Power grid fault diagnosis method and system based on FFT-CNN-GCN hybrid network

Abstract

The invention discloses a power grid fault diagnosis method and system based on an FFT-CNN-GCN hybrid network. The method comprises the steps that fault type judgment, fault line positioning and fault distance judgment of a power grid line are achieved at the same time; the method comprises the following steps of: performing time domain and frequency domain decomposition on voltage and current signals through fast Fourier transform (FFT), and extracting fundamental frequency and phase of the signals; performing time sequence feature extraction on the decomposed data by using CNN, and introducing a layer normalization technology to enhance the stability and performance of the model; and processing a spatial topological structure of the power grid by combining with the GCN, and extracting and integrating spatial features. Under the influence of different tasks, sampling intervals and different noise, the accuracy of the method is obviously superior to that of an existing method based on artificial intelligence, and the method has good generalization performance.

Description

Technical Field

[0001] The invention belongs to the field of FFT-CNN-GCN hybrid network feature extraction and fault diagnosis, and specifically relates to a power grid fault diagnosis method and system based on the FFT-CNN-GCN hybrid network. Background Art

[0002] Faults often occur in power grid systems. In order to determine the fault type and location for maintenance, the causes of power grid faults vary. Distribution network line faults mainly include single-phase grounding fault (LG), two-phase short circuit fault (LL), two-phase grounding fault (LLG), three-phase short circuit fault (LLL), etc.

[0003] In order to diagnose fault types more quickly and locate them accurately, Petri nets, multi-source information fusion technology, and optimization methods have been widely used in recent years. A study proposed a fault segment identification method for intelligent distribution networks based on fuzzy Petri nets and multi-source data to realize fault area identification in intelligent distribution networks. A study transformed the fault diagnosis problem into a multi-objective optimization problem and used a non-dominated sorting genetic algorithm to solve it, avoiding the problem of poor adaptability of the solution caused by artificially set weights. Some other studies have established a linear integer programming model to locate the circuit breaker fault section. However, the above method is only applicable to power grids with a small number of grid nodes.

[0004] For power grids with a large number of nodes, machine learning and deep learning-based methods have received widespread attention because they can learn from data and automatically adjust weights and parameters through training. Some studies have combined the improved adaptive genetic algorithm with the BP neural network to locate cable break faults, improving the accuracy of cable fault location. Some studies have used graph attention networks to locate small current grounding faults in distribution networks, considering the network topology and the influence between adjacent nodes from a spatial perspective, and improving the ability of fault location models to extract features. Some studies have also achieved fault type discrimination and fault location based on the SVM model by obtaining the phase voltage fundamental and 2-7 harmonic components of each measurement point. However, the above models cannot comprehensively consider the time and space characteristics when processing actual power system data, and have certain limitations, resulting in low accuracy. Existing studies have proposed a deep learning model with spatiotemporal convolution blocks, which comprehensively considers the time and space characteristics and realizes a learning framework for fault location and fault type classification. However, the models proposed in the above literature have not yet formed a complete fault diagnosis system, and cannot simultaneously achieve the three key tasks of fault type classification, fault line location, and fault distance judgment.

[0005] To address the above problems, there is an urgent need for a fault diagnosis model that can simultaneously achieve the three key tasks of fault type classification, fault line location and fault distance judgment. The model should be able to adapt to multi-node systems, have the ability to process large-scale data, automatically learn features from data, and be able to extract features in the time domain, frequency domain, and time and space dimensions. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention proposes a power grid fault diagnosis method and system based on an FFT-CNN-GCN hybrid network, and proposes a new fault diagnosis model - FFT-CNN-GCN, which constructs a complete fault diagnosis framework that can simultaneously realize the three key tasks of fault type classification, fault line location and fault distance judgment. This model combines the feature extraction capability of CNN and the advantages of GCN in graph structure data, can adapt to large-scale data processing in the case of a large number of nodes, automatically learn features from data, and reduce dependence on parameter accuracy. By extracting features in the time domain, frequency domain, and time and space dimensions, the accuracy and efficiency of fault diagnosis are improved.

[0007] To achieve the above object, the present invention provides the following solutions:

[0008] A power grid fault diagnosis method based on FFT-CNN-GCN hybrid network includes the following steps:

[0009] Obtain the original three-phase voltage and three-phase current data of the power grid;

[0010] Build the FFT-CNN-GCN model;

[0011] Based on the original three-phase voltage and three-phase current data, the FFT-CNN-GCN model is used to process the power grid topology, capture the spatial correlation between power grid nodes, and finally determine the fault type and the specific location of the fault.

[0012] Preferably, the method for obtaining original three-phase voltage and three-phase current data of the power grid includes:

[0013] With sampling frequency f s =1000HZ collects the three-phase voltage v of node j j,a (n),v j,b (n),v j,c (n) and three-phase current i j,a (n),i j,b (n),i j,c (n) is a discrete time signal, the sampling time is T = 0.3s, and the number of sampling points for one sampling is N = f s ×T=300.

[0014] Preferably, based on the original three-phase voltage and three-phase current data, the FFT-CNN-GCN model is used to process the power grid topology structure, capture the spatial correlation between the power grid nodes, and finally determine the fault type and the specific location of the fault. The method includes:

[0015] The original three-phase voltage and three-phase current data are decomposed in time domain and frequency domain by fast Fourier transform (FFT) to extract the fundamental frequency and phase of the original data;

[0016] Use CNN to extract time series features of fundamental frequency and phase of raw data;

[0017] Introduce layer normalization to process the extracted temporal features;

[0018] Combined with GCN, the spatial topological structure of the power grid is processed, the spatial characteristics between power grid nodes are extracted and integrated, and finally the fault type and the specific location of the fault are determined.

[0019] Preferably, the method of performing time domain and frequency domain decomposition on the original three-phase voltage and three-phase current data by fast Fourier transform FFT to extract the fundamental frequency and phase of the original data includes:

[0020] The frequency domain representation of the original three-phase voltage and three-phase current data is calculated using the fast Fourier transform FFT, that is, V j,a (k),V j,b (k),V j,c (k) and I j,a (k),I j,b (k),I j,c (k);

[0021] Calculate the amplitude and phase angle of the fundamental wave; the calculation method is: the index k1 corresponding to the fundamental wave frequency = f base ·N' / f s ;

[0022] Among them, amplitude calculation: the fundamental wave amplitude is calculated by the output of fast Fourier transform FFT, and the expression is:

[0023] Among them, phase angle calculation: the fundamental wave phase angle is obtained from the phase part of the fast Fourier transform FFT, and the expression is: θ=arg(X(k)).

[0024] Preferably, the method of extracting time series features from fundamental frequency and phase of original data using CNN includes: CNN is composed of an input layer, a hidden layer and an output layer;

[0025] In the CNN model, the amplitude and phase angle of the voltage and current of all buses within a short period of time after the system fault are transformed by fast Fourier transform FFT as the input of CNN, with fault type, fault line and fault distance as labels; the complete data in the observation window is arranged as the formula:

[0026]

[0027] Where N is the number of nodes, j = 1, 2, ..., N, and the subscripts a, b, and c represent the A, B, and C phases in the power grid data, respectively. j,a ,I j,b ,I j,c Respectively represent the fundamental amplitude of the current data of phase A, phase B, and phase C of the jth node, Respectively represent the phase angle of the j-th node A phase current data, V j,a ,V j,b ,V j,c Respectively represent the fundamental amplitude of the voltage data of phase A, phase B, and phase C of the jth node, They respectively represent the phase angle of the A-phase voltage data of the j-th node;

[0028] For the convolutional layer in the hidden layer, use H i represents the feature map of the i-th layer; H i The convolution operation is performed through the convolution kernel and processed by the activation function, that is,

[0029]

[0030] Where: f(x) is the ReLU activation function; represents the convolution operation; W i is the weight matrix of the i-th convolution kernel; b i is the bias vector of the i-th layer;

[0031] For the pooling layer in the hidden layer, the calculation process is expressed as:

[0032] H i =Maxpool(H i-1 )

[0033] Among them: Maxpool represents the maximum pooling operation;

[0034] For the fully connected layer in the hidden layer, the feature map of the previous layer is convolved through the convolution kernel and processed by the activation function to obtain the output feature map, which is expressed as:

[0035] H i =f(H i-1 W i +b i );

[0036] The output feature map is used to train the CNN model to extract temporal features. Specifically, based on the output feature map, high-level features are first gradually extracted through three layers of convolution-pooling operations, and then the features are processed through global average pooling to reduce the dimension of each feature map to a single scalar value.

[0037] Preferably, the method of introducing layer normalization to process the extracted temporal features includes:

[0038] The layer normalization layer calculates the mean μ and variance σ of each sample along the channel dimension 2 ;

[0039] Normalize the input feature map by subtracting the mean and dividing by the variance;

[0040] Using the learnable parameters γ and β, the normalized feature map is affine transformed to obtain X i ; The specific calculation method is:

[0041]

[0042] Among them, S is the number of samples in a training batch, H is the number of channels of the hidden layer feature map, L and W are the length and width of the hidden layer feature map respectively, and the result after layer normalization is X∈R 7 × 384 .

[0043] Preferably, the method of combining GCN to process the spatial topological structure of the power grid, extracting and integrating the spatial features between the power grid nodes, and finally determining the fault type and the specific location where the fault occurs includes:

[0044] The graph convolution layer is used to process the spatial topological structure of the features. The specific two-layer convolution process is as follows:

[0045]

[0046] in, I is the identity matrix, yes The degree matrix of (0) and W (1) Respectively represent the trainable weight matrices of the first and second layers of GCL; X (0) =X is the input feature matrix, X (2) is the output of the second layer GCL of GCN, ReLU(·) is the activation function;

[0047] In the process of graph convolution, the association information between nodes is fully utilized to update and aggregate the voltage and current features between nodes;

[0048] The features extracted by convolution are combined with the graph structure information and mapped to the normalized probabilities of the labels of fault type, fault line and fault distance, and finally the labels are output by the output layer.

[0049] The present invention also provides a power grid fault diagnosis system based on FFT-CNN-GCN hybrid network, comprising: an acquisition module, a construction module and a judgment module;

[0050] The acquisition module is used to acquire original three-phase voltage and three-phase current data of the power grid;

[0051] The building block is used to build the FFT-CNN-GCN model;

[0052] The judgment module is used to process the power grid topology structure based on the original three-phase voltage and three-phase current data using the FFT-CNN-GCN model, capture the spatial correlation between power grid nodes, and ultimately determine the fault type and the specific location where the fault occurs.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] This paper constructs an FFT-CNN-GCN model that can simultaneously realize fault type judgment, fault line location and fault distance judgment, overcoming the problems of existing methods in terms of high parameter precision requirements, insufficient accuracy and poor adaptability. Compared with existing artificial intelligence-based methods, this model combines the advantages of signal processing, convolutional neural networks and graph convolutional networks, extracts features from data in the time domain and frequency domain, as well as in time and space, with high accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0056] Figure 1 is a flow chart of a power grid fault diagnosis method based on a FFT-CNN-GCN hybrid network in an embodiment of the present invention;

[0057] Figure 2 is the CNN structure in the embodiment of the present invention;

[0058] Figure 3 The structure of GCN in fault location in an embodiment of the present invention;

[0059] Figure 4 Schematic diagram of the GCL structure in an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0062] Embodiment 1

[0063] The present invention proposes a power grid fault diagnosis method based on FFT-CNN-GCN hybrid network. Figure 1 As shown, the specific steps include:

[0064] Obtain the original three-phase voltage and three-phase current data of the power grid;

[0065] Build the FFT-CNN-GCN model;

[0066] Based on the original three-phase voltage and three-phase current data, the FFT-CNN-GCN model is used to process the power grid topology, capture the spatial correlation between power grid nodes, and finally determine the fault type and the specific location of the fault.

[0067] Specifically, based on the original three-phase voltage and three-phase current data, the FFT-CNN-GCN model is used to process the power grid topology structure, capture the spatial correlation between power grid nodes, and finally determine the fault type and the specific location of the fault. The method includes:

[0068] The original three-phase voltage and three-phase current data are decomposed in time domain and frequency domain by fast Fourier transform (FFT) to extract the fundamental frequency and phase of the original data;

[0069] Use CNN to extract time series features of fundamental frequency and phase of raw data;

[0070] Introduce layer normalization to process the extracted temporal features;

[0071] Combined with GCN, the spatial topological structure of the power grid is processed, the spatial characteristics between power grid nodes are extracted and integrated, and finally the fault type and the specific location of the fault are determined.

[0072] Specifically, obtaining the original three-phase voltage and three-phase current data, and using FFT to extract the fundamental frequency and fundamental phase of the original data include:

[0073] Collect signals and obtain original three-phase voltage and three-phase current data;

[0074] Apply FFT to calculate the frequency domain representation of the acquired signal;

[0075] Fundamental amplitude and phase angle calculation.

[0076] Specifically, collect signals to obtain original three-phase voltage and three-phase current data:

[0077] The IEEE39-bus system model was simulated using PSCAD. To obtain sample data, each experiment assumes that one of the buses 15, 16, 17, 19, 21, and 24 fails at 0.35s and lasts for 0.15s. The measuring device samples the three-phase current and three-phase voltage values ​​at intervals of 0.001s from 0.3s to 0.6s. In order to accurately locate faults at different locations, this paper simulates all ten types of faults at nine different locations of the transmission line under study (these locations are 10%, 20%, ..., 90% away from the bus node, respectively).

[0078] Furthermore, FFT is applied to calculate the frequency domain representation of the acquired signal:

[0079] FFT is an algorithm for efficiently calculating the frequency components of discrete signals:

[0080]

[0081] The basic idea is to decompose a signal x(n) of length N into two sub-signals x(2n) and x(2n+1) of length N / 2, and then calculate the FFT of each sub-signal, and finally convert it into the frequency domain representation X[k]. In formula (1), X[k] is the kth frequency component of the frequency domain signal. It is the basis function of discrete Fourier transform and is used to adjust the phase of frequency components.

[0082] Going further, calculate the amplitude and phase angle of the fundamental wave:

[0083] Use FFT to extract the fundamental amplitude and phase angle of three-phase voltage and current. The specific steps are as follows:

[0084] Acquisition signal: at sampling frequency f s =1000HZ collects the three-phase voltage v of node j j,a (n),v j,b (n),v j,c (n) and three-phase current i j,a (n),i j,b (n),i j,c(n) discrete time signal, where n = 1, 2, ..., N'. The sampling time is T = 0.3s, so the number of sampling points in one sampling is N' = f s ×T=300.

[0085] According to formula (1), FFT is applied to the collected signals to obtain the frequency domain representation of the three-phase voltage and three-phase current, that is, V j,a (k),V j,b (k),V j,c (k) and I j,a (k),I j,b (k),I j,c (k).

[0086] Calculation of the amplitude and phase angle of the fundamental wave: The standard frequency of the IEEE39 node system is 60HZ, so in the frequency domain signal of this article, the frequency corresponding to the fundamental wave is f base =60Hz. Extract the amplitude and phase angle of the corresponding fundamental frequency. The index k1 corresponding to the fundamental frequency = f base ·N' / f s .

[0087] Amplitude calculation: The fundamental amplitude can be calculated from the output of FFT, that is, For the voltage signal of phase a at the jth node, the fundamental amplitude |V j,a | can be expressed as

[0088] Phase angle calculation: The fundamental phase angle can be obtained from the phase part of FFT, that is, θ = arg(X(k)). For the voltage signal V of phase a at the jth node j,a , and its phase angle can be expressed as

[0089] Furthermore, an FFT-CNN model is constructed to extract time series features:

[0090] A typical CNN consists of an input layer, a hidden layer, and an output layer, such as Figure 2 As shown in the figure, the input layer receives the original signal or preprocessed data. The hidden layer extracts features and reduces the dimension of the data received by the input layer, and is the core component of the convolutional neural network. The hidden layer includes convolutional layer, activation layer, pooling layer, and fully connected layer: the convolutional layer reduces the number of parameters through sparse connections and parameter sharing, reduces the complexity of training, and alleviates the problems of overfitting and dimensionality disaster; the pooling layer reduces the number of features, improves the robustness of the model to slight deformations, and enhances the generalization ability. The output layer performs probability calculations based on the features extracted by the hidden layer and classifies the input data into the corresponding categories.

[0091] In the CNN model, the amplitude and phase angle of the voltage and current of all buses within a short period of time after the system fault are selected as the input of CNN after FFT transformation, and the fault type, fault line and fault distance are used as labels. These data are first preprocessed into a grid topology similar to an image and suitable for CNN. The data window is 50 (each fault sample contains 50 time points) and the sliding step is 5. Then, when a fault occurs at each fault location, samples. These 50 time points mean that each fault sample used for fault diagnosis contains 0.02s of fault data. In this case, the proposed model can use any 0.02s of data within 0.15 seconds after the fault occurs to determine the fault type, fault branch, and fault distance. The length of the sampling window reflects the fault information contained in each sample. The longer the data window, the more fault information contained in the data sample, and the lower the difficulty of fault diagnosis. The complete data in the observation window is arranged as formula (2):

[0092]

[0093] Where N is the number of nodes, j = 1, 2, ..., N, and the subscripts a, b, and c represent the A, B, and C phases in the power grid data, respectively. j,a ,I j,b ,I j,c Respectively represent the fundamental amplitude of the current data of phase A, phase B, and phase C of the jth node, Respectively represent the phase angle of the j-th node A phase current data, V j,a ,V j,b ,V j,c Respectively represent the fundamental amplitude of the voltage data of phase A, phase B, and phase C of the jth node, They respectively represent the phase angle of the A-phase voltage data of the j-th node.

[0094] For the convolutional layer in the hidden layer, use H i H represents the feature map of the i-th layer. i The convolution operation can be performed through the convolution kernel and processed by the activation function, that is,

[0095]

[0096] Where: f(x) is the ReLU activation function; represents the convolution operation; W i is the weight matrix of the i-th convolution kernel; b i is the bias vector of the i-th layer.

[0097] For the pooling layer in the hidden layer, its calculation process can be expressed as:

[0098] H i=Maxpool(H i1 )(4)

[0099] Where: Maxpool represents the maximum pooling operation.

[0100] For the fully connected layer in the hidden layer, the feature map of the previous layer is convolved through the convolution kernel and processed by the activation function to obtain the output feature map, which can be expressed as:

[0101] H i =f(H i-1 W i +b i ) (5)

[0102] The training goal of CNN is to minimize the loss function. When used for multi-classification problems, the cross entropy loss function is used, which can be expressed as:

[0103]

[0104] In the formula, C represents the number of categories, y ij is the true label, is the prediction probability. During the training process, the Adam (adaptive moment estimation) algorithm, a variant of the gradient descent method, is used to improve the training effect by adjusting the learning rate of each parameter.

[0105] Further, the data is processed through a normalization layer:

[0106] In order to reduce the internal offset covariance, speed up the convergence of subsequent models, and improve the robustness and generalization ability of the model, a normalization layer is added. This paper adopts the layer normalization normalization method. First, the layer normalization layer calculates the mean μ and variance σ of each sample along the dimension H of the channel. 2 ; Then, the input feature map is normalized, that is, the mean is subtracted and divided by the variance; finally, the normalized feature map is affine transformed using the learnable parameters γ and β to obtain X i The specific calculation process is:

[0107]

[0108] Where S is the number of samples in a training batch, H is the number of channels in the hidden layer feature map, L and W are the length and width of the hidden layer feature map respectively. The result after layer normalization is X∈R 7 × 384 .

[0109] Finally, the FFT-CNN-GCN model is constructed, and the GCN's processing ability for the power grid topology is used to capture the spatial correlation between power grid nodes, and finally determine the fault type and the specific location of the fault:

[0110] This module first performs simple processing on the data processed by CNN, and then applies graph convolution operation to perform feature convolution. In the process of graph convolution, it makes full use of the correlation information between nodes, updates and aggregates the voltage and current features between nodes, and finally designs the loss function to improve the accuracy of fault diagnosis by optimizing the loss function of the model.

[0111] GCN is a deep learning model for processing graph data. The overall structure of GCN in this paper includes input layer, graph convolution layer (GCL), activation layer, dropout layer and output layer, such as Figure 3 As shown. GCN generally has two inputs: one input is the feature matrix X = (x1, x2, ..., x i ,…,x N ) T , where N is the total number of nodes in the power grid; x i =(x 1i ,...,x ri ,...,x Ri ) is the feature vector of node i, and R is the number of features. Another input represents the adjacency matrix A of the topology. A∈R N×N , if A ij =1, then there is an edge between nodes i and j; if A ij =0, then there is no edge between nodes i and j.

[0112] The GCN contains two layers of GCL, such as Figure 4 As shown. The first layer of GCL processes the input data and uses the ReLU function as the activation function; the second layer of GCL processes the output of the first layer of GCL to obtain the updated feature matrix. The specific two-layer convolution process is:

[0113]

[0114] in, I is the identity matrix, yes The degree matrix of (0) and W (1) Respectively represent the trainable weight matrices of the first and second layers of GCL; X (0) =X is the input feature matrix, X (2) is the output of the second layer GCL of GCN, ReLU(·) is the activation function.

[0115] After the convolutional layer processing, the node feature X is processed by a logarithmic softmax function. (2) Normalize and obtain the probability of failure, that is:

[0116]

[0117] Among them, p i =[p i,0 ,p i,1 ,...,p i,k ,...,p i,K-1 ] represents the probability vector predicted by the i-sample model, p i,k represents the probability that the sample belongs to category k, K represents a total of K categories. For each sample, the model will select the category with the largest predicted probability as the classification result of the sample, that is, where y i represents the predicted category label of the i-th sample, y i ∈{0,1,...,K-1}.

[0118] In order to measure the difference between the model predicted label and the actual label, the following negative log-likelihood loss function is defined:

[0119]

[0120] where y i is the true label of sample i, y i ∈{0,1,...,K-1} guides the updating of the model weight matrix and node feature matrix by optimizing the loss function, thereby improving the accuracy of the fault location model.

[0121] Embodiment 2

[0122] The present invention also provides a power grid fault diagnosis system based on FFT-CNN-GCN hybrid network, comprising: an acquisition module, a construction module and a judgment module;

[0123] The acquisition module is used to obtain the original three-phase voltage and three-phase current data of the power grid;

[0124] The building module is used to build the FFT-CNN-GCN model;

[0125] The judgment module is used to process the power grid topology structure based on the original three-phase voltage and three-phase current data using the FFT-CNN-GCN model, capture the spatial correlation between power grid nodes, and ultimately determine the fault type and the specific location of the fault.

[0126] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A power grid fault diagnosis method based on FFT-CNN-GCN hybrid network, characterized in that: The following steps are involved: Obtain the original three-phase voltage and three-phase current data of the power grid; Build the FFT-CNN-GCN model; Based on the original three-phase voltage and three-phase current data, the FFT-CNN-GCN model is used to process the power grid topology, capture the spatial correlation between power grid nodes, and finally determine the fault type and the specific location of the fault.

2. The power grid fault diagnosis method based on FFT-CNN-GCN hybrid network according to claim 1 is characterized in that: Methods for obtaining original three-phase voltage and three-phase current data of the power grid include: With sampling frequency f s =1000HZ collects the three-phase voltage v of node j j,a (n),v j,b (n),v j,c (n) and three-phase current i j,a (n),i j,b (n),i j,c (n) is a discrete time signal, the sampling time is T = 0.3s, and the number of sampling points for one sampling is N = f s ×T=300.

3. The power grid fault diagnosis method based on FFT-CNN-GCN hybrid network according to claim 2 is characterized in that: Based on the original three-phase voltage and three-phase current data, the FFT-CNN-GCN model is used to process the power grid topology structure, capture the spatial correlation between power grid nodes, and finally determine the fault type and the specific location of the fault. The method includes: The original three-phase voltage and three-phase current data are decomposed in time domain and frequency domain by fast Fourier transform (FFT) to extract the fundamental frequency and phase of the original data; Use CNN to extract time series features of fundamental frequency and phase of raw data; Introduce layer normalization to process the extracted temporal features; Combined with GCN, the spatial topological structure of the power grid is processed, the spatial characteristics between power grid nodes are extracted and integrated, and finally the fault type and the specific location of the fault are determined.

4. The power grid fault diagnosis method based on FFT-CNN-GCN hybrid network according to claim 3 is characterized in that: Methods for decomposing the original three-phase voltage and three-phase current data in the time domain and frequency domain by fast Fourier transform FFT and extracting the fundamental frequency and phase of the original data include: The frequency domain representation of the original three-phase voltage and three-phase current data is calculated using the fast Fourier transform FFT, that is, V j,a (k),V j,b (k),V j,c (k) and I j,a (k),I j,b (k),I j,c (k); Calculate the amplitude and phase angle of the fundamental wave; the calculation method is: the index k1 corresponding to the fundamental wave frequency = f base ·N' / f s ; Among them, amplitude calculation: the fundamental wave amplitude is calculated by the output of fast Fourier transform FFT, and the expression is: Among them, phase angle calculation: the fundamental wave phase angle is obtained from the phase part of the fast Fourier transform FFT, and the expression is: θ=arg(X(k)).

5. The power grid fault diagnosis method based on FFT-CNN-GCN hybrid network according to claim 3 is characterized in that: The method of extracting time series features of fundamental frequency and phase of original data by using CNN includes: CNN is composed of an input layer, a hidden layer and an output layer; In the CNN model, the amplitude and phase angle of the voltage and current of all buses within a short period of time after the system fault are transformed by fast Fourier transform FFT as the input of CNN, with fault type, fault line and fault distance as labels; the complete data in the observation window is arranged as the formula: Where N is the number of nodes, j = 1, 2, ..., N, and the subscripts a, b, and c represent the A, B, and C phases in the power grid data, respectively. j,a ,I j,b ,I j,c Respectively represent the fundamental amplitude of the current data of phase A, phase B, and phase C of the jth node, Respectively represent the phase angle of the j-th node A phase current data, V j,a ,V j,b ,V j,c Respectively represent the fundamental amplitude of the voltage data of phase A, phase B, and phase C of the jth node, They respectively represent the phase angle of the A-phase voltage data of the j-th node; For the convolutional layer in the hidden layer, use H i represents the feature map of the i-th layer; H i The convolution operation is performed through the convolution kernel and processed by the activation function, that is, Where: f(x) is the ReLU activation function; represents the convolution operation; W i is the weight matrix of the i-th convolution kernel; b i is the bias vector of the i-th layer; For the pooling layer in the hidden layer, the calculation process is expressed as: H i =Maxpool(H i-1 ) Among them: Maxpool represents the maximum pooling operation; For the fully connected layer in the hidden layer, the feature map of the previous layer is convolved through the convolution kernel and processed by the activation function to obtain the output feature map, which is expressed as: H i =f(H i-1 W i +b i ); The output feature map is used to train the CNN model to extract temporal features. Specifically, based on the output feature map, high-level features are first gradually extracted through three layers of convolution-pooling operations, and then the features are processed through global average pooling to reduce the dimension of each feature map to a single scalar value.

6. The power grid fault diagnosis method based on FFT-CNN-GCN hybrid network according to claim 3 is characterized in that: Methods for introducing layer normalization to process extracted temporal features include: The layer normalization layer calculates the mean μ and variance σ of each sample along the channel dimension 2 ; Normalize the input feature map by subtracting the mean and dividing by the variance; Using the learnable parameters γ and β, the normalized feature map is affine transformed to obtain X i ; The specific calculation method is: Among them, S is the number of samples in a training batch, H is the number of channels of the hidden layer feature map, L and W are the length and width of the hidden layer feature map respectively, and the result after layer normalization is X∈R 7×384 .

7. The power grid fault diagnosis method based on FFT-CNN-GCN hybrid network according to claim 3 is characterized in that: The methods for processing the spatial topological structure of the power grid by combining GCN, extracting and integrating the spatial features between the power grid nodes, and finally determining the fault type and the specific location of the fault include: The graph convolution layer is used to process the spatial topological structure of the features. The specific two-layer convolution process is as follows: in, I is the identity matrix, yes The degree matrix of (0) and W (1) Respectively represent the trainable weight matrices of the first and second layers of GCL; X (0) =X is the input feature matrix, X (2) is the output of the second layer GCL of GCN, ReLU(·) is the activation function; In the process of graph convolution, the association information between nodes is fully utilized to update and aggregate the voltage and current features between nodes; The features extracted by convolution are combined with the graph structure information and mapped to the normalized probabilities of the labels of fault type, fault line and fault distance, and finally the labels are output by the output layer.

8. A power grid fault diagnosis system based on FFT-CNN-GCN hybrid network, characterized in that: include: Acquisition module, construction module and judgment module; The acquisition module is used to acquire original three-phase voltage and three-phase current data of the power grid; The building block is used to build the FFT-CNN-GCN model; The judgment module is used to process the power grid topology structure based on the original three-phase voltage and three-phase current data using the FFT-CNN-GCN model, capture the spatial correlation between power grid nodes, and ultimately determine the fault type and the specific location where the fault occurs.

Images