi-phase short-circuit fault type that had never occurred before through the generated virtual fault data, and accurately located the fault with the support of physical calculations. The alarm module was quickly activated and automatically isolated the fault line, avoiding the occurrence of a large-scale power outage.

[0126] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code that includes one or more executable instructions for implementing the steps of a specific logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present invention belong.

[0127] It should be understood that each part of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution device.

[0128] A person of ordinary skill in the art may understand that all or part of the steps of the method for implementing the above-mentioned embodiment may be completed by instructing the relevant hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one of the steps of the method embodiment or a combination thereof.

[0129] The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc.

[0130] In the description of this specification, the description with reference to the terms "an embodiment", "some embodiments", "example", "specific example", or "embodiment" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0131] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

[0132] The power grid fault detection system based on the generative adversarial neural network of the present invention has the following technical advantages: (1) Improving detection sensitivity: GAN is used to generate a large amount of high-quality simulated fault data to make up for the deficiency of actual fault data, so that the detection model can still maintain high sensitivity and accuracy when facing rare or complex fault types. (2) Expanding the coverage of the data set: By generating diversified simulated data, the training data set of the fault detection model is expanded so that it can cover more potential fault scenarios and improve the generalization ability of the model. (3) Enhancing the ability to distinguish faults: By combining the generated simulated data with the actual fault data and incorporating the loss calculation of physical characteristics to optimize the feature extraction and classification capabilities of the fault detection model, it can accurately distinguish similar types of faults and improve the accuracy of fault handling.

[0133] In this specification, the present invention has been described with reference to specific embodiments thereof. However, it is apparent that various modifications and variations may be made without departing from the spirit and scope of the present invention. Therefore, the specification and drawings should be regarded as illustrative rather than restrictive.

Claims

1. A power grid fault detection system based on generative adversarial neural network, characterized in that: The system comprises: The data acquisition and preprocessing module is used to collect real-time operation data and historical storage data from the power grid bus and various databases, obtain data sets of different fault types, and perform denoising and standardization processing on them; A data generation module, connected to the data acquisition and preprocessing module, for generating virtual fault data similar to actual fault data based on a generative adversarial neural network (GAN); A fault detection and analysis module, connected to the data generation module and the data acquisition and preprocessing module, is used to integrate physical computing features into the fault detection system built by the deep learning model, and to efficiently detect and distinguish different types of faults through model training; An alarm and response module, connected to the fault detection and analysis module, is used to trigger an alarm mechanism after a fault is detected and the fault type is determined, and to provide corresponding processing suggestions or automatically perform corresponding fault processing operations according to the fault type; and The data storage and management module is connected to the data acquisition and preprocessing module and is used to store all fault data, generated data and test results, and provide historical data support for the system for future system optimization and update.

2. The power grid fault detection system based on generative adversarial neural network according to claim 1 is characterized in that: The generative adversarial neural network GAN consists of a generator and a discriminator, wherein the generator is used to input random noise or normal operation data to generate fault-like data; the discriminator is used to evaluate the generated data and compare it with the real fault data, so as to guide the generator to generate more realistic data.

3. The power grid fault detection system based on generative adversarial neural network according to claim 2 is characterized in that: The generator is composed of multiple layers of fully connected layers and deconvolution layers, and specifically performs the following processing: (1.1) Input random noise z or normal operation data X normal Feature extraction is performed through the fully connected layer, and the output is a high-dimensional vector; The high-dimensional vector described in (1.2) is gradually upsampled through the deconvolution layer to generate simulated data X with the same dimension as the actual fault data fake ; (1.3) Batch Normalization technology is used and Leaky ReLU activation function is added to each layer to make the generated simulation data as close as possible to the actual fault data in the feature space.

4. The power grid fault detection system based on generative adversarial neural network according to claim 3 is characterized in that: The discriminator is composed of multiple layers of fully connected layers and deconvolution layers, and specifically performs the following processing: The discriminator input described in (2.1) includes the real fault data X real and the simulated fault data X generated by the generator fake ; (2.2) The convolution layer extracts local features from the data through filters and downsamples through the Max Pooling layer to reduce the amount of computation while retaining the main features; The output of the discriminator described in (2.3) is a multidimensional vector, each data represents the probability that the input corresponding fault category is a real fault data, and the Sigmoid activation function is used to make the output value between 0 and 1, and the Dropout technology is used to prevent overfitting; (2.4) By distinguishing the authenticity of the input data, the classification ability of real fault data and simulated fault data is optimized, and the error loss between categories is maximized, so as to enhance the discrimination of simulated samples and the difference relationship between physical characteristics.

5. The power grid fault detection system based on generative adversarial neural network according to claim 4 is characterized in that: The generative adversarial neural network GAN performs model training on the data generation module in the following manner: (a) using the original real fault data to train the discriminator and maximize the output of the discriminator for the real data; (b) training the discriminator using the simulated fault data generated by the generator, minimizing the output error of the discriminator to the generated data, and maximizing the classification loss of samples of different categories; (c) fixing the discriminator and back-propagating the parameters required to update the generator so that the data generated by the generator can deceive the discriminator; (d) Repeat the above steps until the GAN network model converges, that is, the discriminator cannot distinguish between real data and generated data, so that the generator can generate simulated data that is close to the distribution of real fault data.

6. The power grid fault detection system based on generative adversarial neural network according to claim 5 is characterized in that: The GAN network model expresses the objective function in the following way: Where x represents the real fault data, G(z) represents the simulated fault data generated by the generator, D(x) and D(G(z)) represent the output probabilities of the discriminator for the real fault data and the generated data, respectively, V(D,G) represents the objective function of the generative adversarial network, x~p data (x) represents the distribution of real data p data Sampling x in the middle, that is, taking real data points, z~p z (z) represents the noise distribution p z The sampling noise z in (z) is 7. The power grid fault detection system based on generative adversarial neural network according to claim 5, characterized in that: The fault detection and analysis module specifically performs the following processing: (e) Extracting the temporal features of the input power grid data through the convolutional layer; (f) The physical calculation features of the power grid are fused with the high-level features extracted by the convolutional layer through the physical feature fusion layer; (g) The fusion features are further extracted through the fully connected layer for the input of the classifier, thereby outputting the probability distribution of the fault type and distinguishing the fault type.

8. The power grid fault detection system based on generative adversarial neural network according to claim 7 is characterized in that: The fault detection and analysis module minimizes the difference between the actual fault type and the predicted fault type through a customized multi-objective loss function, specifically: the following method is used to weight the cross entropy loss L CE and mean square error L MSE Jointly optimize the multi-objective loss function L: L=α·L CE +β·L MSE Among them, α and β are weight coefficients, and the cross entropy loss calculation formula is: Among them, Y true To minimize the actual fault type, Y pred To predict the fault type, i represents the i-th category label, represents the predicted probability of the i-th category.

Images