Neutral point small current grounding power distribution network fault section positioning method based on space-time diagram wavelet neural network

Abstract

The invention relates to the technical field of power distribution network fault detection, and provides a neutral point small current grounding power distribution network single-phase grounding fault section positioning method based on a space-time diagram wavelet neural network. The method comprises the following steps: constructing a training data set covering multiple types of single-phase earth fault scenes, and carrying out time sequence electrical quantity sampling; data preprocessing is completed through operations such as data cleaning and missing completion; constructing a spatial-temporal feature extraction module fusing a graph wavelet neural network and a gating circulation unit network, and carrying out spatial-temporal joint modeling on node voltage and current time sequence information; and outputting a node-level fault positioning probability through a full connection layer and a Softmax classifier. According to the method, the response capability of the model to topological disturbance is enhanced through frequency domain wavelet filtering, the information transmission efficiency is improved in combination with a residual mechanism, the fault section identification precision under the conditions of sparse observation data and weak electrical disturbance characteristics is effectively improved, and the method is suitable for a small current grounding fault positioning task in a complex operation environment of a power distribution network.

Description

Technical Field

[0001] The present invention belongs to the technical field of distribution network fault detection, and in particular relates to a method for locating a fault section of a neutral point low current grounded distribution network based on a space-time graph wavelet neural network. Background Art

[0002] In modern power systems, distribution networks, as the terminal link in power transmission, play a critical role in distributing power from the high-voltage transmission network to end users. The neutral-point grounding methods in my country's distribution networks can be divided into high-current grounding systems and low-current grounding systems, depending on the magnitude of the fault current after a fault occurs. Low-current grounding distribution networks are widely used in urban and rural distribution networks because they can maintain short-term power supply during faults.

[0003] However, there are many technical difficulties in locating the fault section of a low-current grounded distribution network: first, the fault characteristics of a single-phase grounding fault are weak, making it difficult to accurately set the fault judgment threshold, which can easily lead to misjudgment or missed judgment; second, with the large-scale access of renewable energy sources such as photovoltaic and wind power to the distribution network, the direction and distribution of the fault current have changed significantly, and the applicability of traditional positioning methods based on steady-state models has decreased; third, in remote areas or rural areas, distribution monitoring equipment is sparsely deployed, and the available fault data is limited, further increasing the difficulty of fault location.

[0004] Currently, methods for locating fault sections in distribution networks mainly include mathematical methods based on fault feature modeling and data-driven artificial intelligence methods. Mathematical methods based on fault feature modeling are highly dependent on network parameters and fault models, and their positioning accuracy, security, and generalization capabilities are difficult to guarantee under complex working conditions. In recent years, although artificial intelligence methods such as deep learning and graph neural networks have been applied to distribution network fault location, the technologies described in patents such as CN118914759A, CN118569844A, and CN115659258A have not effectively solved the problem of the effectiveness of artificial intelligence models in fault location under sparse data scenarios.

[0005] To address these issues, this paper proposes a method based on a spatiotemporal graph wavelet neural network to address the task of locating fault sections in low-current grounded distribution networks under sparse monitoring scenarios. By incorporating graph wavelet transform technology and spatiotemporal networks, the model can, to a certain extent, fill the gaps left by the lack of neighbor features and fully exploit the spatial and temporal characteristics of known data, thereby achieving precise positioning. Summary of the Invention

[0006] The purpose of the present invention is to provide a fault section location method for a neutral point low current grounding distribution system based on a space-time graph wavelet neural network, which can realize the accurate and rapid location of faults in the distribution network under the scenario of limited observable data nodes, thereby ensuring the robustness and power supply reliability of the distribution system.

[0007] To achieve the above-mentioned purpose, the present invention is implemented through the following technical solutions:

[0008] Step S1: Obtain a dataset for a distribution network with a low-current grounded neutral point fault. Obtain operational data from observable nodes in the grid under different fault scenarios using a real distribution system or a distribution network simulation model. Select equally spaced time segments and extract time-series electrical quantity information covering the entire steady-state, transient, and new steady-state process from the sampled data. This information serves as input data for subsequent modeling.

[0009] Step S2: Preprocess the collected data, including data cleaning, missing value processing, timestamp processing, and data labeling.

[0010] Step S3: Construct a fault section location model for the neutral point small current grounding distribution network based on the spatiotemporal graph wavelet neural network. The proposed model is divided into two steps: spatiotemporal feature extraction and classification.

[0011] Step S4: Model training and testing. The model is trained using the training data preprocessed in Step 1. After the model's loss function converges, the network parameters are saved. The model's positioning accuracy is tested using the test data to calculate the location of the distribution network ground fault.

[0012] As a preferred solution of the present invention, the method for generating a single-phase grounding fault dataset of a neutral point low current grounding distribution network in step S1 includes:

[0013] Set different low-current grounding methods for the neutral point of the distribution network. These include: ungrounded neutral point, high-resistance grounded neutral point, and grounded neutral point via arc suppression coil. During the simulation experiment, the neutral point line grounding method should be dynamically adjusted to ensure the objective authenticity and effective generalization of the data sample.

[0014] We set up different single-phase grounding fault scenarios for the distribution network. This involved dynamically adjusting the fault location, the grounding method, the grounding resistor value, and the ratio of overhead lines to cables. We conducted numerous experiments and collected fault sample data for each of these scenarios.

[0015] All sample data of different single-phase grounding faults occurring under different distribution network neutral point grounding modes are collected respectively, and a neutral point small current grounding distribution network fault dataset is constructed. The dataset is divided into training set, validation set and test set in the ratio of 7:1:2.

[0016] Furthermore, the fault sample data obtained in step S1 includes three-phase voltage values, three-phase current values, zero-sequence voltage values, and zero-sequence current values ​​of the observation node.

[0017] As a preferred solution of the present invention, the method of data preprocessing in step S2 includes:

[0018] Data cleaning: Based on the outlier detection method of statistical analysis, the interquartile range of the running data of each node is calculated to identify and eliminate abnormal data points that are beyond the normal range.

[0019] Missing value processing: Taking advantage of the continuity of time series data, missing data is filled by combining linear interpolation and cubic spline interpolation.

[0020] Timestamp processing: Ensures accurate alignment of timestamps for all data samples, performs synchronization correction on data with time deviations, unifies time resolution, and eliminates time inconsistencies caused by acquisition devices or simulation processes.

[0021] Data labeling: Using fault location and fault type as key information, accurately label data samples and establish a clear labeling system to ensure a one-to-one correspondence between labels and data samples, providing a clear classification basis for model training.

[0022] As a preferred solution of the present invention, in step S3, the method for locating the fault section of the neutral point low current grounding distribution network based on the space-time graph wavelet neural network includes:

[0023] A spatiotemporal feature extraction module is designed, which combines a graph wavelet neural network (GWNN) and two gated recurrent unit networks (GRU) to extract the complex spatiotemporal dependencies between the distribution network topology and the node time series electrical quantities.

[0024] The graph wavelet neural network module is based on the graph Laplacian matrix generated by the distribution network topology structure, and further constructs a multi-scale graph wavelet filter to capture spatial features at different frequency scales and perform sparse domain convolution extraction on node features.

[0025] Specifically, let the input tensor

[0026]

[0027] Where t represents the time step and m represents the number of input feature channels. After being processed by a layer of graph wavelet convolutional neural network (GWNN), its output feature It can be expressed as:

[0028]

[0029] Among them, ψ s represents the wavelet transform kernel, is its approximate inverse transform (in practice, its transpose is often taken), is the trainable frequency domain filter weight, and ReLU is the activation function. This operation means that each input feature dimension is first subjected to an inverse wavelet transform, then weighted by the frequency domain filter, and then activated by a forward wavelet transform back to the image domain.

[0030] Among them, the wavelet transform kernel ψ s , which is defined as:

[0031] ψ s =PA s P T

[0032] is an orthogonal matrix composed of the eigenvectors of the graph Laplacian matrix L;

[0033] A s =diag(g(sλ1),…,g(sλ n ))

[0034] is the spectral domain wavelet kernel diagonal matrix with scale parameter s, λ i is the Laplace eigenvalue, and g(·) is the given spectral filter function.

[0035] For time series dimension feature extraction, the gated recurrent unit (GRU) is used to model the multi-time step features extracted by the graph wavelet module, extract the time series evolution features of each node, and retain important transient disturbance patterns.

[0036] The input of the spatiotemporal feature extraction module is the preprocessed three-phase current, voltage, zero-sequence current and voltage tensors of the observable distribution nodes, whose spatial structure is defined by the physical topology of the distribution network, and the output is the spatiotemporal fusion feature vector of each node.

[0037] Subsequently, the time-space fusion feature vector is nonlinearly mapped and normalized through the multi-layer perceptron and Softmax layer, and the probability of each node being a fault section is output to achieve accurate positioning of the fault section of the distribution network.

[0038] As a preferred solution of the present invention, the content of model training and testing in step S4 includes:

[0039] The model adopts the node-level cross entropy loss function in the training phase. Considering the imbalanced number of actual fault samples, a category weight adjustment mechanism is introduced to improve the fault recognition accuracy. The loss function is constructed as follows:

[0040]

[0041] Among them, y i,c represents the true label of the i-th node in the c-th category, represents its predicted probability, ω c are the importance weight parameters for different categories.

[0042] Furthermore, to prevent the model from overfitting, the L2 regularization term is further introduced into the loss function to constrain the model. The final loss function is constructed as follows:

[0043]

[0044] Among them, β is the regularization coefficient and Θ is the set of all trainable parameters.

[0045] Furthermore, after the model training results converge, the accuracy of the model positioning is tested using the data from the test set.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] By introducing a graph decomposition mechanism, this invention achieves selective response to different topological features in the frequency dimension, effectively extracting the local disturbance characteristics and spatial structural variation patterns of each node in the distribution network. Furthermore, combined with the gated recurrent unit network's ability to dynamically model node time-series electrical quantities, this method possesses strong fault feature capture and segment location capabilities even in scenarios with sparse observation nodes and incomplete time-series information. This improves the model's robustness, adaptability, and generalization capabilities for locating low-current ground faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a flow chart of a method for locating a fault section in a neutral point low current grounded distribution network based on a space-time graph wavelet neural network according to an embodiment of the present invention;

[0049] Figure 2 1 is a network structure diagram of a space-time graph wavelet neural network according to an embodiment of the present invention;

[0050] Figure 3 1 is a module structure diagram of the spatiotemporal feature extraction module G3 shown in an embodiment of the present invention; Specific implementation plan

[0051] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described below with reference to the accompanying drawings and embodiments.

[0052] A method for locating a fault section in a neutral point low current grounding distribution network based on a space-time graph wavelet neural network, the method comprising the following steps:

[0053] Step S1: By collecting real distribution system operation data or building a distribution network simulation model to generate simulation data, a neutral point small current grounding distribution network fault section location data set is constructed.

[0054] Step S2: Preprocess the collected data separately, including data cleaning, missing value processing, timestamp processing and data labeling.

[0055] Step S3: Construct a fault section location model for neutral point low current grounding distribution network based on spatiotemporal graph wavelet neural network. The proposed model is divided into two steps: spatiotemporal feature extraction and classification. The temporal feature extraction uses gated recurrent unit (GRU), the spatial feature extraction uses graph wavelet neural network (GWNN), and the fault location result is obtained using Softmax classifier.

[0056] Step S4: Model training and testing. The model is trained using the training data preprocessed in Step 1. After the model's loss function converges, the network parameters are saved. The model's positioning accuracy is tested using the test data to calculate the location of the distribution network ground fault.

[0057] Example 1

[0058] Reference Figure 1 As shown, this embodiment specifically illustrates the implementation details and key technical steps of the method for locating the fault section of a neutral point small current grounded distribution network based on a space-time graph wavelet neural network proposed in the present invention through further refined process descriptions and mathematical expressions, so as to facilitate a comprehensive understanding and technical reproduction of the technical solution of the present invention.

[0059] Step S1: constructing a data set for locating the fault section of a distribution network with a low neutral current grounding.

[0060] To build a high-quality distribution network fault dataset suitable for deep learning model training and testing, we first need to design and build a distribution network simulation environment for the neutral point low current grounding method, or select actual operating data with typical operating conditions. By systematically changing the neutral point grounding method, selecting a variety of typical distribution network topologies, setting different single-phase grounding point locations and various fault types, we can comprehensively cover a variety of possible fault conditions. For details on various experimental variables and their combination schemes, see

[0061] Listed in Table 1, it is used to ensure the breadth, representativeness and generalization ability of the data sample.

[0062]

[0063]

[0064] Based on the experimental variable factors described in the table, 14,400 (3 × 4 × 3 × (32 + 68) × 4) experiments are required. During the simulation, the following electrical quantities should be collected at each observation node, including but not limited to: three-phase current, three-phase voltage, zero-sequence current, and zero-sequence voltage. The sampling frequency is 2 kHz to ensure that the detailed characteristics of the transient steady state are fully recorded. Each sample data should cover the entire process from the steady state before the fault, the moment of fault occurrence, to the new steady state shortly after the fault, with a typical data length of 100 ms.

[0065] To ensure the model generalization ability and evaluation effect in the training, verification and testing stages, all collected fault sample data sets are divided into training set, verification set and test set according to the ratio of 70%, 10% and 20% respectively. The fault category, fault occurrence time and actual fault section label corresponding to each group of samples are recorded for subsequent model supervised learning.

[0066] Step S2: Preprocessing of the data set for locating the fault section of the neutral point small current grounding distribution network

[0067] In this embodiment, to ensure the readability and stability of the fault data input to the model and improve the training efficiency and generalization performance of the subsequent model, the original fault sample data constructed in step S1 is fully preprocessed. The data preprocessing process includes four steps: data cleaning, missing value filling, timestamp alignment, and data labeling.

[0068] Data cleaning:

[0069] To address simulation errors or noise from field measurements, a statistically distributed outlier detection mechanism is employed. Specifically, the mean and interquartile range of the voltage, current, and zero-sequence quantity time series data for each observation node are calculated, and a reasonable anomaly threshold range is defined. Data points exceeding the upper or lower thresholds are considered outliers and removed or replaced with a sliding average to smooth out abnormal fluctuations and eliminate the impact of systematic disturbances.

[0070] Furthermore, during the data cleaning process, a robust outlier detection strategy based on time windows and contextual information was implemented to prevent key disturbance features during ground fault transients from being misidentified as outliers. For sections suspected of transient anomalies, a retention mechanism was implemented and the model was used to further learn their dynamic features, ensuring that critical fault information was not mistakenly deleted.

[0071] Missing value filling:

[0072] Taking into account the deployment of the actual distribution network monitoring system or the system errors or operational errors that may exist in the simulation data recording process, some observation nodes may experience data collection interruption or data transmission loss in some time periods. In response to the above problems, the present invention adopts an interpolation completion strategy based on the continuity of time series to repair missing data and ensure the integrity and availability of the input sequence. Among them, for the case where the missing time span is short, linear interpolation is used to recover the data; for the section with a longer missing time, the cubic spline interpolation method is introduced to better maintain the smoothness and trend consistency of the data, thereby providing a stable and reliable input basis for the time series modeling and feature learning of the subsequent model.

[0073] 3. Timestamp alignment:

[0074] Data collected at each node in the distribution network may exhibit slight time deviations. Directly inputting this data into the model can distort time series characteristics. Therefore, during the preprocessing phase, the timestamps of all observations are aligned to the same time base. A time window sliding matching algorithm based on a reference synchronization node ensures that all data samples are strictly synchronized at every time step, improving the accuracy of subsequent time series modeling.

[0075] 4. Data annotation:

[0076] To enable supervised learning training, the present invention accurately labels each sample sequence. This labeling information includes whether the sample has a fault, and if so, the fault location, fault type, and ground resistance category. This labeling information is stored in a structured format, corresponding one-to-one with the sample data, ensuring that the corresponding supervisory information can be used for loss calculation and accuracy assessment during model training and evaluation.

[0077] Through the above preprocessing steps, a distribution network fault sample dataset with standardized structure, consistent dimensions, unified time series and complete labels is obtained, providing a high-quality input foundation for the model in subsequent feature extraction and segment identification tasks.

[0078] Step S3: constructing a fault section location model for a neutral point low current grounding distribution network based on a space-time graph wavelet neural network.

[0079] In this embodiment, in order to fully explore the dynamic characteristics of node electrical quantities in the spatial structure and time series dimensions during the distribution network fault process, the present invention constructs a distribution network fault section location model based on the Spatio-Temporal Graph Wavelet Neural Network (ST-GWNN).

[0080] The network structure of the space-time graph wavelet neural network is as follows Figure 3As shown in Figure 3, the model consists of two parts: a spatiotemporal feature extraction module and a detection output module.

[0081] 1. Construction of spatiotemporal feature extraction module

[0082] The spatiotemporal feature extraction module is named G3 module, and its network structure is shown in the following figure: Figure 2 As shown. The G3 module contains the following core substructures:

[0083] (1) Graph Wavelet Neural Network (GWNN):

[0084] Based on the physical topology of the distribution network, the GWNN constructs a Laplacian matrix L and obtains its eigenvector matrix P and eigenvalue diagonal matrix through spectral decomposition. Combined with the wavelet basis function g(sλ), a spectral domain wavelet kernel with adjustable frequency response is constructed. Multi-scale spatial filtering is performed on node features to extract spatial structural features in different frequency bands. Assuming the input tensor dimension is T×N×F, where T is the number of time steps, N is the number of nodes, and F is the number of input channels, the GWNN output feature X′ is expressed by the following formula:

[0085] X′=σ(Ψ T ·A·Ψ·X)

[0086] Here, Ψ represents the wavelet transform kernel matrix, A is the diagonal matrix of trainable spectral filter weights, and σ is a nonlinear activation function (such as ReLU). This process achieves the inverse transformation of information from the frequency domain to the image domain and combines learnable weights to enhance the structural response capabilities at different frequencies.

[0087] Furthermore, Ψ is defined as:

[0088] ψ s =PA s P T

[0089] Where P is an orthogonal matrix composed of the eigenvectors of the graph Laplace matrix L; A is a spectral domain wavelet kernel diagonal matrix with a scale parameter of s, and λ i is the Laplace eigenvalue, and g(·) is the given spectral filter function.

[0090] Furthermore, the spectral domain wavelet kernel diagonal matrix A is expressed as:

[0091] A=diag(g(sλ1),…,g(sλ n ))

[0092] Specifically, let the input distribution network single-phase grounding fault feature tensor be I. After being processed by a layer of graph wavelet convolutional neural network (GWNN), its output feature can be expressed as:

[0093]

[0094] Among them, ψ s represents the wavelet transform kernel, is its approximate inverse transform (in practice, its transpose is often taken), is the trainable frequency domain filter weight, and ReLU is the activation function. This operation means that each input feature dimension is first subjected to an inverse wavelet transform, then weighted by the frequency domain filter, and then activated by a forward wavelet transform back to the image domain.

[0095] Gated Recurrent Unit (GRU):

[0096] In order to effectively model the evolution relationship of electrical quantities of distribution network nodes in the time dimension, the present invention further organizes the spatial structural features according to the time series and inputs them into the gated recurrent unit network (GRU) for time series feature modeling after the graph wavelet neural network (GWNN) extracts them. GRU is a variant of recurrent neural network with a streamlined structure and excellent performance. It has the ability to model long-term and short-term dependencies and is particularly suitable for scenarios where time series data contains significant transient disturbances and steady-state transition processes. In the present invention, GRU can capture the transient and steady-state changes of voltage and current signals at different time scales, and dynamically retain or forget historical information according to the gating mechanism. For the feature vector sequence of each node, its time series evolution is described by the following formula:

[0097] z t =σ(W z x t +U z h t-1 +b z )

[0098] r t =σ(W r x t +U r h t-1 +b r )

[0099]

[0100] The above four formulas represent the update gate calculation, reset gate calculation, candidate state calculation and final state update of GRU respectively. t represents the input of the tth time step, h t is the output hidden state, z t 、r t They are update gate and reset gate respectively.

[0101] Through the above mechanism, GRU can independently determine which historical information should be retained for long-term modeling and which can be ignored based on the changes in the characteristics of the current input, thereby enhancing the model's ability to robustly capture transient disturbances and steady-state transitions, and adapting to the characteristics of significant but short-lived fluctuations in node electrical quantities during ground faults.

[0102] Furthermore, to improve the model's expression stability and the information transfer efficiency of its multi-layer structure, the present invention introduces a residual connection mechanism in each spatiotemporal feature extraction module G3. Specifically, the module input and output are aligned in their feature dimensions and then element-by-element addition is performed to form a residual fusion feature. This mechanism helps preserve low-order feature information in the original topology, enhancing the model's responsiveness to weak disturbances in low-current fault identification scenarios. It also alleviates the gradient degradation problem during model training, improving the accuracy and robustness of positioning results.

[0103] 2. Detection output module construction

[0104] After extracting the spatiotemporal features, a fused feature vector corresponding to each node is obtained. This feature vector is then input into a classification network consisting of a multi-layer perceptron (MLP) and a softmax layer to achieve node-level fault identification.

[0105] The classifier outputs the probability that each node is a fault segment, and the final positioning result is the node index or segment number corresponding to the maximum probability value. The output of the Softmax layer is expressed as follows:

[0106]

[0107] Where C is the number of categories (can be binary or multi-category), z i is the output of the perceptron.

[0108] Step S4: training and testing a fault section location model for a neutral point low current grounding distribution network based on a space-time graph wavelet neural network.

[0109] In this embodiment, in order to realize the effective application of the proposed space-time graph wavelet neural network model in the fault location task of small current grounding distribution network, the model is trained and tested, and its performance is evaluated using quantitative indicators.

[0110] 1. Model training process

[0111] The training process is based on the constructed training set and validation set, and uses a supervised learning strategy to iteratively optimize the model parameters. The main process is as follows:

[0112] (1) Loss function design:

[0113] Considering the imbalance between different categories in distribution network fault samples, in order to improve the model's ability to identify the minority class (i.e., faulty nodes), a weighted cross-entropy loss function (Weighted Cross-Entrs) is used, which is defined as follows:

[0114]

[0115] Among them, N is the number of nodes, C is the number of categories, and y i,c is the true label of the i-th node under the c-th category, Its corresponding predicted probability, ω c is the loss weight of class c.

[0116] (2) Regularization constraints:

[0117] In order to suppress the overfitting problem of the model, the L2 regularization term is introduced to constrain the set of all trainable parameters Θ. The final loss function is defined as:

[0118]

[0119] Among them, λ is the regularization coefficient, which is usually adjusted according to the performance of the validation set.

[0120] (3) Training parameter configuration:

[0121] The Adam optimizer is used during training, with an initial learning rate of 0.00 and dynamic adjustment based on validation set performance. The batch size Θ is capped at 100 training epochs or until validation set loss converges.

[0122] (4) Early stopping mechanism:

[0123] If the accuracy of the validation set does not improve for several consecutive rounds, the training is terminated early to avoid overfitting.

[0124] 2. Model testing and performance evaluation

[0125] After training is complete, load the optimal parameters and perform model inference on an independent test set to evaluate the accuracy of fault section location. Common evaluation indicators include:

[0126] Accuracy: the proportion of correctly classified nodes;

[0127] Recall: The ratio of correctly identified faulty nodes to all true faulty nodes.

[0128] F1-score: The harmonic mean of precision and recall, used to comprehensively evaluate model performance;

[0129] Positioning error distance (if the data contains geographic / topological information): the shortest path distance or physical distance between the predicted node and the actual fault node.

[0130] The output result of the testing phase is the probability that each node belongs to the fault segment. The predicted label is selected through the threshold judgment mechanism and compared with the real label to evaluate the performance.

[0131] It should be noted that the embodiments described are only some of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0132] Those skilled in the art will understand that the accompanying drawings are only a schematic diagram of a preferred embodiment, and the serial numbers of the embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

Claims

1. A method for locating fault sections in neutral point low current grounding distribution network based on space-time graph wavelet neural network, characterized in that: The method comprises the following steps: S1: Construct a neutral point low current grounding distribution network fault dataset, which is generated using real distribution network operation data or simulation models and contains electrical quantity information under multiple fault scenarios; S2: Preprocessing the fault data, including data cleaning, missing value filling, timestamp alignment and data labeling; S3: A distribution network fault section location model is constructed based on a spatiotemporal graph wavelet neural network. This model consists of a spatiotemporal feature extraction module and a detection output module. The spatiotemporal feature extraction module is used to jointly extract the spatial topological structure characteristics and time series evolution characteristics of electrical quantities; the detection output module outputs the probability of each node being a fault point. S4: The model is trained based on the cross-entropy loss function and the category weight strategy to ultimately achieve accurate positioning of the fault section in the distribution network.

2. The method according to claim 1, characterized in that The method of constructing the data set in step S1 includes the following operations: Set the neutral point grounding mode of the distribution network to be ungrounded, high-resistance grounded or grounded via arc suppression coil; Set different single-phase ground fault locations, ground resistance values, ground fault types, fault occurrence times and line combination types respectively; All sample data obtained are divided into training set, validation set and test set according to the ratio.

3. The method according to claim 1, characterized in that The data preprocessing method in step S2 includes: Data cleaning: Based on the outlier detection method of statistical analysis, the interquartile range of the running data of each node is calculated to identify and eliminate abnormal data points that are outside the normal range; Missing value processing: Taking advantage of the continuity of time series data, missing data is filled by combining linear interpolation with cubic spline interpolation; Timestamp processing: ensures accurate alignment of timestamps for all data samples, performs synchronization correction on data with time deviations, unifies time resolution, and eliminates time inconsistencies caused by acquisition devices or simulation processes. Data labeling: Using fault location and fault type as key information, accurately label data samples and establish a clear labeling system to ensure a one-to-one correspondence between labels and data samples, providing a clear classification basis for model training.

4. The method according to claim 1, wherein The graph wavelet neural network is based on the graph Laplace matrix generated by the distribution network topology structure to construct the graph wavelet transform kernel. Where P is the eigenvector matrix of the Laplace matrix, A s is the spectral domain wavelet kernel diagonal matrix, and the convolution output of the graph wavelet neural network is: Among them, ψ s represents the wavelet transform kernel, is its approximate inverse transform (in practice, its transpose is often taken), is the trainable frequency domain filter weight, and ReLU is the activation function.

5. The method according to claim 1, wherein The gated recurrent unit (GRU) is used to model the temporal features of multiple time steps of a node, and the hidden state of its output is a fusion of features extracted by graph wavelet according to the time series.

6. The method according to claim 1, characterized in that The Softmax classifier is used to map the fused features to the fault probability distribution at the node level to achieve node classification.

7. The method according to claim 1, characterized in that The loss function is a weighted cross entropy function, defined as: Among them, y i,c represents the true label of the i-th node in the c-th category, represents its predicted probability, ω c are the importance weight parameters for different categories; To prevent the model from overfitting, the L2 regularization term is further introduced into the loss function to constrain the model. The final loss function can be expressed as: Among them, β is the regularization coefficient and Θ is the set of all trainable parameters.

Images