Multi-branch power distribution network fault positioning method and system based on multi-dimensional fault characteristics

Abstract

The invention discloses a multi-branch power distribution network fault positioning method and system based on multi-dimensional fault characteristics. The method comprises the following steps: acquiring fault voltage traveling wave waveform data; performing phase-mode transformation on the waveform data to extract a line-mode component, extracting wave head coordinates, amplitude and slope information of a fault traveling wave in a set time window, and after redundancy elimination, determining a feature matrix, an adjacent matrix and an edge feature tensor based on a wave head information space so as to construct a multi-dimensional fault feature space; constructing a section identification model and a precise positioning model, and performing hyper-parameter tuning on the models by adopting a parameter-setting-free harmony search algorithm; and whether fault section positioning is carried out is judged based on the multi-dimensional fault feature space, if yes, the multi-dimensional fault feature space is input into a section recognition model to obtain a section recognition result, and if not, the multi-dimensional fault feature space is input into a precise positioning model to obtain a positioning result. And dual targets of power distribution network fault section identification and accurate positioning are realized from two aspects of multi-dimensional fault feature space extraction and a power distribution network fault positioning model.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of fault location of power distribution network, and in particular to a multi-branch power distribution network fault location method and system based on multi-dimensional fault features. BACKGROUND

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute the prior art.

[0003] As the terminal link of the power system, the safe and reliable operation of the power distribution network directly affects the quality of power supply, and accurate and rapid fault location is the key to ensuring its reliable operation. Due to the complex structure, multiple branch nodes and extensive coverage of the power distribution network, traditional fault detection techniques are difficult to accurately locate the fault of the multi-branch line, and cannot meet the actual needs of the operation and maintenance of the power distribution network. Data-driven deep learning algorithms have shown significant advantages in fault information mining tasks due to their strong feature extraction capabilities, and have been gradually introduced into the fields of new energy station, power transmission network and power distribution network fault location.

[0004] The power distribution network has numerous branches, and the characteristics after a short-circuit fault are more complex, which puts higher requirements on the feature mining ability and generalization performance of the power distribution network fault location model. With the continuous improvement of information coupling degree of new power distribution network, the fault features show complex nonlinear characteristics, and deep learning technology realizes online rapid decision of power distribution network fault section location through offline training of nonlinear fitting between fault features and positioning results.

[0005] Data-driven deep learning algorithms provide a new way for power distribution network fault detection, and their hierarchical feature learning mechanism effectively breaks through the limitations of traditional methods in model learning ability, data processing efficiency and generalization performance. However, existing simple data-driven models cannot fully capture the multi-dimensional correlation of fault features, and are mostly limited to section positioning. Moreover, the parameters of neural network algorithms are difficult to adjust, and the randomness of manually configured parameters is large, so direct input into the positioning model often cannot achieve high-precision positioning results. SUMMARY

[0006] To overcome the shortcomings of the prior art, the present application provides a multi-branch power distribution network fault location method and system based on multi-dimensional fault features, which realizes the dual goals of power distribution network fault section identification and accurate positioning from the aspects of multi-dimensional fault feature space extraction and power distribution network fault location model.

[0007] To achieve the above purpose, one or more embodiments of the present application provide the following technical solutions: In a first aspect, the present application provides a multi-branch power distribution network fault location method based on multi-dimensional fault features, comprising: obtaining fault voltage traveling wave waveform data of the multi-branch power distribution network; The line mode component is extracted by performing phase-mode transformation on the fault voltage traveling wave waveform data, the fault traveling wave head coordinate, amplitude and slope information are extracted within a set time window, and a head information space is formed; after redundancy elimination, a feature matrix, an adjacency matrix and an edge feature tensor are determined based on the head information space, and a multi-dimensional fault feature space is constructed based on the feature matrix, the adjacency matrix and the edge feature tensor; A section identification model and a precise positioning model are constructed, and a parameter-free setting and a sound search algorithm are used to optimize the hyperparameters of the models; Based on the feature matrix of the multi-dimensional fault feature space, it is determined whether to perform fault section identification, if yes, the multi-dimensional fault feature space is input into the section identification model to obtain a section identification result, and if not, the multi-dimensional fault feature space is input into the precise positioning model to obtain a positioning result.

[0008] In a further technical solution, the traveling wave information characteristics of a multi-branch power distribution network at different fault locations and fault branches are analyzed to obtain the mapping relationship between the wave head time sequence, polarity and amplitude characteristics and the fault location.

[0009] In a further technical solution, the line mode component is extracted by performing Karenbauer phase-mode transformation on the fault voltage traveling wave waveform data, and the fault traveling wave head coordinate, amplitude and slope information are extracted based on a straight line detection algorithm.

[0010] In a further technical solution, the section identification model and the precise positioning model use a hybrid network model, including an edge propagation graph convolutional neural network and a bidirectional gated recurrent unit.

[0011] In a further technical solution, the edge propagation graph convolutional neural network constructs edge embedding by fusing node pairs and edge features, and introduces a self-attention mechanism to assign weights, updates the node features after weighted aggregation and residual connection, and extracts spatiotemporal features containing topology and edge attributes.

[0012] In a further technical solution, the bidirectional gated recurrent unit uses a time-aligned hidden state splicing strategy to splice features to obtain output features.

[0013] In a further technical solution, the parameter-free setting and the sound search algorithm adjust the sound memory consideration rate and the pitch adjustment rate simultaneously.

[0014] In a second aspect, the present application provides a multi-branch power distribution network fault location system based on multi-dimensional fault features, comprising: A data acquisition module configured to acquire fault voltage traveling wave waveform data of a multi-branch power distribution network; a feature construction module configured to: perform phase-mode transformation on the fault voltage traveling wave waveform data to extract line mode components, extract fault traveling wave head coordinates, amplitude, and slope information within a set time window to form a wave head information space; after redundancy elimination, determine a feature matrix, an adjacency matrix, and an edge feature tensor based on the wave head information space, and construct a multi-dimensional fault feature space based on the feature matrix, the adjacency matrix, and the edge feature tensor; a model construction module configured to: construct a section identification model and a precise positioning model, and use parameter-free setting and a sound search algorithm to optimize the hyperparameters of the models; a fault location module configured to: determine whether to perform fault section identification based on the feature matrix of the multi-dimensional fault feature space, if yes, input the multi-dimensional fault feature space into the section identification model to obtain a section identification result, and if no, input the multi-dimensional fault feature space into the precise positioning model to obtain a positioning result.

[0015] In a third aspect, the present application provides a computer readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the multi-branch power distribution network fault location method based on multi-dimensional fault features as described in the first aspect.

[0016] In a fourth aspect, the present application provides a computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps in the multi-branch power distribution network fault location method based on multi-dimensional fault features as described in the first aspect when executing the program.

[0017] The above one or more technical solutions have the following beneficial effects: The present application realizes the dual goals of fault section identification and precise positioning of the power distribution network from two aspects of multi-dimensional fault feature space extraction and power distribution network fault location model. First, the traveling wave information characteristics of the multi-branch power distribution network at different fault locations and fault branches are analyzed, the wave head feature straight line segment in the image is detected by a wave head calibration method based on linear feature detection (LSD), the fault traveling wave head coordinates, amplitude, and slope information are extracted, after redundancy elimination, a graph structure representation consistent with the topology structure of the power distribution network is constructed based on the node feature matrix, the adjacency matrix, and the edge feature tensor, and a multi-dimensional fault feature space is formed. Second, an EPGCN-BiGRU fault location model is built, and the hyperparameters of the model are optimized using parameter-free setting and sound search algorithm (PHS), the local spatial features are extracted by edge propagation graph convolutional neural network, and the correlation between the time series features, amplitude features, and fault location is fully mined by combining the bidirectional gate recurrent unit to capture the time series correlation characteristics. Then, the fault section identification and fault precise positioning are realized through classification task and regression task respectively.

[0018] The application is based on a wave head calibration technology of LSD straight line detection, extracts the coordinate, amplitude and slope information of multiple fault traveling wave wave heads in a reasonable time window, and after redundancy elimination, based on the node feature matrix, the adjacency matrix and the edge feature tensor, a graph structure representation consistent with the topology structure of the distribution network is constructed, a multi-dimensional fault feature space is formed, the key information of the fault section and the fault point is effectively fused, and the model input dimension is greatly reduced; an EPGCN-BiGRU fault location model is built, the local spatial features are extracted through the edge propagation graph convolutional neural network, the time sequence correlation characteristics are captured in combination with the bidirectional gate recurrent unit, and the correlation between the time sequence features, the amplitude features and the fault position is fully mined; the PHS algorithm is introduced to optimize the model hyperparameters, the key parameters can be dynamically adjusted, the limitation of the fixed parameters of the traditional method is overcome, so that the hyperparameter optimization is more effectively carried out, and the fault location precision is improved. The EPGCN-BiGRU fault section identification model and the EPGCN-BiGRU fault precise positioning model are comprehensively used, the fault features are extracted by using the LSD algorithm, the PHS algorithm is introduced to optimize the model hyperparameters, a PHS-EPGCN-BiGRU multi-branch distribution network fault location method based on multi-dimensional fault feature extraction is provided, and the dual goals of distribution network fault section identification and precise positioning are realized, so that an efficient and reliable solution is provided for the distribution network fault section identification and the fault precise positioning. BRIEF DESCRIPTION OF DRAWINGS

[0019] The drawings accompanying the specification of this application form a part thereof, serve to provide further understanding of the application, and together with the description, make up an embodiment of the application. The embodiment of the application and its description are used to explain the application, and do not constitute an improper limitation on the application.

[0020] Figure 1 It is a flowchart of the multi-branch distribution network fault location method based on multi-dimensional fault features of the embodiment of the application; Figure 2 It is a multi-branch distribution network fault topology structure diagram of the embodiment of the application; Figure 3 It is a F1 and F2 fault traveling wave propagation diagram of the embodiment of the application; Figure 4 It is a F1 and F3 fault traveling wave propagation diagram of the embodiment of the application; Figure 5 It is a wave head calibration result diagram under a reasonable time window of the embodiment of the application; Figure 6 It is an edge propagation graph convolutional neural network structure diagram of the embodiment of the application; Figure 7 It is a GRU structure diagram of the embodiment of the application; Figure 8 It is a BiGRU structure diagram of the embodiment of the application. DETAILED DESCRIPTION

[0021] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0022] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0023] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0024] Example 1 like Figure 1 As shown, this embodiment discloses a multi-branch distribution network fault location method based on multi-dimensional fault characteristics, which includes the following steps: S1: Analyze the characteristics of traveling wave information in the multi-branch distribution network at different fault locations and fault branches, and derive the mapping relationship between the wave head timing, polarity and amplitude characteristics and the fault location; S101: Analyze the internal fault characteristics of the same line.

[0025] like Figure 2 As shown in the topological structure diagram of the multi-branch power grid, traveling wave acquisition devices are set at both ends of the trunk line and the end of each branch line, and and Setting up faults . Fault point and Located on branch road , respectively 1km and 2km away from the M node, the fault point Located on branch road ,distance Branch node 1km.

[0026] like Figure 3 As shown, select f 1 and f 2 Two typical fault points are taken as analysis objects, and a traveling wave propagation grid diagram is constructed based on the traveling wave transmission theory. Considering the complex refraction and reflection characteristics of the distribution network, the traveling wave transmission between the fault point and the detection device is characterized, and the key structural characteristics of the local network are presented. The three characteristic wave heads (initial traveling wave , Fault point reflected wave , the reflected wave of the opposite bus ), systematically analyze the timing differences of traveling wave propagation at different fault locations.

[0027] Based on the fault traveling wave time domain analysis method, the traveling wave signal received by the M-end detection device is essentially a composite waveform formed by the initial traveling wave of the fault undergoing refraction and reflection in the line, superimposed in a specific time sequence. Due to the different spatial locations of the fault points, the transmission path and refraction and reflection process of the initial traveling wave will be different, which will eventually manifest as significant differences in the amplitude and waveform characteristics of the detection signal. Combined with the traveling wave propagation path, it can be seen that and The first three traveling wave transmission trajectories of the fault point have similar characteristics.

[0028] In order to quantitatively analyze the difference in the traveling wave responses of the two, it can be derived that and The time domain expression of the characteristic wave head of the voltage traveling wave at the fault point is:

[0029] in, 、 and Respectively represent The corresponding initial traveling wave, reflected wave at the fault point and reflected wave at the opposite busbar; Fault point applied step voltage; 、 and are the traveling wave reflection coefficients at M, F1 and T1 respectively; is the traveling wave refraction coefficient at F1; is the line propagation coefficient; and represent the distances of the MF1 segment and the F1T1 segment, respectively; represents the speed of traveling wave propagation.

[0030]

[0031] in, 、 and Respectively represent The corresponding initial traveling wave, reflected wave at the fault point and reflected wave at the opposite busbar; Fault point applied step voltage; and represent the traveling wave reflection coefficients at F2 and T1 respectively; is the traveling wave refraction coefficient at F2; and Represent the distances of the MF2 segment and the F2T1 segment respectively.

[0032] Although And The transition resistance, initial phase angle and other electrical parameters of the fault point remain consistent, and are located in the same feeder branch, but due to the difference in spatial position of the fault point, the initial traveling wave produces significant time difference characteristics in the process of fold reflection, and the arrival timing of the three characteristic wave heads is obviously different, thereby forming distinguishable fault characteristics on the time domain waveform.

[0033] On the other hand, from the frequency domain analysis dimension, the frequency response characteristics of the traveling wave signal in the distribution network are affected by two factors: first, the transmission function has distance-frequency coupling characteristics, and the attenuation factor is exponentially related to the propagation distance, and the attenuation rate of high-frequency components is significantly higher than that of low-frequency components; second, the fold reflection coefficient matrix shows frequency dependence, and the reflectivity of different frequency components at the wave impedance discontinuity point is different. The characteristics of the traveling wave signal transmission function and the fold reflection coefficient are as follows:

[0034]

[0035]

[0036]

[0037]

[0038] wherein, is the transmission function; is the propagation coefficient; is the reflection coefficient; is the reflection coefficient; and are the wave impedances on both sides of the line; is the wave impedance function of the line end; , , and are the resistance, inductance, conductance and capacitance parameters per unit length. is the resistance component, is the reactance component, is the imaginary unit, is the angular frequency.

[0039] S102: Analyze the fault characteristics of different sections.

[0040] Two typical fault points are selected and as the analysis objects, and a traveling wave propagation grid diagram is constructed based on the traveling wave transmission theory. The fault points The time domain expression of the voltage traveling wave characteristic wave head is:

[0041] wherein, , and respectively represent corresponding initial traveling wave, fault point reflected wave and opposite bus reflected wave; is the step voltage applied at the fault point ; and are the traveling wave reflection coefficients at F3 and P1 respectively; and are the traveling wave refraction coefficients at T1 and F3 respectively; , , and respectively represent the distances of MT1 section, T1F3 section, T1P1 section and F3P1 section.

[0042] It can be seen from Figure 4 that the initial traveling wave at fault point and the fault point reflected wave propagation path difference is , the initial traveling wave at fault point and the fault point reflected wave propagation path difference is , since the two fault points distance wave impedance inequality point distance is same, therefore and the first two wave head time difference is same; and the two fault points applied step voltage is same, only rely on the first two wave head time domain information and can not distinguish two kinds of situation. But the first two wave head of fault point compared to refraction, therefore, and traveling wave signal difference mainly comes from refraction reflection coefficient, essentially is the different of refraction process. Since the two fault points are located in different branch, and the length of branch is different, this also affects the opposite bus reflected wave arrival time has a significant difference.

[0043] Based on the above theoretical analysis, the fault traveling wave characteristics can be summarized as follows: for the same section and different location faults, under similar fault conditions, the traveling wave propagation mechanism is mainly affected by the transmission function and the reflection and refraction coefficient; due to the similarity of the reflection and refraction process of the traveling wave under the same line parameters, the waveform profile usually presents basic similarity characteristics, but with the increase of the propagation distance, the difference in the arrival time of the wave head will cause the phase shift of the local waveform. For the fault scene of different sections, the refraction coefficient will cause the reflection and refraction process of the traveling wave to produce characteristic differences, and the differences produced thereby can be reflected in the polarity characteristics and amplitude characteristics of the wave head. By constructing a multi-dimensional feature space containing the wave head timing, polarity and amplitude characteristics, the mapping relationship between the multi-dimensional fault feature space of the detection point of the multi-branch distribution network and the fault location can be effectively established, so as to realize the accurate positioning of the fault point of the distribution network.

[0044] S2: Collecting the fault voltage traveling wave waveform data of the multi-branch distribution network, performing Karenbauer phase mode transformation on the fault voltage traveling wave waveform data to extract the line mode component, extracting the fault traveling wave head coordinates, amplitude and slope information in a reasonable time window to form a wave head information space; after eliminating redundancy, determining the feature matrix, the adjacency matrix and the edge feature tensor based on the wave head information space, constructing a multi-dimensional fault feature space based on the node feature matrix, the adjacency matrix and the edge feature tensor, and forming a training set and a test set; the constructed multi-dimensional fault feature space is a graph structure representation consistent with the topology structure of the distribution network, which is used as the input data of the edge propagation graph convolutional neural network (EPGCN).

[0045] Based on the fault traveling wave head coordinates, amplitude and slope information, a multi-dimensional fault feature space is constructed, and a training set and a test set are formed; S201: Extracting multi-dimensional fault features based on a straight line detection algorithm.

[0046] As shown in Figure 5 , the wave head calibration technology using the straight line detection algorithm (Line Segment Detector, LSD) can effectively extract the transient characteristics of the signal, and can extract the endpoint coordinates, length and slope of the fault traveling wave head in the waveform graph. In a certain reasonable time window, the first fault traveling wave head information detected can be expressed as:

[0047] wherein, is the feature vector of the first wave head; is the initial endpoint horizontal coordinate of the first wave head; is the initial endpoint vertical coordinate of the first wave head; is the length of the first the horizontal coordinate of the end point of the wave front; the vertical coordinate of the end point of the wave front; the amplitude of the wave front; the slope of the wave front. the vertical coordinate of the end point of the wave front. the amplitude of the wave front. the slope of the wave front.

[0048] S202: Construct a multi-dimensional fault feature space based on multi-dimensional fault features.

[0049] The fault traveling wave front information can be combined into a wave front information space , and the expression is:

[0050] wherein, N is the total number of detected wave fronts.

[0051] The wave front information space contains redundant feature information. After removing the redundant information, in order to meet the structured requirements of the edge propagation graph convolutional neural network (EPGCN) for input data, the processed samples are further converted into a graph structure. Specifically, let the node set of the power distribution network be V , the edge set determined by the primary equipment topological relationship is E , and based on this, three types of input data of node features, adjacency relationships and edge features are constructed to ensure that the model can utilize node attributes and topological structure information for feature learning at the same time.

[0052] (1) Feature matrix , and the expression is:

[0053] (2) Adjacency matrix , which is determined by the connection relationship of the actual power distribution network and the traveling wave reachability, and takes binary adjacency.

[0054] (3) Edge feature tensor , each edge is given:

[0055] wherein, is the segment length, is the wave impedance.

[0056] Based on the above three types of input data, a graph structure representation consistent with the topology of the power distribution network is constructed, thereby forming a multi-dimensional fault feature space , thereby providing support for the subsequent construction of the edge propagation graph convolutional neural network under the unified representation of node attributes and network connection relationships.

[0057] S3: Build EPGCN-BiGRU section identification model and EPGCN-BiGRU accurate positioning model and input training set for training, and use PHS algorithm to optimize its hyperparameters; In the embodiment, the section identification model adopts a hybrid network model, i.e., an EPGCN-BiGRU model.

[0058] S301: Build an edge propagation graph convolutional neural network model.

[0059] Edge propagation graph convolutional neural network (EPGCN) is a convolutional network for graph structure data, and its input needs to include node features, topological relationships and edge features. Therefore, based on the node feature matrix, the adjacency matrix and the edge feature tensor, the graph structure representation consistent with the topology structure of the power distribution network is constructed, thereby forming a multi-dimensional fault feature space , wherein the node feature matrix Collect the wave head arrival time, amplitude and slope information of each monitoring end in a reasonable time window, the adjacency matrix Determined by the topological structure of the primary equipment and the traveling wave reachability, the edge feature tensor Then, the edge attributes such as the coding section length and the wave impedance are encoded. Subsequently, the EPGCN realizes feature propagation and extraction under topological-physical constraints through self-attention weighted neighborhood aggregation and edge embedding.

[0060] The network is based on the topological connection relationship of nodes, aggregates neighbor node information in a graph convolution manner, and introduces edge attributes as weight control factors to enhance the spatial structure expression ability, thereby realizing deep modeling of the features of each node in the graph. A typical EPGCN network structure is as shown in Figure 6 .

[0061] In the EPGCN, each layer of convolution operation not only depends on the features of adjacent nodes, but also explicitly considers the input features of edges. The source node, target node and edge features are spliced through an edge embedding function, and then mapped to edge representation through linear transformation and activation function, and further used for node update. The calculation expression of the process is as follows:

[0062] , wherein, is the edge feature embedding between the node and the node in the i-th layer; is a linear mapping matrix; and concat represents a feature splicing operation; and are the node and the node in the i-th layer, respectively. and nodes Node characteristics; is the edge feature.

[0063] Next, by introducing the self-attention mechanism, we learn the attention weight of each edge to measure its importance in feature aggregation. The normalized attention coefficient and edge embedding jointly participate in the node update operation. The node update expression is as follows:

[0064] Among them, ReLU is the activation function; is the linear transformation matrix of the node self-mapping weight; For the Nodes in the layer Node characteristics; For nodes The set of neighbors of is the edge attention weight; is the Hadamard product, indicating element-wise weighted fusion.

[0065] By adding a time series modeling structure to the end of EPGCN, the dynamic evolution of fault propagation can be further captured, improving the model's perception of joint spatiotemporal features. This structure can jointly model the local features of nodes in the graph with global topological relationships and edge characteristics, achieving structured and efficient feature extraction and representation optimization of input graph data.

[0066] S302: Build a bidirectional gated recurrent unit.

[0067] The Bidirectional Gated Recurrent Unit (BiGRU) effectively alleviates the vanishing or exploding gradient problem of traditional recurrent neural networks in processing long sequences of data. It achieves dynamic information regulation by introducing a gating mechanism, where the update gate controls the degree of historical information retention, and the reset gate determines the fusion of current input and past states. This dynamic regulation strategy can capture dependencies at different time scales. The specific structure is as follows: Figure 7 shown.

[0068] The update gate realizes the dynamic fusion control of historical information and current input, which depends on the nonlinear combination of current input features and the hidden state at the previous moment:

[0069] in, for t Update the output of the gate at all times; is the Sigmoid activation function; To update the gate weight; yes t -1 hidden layer output; yes t Input at the moment; is the bias of the update gate.

[0070] The reset gate achieves selective forgetting of historical information through a dynamic control mechanism. It uses a nonlinear combination of input features and the state at the previous moment to determine the degree of forgetting of the historical state, thereby helping the model discard historical information and focus on the current state. Its mathematical expression is defined as:

[0071] in, for t Reset the output of the gate at all times; To reset the gate weight; is the bias for the reset gate.

[0072] The candidate hidden state represents the potential updated state after the historical state is selectively forgotten by the reset gate. It is generated by the nonlinear combination of the current input feature and the historical state modulated by the reset gate. Its mathematical expression is:

[0073] in, is the candidate hidden state output; is the hyperbolic tangent activation function; is the candidate hidden weight; is the candidate state bias; is the element-wise product.

[0074] The final hidden state is generated by updating the gate and weighting the candidate hidden state. The mathematical expression is:

[0075] in, for t The final hidden layer output at time t.

[0076] The core of BiGRU lies in its bidirectional time series modeling architecture, which consists of two parallel GRU layers, which perform forward and reverse feature extraction on time series data respectively. Through this structural design, the model can simultaneously capture the bidirectional dependencies of sequence data, thereby significantly improving the utilization efficiency of past and present information of data. The specific structure is as follows Figure 8 shown.

[0077] In terms of feature fusion mechanism, BiGRU adopts a time-aligned hidden state splicing strategy. t The state is integrated through feature concatenation:

[0078] in, It is the feature aggregation from the starting point of the data to the current moment; It is the feature aggregation from the current moment to the starting point of the data.

[0079] S303: Select an evaluation index to determine the accuracy of the positioning result.

[0080] The performance of fault branch identification is evaluated using a confusion matrix. As a core evaluation tool for classification models, the confusion matrix uses a two-dimensional table structure to intuitively display the correspondence between predicted results and actual classifications. The rows of this matrix represent the true classification of the samples, while the columns correspond to the model's predicted classification results. Its mathematical expression is shown in Table 1.

[0081] Table 1 Binary classification confusion matrix

[0082] As shown in Table 1, in a binary classification problem, TP and TN are the number of positive and negative samples that are correctly identified, respectively, while FN and FP are the number of positive and negative samples that are misclassified, respectively. In model performance evaluation, classification accuracy is a core metric, defined as the ratio of samples correctly identified by the model to all samples. The mathematical expression is:

[0083] in, Represents the accuracy of the classification model.

[0084] For fault location tasks, the root mean square error (RMSE) is used. , mean absolute error (MAE) and coefficient of determination (R2) To conduct an assessment:

[0085]

[0086]

[0087] in, n is the number of samples in the test set; and Respectively The true value and predicted value of the fault distance of each sample; is the average fault distance.

[0088] The smaller the values of the root mean square error and the mean absolute error, the higher the accuracy of fault location; the larger the value of the determination coefficient, the better the fitting effect.

[0089] S304: The output error of the positioning result is used as the target function, and the PHS algorithm is used to optimize the hyperparameters of the EPGCN-BiGRU model.

[0090] To improve the positioning accuracy of the EPGCN-BiGRU model and overcome the problem of neural network hyperparameter optimization, a parameter-free harmony search algorithm (PHS) is proposed to collaboratively optimize the hyperparameters of the EPGCN-BiGRU. In terms of model architecture setting, the EPGCN module uses ReLU as the activation function, and the BiGRU module uses the Selu function to activate the neurons. The Adam optimizer is used to drive parameter updates during the optimization process, and the training is completed in 80 iterations. Due to the differentiated influence of model parameters on positioning performance, the core hyperparameters that significantly affect the efficiency of the prediction model are selected for optimization search. The optimization range of the EPGCN edge embedding dimension is set to [8, 128], the optimization range of the BiGRU neuron number is set to [8, 128], the optimization range of the learning rate is set to [0.001, 0.02], and the optimization range of the batch size is set to [20, 50].

[0091] In the harmony search (HS) algorithm, the harmony memory consideration rate ( R HMC ) and the pitch adjustment rate ( R PA ) have a significant impact on its effectiveness. These two parameters are crucial in determining whether to use the best value calculated previously in the harmony memory or to adjust it and use it again. There is a limitation in the HS method, which is to use only one value after a certain stage of the harmony search process. To address this issue, an improved parameter-free harmony search algorithm (PHS) is introduced, which uses a synchronous adjustment setting that can adjust R HMC and R PA values.

[0092] The formulas for determining R HMC and R PA are as follows:

[0093]

[0094] in, Indicates the current iteration number, n represents the maximum number of iterations, v Indicates the number of variables that can be modified.

[0095] The steps for parameter-free harmony search algorithm are as follows: Step 1: Problem modeling and parameter pre-configuration. Clearly define the problem to be solved as minimizing the objective function ,in X is composed of D decision variables Initialize the harmony creation times ( T max), harmony memory size ( S HM ) and pitch adjustment bandwidth ( W B ) Fixed parameters; Initialize harmony memory consideration rate ( R HMC ), Pitch Adjustment Rate( R PA ) can adjust parameters and define synchronization adjustment rules.

[0096] Step 2: Construct the initial harmony memory library. Generate by random sampling in the solution space S HM The harmonic vectors are stored in the harmonic memory as the starting point for iterative optimization of the algorithm.

[0097] Step 3: Dynamic parameter collaborative update. Enter the iterative loop and update according to the preset synchronization adjustment strategy R HMC and R PA .

[0098] Step 4: New harmony generation strategy. For each new harmony to be generated in each iteration, perform the following operations on each dimension of the variable: R HMC The probability of the value is, randomly select a variable from the harmony memory bank; with 1- R HMC If the selected variables are from the harmony memory library, then further use the current R PA The probability of W B Fine-tune the variable within a limited range to obtain the updated variable; otherwise, keep the variable unchanged and complete the construction of the new harmony.

[0099] Step 5: Update the harmony memory. Calculate the new harmony in the objective function The fitness value of the new harmony is compared with the worst harmony in the harmony memory. If the new harmony has a better fitness value, the worst harmony in the harmony memory is replaced by the new harmony; otherwise, the original content of the harmony memory is kept.

[0100] Step 6: Terminate the iteration judgment. Check whether the current iteration number reaches the preset T max or whether the algorithm meets the convergence condition, such as the variation amplitude of the optimal solution after continuous iterations being less than a specific threshold. If the termination condition is met, the iteration is stopped, and the optimal solution in the current harmony memory is output; otherwise, the third step is returned to continue the next round of iteration calculation.

[0101] In the embodiment, the example steps are as follows: (1) build a multi-branch power grid fault model: based on the MATLAB / SIMULINK simulation software platform, a multi-branch power grid simulation model as shown in FIG. 1 is built, and simulation fault points are set on each branch at an interval of every 100 meters, totaling 280 test points. Figure 1 (2) batch simulation: different fault types and transition resistance parameter combinations are set for each fault test point to carry out batch simulation, and the traveling wave detection device is used to obtain fault voltage traveling wave waveform data under a 2 ms long time window. (3) phase-mode transformation: the fault voltage traveling wave waveform data is subjected to Karenbauer phase-mode transformation to extract line-mode components. (4) time window selection: the same reasonable time window is selected for the fault traveling waves collected at each end, and a number of data points before the arrival of the initial traveling wave and 6 τ(5) Fine calibration of fault traveling wave head: Under a reasonable time window, the LSD algorithm is used for fine calibration to detect the straight line segment containing the wave head feature in the image. (6) Wave head information conversion: The straight line segment information containing the wave head feature in the image is converted into fault traveling wave head coordinates, amplitude and slope information. (7) Construction of multi-dimensional fault feature space: After removing the redundancy of the wave head features extracted from each sampling point, the node features are organized, and the adjacent relationship and edge features generated by the power distribution network topology are combined to form a graph structure consistent with the network structure, which is used to construct a multi-dimensional fault feature space. (8) Fault feature database: In the data storage link, a fault feature database covering multiple detection points is established, a fault branch identification label and a fault point position label are designed, and a fault feature space sample set convenient for updating and reuse is constructed. (9) Database division: In order to ensure the scientificity of model training, the total fault feature data set is divided into training set and test set in the proportion of 7:3 by using random sampling method. (10) Determine whether to perform fault section identification: If fault section identification is performed, go to step (11); if fault section identification is not performed, go to step (12). (11) PHS-EPGCN-BiGRU section identification: First, build an EPGCN-BiGRU section identification model, initialize each layer of hidden feature dimension, attention weight linear mapping parameter, edge feature dimension and aggregation method, BiGRU neuron number, learning rate and batch size; second, use the PHS algorithm to optimize the hyperparameters of the EPGCN-BiGRU model, and complete the model training; finally, input the test set sample into the trained PHS-EPGCN-BiGRU section identification model to identify the fault branch. (12) PHS-EPGCN-BiGRU accurate positioning: First, build an EPGCN-BiGRU accurate positioning model, initialize each layer of hidden feature dimension, attention weight linear mapping parameter, edge feature dimension and aggregation method, BiGRU layer number and learning rate; second, use the PHS algorithm to optimize the hyperparameters of the EPGCN-BiGRU model, and complete the model training; finally, input the test set sample into the trained PHS-EPGCN-BiGRU accurate positioning model to obtain the fault distance.

[0102] S4: Determine whether to perform fault section identification based on the feature matrix of the multi-dimensional fault feature space, if yes, input the multi-dimensional fault feature space into a section identification model to obtain a section identification result, if no, input the multi-dimensional fault feature space into an accurate positioning model to obtain a positioning result.

[0103] Embodiment Two The embodiment discloses a multi-branch power distribution network fault locating system based on multi-dimensional fault features, comprising: a data acquisition module configured to acquire fault voltage traveling wave waveform data of the multi-branch power distribution network; a feature construction module configured to perform phase-mode transformation on the fault voltage traveling wave waveform data to extract line-mode components, extract fault traveling wave head coordinates, amplitude and slope information within a set time window to form a head information space, determine a feature matrix, an adjacency matrix and an edge feature tensor based on the head information space after redundancy elimination, and construct a multi-dimensional fault feature space based on the feature matrix, the adjacency matrix and the edge feature tensor; a model construction module configured to construct a section identification model and a precise positioning model, and perform hyperparameter optimization on the models by using a parameter-free setting and a sound search algorithm; a fault locating module configured to determine whether to perform fault section identification based on the feature matrix of the multi-dimensional fault feature space, input the multi-dimensional fault feature space into the section identification model to obtain a section identification result if yes, or input the multi-dimensional fault feature space into the precise positioning model to obtain a positioning result if no.

[0104] Embodiment three The purpose of the embodiment is to provide a computing device, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the method of embodiment one when executing the program.

[0105] Embodiment four The purpose of the embodiment is to provide a computer-readable storage medium, a computer-readable storage medium having a computer program stored thereon, wherein the program is executed by a processor to perform the steps of the method of embodiment one.

[0106] The steps involved in the devices of embodiments three and four above correspond to the method of embodiment one, and the specific embodiments can be referred to the relevant description part of embodiment one. The term "computer-readable storage medium" should be understood as including a single medium or multiple media of one or more instruction sets; it should also be understood as including any medium capable of storing, encoding or carrying instruction sets for execution by a processor and causing the processor to perform any method in the present application.

[0107] Those skilled in the art should understand that each module or step of the present application described above can be realized by a general computer device, alternatively, they can be realized by program codes executable by a computing device, so that they can be stored in a storage device for execution by a computing device, or they can be respectively manufactured into individual integrated circuit modules, or a plurality of modules or steps among them can be manufactured into a single integrated circuit module to realize. The present application is not limited to any specific combination of hardware and software.

[0108] The above merely provides the preferred embodiments of the present application, but not intended to limit the present application. The present application can be variously changed and modified by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall fall within the scope of the present application.

[0109] The above merely provides the preferred embodiments of the present application, but not intended to limit the present application. The present application can be variously changed and modified by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall fall within the scope of the present application.

Claims

1. A multi-distributed power grid fault location method based on multi-dimensional fault characteristics, characterized in that: include: Obtaining fault voltage traveling wave waveform data of multi-branch distribution grids; Performing phase mode transformation on the fault voltage traveling wave waveform data to extract the line mode component, extracting the coordinates, amplitude, and slope information of the fault traveling wave head within a set time window to form a wave head information space; after redundancy elimination, determining the characteristic matrix, adjacency matrix, and edge feature tensor based on the wave head information space, and constructing a multidimensional fault feature space based on the characteristic matrix, adjacency matrix, and edge feature tensor; Build segment recognition model and precise positioning model, and use parameter-free harmony search algorithm to tune the model hyperparameters; Based on the characteristic matrix of the multidimensional fault feature space, it is determined whether to perform fault section identification. If so, the multidimensional fault feature space is input into the section identification model to obtain the section identification result. Otherwise, the multidimensional fault feature space is input into the precise positioning model to obtain the positioning result.

2. The multi-branch distribution network fault location method based on multi-dimensional fault characteristics according to claim 1, characterized in that: The traveling wave information characteristics of the multi-branch distribution network at different fault locations and fault branches are analyzed, and the mapping relationship between the wave head timing, polarity and amplitude characteristics and the fault location is obtained.

3. The multi-branch distribution network fault location method based on multi-dimensional fault characteristics according to claim 1, characterized in that: Karenbauer phase mode transformation is performed on the fault voltage traveling wave waveform data to extract the line mode component, and the coordinates, amplitude and slope information of the fault traveling wave head are extracted based on the straight line detection algorithm.

4. The multi-branch distribution network fault location method based on multi-dimensional fault characteristics according to claim 1, characterized in that: The segment recognition model and precise positioning model adopt a hybrid network model, including an edge propagation graph convolutional neural network and a bidirectional gated recurrent unit.

5. The multi-branch distribution network fault location method based on multi-dimensional fault characteristics according to claim 4, characterized in that: The edge propagation graph convolutional neural network constructs edge embedding by fusing node pairs and edge features, and introduces a self-attention mechanism to assign weights. The node features are updated after weighted aggregation and residual connection, and the spatiotemporal features containing topology and edge attributes are extracted.

6. The multi-branch distribution network fault location method based on multi-dimensional fault characteristics according to claim 4, characterized in that: The bidirectional gated recurrent unit adopts a time-aligned hidden state splicing strategy to perform feature splicing to obtain output features.

7. The multi-branch distribution network fault location method based on multi-dimensional fault characteristics according to claim 1, characterized in that: The parameter-free harmony search algorithm synchronously adjusts the harmony memory consideration rate and the pitch adjustment rate.

8. A multi-branch distribution network fault location system based on multi-dimensional fault characteristics, characterized in that: include: A data acquisition module is configured to: acquire fault voltage traveling wave waveform data of a multi-branch distribution network; A feature construction module is configured to: perform phase mode transformation on the fault voltage traveling wave waveform data to extract the line mode component, extract the coordinates, amplitude, and slope information of the fault traveling wave head within a set time window to form a wave head information space; after redundancy elimination, determine a feature matrix, an adjacency matrix, and an edge feature tensor based on the wave head information space, and construct a multidimensional fault feature space based on the feature matrix, the adjacency matrix, and the edge feature tensor; A model building module is configured to: build a segment recognition model and a precise positioning model, and use a parameter-free harmony search algorithm to perform hyperparameter tuning on the models; The fault location module is configured to: determine whether to identify the fault section based on the characteristic matrix of the multidimensional fault feature space; if so, input the multidimensional fault feature space into the section identification model to obtain the section identification result; otherwise, input the multidimensional fault feature space into the precise positioning model to obtain the positioning result.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the multi-branch distribution network fault location method based on multi-dimensional fault characteristics as claimed in any one of claims 1 to 7 are implemented.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the multi-branch distribution network fault location method based on multi-dimensional fault characteristics are implemented as described in any one of claims 1 to 7.

Images