Power distribution network feeder line fault detection method, device and system, and storage medium

Abstract

The invention discloses a power distribution network feeder fault detection method, device and system, and a storage medium. The method comprises the steps of obtaining topological data of a power distribution network, and real-time data, historical data and fault data collected by an intelligent electric meter and a sensor; obtaining heterogeneous graph characteristics of topological data, real-time data, historical data and fault data at each moment through a heterogeneous graph time sequence convolutional network; performing feature aggregation and fusion on the heterogeneous time sequence diagram features to obtain power distribution network feeder features; training an LGNN network according to the feeder characteristics of the power distribution network to obtain a feeder fault detection model; wherein the trained feeder line fault detection model is used for feeder line state detection. By adopting the technical scheme of the invention, the high-efficiency detection of the feeder line fault of the power distribution network is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of distribution network fault detection, and in particular relates to a distribution network feeder fault detection method and device, system, and storage medium. Background Art

[0002] As an important part of the power system, the stability and reliability of the distribution network are crucial to the safe operation of the entire power grid. Feeder fault detection is a key link to ensure the stable operation of the distribution network. With the development of intelligent sensing technology and information and communication technology, the distribution network can collect a large amount of real-time operation data. However, how to accurately and quickly identify feeder faults from this data still faces serious challenges. Traditional fault detection methods mainly rely on set thresholds and empirical rules, which have limitations in dealing with feeder fault detection under complex distribution network conditions.

[0003] At present, distribution network feeder fault detection mainly adopts some classic machine learning algorithms, such as support vector machines, random forests, etc. These algorithms are trained on historical data sets to learn fault features to achieve fault classification and location. Although these methods have achieved certain results in some cases, they usually require experts to perform tedious feature engineering and are difficult to adapt to the dynamic changes in power grid operating conditions. In addition, these models also have deficiencies in generalization ability, real-time and interpretability, which limits their application effect in actual power grid fault detection. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a distribution network feeder fault detection method and device, system, and storage medium to achieve efficient detection of distribution network feeder faults.

[0005] To achieve the above object, the present invention adopts the following technical solution:

[0006] A method for detecting a distribution network feeder fault, comprising:

[0007] Step S1, obtaining topological data of the distribution network, real-time data, historical data and fault data collected by smart meters and sensors;

[0008] Step S2: Obtain heterogeneous graph features of topological data, real-time data, historical data, and fault data at each moment through a heterogeneous graph temporal convolutional network;

[0009] Step S3, performing feature fusion on the heterogeneous time series diagram features to obtain distribution network feeder features;

[0010] Step S4: training the LGNN network according to the distribution network feeder characteristics to obtain a feeder fault detection model; wherein the trained feeder fault detection model is used to perform feeder status detection.

[0011] Preferably, step S3 comprises:

[0012] Based on RCA, the heterogeneous timing graph features on fault data, real-time data and historical data are aggregated to obtain the heterogeneous timing graph feature representation.

[0013] Heterogeneous timing diagram features Send it to the graph convolutional network for training and learning to obtain the distribution network feeder characteristics Right now,

[0014]

[0015] Among them, GCN represents the graph convolution operation, Respectively represent the extracted features of fault, real-time, and historical data sources.

[0016] Preferably, the loss function of the feeder fault detection model is:

[0017]

[0018] Among them, x i Representative Matrix The feature vector of each node in .

[0019] The present invention also provides a distribution network feeder fault detection device, comprising:

[0020] The acquisition module is used to obtain the topological data of the distribution network, the real-time data collected by smart meters and sensors, historical data and fault data;

[0021] The first processing module is used to obtain heterogeneous graph features of topological data, real-time data, historical data and fault data at each moment through a heterogeneous graph temporal convolutional network;

[0022] The second processing module is used to perform feature fusion on the heterogeneous time sequence diagram features to obtain the distribution network feeder features;

[0023] The detection module is used to train the LGNN network according to the distribution network feeder characteristics to obtain a feeder fault detection model; wherein the trained feeder fault detection model is used to perform feeder status detection.

[0024] Preferably, the second processing module comprises:

[0025] The first processing unit is used to perform aggregation operations on the heterogeneous timing graph features on the fault data, real-time data and historical data based on RCA to obtain the heterogeneous timing graph feature representation

[0026] The second processing unit is used to convert the heterogeneous timing diagram features Send it to the graph convolutional network for training and learning to obtain the distribution network feeder characteristics Right now,

[0027]

[0028] Among them, GCN represents the graph convolution operation, Respectively represent the extracted features of fault, real-time, and historical data sources.

[0029] Preferably, the loss function of the feeder fault detection model is:

[0030]

[0031] Among them, x i Representative Matrix The feature vector of each node in .

[0032] An embodiment of the present invention further provides a distribution network feeder fault detection system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a distribution network feeder fault detection method when executed by the processor.

[0033] An embodiment of the present invention further provides a storage medium, on which a computer program is stored, and the computer program executes the distribution network feeder fault detection method when running.

[0034] The present invention first collects key electrical parameters and environmental data through intelligent sensors, and ensures data quality through cleaning and standardization. Then, the HTGCN model is constructed to fuse multi-source time series data to form a heterogeneous time series diagram to automatically extract feeder fault characteristics. Next, the RCA technology is used to aggregate the features of fault, real-time and historical data to enhance the model's spatiotemporal information characterization capabilities. By training on historical fault and normal data, the present invention not only achieves rapid and accurate analysis of real-time monitoring data, but also improves the ability to distinguish different feeder faults; it not only improves the accuracy and efficiency of feeder fault detection, but also enhances the generalization and interpretability of the model; it provides strong technical support for the stable operation and safe management of the distribution network. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0036] Figure 1 This is a flow chart of a method for detecting a feeder fault in a distribution network according to an embodiment of the present invention;

[0037] Figure 2 Schematic diagram of the aggregation process of heterogeneous timing graph features using RCA. DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0039] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0040] Embodiment 1:

[0041] The embodiment of the present invention provides a distribution network feeder fault detection method, which uses a heterogeneous temporal graph convolutional network (HTGCN) and residual compression aggregation (RCA) technology to achieve efficient detection of distribution network feeder faults. Figure 1 As shown, including:

[0042] Step 1: Data preparation

[0043] 1) Fault data generation

[0044] In order to solve the problem of insufficient distribution network feeder fault data, the present invention uses the fault data generated by the determinant point process (DPP: Determinant Point Process) model to simulate various types of fault conditions that may occur in the real distribution network, such as single-phase grounding faults, two-phase short circuit faults, three-phase short circuit faults, etc. of the feeder. This data generation process can expand the diversity and coverage of the training set, ensuring that the feeder fault detection model can also show good performance when facing rare or unknown fault types. The generated fault data mainly includes the exact time point, location, duration, distribution network status and fault characteristics of the fault. In this way, the generated feeder fault data can enhance the training effect of the feeder fault detection model and improve its adaptability and detection accuracy to real feeder fault events in the distribution network. The process of generating distribution network feeder fault data using the DPP model includes defining the fault feature space, constructing a similarity matrix, calculating the DPP kernel matrix, generating a fault data set, and evaluating and iteratively optimizing the fault data set, etc., which are described in detail as follows:

[0045] (1) Defining fault feature space: When generating a distribution network feeder fault, it is first necessary to define the feature space of the feeder fault. In the present invention, the fault feature space F is defined as F = {f1, f2, ..., f i ,…,f n}, where f i Indicates any possible fault characteristic.

[0046] (2) Constructing a similarity matrix: When constructing the similarity matrix S, the cosine similarity formula (1) is used to calculate the fault feature f i and f j The similarity between them is used to obtain the element s of S ij .

[0047]

[0048] (3) Calculate the DPP kernel matrix: The kernel matrix K can be calculated by the similarity matrix S. Each element k in K ij Indicates the fault characteristics f i and f j Interactions between:

[0049]

[0050] Among them, γ is the parameter controlling the interaction, s ij represents any element in S.

[0051] (4) Generate a fault data set: That is, use the DPP model to select a set of fault features from the feature space to generate a fault data set. The specific set of fault features selected is determined by the DPP sampling probability, that is, for any subset in the kernel matrix K Its sampling probability P(S) is given by the following formula:

[0052]

[0053] Among them, |·| represents the determinant operation, K S is the submatrix in the kernel matrix K corresponding to the subset S.

[0054] (5) Evaluation and iterative optimization: that is, evaluate whether the generated fault data set GF meets the expected value. If so, directly output the fault data set; otherwise, adjust the parameter γ of DPP according to the evaluation result and iteratively optimize.

[0055] 2) Integration of real-time data and historical data

[0056] In the distribution network feeder fault detection system, it is necessary to integrate real-time sensor data and historical data to achieve accurate fault diagnosis. First, the timestamp synchronization technology is used to ensure that the real-time data collected from smart meters, sensors and other devices are accurately aligned with the historical records. Then, data fusion algorithms such as Kalman filtering or sequence-to-sequence models in machine learning are used to fuse high-frequency real-time data and low-frequency historical data to form a continuous time series. Finally, feature extraction techniques such as wavelet transform or short-time Fourier transform are used to extract key features from the fused data, such as sudden changes in current and voltage, frequency deviation, etc., so that HTGCN can learn the timing pattern and feature evolution of fault occurrence and improve the accuracy and reliability of fault detection.

[0057] 3) Distribution network topology data sorting

[0058] The topology of the distribution network is crucial for fault detection, so the accuracy and up-to-dateness of the topology data must be ensured. Specifically, it is necessary to accurately collect and update the geographic information of the feeders, the connection methods of the lines, the layout of the substations and user nodes, and the parameters of the protection devices and switchgear. These topological data are used to construct a graph structure of the distribution network topology as the input of HTGCN. In this way, HTGCN is able to simulate the propagation process of faults in the power grid, identify the fault path, and associate the fault characteristics with the physical structure of the power grid, so as to quickly and accurately locate the fault location and provide decision support for fault isolation and repair.

[0059] 4) Data preprocessing

[0060] Distribution network data usually contains noise and incomplete information, so it needs to be preprocessed before inputting into HTGCN, such as denoising, interpolation and normalization. It is also necessary to extract features useful for fault detection from the raw data, such as sudden changes in current and voltage, frequency deviation, energy change, etc. These features will be used as inputs of HTGCN to improve the fault recognition rate of the model. The following lists some denoising, interpolation and normalization methods used in the present invention:

[0061] (1) Denoising method

[0062] a) Use moving average filter for denoising, the formula is as follows:

[0063]

[0064] Among them, y t is the filtered signal at time t, x t is the original signal and n is the size of the filter window.

[0065] b) Use continuous wavelet transform for denoising. First calculate the wavelet coefficients W(f,t):

[0066]

[0067] Where x(t) is the original signal, is the wavelet function, f is the frequency, t is the time, f0 is the center frequency of the wavelet, and t0 is the time shift. Then, the denoised signal of x(t) is

[0068]

[0069] in, H is the threshold function, which is used to reduce the wavelet coefficients to the threshold range; is the complex conjugate wavelet function, f and t represent the frequency and translation parameters respectively.

[0070] c) Use the reconstruction error of the autoencoder for denoising, the formula is:

[0071]

[0072] Among them, ψ(·) and σ(·) are the decoding function and activation function of the autoencoder, W and b are the weight and bias parameters, X is the original data, is the denoised data.

[0073] (2) Interpolation method

[0074] For missing data points, polynomial interpolation can be used:

[0075] x t =a0+a1(t-t0)+a2(t-t0) 2 +…+a k (t-t0) k (8)

[0076] Among them, x t are the interpolated data points, a0, a1, a2, …, a k are the coefficients of the interpolation polynomial, and t0 is the reference time point.

[0077] (3) Normalization method

[0078] The present invention adopts the maximum-minimum normalization method to scale all data to the range of [0,1] to avoid the influence of dimension. The formula is:

[0079]

[0080] Among them, x′ norm is the normalized data, x is the original data, and x max and x minare the maximum and minimum values ​​in the data set, respectively.

[0081] Step 2: Heterogeneous temporal graph convolutional network construction

[0082] Heterogeneous Temporal Graph Convolutional Network (HTGCN) is an extension of graph neural network, which is specially designed to process graph structured data containing multiple types of nodes and edges, and can capture the changes of node states over time.

[0083] First, integrate the data and build a heterogeneous time series graph structure

[0084] The present invention integrates topological data, real-time data, historical data and fault data through interpolation, denoising and normalization to form a unified data set. Topological data describes the structure of the power grid, real-time data and historical data provide status information of the power grid operation, and fault data records fault events that have occurred in the past. According to the topological data, a graph structure of the power grid is constructed, in which nodes can represent equipment such as substations, transformers, sensors, etc., and edges represent electrical connections between devices. Features such as voltage, current, power, frequency, etc. are extracted from real-time data and historical data, and these features can be used as attributes of the nodes. Fault features are extracted from fault data, such as the time and location of the fault, the type of fault, etc. For the convenience of description, the present invention constructs the following heterogeneous time series graph set based on the graph structure G=(V,E) (V represents a finite non-empty set of nodes, and E represents a finite set of node pairs in V):

[0085]

[0086] Where T = {t i |i=1,2,...,n} represents time series; represents the graph set consisting of fault data on T, represents the atlas consisting of real-time data on T, Represents the atlas consisting of historical data on T; Indicates t i Fault diagram at the moment, Indicates t i Real-time graph of the moment, Indicates t i A historical graph of the moment.

[0087] Secondly, HTGCN is used to extract heterogeneous graph feature representation

[0088] This paper proposes an innovative heterogeneous graph temporal convolutional network (HTGCN), which aims to learn the dynamic representation of each node in a heterogeneous graph that evolves over time. The network captures the characteristics of the neighboring area of ​​each target node through convolution operations, and effectively integrates the feature information and spatial structure information from different types of nodes. In this framework, first, the adjacency matrix of the heterogeneous graph is defined as:

[0089] A HG ={A ij |i=1,2,...,m; j=1,2,...,m;} (11)

[0090] Among them, A ij ={a (ix,jy)} is a two-dimensional matrix, when the node v ix With v jy When there is an edge between (ix,jy) =1, otherwise a (ix,jy) =0; i and j represent different node types respectively, and m represents the number of node types.

[0091] Next, we can use the degree matrix D of the graph ij =∑ j A ij To obtain the degree matrix D of the heterogeneous graph HG and the Laplace matrix L HG as follows:

[0092] D HG =diag(D ij ), L HG =D HG -A HG (12)

[0093] Then, the heterogeneous graph convolutional network can be expressed by formula (13):

[0094]

[0095] in, W i represents the weight of the i-th convolution layer, and σ is the activation function. In this way, we can get the value of the dataset ξ at a certain time t i Heterogeneous Graph Representation

[0096] Finally, the heterogeneous graph feature representation of the dataset ξ at each moment can be obtained:

[0097]

[0098] Step 3: Feature fusion based on RCA

[0099] First, the aggregation of heterogeneous timing graph features is completed based on RCA. The specific process is as follows: Figure 2 shown.

[0100] Figure 2 middle represents the extended Hadamard product (i.e., Hadamard product), In actual operation, the present invention performs aggregation operations on the heterogeneous time series graph features on fault data, real-time data and historical data based on RCA, thereby obtaining the heterogeneous time series graph feature representation on all data sets.

[0101] Secondly, the heterogeneous time series graph features on each data set are fused

[0102] In order to better learn the spatiotemporal information characteristics of the distribution network feeder, the present invention combines the heterogeneous timing diagram characteristics Then it is sent to the graph convolutional network for training and learning, and finally the distribution network feeder characteristics are obtained

[0103]

[0104] Among them, GCN represents the graph convolution operation, Respectively represent the extracted features of fault, real-time, and historical data sources.

[0105] Step 4: Feeder Fault Detection

[0106] The LGNN (Label-inputted Graph Neural Networks) network is trained according to the distribution network feeder characteristics to obtain a feeder fault detection model; when the feeder fault detection model performs the distribution network feeder fault detection task, the loss function is a key indicator to measure the difference between the model fault detection and the actual label, and plays a vital role in the convergence and detection accuracy of the model. Therefore, the loss function is defined in the present invention:

[0107]

[0108] Where C represents the label set of all feeder faults, c i represents the knowledge probability of failure of the feeder to which the node belongs, represents the detection probability of the feeder fault to which the node belongs; Γ C is the set of permutations and combinations of C, is Γ C Any permutation of ; When calculating, it will be assigned an "infinite value" within a limited range. Assume is the node feature matrix The mapping function to the real feeder fault C is: is the node feature matrix Feeder fault detection model The final loss function formula is:

[0109]

[0110] Among them, x i Representative Matrix The feature vector of each node in .

[0111] Loss function f Loss The goal is to find the minimum cross entropy loss value among all possible category label arrangements to guide the optimization of feeder fault detection model parameters; use the trained feeder fault detection model to perform feeder status detection.

[0112] The HTGCN model of the present invention utilizes the powerful expression ability of graph neural networks, and by analyzing the time series data of feeder current and voltage, it can capture abnormal fluctuations in real time when a fault occurs. The model can not only process static distribution network structure data, but also capture dynamically changing time series characteristics, thereby more accurately identifying feeder faults. In addition, the automatic feature extraction mechanism of HTGCN reduces the reliance on expert knowledge, reduces the complexity of feature engineering, and improves the practicality and flexibility of the model. The HTGCN model of the present invention effectively integrates multi-source time series data in the distribution network by constructing a comprehensive graph model. The model not only considers the time series characteristics of the data, but also learns the relationship between different sensor data, thereby improving the ability to extract feeder fault features. By simulating the propagation process of feeder faults in the distribution network, HTGCN can identify the fault path and associate the feeder fault characteristics with the physical structure of the distribution network to achieve fast and accurate feeder fault detection. The HTGCN model of the present invention has adaptive learning capabilities and can learn the time series changes of the feeder topology of the distribution network. This capability enables the model to dynamically adjust parameters and maintain the accuracy and effectiveness of fault detection in the face of distribution network upgrades, equipment failures, or load fluctuations. This adaptability is not available in traditional fault detection methods, which greatly improves the robustness of the model. The HTGCN model of the present invention improves the generalization ability of the model by deeply learning the timing characteristics and topological structure of the distribution network data, so that it not only performs well on the training data, but also can adapt to new and unseen distribution network environments. In addition, the model of the present invention provides a clear fault detection logic, so that the fault detection results have high interpretability, and enhances the user's trust in the model detection results. In order to improve the generalization ability and adaptability of the model, the present invention uses a determinant point process (DPP) model to generate fault data, simulating various types of fault conditions that may occur in a real distribution network; this method not only improves the diversity of training data, but also increases the adaptability of the model to rare or unknown fault types; at the same time, the present invention also includes data preprocessing techniques such as denoising, interpolation and normalization, which ensures the quality and accuracy of the model input data and provides a basis for the high performance of the model.

[0113] Embodiment 2:

[0114] The embodiment of the present invention further provides a distribution network feeder fault detection device, comprising:

[0115] The acquisition module is used to obtain the topological data of the distribution network, the real-time data collected by smart meters and sensors, historical data and fault data;

[0116] The first processing module is used to obtain heterogeneous graph features of topological data, real-time data, historical data and fault data at each moment through a heterogeneous graph temporal convolutional network;

[0117] The second processing module is used to perform feature fusion on the heterogeneous time sequence diagram features to obtain the distribution network feeder features;

[0118] The detection module is used to train the LGNN network according to the distribution network feeder characteristics to obtain a feeder fault detection model; wherein the trained feeder fault detection model is used to perform feeder status detection.

[0119] As an implementation of the embodiment of the present invention, the second processing module includes:

[0120] The first processing unit is used to perform aggregation operations on the heterogeneous timing graph features on the fault data, real-time data and historical data based on RCA to obtain the heterogeneous timing graph feature representation

[0121] The second processing unit is used to convert the heterogeneous timing diagram features Send it to the graph convolutional network for training and learning to obtain the distribution network feeder characteristics Right now,

[0122]

[0123] Among them, GCN represents the graph convolution operation, Respectively represent the extracted features of fault, real-time, and historical data sources.

[0124] As an implementation method of the embodiment of the present invention, the loss function of the feeder fault detection model is:

[0125]

[0126] Among them, x i Representative Matrix The feature vector of each node in .

[0127] Embodiment 3:

[0128] An embodiment of the present invention further provides a distribution network feeder fault detection system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a distribution network feeder fault detection method when executed by the processor.

[0129] Embodiment 4:

[0130] An embodiment of the present invention further provides a storage medium, on which a computer program is stored, and the computer program executes the distribution network feeder fault detection method when running.

[0131] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for detecting a distribution network feeder fault, characterized in that: include: Step S1, obtaining topological data of the distribution network, real-time data, historical data and fault data collected by smart meters and sensors; Step S2: Obtain heterogeneous graph features of topological data, real-time data, historical data, and fault data at each moment through a heterogeneous graph temporal convolutional network; Step S3, performing feature fusion on the heterogeneous time series diagram features to obtain distribution network feeder features; Step S4: training the LGNN network according to the distribution network feeder characteristics to obtain a feeder fault detection model; wherein the trained feeder fault detection model is used to perform feeder status detection.

2. The method for detecting a distribution network feeder fault according to claim 1, wherein: Step S3 includes: Based on RCA, the heterogeneous timing graph features on fault data, real-time data and historical data are aggregated to obtain the heterogeneous timing graph feature representation. Heterogeneous timing diagram features Send it to the graph convolutional network for training and learning to obtain the distribution network feeder characteristics Right now, Among them, GCN represents the graph convolution operation, Respectively represent the extracted features of fault, real-time, and historical data sources.

3. The method for detecting a distribution network feeder fault according to claim 2, wherein: The loss function of the feeder fault detection model is: Among them, x i Representative Matrix The feature vector of each node in .

4. A distribution network feeder fault detection device, characterized in that: include: The acquisition module is used to obtain the topological data of the distribution network, the real-time data collected by smart meters and sensors, historical data and fault data; The first processing module is used to obtain heterogeneous graph features of topological data, real-time data, historical data and fault data at each moment through a heterogeneous graph temporal convolutional network; The second processing module is used to perform feature fusion on the heterogeneous time sequence diagram features to obtain the distribution network feeder features; The detection module is used to train the LGNN network according to the distribution network feeder characteristics to obtain a feeder fault detection model; wherein the trained feeder fault detection model is used to perform feeder status detection.

5. The distribution network feeder fault detection device according to claim 4, characterized in that: The second processing module includes: The first processing unit is used to perform aggregation operations on the heterogeneous timing graph features on the fault data, real-time data and historical data based on RCA to obtain the heterogeneous timing graph feature representation The second processing unit is used to convert the heterogeneous timing diagram features Send it to the graph convolutional network for training and learning to obtain the distribution network feeder characteristics Right now, Among them, GCN represents the graph convolution operation, Respectively represent the extracted features of fault, real-time, and historical data sources.

6. The distribution network feeder fault detection device according to claim 5, characterized in that: The loss function of the feeder fault detection model is: Among them, x i Representative Matrix The feature vector of each node in .

7. A distribution network feeder fault detection system, characterized in that: include: A memory and a processor, wherein the memory stores a computer program executed by the processor, and when the computer program is executed by the processor, the method for detecting feeder faults in a distribution network as claimed in any one of claims 1 to 3 is executed.

8. A storage medium, characterized in that: The storage medium stores a computer program, which executes the distribution network feeder fault detection method according to any one of claims 1 to 3 when running.

Images