Power distribution network traveling wave fault positioning method, apparatus and device, and storage medium

Abstract

The invention discloses a power distribution network traveling wave fault positioning method and device, equipment and a storage medium, which are used for solving the technical problem of low detection accuracy of weak faults. The method comprises the following steps: acquiring a voltage traveling wave signal after a fault occurs from a single-end detection point of a power distribution network, wherein the voltage traveling wave signal comprises a first wave head signal and a secondary wave head signal; converting the voltage traveling wave signal to obtain a time-frequency matrix of a panoramic waveform; inputting the time-frequency matrix into a pre-trained CNN-LSTM fusion model to obtain fault branch discrimination probability distribution and a fault distance; and positioning the traveling wave fault of the power distribution network according to the fault branch discrimination probability distribution and the fault distance.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution networks, and in particular, to a method, device, equipment and storage medium for traveling wave fault location in a distribution network. Background Art

[0002] Currently, there are mainly three mainstream methods for distribution network fault location: impedance method, distribution automation method and traveling wave method. The location accuracy of the impedance method is easily affected by factors such as the transition resistance at the fault point and the neutral grounding method, and in the case of many branches, there is a problem of misjudging the fault point easily. The distribution automation method has strong adaptability to complex branch lines. Although nowadays, automated methods such as the traveling wave method and the impedance method can locate the fault section, there are still challenges in achieving accurate fault point location.

[0003] In the existing medium and low voltage distribution network, a small current grounding system is generally adopted, which is also called a non-directly grounded neutral system. This system is affected by factors such as the wiring method of the distribution network neutral point and the aging of line insulation, and there are some conditions where it is difficult to trigger the protection device for faults. For example: faults near the zero crossing of the phase voltage in an ineffective grounding system, high-resistance grounding faults of overhead lines, arc faults of underground cables, etc. These are collectively called weak faults. After a weak fault occurs, the system is allowed to operate in the fault state for 1 to 2 hours. However, if these faults develop to a certain stage, they may cause an interphase short circuit, resulting in an expansion of the fault range, and ultimately causing the protection device to operate and trip, resulting in unplanned power outages, and thus causing significant economic losses. Therefore, it is particularly urgent to study an accurate and reliable distribution network fault location method, which helps to detect and eliminate weak faults in time and reduce the risk of system power outages. Summary of the Invention

[0004] The present invention provides a method, device, equipment and storage medium for traveling wave fault location in a distribution network, which is used to solve the technical problem of low detection accuracy of weak faults.

[0005] The present invention provides a method for traveling wave fault location in a distribution network, including:

[0006] Collecting a voltage traveling wave signal after a fault occurs at a single-end detection point of the distribution network, where the voltage traveling wave signal includes a first wave head signal and a second wave head signal;

[0007] Transforming the voltage traveling wave signal to obtain a time-frequency matrix of the panoramic waveform;

[0008] Inputting the time-frequency matrix into a pre-trained CNN-LSTM fusion model to obtain a fault branch discrimination probability distribution and a fault distance;

[0009] Locate the traveling wave fault of the distribution network according to the fault branch discrimination probability distribution and the fault distance.

[0010] Optionally, the step of transforming the voltage traveling wave signal to obtain the time-frequency matrix of the panoramic waveform includes:

[0011] Perform phase-mode transformation on the first wavehead signal to extract the line-mode component of the fault traveling wave;

[0012] Perform continuous wavelet transform on the line-mode component to obtain the time-frequency matrix of the panoramic waveform.

[0013] Optionally, the step of inputting the time-frequency matrix into the pre-trained CNN-LSTM fusion model to obtain the fault branch discrimination probability distribution and the fault distance includes:

[0014] Input the time-frequency matrix into the preset CNN network to extract the sub-wavehead frequency distribution characteristics of each sub-wavehead signal in the panoramic waveform;

[0015] Based on the sub-wavehead frequency distribution characteristics, extract the time correlation characteristics of the typical wavehead signal through the preset LSTM;

[0016] Perform weighted fusion on the sub-wavehead frequency distribution characteristics and the time correlation characteristics through the attention mechanism to obtain the context vector;

[0017] Map the context vector to the classification space through the classifier to obtain the fault branch discrimination probability distribution;

[0018] Map the context vector to the distance space to obtain the fault distance in the panoramic waveform.

[0019] Optionally, the step of performing weighted fusion on the sub-wavehead frequency distribution characteristics and the time correlation characteristics through the attention mechanism to obtain the context vector includes:

[0020] Use the sub-wavehead frequency distribution characteristics as the hidden layer eigenvalue of each moment of the LSTM;

[0021] Extract feature vectors by performing feature extraction on the time-frequency matrix;

[0022] Perform linear transformation on the feature vectors to generate query vectors;

[0023] Use the hidden layer eigenvalue and the query vector to generate the attention score of the time step;

[0024] Calculate the attention weight of the time step according to the attention score;

[0025] Use the hidden layer eigenvalue and the attention weight to generate the context vector.

[0026] The present invention also provides a traveling wave fault location device for a distribution network, comprising:

[0027] A voltage traveling wave signal acquisition module, configured to acquire voltage traveling wave signals after a fault occurs from a single - end detection point of the distribution network, where the voltage traveling wave signals include a first wave - head signal and a second wave - head signal;

[0028] A time - frequency matrix acquisition module, configured to transform the voltage traveling wave signals to obtain a time - frequency matrix of the panoramic waveform;

[0029] A fault branch discrimination probability distribution and fault distance acquisition module, configured to input the time - frequency matrix into a pre - trained CNN - LSTM fusion model to obtain a fault branch discrimination probability distribution and a fault distance;

[0030] A fault location module, configured to locate the traveling wave fault of the distribution network according to the fault branch discrimination probability distribution and the fault distance.

[0031] Optionally, the time - frequency matrix acquisition module includes:

[0032] A phase - mode transformation sub - module, configured to perform phase - mode transformation on the first wave - head signal to extract the line - mode component of the fault traveling wave;

[0033] A continuous wavelet transform sub - module, configured to perform continuous wavelet transform on the line - mode component to obtain a time - frequency matrix of the panoramic waveform.

[0034] Optionally, the fault branch discrimination probability distribution and fault distance acquisition module includes:

[0035] A second wave - head frequency distribution feature extraction sub - module, configured to input the time - frequency matrix into a preset CNN network to extract the second wave - head frequency distribution features of each second wave - head signal in the panoramic waveform;

[0036] A time - correlation feature extraction sub - module, configured to extract the time - correlation features of typical wave - head signals through a preset LSTM based on the second wave - head frequency distribution features;

[0037] A weighted fusion sub - module, configured to perform weighted fusion on the second wave - head frequency distribution features and the time - correlation features through an attention mechanism to obtain a context vector;

[0038] A fault branch discrimination probability distribution acquisition sub - module, configured to map the context vector to a classification space through a classifier to obtain a fault branch discrimination probability distribution;

[0039] A fault distance acquisition sub - module, configured to map the context vector to a distance space to obtain the fault distance in the panoramic waveform.

[0040] Optionally, the weighted fusion sub-module includes:

[0041] A hidden layer eigenvalue determination unit, configured to use the sub-wavehead frequency distribution feature as the hidden layer eigenvalue at each moment of the LSTM;

[0042] A feature extraction unit, configured to extract features from the time-frequency matrix to obtain a feature vector;

[0043] A linear transformation unit, configured to perform a linear transformation on the feature vector to generate a query vector;

[0044] An attention score generation unit, configured to generate an attention score for the time step by using the hidden layer eigenvalue and the query vector;

[0045] An attention weight calculation unit, configured to calculate an attention weight for the time step according to the attention score;

[0046] A context vector generation unit, configured to generate a context vector by using the hidden layer eigenvalue and the attention weight.

[0047] The present invention also provides an electronic device, which includes a processor and a memory:

[0048] The memory is configured to store program code and transmit the program code to the processor;

[0049] The processor is configured to execute the distribution network traveling wave fault location method according to any one of the above according to the instructions in the program code.

[0050] The present invention also provides a computer-readable storage medium, which is configured to store program code, and the program code is used to execute the distribution network traveling wave fault location method according to any one of the above.

[0051] As can be seen from the above technical solutions, the present invention has the following advantages: The present invention provides a distribution network traveling wave fault location method, and specifically discloses: collecting a voltage traveling wave signal after a fault occurs at a single-end detection point of the distribution network, where the voltage traveling wave signal includes a first wavehead signal and a sub-wavehead signal; performing a transformation on the voltage traveling wave signal to obtain a time-frequency matrix of a panoramic waveform; inputting the time-frequency matrix into a pre-trained CNN-LSTM fusion model to obtain a fault branch discrimination probability distribution and a fault distance; and locating the distribution network traveling wave fault according to the fault branch discrimination probability distribution and the fault distance. The present invention improves the accuracy and efficiency of fault location through the deep learning technology CNN-LSTM, and at the same time enhances the adaptability and robustness of the model, providing strong support for the stable operation of the power system. Description of the Drawings

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0053] Figure 1 It is a flowchart of the steps of a traveling wave fault location method for a distribution network provided by an embodiment of the present invention;

[0054] Figure 2 It is a traveling wave transmission grid diagram during a fault provided by an embodiment of the present invention;

[0055] Figure 3 It is a schematic diagram of a CNN-LSTM hybrid neural network model provided by an embodiment of the present invention;

[0056] Figure 4 It is a schematic diagram of the cumulative distribution of the absolute errors of fault location for 4 models;

[0057] Figure 5 It is a structural block diagram of a traveling wave fault location device for a distribution network provided by an embodiment of the present invention. Detailed implementation manners

[0058] The embodiments of the present invention provide a traveling wave fault location method, device, equipment and storage medium for a distribution network, which are used to solve the technical problem of low detection accuracy for weak faults.

[0059] To make the invention purpose, features and advantages of the present invention more obvious and understandable, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the embodiments described below are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0060] Please refer to Figure 1 , Figure 1 It is a flowchart of the steps of a traveling wave fault location method for a distribution network provided by an embodiment of the present invention.

[0061] A traveling wave fault location method provided by the present invention may specifically include the following steps:

[0062] Step 101, collect the voltage traveling wave signal after the fault occurs at the single-end detection point of the distribution network. The voltage traveling wave signal includes the first wavehead signal and the second wavehead signal;

[0063] In an embodiment of the present invention, a single - end detection point may be set at the distribution network bus, and the voltage traveling - wave signal passing through this point is detected through the single - end detection point. Among them, the voltage traveling - wave signal may include a first wave - head signal and a second wave - head signal; the second wave - head signal is formed by the superposition of a forward traveling - wave and a backward traveling - wave.

[0064] Step 102: Transform the voltage traveling - wave signal to obtain a time - frequency matrix of the panoramic waveform;

[0065] In an embodiment of the present invention, step 102 may include the following sub - steps:

[0066] S21: Perform phase - mode transformation on the first wave - head signal to extract the fault traveling - wave line - mode component;

[0067] S22: Perform continuous wavelet transform on the line - mode component to obtain a time - frequency matrix of the panoramic waveform.

[0068] In a specific implementation, the time when the voltage traveling - wave signal is collected at the single - end detection point can be calibrated. Taking the calibrated time as the center, a time window with a width of 0.5 ms is used to intercept the traveling - wave first wave - head signal. The Karenbauer phase - mode transformation is used to extract the fault traveling - wave line - mode component, and then the continuous wavelet transform (CWT) is used to draw the traveling - wave time - frequency domain waveform within a time window, which is defined as the traveling - wave panoramic waveform. The traveling - wave panoramic waveform is a visual representation of the traveling - wave time - frequency distribution, which not only retains the time information of the traveling - wave waveform at the detection point but also intuitively reflects the frequency distribution at each moment.

[0069] According to the traveling - wave transmission characteristics, a fault traveling - wave transmission grid diagram of the distribution network during a fault can be drawn, as Figure 2 shown. According to the different reflection and refraction propagation paths of the traveling - wave, the voltage traveling - wave detected by the mutual inductor is divided into five categories: ① the initial fault traveling - wave, ② the reflected wave at the fault point, ③ the reflected wave at the end - bus, ④ the reflected wave after reflection from the upstream line and then passing through the fault point, ⑤ the refracted wave after reflection from the downstream line and then passing through the fault point. The expression of the reverse line - mode voltage traveling - wave reaching R1 is as follows:

[0070]

[0071] Among them, represents the voltage traveling - wave at the fault point, represents the voltage at the mutual - inductor installation location, x represents the distance from the fault point to the mutual - inductor installation location, L1, L2, and L3 respectively represent the lengths of lines PM, MN, and NQ, represents the refraction coefficient at point N, represents the reflection coefficient at point Q, represents the refraction coefficient at point M, represents the reflection coefficient at point P, represents the refraction coefficient at point f, Denote the reflection coefficient at point f Denote the complex propagation constant of the line, which can usually be expressed as , where α is the attenuation constant, reflecting the attenuation degree of the traveling wave amplitude with distance; β is the phase constant, reflecting the change of the traveling wave phase with distance.

[0072] It can be inferred from the above formula that the differences in traveling wave signals are mainly attributed to the reflection and transmission coefficients, which reflect different characteristics of the refraction and reflection processes. Since the reflection and transmission coefficients depend on frequency, different frequency components have different attenuation degrees during the refraction and reflection processes, which will affect the frequency distribution of each wave head. Therefore, theoretically, under different fault paths, the frequency distributions of each wave head in the full-view waveform of the traveling wave will show significant differences.

[0073] Step 103: Input the time-frequency matrix into the pre-trained CNN-LSTM fusion model to obtain the fault branch discrimination probability distribution and the fault distance;

[0074] In the embodiment of the present invention, the time-frequency matrix can be input into the pre-trained CNN-LSTM fusion model to obtain the fault branch discrimination probability distribution and the fault distance.

[0075] In one example, step 103 may include the following sub-steps:

[0076] S31: Input the time-frequency matrix into the preset CNN network to extract the sub-wave head frequency distribution characteristics of each sub-wave head signal in the panoramic waveform;

[0077] S32: Based on the sub-wave head frequency distribution characteristics, extract the time correlation characteristics of the typical wave head signal through the preset LSTM;

[0078] S33: Perform weighted fusion on the sub-wave head frequency distribution characteristics and the time correlation characteristics through the attention mechanism to obtain the context vector;

[0079] S34: Map the context vector to the classification space through the classifier to obtain the fault branch discrimination probability distribution;

[0080] S35: Map the context vector to the distance space to obtain the fault distance in the panoramic waveform.

[0081] Typical wave heads refer to several key traveling wave heads with representativeness and obvious characteristics during the transmission of traveling wave signals in the distribution network fault. These wave heads usually include the initial traveling wave head, the fault point reflection wave head, and the far-end reflection wave head, etc.

[0082] CNN (Convolutional Neural Network) is a method that can effectively extract the spatial features of images. Convolution operation is the key to feature mining in CNN. The convolutional layer uses multiple learnable convolutional kernels to perform convolution on the feature map of the previous layer, and then passes through the activation function to obtain the feature map of the next layer. The more convolutional kernels there are, the more comprehensive the extracted features are. The feature matrix can be expressed by the following calculation formula:

[0083]

[0084] Among them, is the weight matrix of the k-th convolutional kernel in the i-th layer; is the bias vector; " " represents the convolution operation; is the activation function.

[0085] LSTM (Long Short-Term Memory) is a special recurrent neural network (RNN), which is specifically designed to handle the long-term dependence problems in sequential data.

[0086] LSTM consists of three multiplicative gates and an internal memory unit. The multiplicative gates include: forget gate, input gate, and output gate. At any time t, LSTM updates the hidden state h t-1 by calculating the previous hidden state h t and the current external input x t , that is, h t = g(x t , h t-1 ) where g is the implicit function of state calculation, which is jointly determined by the activation of the three gates and the memory unit. The state h t is related to the inputs (x1, x2, ···, x t ) at each time. This benefits from the flexible control and cooperation of the multiplicative gates and the memory unit. Simply put, the calculation rules of each variable in LSTM are shown in the following formula. By continuously updating the weight parameters of the multiplicative gates and the memory unit, LSTM can learn the non-linear function relationship between the hidden state h t and the inputs (x1, x2, ···, x t ) at each time, and has the long-term memory ability for effective information.

[0087]

[0088]

[0089]

[0090] Among them: , , , , , , , are the weight matrices corresponding to the inputs respectively; , , bc are the bias vectors; is the memory cell unit; is the activation function; is the external input; is the Forget Gate; is the Input Gate; is the Output Gate.

[0091] In the embodiments of the present invention, a hybrid neural network model CNN-LSTM fusion model is constructed by jointly using CNN and LSTM. As Figure 3 shown, among them, the CNN model is built with reference to the classic lightweight neural network LeNet-5, and has the advantages of shallow depth, fast training speed, and high accuracy. This hybrid neural network has a total of 11 layers of structure, including 2 convolutional layers, 2 max-pooling layers, 1 Dropout layer, 1 LSTM layer, 1 fully connected layer, 2 parallel fully connected layers, and 2 max-pooling layers. One of the parallel output layers is used as the classification output for outputting the faulty branch, with 13 neurons set, corresponding to 13 branches of the IEEE 14-node distribution network model, and the activation function is softmax; the other is used as the regression output for outputting the fault distance, with 1 neuron set.

[0092] The input of the model is a panoramic waveform time-frequency matrix with a sequence length of N and a dimension of M. For the classification problem, the CNN-LSTM fusion model will map the sub-wavehead frequency distribution features extracted by CNN to a fault branch discrimination probability distribution y1 in the form of the hidden layer feature values h t of each moment of LSTM through the fully connected layer and the softmax classifier to achieve the discrimination of the faulty branch:

[0093]

[0094] Among them, and are the weight and bias term of the softmax classifier respectively; is the probability that the network fault data is recognized as the i-th branch, and k is the total number of branches.

[0095] For regression problems, the CNN-LSTM fusion model can use the weight and bias term parameters of the fully connected layer to map the feature vector h t to the fault point location y2:

[0096]

[0097] To improve the training accuracy and the accuracy of positioning, during the CNN feature extraction and LSTM time series modeling processes, an attention layer can be used to assign higher weights to important information, thereby achieving the weighted fusion of the sub-wavehead frequency distribution features and the time correlation features.

[0098] In one example, the process of weighted fusion of the sub-wavehead frequency distribution features and the time correlation features can specifically include the following sub-steps:

[0099] S331, Take the sub-wavehead frequency distribution features as the hidden layer feature values at each moment of the LSTM;

[0100] S332, Extract features from the time-frequency matrix to obtain a feature vector;

[0101] S333, Perform a linear transformation on the feature vector to generate a query vector;

[0102] S334, Use the hidden layer feature values and the query vector to generate the attention scores for the time steps;

[0103] S335, Calculate the attention weights for the time steps according to the attention scores;

[0104] S336, Use the hidden layer feature values and the attention weights to generate a context vector.

[0105] The hidden layer feature values, that is, the hidden state h of the time steps of the time-frequency matrix t include the following information: the time position of the wavehead in the panoramic waveform, the signal intensity difference at each time step, and the frequency component of the panoramic waveform at time step t. The hidden state h t is a compressed representation of the time step features learned by the LSTM, reflecting the temporal correlation in the sequence.

[0106] In a specific implementation, the attention mechanism, by allocating weights, enhances the model's attention to the important parts of the input features while suppressing irrelevant or redundant information. For temporal tasks, the attention mechanism mainly acts on the output of the LSTM, enabling the model to dynamically focus on certain time steps of the input sequence.

[0107] First, an attention layer is added after the LSTM module to calculate the attention score e t for the hidden state h t at time step t and the query vector q:

[0108]

[0109] The attention score function is the core part of the attention mechanism, which is used to calculate the importance weights of each time step in the input sequence. For each time step t, the attention score e t is calculated based on the current hidden state h t and the query vector q. In this paper, the LSTM and CNN modules are weighted in a learnable weight form:

[0110]

[0111] where and are learnable weight matrices, is the bias term, is the learnable vector for linear transformation.

[0112] After weighting, the attention weights are normalized. The core of attention weight normalization is to convert the attention scores e t at different time steps into a probability distribution through the softmax function, so that the model can clarify the importance of each time step to the overall output. This normalization process ensures that the sum of the weights of all time steps is 1, thus providing a clear weight allocation mechanism for the model. For time step t, its attention weight is calculated by the following formula:

[0113]

[0114] where is the normalized weight of the t-th feature, T is the length of the time series, and e t is the attention score at time step t.

[0115] The score e t of each time step can be calculated in dot product form or additive form, and then the softmax function is used. The score e t of each time step is exponentiated, which can ensure that all values are positive, and larger scores will be amplified while smaller scores will be compressed.

[0116] After obtaining the attention weights , the context vector calculation is carried out. The context vector is the core product of the attention mechanism, representing the important information in the input sequence that the model focuses on. It is obtained by weighted summation of the hidden state h t using the attention weights It is transformed into a set of comprehensive representations focusing on key features for subsequent classification or regression tasks, and the formula is as follows:

[0117]

[0118] where h t is the hidden state at time step t, capturing the information of the sequence at this time step. Context is the context vector, which is a weighted comprehensive representation of the input sequence. By adding up the weighted results of all time steps, the context vector Context is obtained, which is a comprehensive sequence representation. The attention weights amplify the information of key time steps and suppress the interference of unimportant time steps, ensuring that the context vector more concentratedly expresses the useful features in the global sequence. If certain time steps in the input sequence contain key information (such as important waveheads or specific signals), the attention mechanism will automatically assign higher weights , making the contributions of these time steps to the context vector greater.

[0119] Furthermore, the embodiment of the present invention adopts the CBAM attention mechanism to perform weighted fusion on the above features, combines channel and spatial attention for feature map processing, to more finely focus on important regions and improve accuracy. Specifically, the CBAM mechanism uses channel attention to adjust the channel feature map, highlighting the most important channels by calculating the importance of each channel. At the same time, it uses spatial attention to weight the original feature map to highlight the most important spatial positions. This attention mechanism can help the network better understand the features in the image and improve the accuracy of image classification and detection.

[0120] The embodiment of the present invention uses the context vector Context extracted by the attention mechanism for classification tasks and regression tasks. The formula for the classification task is as follows:

[0121]

[0122] where Context is the context vector, containing the global information of the sequence, is the weight matrix of the classification layer, mapping the context vector to the classification space, is the bias term to adjust the output of the classification layer, is the classification result representing the probability distribution of different faulty branches.

[0123] In this process, the context vector needs to be input first. Context is a comprehensive representation of the time series. After being input into the classification layer, it is used to predict the faulty branch, and then a linear transformation is performed. The classification layer passes through the weight matrix and the bias Perform a linear transformation on the context vector, map it to the classification space, and finally perform a softmax activation: The softmax function converts the output of the classification layer into a probability distribution, representing the probability that each branch is the faulty branch.

[0124] The regression task represents the fault distance, and the formula is as follows:

[0125]

[0126] Among them, represents the weight matrix of the regression layer, which maps the context vector to the distance space, represents the bias term, which adjusts the output of the regression layer, is the regression result, representing the fault distance. In this process, Context is the comprehensive representation of the time series, which is input into the regression layer and used to predict the location of the fault point. The regression layer performs a linear transformation on the context vector through the weight matrix and the bias to map it to the real number space of the fault distance, and the output is a scalar, representing the actual distance from the predicted fault point to the detection point.

[0127] Step 104: Locate the traveling wave fault of the distribution network according to the fault branch discrimination probability distribution and the fault distance.

[0128] After obtaining the fault branch discrimination probability distribution and the fault distance, the traveling wave fault of the distribution network can be located according to the fault branch discrimination probability distribution and the fault distance.

[0129] The present invention improves the accuracy and efficiency of fault location through the deep learning technology CNN-LSTM, and at the same time enhances the adaptability and robustness of the model, providing strong support for the stable operation of the power system.

[0130] For easy understanding, the following uses a specific example to illustrate the embodiments of the present invention:

[0131] The sampling frequency is set to 500 kHz to ensure that the time-frequency characteristics of traveling waves can be captured in detail. To comprehensively present the panoramic waveform of traveling waves, the sampling time window of 100 microseconds is selected in the embodiments of the present invention. According to the fault parameter settings shown in Table 1, a total of 60 fault points are simulated in the embodiments of the present invention. At each simulated fault point, the settings are made according to the parameters listed in Table 1. By systematically arranging and combining these parameters, the embodiments of the present invention have comprehensively traversed all possible fault parameter configurations. There are a total of 38,400 fault samples. In this study, a data matrix with a dimension of 224×224 is constructed as the sample set, and the sample set is divided into a training set and a test set according to a ratio of 4:1. Among the total 7,680 test samples, there are 5,376 samples with a transition resistance exceeding 1,000 Ω, and 1,920 samples with a fault initial phase angle of 1.5° or 181.5°. Such a test set design can effectively evaluate the adaptability and effectiveness of the proposed method for subtle fault situations such as high-resistance grounding faults and faults close to the zero-crossing point.

[0132] The CNN-LSTM fusion model proposed in the embodiments of the present invention is compared with a convolutional neural network (CNN), a long short-term memory network (LSTM), and a hybrid model of LSTM and CNN. In these comparisons, the parameter configurations of each model follow the settings in Table 1, and the output layer structures of all models are kept consistent. These models are all used to simultaneously predict the branch where the fault occurs and the exact fault location. Through comparison, the performances of these four models in fault branch identification and precise fault location are respectively evaluated.

[0133] Table 1 Traversal Table of Fault Parameters for the Total Sample Set

[0134]

[0135] Among them, the evaluation formula for the discrimination effect of the fault branch is as follows:

[0136]

[0137] The effect of precise fault location is evaluated using the mean absolute error (MAE), and the formula is as follows:

[0138]

[0139] The fault branch discrimination accuracies of the 4 neural network models are shown in Table 2. Among the 4 models, the CNN-LSTM fusion model with an attention mechanism has the highest branch discrimination accuracy, followed by LSTM-CNN, and the accuracy of the single model is lower than that of the hybrid model. From Figure 4It can be observed that the positioning error of the CNN-LSTM fusion model with the attention mechanism is the least affected by the absolute error, showing a relatively significant advantage in precise positioning. Among them, the error range of about 75% of the samples is between 0 and 100 meters. In contrast, the positioning error of the CNN-LSTM model is relatively large. Among the four models, the performance of the CNN model lags significantly behind the other three. Therefore, when the input feature is the traveling wave panoramic waveform, the performance of the CNN-LSTM fusion model with the attention mechanism is better than the CNN, CNN-LSTM, and LSTM-CNN models. This result may be because although all four models can directly process the two-dimensional time-frequency matrix input of the traveling wave panoramic waveform, there are significant differences in the way they process these matrices.

[0140] Table 2 Fault branch discrimination results of different neural network models

[0141]

[0142] Furthermore, a simulation analysis is carried out for different transition resistances. When the transition resistance is high or the fault initial phase angle is close to zero, the amplitude of the initial traveling wave is very small. In a distribution network with many branch nodes, the subsequent wavefronts are severely attenuated due to excessive reflection and refraction, and the fault characteristics are very weak, increasing the positioning difficulty. A single-phase grounding fault is set at 1 km from node 2 on branch 2-6, with a fault initial phase angle of 90°. The positioning results under different transition resistances are simulated and analyzed, as shown in Table 3. The larger transition resistance has a certain attenuation effect on the amplitudes of each frequency band of the traveling wave signal, but the proposed method normalizes different traveling wave panoramic waveform samples, eliminating the differences caused by the traveling wave amplitude to a certain extent, which is beneficial to improving the adaptability of the proposed method to the fault transition resistance.

[0143] Table 3 Positioning results under different transition resistances

[0144]

[0145] The simulation results confirm that the method proposed in the embodiments of the present invention can effectively extract accurate fault location information from the traveling wave panoramic waveform. Specifically, the frequency distribution of each wave head in the waveform corresponds to the branch where the fault occurs, and the arrival time of the typical wave head is associated with the distance to the fault point. Therefore, there is a one-to-one correspondence between the traveling wave panoramic waveform and the fault location, and the traveling wave panoramic waveform of the fault has uniqueness. By analyzing the time-frequency information in the traveling wave panoramic waveform, accurate positioning of the distribution network fault can be achieved. Based on the strategy of "first determining the fault branch and then determining the fault distance", this paper proposes a CNN-LSTM spatio-temporal hybrid model. This model uses the convolutional layer to mine the frequency domain distribution information in the traveling wave panoramic waveform (i.e., the frequency distribution of each wave head), and at the same time uses the long short-term memory layer to mine the time correlation information (i.e., the arrival sequence of the typical wave head), so as to achieve accurate positioning of the distribution network fault.

[0146] Please refer to Figure 5 , Figure 5 which is the structural block diagram of a traveling wave fault location device for a distribution network provided by the embodiments of the present invention.

[0147] The embodiments of the present invention provide a traveling wave fault location device for a distribution network, including:

[0148] A voltage traveling wave signal acquisition module 501, configured to acquire the voltage traveling wave signal after the fault occurs from a single-end detection point of the distribution network, where the voltage traveling wave signal includes a first wave head signal and a secondary wave head signal;

[0149] A time-frequency matrix acquisition module 502, configured to transform the voltage traveling wave signal to obtain the time-frequency matrix of the panoramic waveform;

[0150] A fault branch discrimination probability distribution and fault distance acquisition module 503, configured to input the time-frequency matrix into a pre-trained CNN-LSTM fusion model to obtain the fault branch discrimination probability distribution and the fault distance;

[0151] A fault location module 504, configured to locate the traveling wave fault of the distribution network according to the fault branch discrimination probability distribution and the fault distance.

[0152] In the embodiments of the present invention, the time-frequency matrix acquisition module 502 includes:

[0153] A phase-mode transformation sub-module, configured to perform phase-mode transformation on the first wave head signal to extract the line-mode component of the fault traveling wave;

[0154] A continuous wavelet transform sub-module, configured to perform continuous wavelet transform on the line-mode component to obtain the time-frequency matrix of the panoramic waveform.

[0155] In the embodiments of the present invention, the fault branch discrimination probability distribution and fault distance acquisition module 503 includes:

[0156] The secondary wavefront frequency distribution feature extraction sub-module is used to input the time-frequency matrix into a preset CNN network to extract the secondary wavefront frequency distribution features of each secondary wavefront signal in the panoramic waveform;

[0157] The time correlation feature extraction sub-module is used to extract the time correlation features of typical wavefront signals through a preset LSTM based on the secondary wavefront frequency distribution features;

[0158] The weighted fusion sub-module is used to perform weighted fusion on the secondary wavefront frequency distribution features and the time correlation features through an attention mechanism to obtain a context vector;

[0159] The fault branch discrimination probability distribution acquisition sub-module is used to map the context vector to a classification space through a classifier to obtain the fault branch discrimination probability distribution;

[0160] The fault distance acquisition sub-module is used to map the context vector to a distance space to obtain the fault distance in the panoramic waveform.

[0161] In an embodiment of the present invention, the weighted fusion sub-module includes:

[0162] The hidden layer eigenvalue determination unit is used to use the secondary wavefront frequency distribution features as the hidden layer eigenvalues at each moment of the LSTM;

[0163] The feature extraction unit is used to extract features from the time-frequency matrix to obtain a feature vector;

[0164] The linear transformation unit is used to perform a linear transformation on the feature vector to generate a query vector;

[0165] The attention score generation unit is used to generate the attention score of the time step by using the hidden layer eigenvalues and the query vector;

[0166] The attention weight calculation unit is used to calculate the attention weight of the time step according to the attention score;

[0167] The context vector generation unit is used to generate a context vector by using the hidden layer eigenvalues and the attention weight.

[0168] An embodiment of the present invention also provides an electronic device, which includes a processor and a memory:

[0169] The memory is used to store program code and transmit the program code to the processor;

[0170] The processor is used to execute the traveling wave fault location method for a distribution network in an embodiment of the present invention according to the instructions in the program code.

[0171] An embodiment of the present invention also provides a computer-readable storage medium, which is used to store program codes for executing the traveling wave fault location method for a distribution network according to the embodiment of the present invention.

[0172] Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working processes of the above-described systems, devices, and units can refer to the corresponding processes in the foregoing method embodiments, and will not be described herein again.

[0173] Each embodiment in this specification is described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. The same or similar parts among the various embodiments can be referred to each other.

[0174] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a device, or a computer program product. Therefore, the embodiments of the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the embodiments of the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program codes.

[0175] The embodiments of the present invention are described with reference to the flowcharts and / or block diagrams of the methods, terminal devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram can be implemented by computer program instructions, and the combination of the processes and / or blocks in the flowchart and / or block diagram can also be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal devices generate a device for implementing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0176] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing terminal devices to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured product including an instruction device, and the instruction device implements the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0177] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device, so that a series of operation steps are executed on the computer or other programmable terminal device to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable terminal device provide steps for implementing the functions specified in one process or multiple processes and / or blocks. Figure 1 One process or multiple processes and / or blocks Figure 1 Steps for implementing the functions specified in one block or multiple blocks.

[0178] Although the preferred embodiments of the embodiments of the present invention have been described, those skilled in the art can make additional changes and modifications once they learn the basic creative concepts. Therefore, the appended claims are intended to be construed to include the preferred embodiments and all changes and modifications falling within the scope of the embodiments of the present invention.

[0179] Finally, it should also be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or terminal device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or terminal device. Without further limitation, an element defined by the statement "comprising one..." does not exclude the presence of additional identical elements in the process, method, article or terminal device comprising the said element.

[0180] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for locating traveling wave faults in a distribution network, characterized in that: include: Collecting a voltage traveling wave signal after a fault occurs from a single-end detection point of the distribution network, wherein the voltage traveling wave signal includes a first wave head signal and a second wave head signal; Transforming the voltage traveling wave signal to obtain a time-frequency matrix of a panoramic waveform; The time-frequency matrix is ​​input into the pre-trained CNN-LSTM fusion model to obtain the fault branch discrimination probability distribution and the fault distance; The distribution network traveling wave fault is located according to the fault branch identification probability distribution and the fault distance.

2. The method according to claim 1, characterized in that The step of transforming the voltage traveling wave signal to obtain a time-frequency matrix of a panoramic waveform includes: Performing phase mode transformation on the first wave head signal to extract the fault traveling wave line mode component; The line mode components are subjected to continuous wavelet transform to obtain a time-frequency matrix of the panoramic waveform.

3. The method according to claim 1, characterized in that The step of inputting the time-frequency matrix into the pre-trained CNN-LSTM fusion model to obtain the fault branch discrimination probability distribution and the fault distance includes: Input the time-frequency matrix into a preset CNN network to extract sub-wave head frequency distribution characteristics of each sub-wave head signal in the panoramic waveform; Based on the sub-wave head frequency distribution characteristics, the time correlation characteristics of the typical wave head signal are extracted by presetting LSTM; The sub-wave head frequency distribution feature and the time correlation feature are weightedly fused through an attention mechanism to obtain a context vector; Mapping the context vector to a classification space through a classifier to obtain a fault branch discrimination probability distribution; The context vector is mapped to a distance space to obtain a fault distance in the panoramic waveform.

4. The method according to claim 3, characterized in that The step of performing weighted fusion of the sub-wave head frequency distribution feature and the time association feature through an attention mechanism to obtain a context vector comprises: The sub-wave head frequency distribution characteristics are used as the hidden layer feature values ​​of the LSTM at each moment; Performing feature extraction on the time-frequency matrix to obtain a feature vector; Performing a linear transformation on the feature vector to generate a query vector; Generating an attention score for the time step using the hidden layer feature value and the query vector; Calculate the attention weight of the time step according to the attention score; A context vector is generated using the hidden layer feature value and the attention weight.

5. A distribution network traveling wave fault location device, characterized in that: include: A voltage traveling wave signal acquisition module is used to collect a voltage traveling wave signal after a fault occurs from a single-end detection point of the distribution network, wherein the voltage traveling wave signal includes a first wave head signal and a second wave head signal; A time-frequency matrix acquisition module, used to transform the voltage traveling wave signal to obtain a time-frequency matrix of a panoramic waveform; A fault branch identification probability distribution and fault distance acquisition module, used for inputting the time-frequency matrix into a pre-trained CNN-LSTM fusion model to obtain a fault branch identification probability distribution and a fault distance; The fault location module is used to locate the distribution network traveling wave fault according to the fault branch judgment probability distribution and the fault distance.

6. The device according to claim 5, characterized in that The time-frequency matrix acquisition module includes: A phase mode transformation submodule, used for performing phase mode transformation on the first wave head signal to extract the fault traveling wave line mode component; The continuous wavelet transform submodule is used to perform continuous wavelet transform on the line mode component to obtain a time-frequency matrix of the panoramic waveform.

7. The device according to claim 5, characterized in that The fault branch identification probability distribution and fault distance acquisition module includes: A sub-wave head frequency distribution feature extraction submodule, used for inputting the time-frequency matrix into a preset CNN network to extract the sub-wave head frequency distribution features of each sub-wave head signal in the panoramic waveform; A time-related feature extraction submodule is used to extract the time-related features of a typical wave head signal through a preset LSTM based on the frequency distribution features of the secondary wave head; A weighted fusion submodule, used for weighted fusion of the sub-wave head frequency distribution feature and the time correlation feature through an attention mechanism to obtain a context vector; A fault branch discrimination probability distribution acquisition submodule is used to map the context vector to a classification space through a classifier to obtain a fault branch discrimination probability distribution; The fault distance acquisition submodule is used to map the context vector to the distance space to obtain the fault distance in the panoramic waveform.

8. The device according to claim 7, characterized in that The weighted fusion submodule comprises: A hidden layer feature value determination unit, used to use the sub-wave head frequency distribution feature as the hidden layer feature value of the LSTM at each moment; A feature extraction unit, used to extract features from the time-frequency matrix to obtain a feature vector; A linear transformation unit, used for performing a linear transformation on the feature vector to generate a query vector; an attention score generating unit, configured to generate an attention score for the time step by using the hidden layer feature value and the query vector; an attention weight calculation unit, configured to calculate the attention weight of the time step according to the attention score; A context vector generating unit is used to generate a context vector using the hidden layer feature value and the attention weight.

9. An electronic device, characterized in that: The device comprises a processor and a memory: The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the distribution network traveling wave fault location method according to any one of claims 1-4 according to the instructions in the program code.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store program codes, and the program codes are used to execute the distribution network traveling wave fault location method according to any one of claims 1 to 4.

Images