Pilot protection method, system and equipment for active power distribution network and medium

Abstract

The invention belongs to the technical field of active power distribution networks, and discloses a pilot protection method, system and device for an active power distribution network and a medium. Acquiring three-phase current sampling data of a home terminal protection device after a preset duration as a first current sampling sequence; acquiring three-phase current sampling data of the opposite end protection device after a preset duration as a second current sampling sequence, and converting the first current sampling sequence and the second current sampling sequence into topological structure data; performing convolution operation on the converted topological structure data to obtain a node feature matrix, performing global pooling, and calculating graph convolution sequence similarity; and on the basis of internal and external fault identification criteria, an internal short-circuit fault, a normal operation fault or an external short-circuit fault are judged in combination with graph convolution sequence similarity. The technical problem that a traditional protection principle is prone to misjudgment under strong noise interference is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of active power distribution network technology, and in particular to an active power distribution network longitudinal protection method, system, equipment and medium. Background Technology

[0002] Driven by the digital transformation of power systems and the construction of smart grids, distribution networks are undergoing a profound transformation from traditional passive networks to active distribution networks. The large-scale integration of distributed photovoltaic, wind power, and other new energy devices has resulted in new characteristics in distribution networks, including bidirectional power flow and multi-point inflow of short-circuit current. This fundamental change in network topology and operation presents entirely new challenges to distribution line protection. Traditional protection schemes based on single electrical quantity amplitude or phase criteria face severe tests of adaptability and reliability under complex operating conditions.

[0003] To address the challenges of active distribution network protection, longitudinal protection is considered an effective solution due to its clearly defined protection range and ability to comprehensively utilize information from both ends of the line. Existing research includes several approaches: some scholars have proposed an ADN current differential protection strategy based on positive-sequence fault components, which identifies faults by constructing an amplitude ratio action criterion and a phase difference braking criterion for the currents at both ends of the protected line; others have designed deep learning-based longitudinal protection models that use neural networks to fit complex fault characteristics and construct fault discrimination criteria; still others have proposed a fault component extraction and longitudinal protection scheme based on wavelet transform, which separates fault feature components using wavelet transform and then compares and analyzes the data from both ends.

[0004] However, existing active distribution network longitudinal protection schemes generally suffer from insufficient noise immunity. First, while ADN current differential protection schemes based on positive-sequence fault components can adapt to various distributed power source access scenarios, the extraction accuracy of positive-sequence components is severely affected by strong noise scenarios such as electromagnetic interference and measurement errors in actual distribution network operation. This reduces the reliability of differential criteria and leads to the risk of protection misjudgment. Second, deep learning-based longitudinal protection models are extremely sensitive to noise in the input signal. If the training data does not cover strong noise scenarios, the model's generalization ability will significantly decrease, making it difficult to operate stably in the complex noise environment of actual distribution networks. Furthermore, existing deep learning methods mostly use fully connected or convolutional networks to process time-series data, failing to fully exploit the topological features and temporal correlation information of the current sequence. Third, in wavelet transform-based schemes, the selection of wavelet bases depends on engineering experience and has limited noise immunity. In strong noise scenarios, fault features are easily masked by noise signals, resulting in a significant reduction in feature extraction accuracy. Furthermore, existing fault feature extraction methods under noise interference lack the ability to deeply characterize the intrinsic structural relationships of current sequences. Traditional similarity calculation methods based on point-to-point comparisons or statistical features are insufficient to effectively address noise pollution in the complex electromagnetic environment of distribution networks, resulting in reduced feature differentiation between faults inside and outside the distribution area. In summary, existing longitudinal protection schemes experience a significant decrease in fault feature extraction accuracy and criterion calculation reliability under the strong noise conditions of actual distribution network operation.

[0005] Therefore, how to provide an active distribution network longitudinal protection scheme that can deeply mine the structured characteristics of current sequences and has stronger noise resistance has become an urgent problem to be solved. Summary of the Invention

[0006] This invention provides an active distribution network longitudinal protection method, system, device, and medium to address the problem that the accuracy of fault feature extraction and the reliability of criterion calculation are significantly reduced in the case of strong noise in actual operation of distribution networks.

[0007] According to a first aspect of the present invention, an active distribution network longitudinal protection method is provided.

[0008] In one embodiment, the active distribution network longitudinal protection method includes: calculating the phase current mutation in real time; when the phase current mutation exceeds a preset multiple of the load current, acquiring three-phase current sampling data after a preset time period at the installation location of the local protection device as a first current sampling sequence, and sending the first current sampling sequence to the peer protection device; acquiring three-phase current sampling data after a preset time period at the installation location of the peer protection device as a second current sampling sequence, and converting the first current sampling sequence and the second current sampling sequence into topology data; performing convolution operation on the converted topology data to obtain a node feature matrix, and performing global pooling on the node feature matrix to calculate the graph convolution sequence similarity; based on the fault identification criteria inside and outside the zone, if the graph convolution sequence similarity is less than a preset threshold, it is judged as an intra-zone short-circuit fault, and a trip command is sent; otherwise, it is judged as normal operation or an extra-zone short-circuit fault, and the protection device resets.

[0009] In one embodiment, calculating the phase current abrupt change includes: calculating the difference between the current value at the current sampling point and the current value at the sampling point corresponding to the previous cycle, and obtaining a first absolute value of the difference by taking the absolute value; calculating the difference between the current value at the sampling point corresponding to the previous cycle and the current values ​​at the sampling points corresponding to the previous two cycles, and obtaining a second absolute value of the difference by taking the absolute value; solving for the difference between the first absolute value of the difference and the second absolute value of the difference, and obtaining the phase current abrupt change by taking the absolute value.

[0010] In one embodiment, converting the data into topological structure data includes mapping the first current sampling sequence and the second current sampling sequence into undirected graphs, specifically including: constructing a node set by using the values ​​of the current sampling points as nodes in the graph based on the first current sampling sequence and the second current sampling sequence; constructing an edge set by determining the connection relationship between nodes through the nearest neighbor rule; establishing a connection between each node and its adjacent sampling points in the nearest neighbor rule; calculating the edge weight between each node in the edge set; obtaining the edge weight by normalizing the sampling point time difference; establishing an adjacency matrix based on the weight values ​​to represent the connection relationship of the graph; and constructing undirected graphs of the first current sampling sequence and the second current sampling sequence according to the node set, the edge set, and the adjacency matrix.

[0011] In one embodiment, the edge weight is obtained by normalizing the time difference between sampling points; calculating the absolute value of the time difference between two sampling points based on their timestamps; dividing the absolute value of the time difference by a time scale parameter and taking a negative value; and performing an exponential operation on the negative result to obtain the edge weight.

[0012] In one embodiment, after establishing the adjacency matrix based on the weight values, the adjacency matrix is ​​further normalized. Specifically, this includes: adding an identity matrix to the adjacency matrix to preserve the node's own characteristics by introducing self-loops; establishing a degree matrix based on the adjacency matrix after adding the identity matrix; and normalizing the adjacency matrix using the degree matrix to eliminate the influence of node degree differences on convolution.

[0013] In one embodiment, global pooling of the node feature matrices to calculate the similarity of graph convolutional sequences includes: obtaining a first node feature matrix and a second node feature matrix by performing graph convolution operations on the first current sampling sequence and the second current sampling sequence, respectively; performing global average pooling on the first node feature matrix and the second node feature matrix to obtain the corresponding global feature vectors of the sequences; and using cosine similarity as a similarity metric for the global features of the sequences to calculate the similarity of the graph convolutional sequences.

[0014] In one embodiment, performing global average pooling on the first node feature matrix and the second node feature matrix respectively includes: obtaining all row vectors of the node feature matrix based on the first node feature matrix and the second node feature matrix respectively; summing the row vectors of each row of the node feature matrix; and dividing the summation result by the total number of rows of the node feature matrix to obtain the global feature vector of the sequence.

[0015] In one embodiment, cosine similarity is used as a similarity metric for global features of a sequence. Calculating the similarity of a graph convolution sequence includes: calculating the dot product of the global feature vector of the first node feature matrix and the global feature vector of the second node feature matrix; calculating the L2 norm of the global feature vector of the first node feature matrix and the global feature vector of the second node feature matrix respectively; and dividing the dot product by the product of the two L2 norms to obtain the graph convolution sequence similarity.

[0016] According to a second aspect of the present invention, an active power distribution network longitudinal protection system is provided.

[0017] In one embodiment, the active distribution network longitudinal protection system includes: a start-up acquisition module, used to calculate the phase current mutation in real time; when the phase current mutation exceeds a preset multiple of the load current, it acquires three-phase current sampling data after a preset time period at the installation location of the local protection device as a first current sampling sequence and sends the first current sampling sequence to the peer protection device; a data conversion module, used to acquire three-phase current sampling data after a preset time period at the installation location of the peer protection device as a second current sampling sequence and convert the first current sampling sequence and the second current sampling sequence into topology data; a feature calculation module, used to perform convolution operation on the converted topology data to obtain a node feature matrix, and perform global pooling on the node feature matrix to calculate the graph convolution sequence similarity; and a fault discrimination module, used to determine whether the graph convolution sequence similarity is less than a preset threshold based on the fault identification criteria inside and outside the zone, and send a trip command if the similarity is less than a preset threshold; otherwise, it determines whether the fault is a normal operation or an external short-circuit fault, and the protection device resets.

[0018] According to a third aspect of the present invention, a computer device is provided.

[0019] In one embodiment, the computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described above.

[0020] According to a fourth aspect of the present invention, a computer-readable storage medium is provided.

[0021] In one embodiment, a computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the steps of the method described above.

[0022] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0023] (1) The active distribution network longitudinal protection method provided by the present invention transforms the current sampling sequence into graph structure data and uses the neighborhood aggregation mechanism of graph convolutional network to deeply mine the spatiotemporal correlation features between sampling points. This method breaks through the limitations of traditional similarity calculation methods based on point-to-point comparison or statistical features in feature representation capabilities and can fully capture the inherent structural relationship and temporal dependency pattern of the current sequence. The method maps the current sequence into a high-dimensional feature vector through multi-layer feature extraction and global pooling operation of graph convolutional layer, and then uses cosine similarity to quantify the similarity of the current sequences at both ends of the line. This achieves a deep-level representation of complex fault features, significantly improves the feature differentiation of faults inside and outside the area, and solves the problems of insufficient fault feature extraction depth and lack of ability to mine the topological structure features of current sequence in the existing longitudinal protection scheme.

[0024] (2) This invention uses graph convolutional networks to perform deep learning and robust characterization of the structured features of the current sequence, enabling the method to maintain high-precision feature extraction capabilities even in the strong noise scenarios of actual power distribution network operation. Graph convolutional networks can effectively suppress noise interference on individual sampling points through neighborhood information aggregation and nonlinear transformation. Experimental verification shows that even under the harsh conditions of 10dB noise superimposed on the current waveform, the protection method proposed in this invention can still correctly identify faults inside and outside the zone, maintaining high sensitivity and reliability. Compared with the shortcomings of traditional wavelet transform methods, such as wavelet basis selection relying on experience and limited noise resistance, and the problem of insufficient generalization ability under strong noise due to the failure of traditional deep learning methods to fully utilize the topological information of the current sequence, this invention significantly enhances the robustness and anti-interference ability of the protection scheme in complex electromagnetic environments through the collaborative modeling of the spatiotemporal features of the current sequence using graph convolutional networks.

[0025] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0026] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0027] Figure 1 This is a detailed implementation diagram illustrating an active distribution network longitudinal protection method according to an exemplary embodiment;

[0028] Figure 2 This is a waveform diagram of the sampling sequence of phase a current on both sides of M and N when a phase-to-phase short-circuit fault occurs between phases A and B in the metallic zone at point f2 in an active distribution network simulation model of an active distribution network longitudinal protection method according to an exemplary embodiment.

[0029] Figure 3 This is a schematic diagram of the normalized adjacency matrix of the current on the M side when a phase-to-phase short-circuit fault occurs between phases A and B in the metallic zone at point f2 in an active distribution network simulation model of an active distribution network longitudinal protection method according to an exemplary embodiment.

[0030] Figure 4 This is a schematic diagram of the normalized adjacency matrix of the N-side current when a phase-to-phase short-circuit fault of phases A and B occurs at point f2 in the metallic zone of an active distribution network longitudinal protection method according to an exemplary embodiment.

[0031] Figure 5This is a schematic diagram of the characteristic norms and similarity values ​​of the a-phase currents on both sides of M and N when a phase-to-phase short-circuit fault occurs at point f2 in the metallic zone of an active distribution network longitudinal protection method according to an exemplary embodiment.

[0032] Figure 6 This is a waveform diagram of the sampling sequence of phase a current on both sides of M and N when a two-phase ground fault occurs outside the metallic zone at point f1 in an active distribution network simulation model of an active distribution network longitudinal protection method according to an exemplary embodiment.

[0033] Figure 7 This is a schematic diagram of the normalized adjacency matrix of the current on the M side when a two-phase ground fault (AB phase) occurs outside the metallic zone at point f1 in an active distribution network simulation model of an active distribution network longitudinal protection method according to an exemplary embodiment.

[0034] Figure 8 This is a schematic diagram of the normalized adjacency matrix of the N-side current when a two-phase ground fault occurs outside the metallic zone (AB phase) at point f1 in an active distribution network simulation model of an active distribution network longitudinal protection method according to an exemplary embodiment.

[0035] Figure 9 This is a schematic diagram of the characteristic norms and similarity values ​​of the a-phase currents on both sides of M and N when a two-phase ground fault occurs outside the metallic zone at point f1 in an active distribution network simulation model of an active distribution network longitudinal protection method according to an exemplary embodiment.

[0036] Figure 10 This is an original waveform diagram of the sampling sequence of phase a current on both sides of M and N when a two-phase ground fault occurs outside the metallic zone at point f1 in an active distribution network simulation model of an active distribution network longitudinal protection method according to an exemplary embodiment.

[0037] Figure 11 This is a waveform diagram of the current sampling of phase a on both sides of M and N after superimposing 30dB Gaussian white noise when a two-phase ground fault occurs outside the metallic zone at point f1 in an active distribution network simulation model of an active distribution network longitudinal protection method according to an exemplary embodiment.

[0038] Figure 12 This is a waveform diagram of the current sampling of phase a on both sides of M and N after superimposing 10dB Gaussian white noise when a two-phase ground fault occurs outside the metallic zone at point f1 in an active distribution network simulation model of an active distribution network longitudinal protection method according to an exemplary embodiment.

[0039] Figure 13 This is a simulation model diagram of an active distribution network according to an exemplary embodiment of a longitudinal protection method for an active distribution network.

[0040] Figure 14 This is a flowchart illustrating an active distribution network longitudinal protection method according to an exemplary embodiment;

[0041] Figure 15 This is a structural block diagram of an active distribution network longitudinal protection system according to an exemplary embodiment;

[0042] Figure 16 This is a schematic diagram of the structure of a computer device according to an exemplary embodiment. Detailed Implementation

[0043] The following description and accompanying drawings fully illustrate specific embodiments described herein to enable those skilled in the art to practice them. Some portions and features of certain embodiments may be included in or replace portions and features of other embodiments. The scope of the embodiments herein includes the entire scope of the claims and all available equivalents thereof. The various embodiments described herein are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments; similar or identical parts between embodiments can be referred to interchangeably.

[0044] The modules in the apparatus or system of this application can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0045] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0046] Figure 1 and Figure 14 An embodiment of an active distribution network longitudinal protection method of the present invention is shown.

[0047] In this optional embodiment, the active distribution network longitudinal protection method includes: real-time calculation of phase current mutation; when the phase current mutation exceeds a preset multiple of the load current, acquiring three-phase current sampling data after a preset time period at the installation location of the local protection device as a first current sampling sequence, and sending the first current sampling sequence to the peer protection device; acquiring three-phase current sampling data after a preset time period at the installation location of the peer protection device as a second current sampling sequence, and converting the first current sampling sequence and the second current sampling sequence into topology data; performing convolution operation on the converted topology data to obtain a node feature matrix, and performing global pooling on the node feature matrix to calculate the graph convolution sequence similarity; based on the fault identification criteria inside and outside the zone, if the graph convolution sequence similarity is less than a preset threshold, it is judged as an intra-zone short-circuit fault, and a trip command is sent; otherwise, it is judged as normal operation or an extra-zone short-circuit fault, and the protection device resets.

[0048] In this optional embodiment, calculating the phase current abrupt change includes: calculating the difference between the current value at the current sampling point and the current value at the sampling point corresponding to the previous cycle, and obtaining the first absolute value of the difference by taking the absolute value; calculating the difference between the current value at the sampling point corresponding to the previous cycle and the current value at the sampling point corresponding to the previous two cycles, and obtaining the second absolute value of the difference by taking the absolute value; solving for the difference between the first absolute value of the difference and the second absolute value of the difference, and obtaining the phase current abrupt change by taking the absolute value.

[0049] In this optional embodiment, the conversion into topological data includes mapping the first current sampling sequence and the second current sampling sequence into undirected graphs, specifically including: constructing a node set by using the values ​​of the current sampling points as nodes in the graph based on the first current sampling sequence and the second current sampling sequence; constructing an edge set by determining the connection relationship between nodes through the nearest neighbor rule; establishing a connection between each node and its adjacent sampling points in the nearest neighbor rule; calculating the edge weight between each node in the edge set; obtaining the edge weight by normalizing the sampling point time difference; establishing an adjacency matrix based on the weight values ​​to represent the connection relationship of the graph; and constructing undirected graphs of the first current sampling sequence and the second current sampling sequence according to the node set, the edge set, and the adjacency matrix.

[0050] In this optional embodiment, the edge weight is obtained by normalizing the time difference between sampling points, including: calculating the absolute value of the time difference between two sampling points based on their timestamps; dividing the absolute value of the time difference by a time scale parameter and taking the negative value; and performing an exponential operation on the negative value to obtain the edge weight.

[0051] The expression for edge weight is:

[0052] w ij =exp(-|t i -t j | / τ);

[0053] In the formula, w ij τ is the edge weight between the i-th sampling point and the j-th sampling point; τ is the time scale parameter; t i t is the timestamp corresponding to the i-th sampling point; j This is the timestamp corresponding to the j-th sampling point.

[0054] In this optional embodiment, after establishing the adjacency matrix based on the weight values, the adjacency matrix is ​​further normalized. Specifically, this includes: adding an identity matrix to the adjacency matrix to preserve the node's own characteristics by introducing self-loops; establishing a degree matrix based on the adjacency matrix after adding the identity matrix; and normalizing the adjacency matrix using the degree matrix to eliminate the influence of node degree differences on convolution.

[0055] In this optional embodiment, the degree matrix is ​​expressed as:

[0056]

[0057] The expression for normalizing the adjacency matrix is:

[0058]

[0059] In the formula, D is the normalized adjacency matrix; D is the degree matrix and I n D is the identity matrix; ii Let A be the i-th diagonal element of the degree matrix D; j be the node index; n be the total number of sampling points; ij I is the element in the i-th row and j-th column of the original adjacency matrix A; n,ij I is an n-order identity matrix n The element in the i-th row and j-th column.

[0060] In this optional embodiment, performing convolution operations on the transformed topological data to obtain the node feature matrix includes: using a spatial domain graph convolution framework based on graph convolutional networks, extracting structured features of the sequence through the local interaction between the convolution kernel and the graph structure to obtain the node feature matrix;

[0061] The expression for the feature matrix of this node is:

[0062]

[0063] In the formula, Z is the node feature matrix; σ is the normalized adjacency matrix; σ(·) is the activation function used to enhance the nonlinear representation of features; X is the initial node feature matrix after the current sampling sequence is converted into a graph structure; W is the convolution kernel weight matrix and F' is the output feature dimension; b is the bias vector.

[0064] In this optional embodiment, global pooling of the node feature matrix to calculate the graph convolution sequence similarity includes: obtaining a first node feature matrix and a second node feature matrix by performing graph convolution operations on the first current sampling sequence and the second current sampling sequence, respectively; performing global average pooling on the first node feature matrix and the second node feature matrix to obtain the corresponding global feature vector of the sequence; and using cosine similarity as a similarity metric for the global features of the sequence to calculate the graph convolution sequence similarity.

[0065] In this optional embodiment, performing global average pooling on the first node feature matrix and the second node feature matrix respectively includes: obtaining all row vectors of the node feature matrix based on the first node feature matrix and the second node feature matrix respectively; summing the row vectors of each row of the node feature matrix; and dividing the summation result by the total number of rows of the node feature matrix to obtain the global feature vector of the sequence.

[0066] The expression for the global feature vector of a sequence is:

[0067]

[0068] In the formula, f m z is the global feature vector of the sequence, and m = 1 or 2; m,k is the k-th row vector of the node feature matrix; n is the total number of rows in the node feature matrix.

[0069] In this optional embodiment, cosine similarity is used as the similarity metric for global features of the sequence. Calculating the similarity of graph convolutional sequences includes: calculating the dot product of the global feature vector of the first node feature matrix and the global feature vector of the second node feature matrix; calculating the L2 norm of the global feature vector of the first node feature matrix and the global feature vector of the second node feature matrix respectively; and dividing the dot product by the product of the two L2 norms to obtain the graph convolutional sequence similarity.

[0070] The expression for the similarity of graph convolutional sequences is:

[0071]

[0072] In the formula, S is the graph convolution sequence similarity; f1 is the sequence global feature vector of the first node feature matrix; and f2 is the sequence global feature vector of the second node feature matrix.

[0073] In this optional embodiment, the value range of the graph convolution sequence similarity is [-1, 1]; the fault identification criteria inside and outside the region include judging the dynamic characteristics of the first current sampling sequence and the second current sampling sequence based on the graph convolution sequence similarity, specifically including: when the value of the graph convolution sequence similarity is greater than or equal to a preset threshold, it is determined that the system is in normal operation or an external short circuit fault has occurred; when the value of the graph convolution sequence similarity is less than the preset threshold, it is determined that an internal short circuit fault has occurred in the system.

[0074] It should be noted that this invention discloses an active distribution network longitudinal protection method. First, it acquires the three-phase current sampling sequences at both ends of the protected distribution line. Then, it converts the current sampling sequences into topological data. Next, it utilizes the neighborhood aggregation characteristic of graph convolutional networks to capture the spatiotemporal correlation between sampling points. Global pooling and cosine similarity are used to quantify the similarity of the current sampling sequences. When two current sampling sequences are similar, the similarity of the graph convolutional sequences approaches 1; otherwise, the similarity approaches -1. Finally, the similarity of the graph convolutional sequences is compared with a preset threshold. When the similarity is less than the preset threshold, it is determined to be an intra-zone short circuit; otherwise, it is normal operation or an inter-zone short circuit. Compared with traditional protection methods, this method has higher sensitivity, reliability, and noise immunity in distribution lines. Even when the current waveform is superimposed with 10dB of noise, the proposed protection method can still operate correctly. The method specifically includes the following steps:

[0075] Step 1: Calculate the phase current surge in real time. Determine whether the surge in phase current meets the requirements for surge start-up. When the surge in any phase current exceeds 0.1 times the load current, the surge start-up is considered to be met. The expression is as follows:

[0076] ||i(n)-i(nN)|-|i(nN)-i(n-2N)||>0.1I L ;

[0077] In the formula, i(n) represents the value of the nth sampling point of a certain phase current; N is the total number of sampling points in one cycle; I L This represents the amplitude of the load current during normal operation.

[0078] Step 2: If the sudden change in the start-up condition is met, the three-phase current sampling data 20ms after the fault at the protection installation location is obtained, and the three-phase current sampling data 20ms after the fault at this end is sent to the protection device at the other end.

[0079] Step 3: After acquiring the three-phase current sampling data 20ms after the fault of the protection device at the other end, convert the current sampling sequence into topology data.

[0080] Specifically, assume two current sampling sequences I1={i 1,1 i 1,2 ,...,i1,n} and I2={i 2,1 i 2,2 ,...,i 2,n}, where n is the total number of current samples. When the sampling frequency is 10kHz, the total number of current samples within 20ms after the fault is n = 200. Each sampling sequence is mapped to an undirected graph G = (V, ε, A). Taking sequence I1 as an example, sequence I1 is transformed into an undirected graph G = (V, ε, A), where V is the node set, V = {v1, v2, ..., v...} n Here, the sampling point current value v is directly taken. i =x i ε is the edge set used to characterize the temporal neighborhood associations between nodes. It adopts the k-nearest neighbor rule (k=2, meaning each node connects to its immediate and next adjacent sampling points). The edge weights are obtained by normalizing the time difference between sampling points, and the expression is:

[0081] w ij =exp(-|t i -t j | / τ);

[0082] In the formula, w ij τ is the edge weight between the i-th and j-th sampling points; τ is the time scale parameter, τ = 100 μs when the sampling frequency is 10 kHz; t i t is the timestamp corresponding to the i-th sampling point, i.e., the acquisition time of that sampling point; j This is the timestamp corresponding to the j-th sampling point, i.e., the sampling time of that sampling point.

[0083] Specifically, the adjacency matrix A represents the connectivity relationships in the graph, satisfying A ij =w ij (if (v i ,v j )∈ε), otherwise A ij =0 and A ii =0.

[0084] Specifically, to eliminate the impact of node degree differences on convolution, the adjacency matrix is ​​normalized, as expressed by:

[0085]

[0086] In the formula, The normalized adjacency matrix; It is an identity matrix (introducing self-loops to preserve the characteristics of the nodes themselves). Let be a degree matrix, satisfying:

[0087]

[0088] In the formula, D iiis the i-th diagonal element of the degree matrix D; j is the node index, corresponding to node i, used to traverse all nodes in the graph structure, representing the "column number in the adjacency matrix / degree matrix"; n is the total number of sampling points; A ij The element in the i-th row and j-th column of the original adjacency matrix A is used to represent the "edge weight between node i and node j" in the graph structure; n,ij I is an n-order identity matrix n The element in the i-th row and j-th column is the identifier element used to "introduce a self-loop to the node".

[0089] Step 4: Next, leverage the neighborhood aggregation properties of graph convolutional networks to capture the spatiotemporal correlations between sampling points. Quantify the similarity of current sampling sequences using global pooling and cosine similarity. First, a spatial domain graph convolution framework is employed, extracting structured features of the sequence through the local interaction between the convolution kernel and the graph structure. The expression for the graph convolution operation is:

[0090]

[0091] In the formula, The node feature matrix output by graph convolution is used to characterize the local-global fusion features of each sampling point in the graph structure; σ is the normalized adjacency matrix; σ(·) is the activation function used to enhance the nonlinear expressive power of features. In this embodiment, the ReLU function σ(x) = max(0,x) is selected; X is the initial node feature matrix after the current sampling sequence is converted into a graph structure. Its core function is to provide the initial node feature input for the graph convolutional network, which is used to extract structured features through neighborhood aggregation in the subsequent process; W is the convolutional kernel weight matrix and F' is the output feature dimension (F' = 16 is used for feature dimensionality increase);

[0092] This is the bias vector.

[0093] Specifically, for the current sampling sequences I1 and I2, feature matrices Z1 and Z2 are obtained through the above convolution operation, respectively. To obtain sequence-level global features, global average pooling is performed on the node feature matrices, expressed as:

[0094]

[0095] In the formula, For sequence I m The global feature vector of the sequence (m = 1, 2); Z is the node feature matrix m The k-th row vector; n is the total number of rows in the node feature matrix.

[0096] Specifically, finally, cosine similarity is used as a similarity metric for global features of the sequence, and its expression is:

[0097]

[0098] In the formula, S is the graph convolution sequence similarity; f1 is the sequence global feature vector of the first node feature matrix; f2 is the sequence global feature vector of the second node feature matrix; · represents the vector dot product; ||·||2 represents the L2 norm; the graph convolution sequence similarity S takes the value range of [-1,1]. The closer the value is to 1, the more similar the graph convolution features of the two sequences are, that is, the more consistent the dynamic characteristics of the two current sequences are. The closer the value is to -1, the greater the difference in graph convolution features between the two sequences.

[0099] Step 5: Fault identification inside and outside the zone:

[0100] During normal operation or external fault conditions, the dynamic characteristics of the two current sequences are nearly identical, with a similarity of 0. <S<1;

[0101] When an intra-regional fault occurs, the similarity between the two current sequences is -1. <S<0。

[0102] If the similarity of the graph convolution sequence is greater than or equal to the preset threshold of 0, it is determined that the operation is normal or there is an external fault, and the protection device is reset.

[0103] If the similarity of the graph convolution sequence is less than the preset threshold of 0, a short circuit fault is detected in the affected area, and a trip command is sent.

[0104] To better illustrate this scheme, an electromagnetic transient simulation system was built using the software PSCAD / EMTDC, as shown below. Figure 13 The simulation model of the medium-voltage active distribution network shown is used to verify the longitudinal protection method of the active distribution network proposed in this embodiment:

[0105] 1) Establish a model

[0106] A simulation model of a medium-voltage active power distribution network built based on PSCAD / EMTDC is shown below. Figure 13 As shown in the figure, the system's rated frequency is 50Hz, and the sampling frequency is 10kHz; the rated capacity of IIDG1, IIDG2, and IIDG3 is 2MW, and all IIDGs adopt PQ, low voltage ride-through, and current limiting control strategies; the apparent power of loads 1-5 are (1+j0.54)MVA, (3+j1.32)MVA, (2+j2.19)MVA, (3+j1.32)MVA, and (2+j2.19)MVA, respectively, and the positive sequence impedance of the line is (0.17+j0.32)Ω / km. The line length is marked in the figure; the preset threshold of the protection is 0. When the similarity of the graph convolution sequence is less than 0, it is judged that an intra-zone short-circuit fault has occurred; otherwise, it is normal operation or an inter-zone short-circuit fault.

[0107] 2) Simulation Analysis

[0108] a) Fault simulation in typical areas

[0109] To verify the effectiveness of the active distribution network longitudinal protection method proposed in this invention, in Figure 13 A phase-to-phase short-circuit fault is set at point f2 within the metallic zone, between phases A and B. Figure 2 The sampled waveforms of the a-phase currents on both sides of the fault are presented. At the moment of the fault, the a-phase currents on both sides show obvious abrupt changes, and the waveform change trends are significantly different. Figure 3 and Figure 4 These are the normalized adjacency matrices of the currents on the M and N sides, respectively. Figure 5 The characteristic norms and similarity values ​​of the a-phase currents on both sides M and N are presented. From the characteristic norms, there is a significant difference in magnitude between the two sides: the characteristic norm on the M side is 0.9011, while the characteristic norm on the N side is 0.3173. From the similarity perspective, the graph convolution sequence similarity values ​​of the a-phase currents on both sides M and N are close to -1, strictly satisfying the action criterion of less than 0, therefore correctly identifying it as an intra-zone short-circuit fault.

[0110] b) Simulation of typical out-of-area faults

[0111] To verify the effectiveness of the active distribution network longitudinal protection method proposed in this invention, in Figure 13 A two-phase ground fault is set at point f1 outside the metallic zone. Figure 6 The sampled waveforms of the a-phase currents on both sides of the fault are presented. At the moment of the fault, the a-phase currents on both sides show obvious abrupt changes, and the waveform changes have the same trend. Figure 7 and Figure 8 These are the normalized adjacency matrices of the currents on the M and N sides, respectively. Figure 9 The characteristic norms and similarity values ​​of the a-phase currents on both sides M and N are presented. Looking at the characteristic norms, the values ​​of the characteristic norms on both sides are close: the characteristic norm on the M side is 0.2758, while the characteristic norm on the N side is 0.2759. In terms of similarity, the graph convolution sequence similarity values ​​of the a-phase currents on both sides M and N are close to 1, strictly satisfying the action criterion of greater than 0. Therefore, it is correctly identified as an external short-circuit fault.

[0112] c) Simulation of different fault locations and fault types

[0113] To further verify the effectiveness of the active distribution network longitudinal protection method proposed in this invention, in Figure 13 Different types of short-circuit faults were set at points f1, f2, and f3. The simulation results are shown in Table 1.

[0114] Table 1 - Simulation results for different fault locations and fault types

[0115]

[0116] As can be seen from the data in the table, at fault locations f1 and f3, the similarity of the graph convolution sequence of phases a, b, and c is 1, indicating that the current characteristics on both sides are highly consistent, consistent with the characteristics of external faults or normal operation. However, at fault location f2, the similarity of the graph convolution sequence of the faulty phases (phases a, b, and c under ABC faults, and phases a and b under AB and ABG faults) is -1, while the similarity of the healthy phases (such as phase c under AB faults) is 1. This is completely consistent with the protection criterion that "simulation results approaching -1 represent faulty phases, and any phase approaching -1 indicates an internal short circuit; simulation results approaching 1 represent healthy phases, and all three phases approaching 1 indicate normal operation or external faults." This fully verifies that the proposed method can accurately distinguish between internal short circuit faults and external faults / normal operation states, and has good adaptability to different fault locations and fault types.

[0117] d) Noise resistance

[0118] by Figure 13 Taking the current waveforms of phase a on both sides when an external metallic two-phase ground fault occurs at point f1 as an example, the original waveform is as follows: Figure 10 As shown. Gaussian white noise with a signal-to-noise ratio of 30dB is superimposed on the original current sampling waveform. The waveform of the current sampling sequence of phase a at both ends is then as follows. Figure 11 As shown, the current waveform has been distorted. Furthermore, Gaussian white noise with a signal-to-noise ratio of 10dB is superimposed on the original current sampling waveform. The waveform of the current sampling sequence at both ends of phase a is then as follows: Figure 12 As shown, the current waveform has been severely distorted, and most traditional protection principles will be unable to correctly identify faults inside and outside the protection zone under high-intensity noise interference.

[0119] To verify the noise immunity of the active distribution network longitudinal protection method based on graph convolution sequence similarity proposed in this embodiment, in Figure 13 Different types of short-circuit faults were set at points f1, f2, and f3. The simulation results with 30dB Gaussian white noise superimposed are shown in Table 2, and the simulation results with 10dB Gaussian white noise superimposed are shown in Table 3.

[0120] Table 2 - Simulation results with 30dB Gaussian white noise superimposed

[0121]

[0122] Table 3 - Simulation results with 10dB Gaussian white noise superimposed

[0123]

[0124] As shown in Table 2, after superimposing 30dB Gaussian white noise, the similarity of the graph convolution sequence of the fault phases (phase a, phase b or phases a, b, and c) at the fault location f2 (three-phase short circuit of ABC, two-phase short circuit of AB, and two-phase ground short circuit of ABG) remains stable at -1, which is completely consistent with the fault characteristics in the noise-free scenario. At the fault locations f1 and f3 outside the area, the similarity of the graph convolution sequence of each phase is close to 1 (e.g., the similarity of phase c under AB fault at point f1 is 0.9999, and the similarity of each phase under each fault type at point f3 is 1 or close to 1), indicating that even with strong noise interference, this method can still accurately distinguish between faults inside and outside the area.

[0125] Further analysis of Table 3 (with 10dB Gaussian white noise superimposed, resulting in higher noise interference intensity) shows that the similarity of the convolutional sequence of the fault phase diagram at fault location f2 within the region remains stable at -1 or close to -1 (e.g., the similarity of phase b under ABC fault is -0.9999, and the similarity of phase a under AB fault is -1); at fault locations f1 and f3 outside the region, the similarity of each phase is still close to 1 (e.g., the similarity of each phase under AB fault at point f1 is 1 or 0.9999, and the similarity of each phase under each fault type at point f3 is 1).

[0126] In summary, regardless of whether 30dB or 10dB Gaussian white noise is superimposed, the active distribution network longitudinal protection method proposed in this invention can accurately identify faults within the zone (fault phase similarity close to -1) and faults outside the zone / normal operation state (healthy phase similarity close to 1), demonstrating excellent noise interference resistance. Even in scenarios where strong noise causes severe distortion of the current waveform (such as... Figure 11 , Figure 12 Despite the waveform distortion shown, this method still maintains reliable fault differentiation performance, effectively solving the technical problem that traditional protection principles are prone to misjudgment under strong noise interference.

[0127] Figure 15 An embodiment of an active distribution network longitudinal protection system according to the present invention is shown.

[0128] In this optional embodiment, the active distribution network longitudinal protection system includes: a start-up acquisition module 201, used to calculate the phase current mutation in real time; when the phase current mutation exceeds a preset multiple of the load current, it acquires three-phase current sampling data after a preset time period at the installation location of the local protection device as a first current sampling sequence and sends the first current sampling sequence to the peer protection device; a data conversion module 202, used to acquire three-phase current sampling data after a preset time period at the installation location of the peer protection device as a second current sampling sequence and convert the first current sampling sequence and the second current sampling sequence into topology data; a feature calculation module 203, used to perform convolution operation on the converted topology data to obtain a node feature matrix, and perform global pooling on the node feature matrix to calculate the graph convolution sequence similarity; and a fault discrimination module 204, used to determine whether the graph convolution sequence similarity is less than a preset threshold based on the fault identification criteria inside and outside the zone, and send a trip command if the similarity is less than a preset threshold; otherwise, it determines whether the fault is a normal operation or an external short-circuit fault, and the protection device resets.

[0129] In this optional embodiment, when the data conversion module 202 converts the data into topological data, it includes mapping the first current sampling sequence and the second current sampling sequence into undirected graphs, respectively. Specifically, it includes: constructing a node set by using the values ​​of the current sampling points as nodes in the graph based on the first current sampling sequence and the second current sampling sequence; constructing an edge set by determining the connection relationship between nodes through the nearest neighbor rule; establishing a connection between each node in the nearest neighbor rule and its adjacent sampling points; calculating the edge weight between each node in the edge set; obtaining the edge weight by normalizing the sampling point time difference; establishing an adjacency matrix based on the weight values ​​to represent the connection relationship of the graph; and constructing undirected graphs of the first current sampling sequence and the second current sampling sequence according to the node set, the edge set, and the adjacency matrix.

[0130] In this optional embodiment, when the feature calculation module 203 performs global pooling on the node feature matrix to calculate the similarity of the graph convolution sequence, it includes: obtaining the first node feature matrix and the second node feature matrix by performing graph convolution operation on the first current sampling sequence and the second current sampling sequence, respectively; performing global average pooling on the first node feature matrix and the second node feature matrix to obtain the corresponding global feature vector of the sequence; and using cosine similarity as the similarity metric of the global features of the sequence to calculate the similarity of the graph convolution sequence.

[0131] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 16As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores static and dynamic information data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0132] Those skilled in the art will understand that Figure 16 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0133] In addition, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0134] In addition, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0135] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0136] This invention is not limited to the structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this invention is limited only by the appended claims.

Claims

1. A method for longitudinal protection of an active distribution network, characterized in that, The active distribution network longitudinal protection method includes: The phase current mutation is calculated in real time. When the phase current mutation exceeds the load current of a preset multiple, the three-phase current sampling data after a preset time at the installation location of the protection device at this end is obtained as the first current sampling sequence and sent to the protection device at the other end. The three-phase current sampling data after a preset time at the installation location of the protection device at the other end is obtained as the second current sampling sequence, and the first current sampling sequence and the second current sampling sequence are converted into topology data. The transformed topological data is subjected to convolution operation to obtain the node feature matrix, and the node feature matrix is ​​then globally pooled to calculate the similarity of the graph convolution sequence. Based on the fault identification criteria inside and outside the zone, if the similarity of the graph convolution sequence is less than the preset threshold, it is judged as a short circuit fault inside the zone and a trip command is sent; otherwise, it is judged as normal operation or a short circuit fault outside the zone, and the protection device resets.

2. The active distribution network longitudinal protection method according to claim 1, characterized in that, The calculation of phase current abrupt changes includes: Calculate the difference between the current value at the current sampling point and the current value at the corresponding sampling point in the previous cycle, and obtain the absolute value of the first difference by taking the absolute value; Calculate the difference between the current value at the sampling point corresponding to the previous cycle and the current value at the sampling point corresponding to the previous two cycles, and obtain the absolute value of the second difference by taking the absolute value; Solve for the difference between the absolute values ​​of the first and second differences, and obtain the phase current abrupt change by taking the absolute value.

3. The active distribution network longitudinal protection method according to claim 1, characterized in that, The conversion into topological data includes mapping the first current sampling sequence and the second current sampling sequence into undirected graphs, specifically including: Based on the first current sampling sequence and the second current sampling sequence, the values ​​of the current sampling points are used as nodes in the graph to construct a node set; The connection relationships between nodes are determined by the nearest neighbor rule, and an edge set is constructed; in the nearest neighbor rule, each node establishes a connection with its preceding and following adjacent sampling points. Calculate the edge weights between nodes in the edge set; these edge weights are obtained by normalizing the time difference of the sampling points. An adjacency matrix is ​​built based on weight values ​​to represent the connectivity relationships in the graph; Based on the node set, edge set, and adjacency matrix, construct undirected graphs for the first current sampling sequence and the second current sampling sequence, respectively.

4. The active distribution network longitudinal protection method according to claim 3, characterized in that, The edge weights are obtained by normalizing the time difference of the sampling points, including: Calculate the absolute value of the time difference between two sampling points based on their timestamps. Divide the absolute value of the time difference by the time scale parameter and take the negative value; The edge weights are obtained by performing an exponential operation on the negative result. The expression for the edge weight is: w ij =exp(-|t i -t j | / τ); In the formula, w ij τ represents the edge weight between the i-th sampling point and the j-th sampling point; τ is the time scale parameter. t i This is the timestamp corresponding to the i-th sampling point; t j This is the timestamp corresponding to the j-th sampling point.

5. The active distribution network longitudinal protection method according to claim 3, characterized in that, After establishing the adjacency matrix based on the weight values, the method further includes normalizing the adjacency matrix, specifically including: Add an identity matrix to the adjacency matrix to preserve the node's own characteristics by introducing self-loops; Based on the adjacency matrix after incorporating the identity matrix, construct the degree matrix; The adjacency matrix is ​​normalized using the degree matrix to eliminate the influence of node degree differences on convolution.

6. The active distribution network longitudinal protection method according to claim 5, characterized in that, The expression for the degree matrix is: The expression for normalizing the adjacency matrix is ​​as follows: In the formula, D is the normalized adjacency matrix; D is the degree matrix and I n D is the identity matrix; ii Let A be the i-th diagonal element of the degree matrix D; j be the node index; n be the total number of sampling points; ij I is the element in the i-th row and j-th column of the original adjacency matrix A; n,ij I is an n-order identity matrix n The element in the i-th row and j-th column.

7. The active distribution network longitudinal protection method according to claim 1, characterized in that, The convolution operation performed on the transformed topological data to obtain the node feature matrix includes: Based on a spatial domain graph convolution framework using graph convolutional networks, the structured features of a sequence are extracted through the local interaction between the convolution kernel and the graph structure, resulting in a node feature matrix. The expression for the feature matrix of this node is: In the formula, Z is the node feature matrix; σ is the normalized adjacency matrix; σ(·) is the activation function used to enhance the nonlinear representation of features; X is the initial node feature matrix after the current sampling sequence is converted into a graph structure; W is the convolution kernel weight matrix and F' is the output feature dimension; b is the bias vector.

8. The active distribution network longitudinal protection method according to claim 1, characterized in that, The step of performing global pooling on the node feature matrix to calculate the similarity of graph convolutional sequences includes: The first node feature matrix and the second node feature matrix are obtained by performing graph convolution operations on the first current sampling sequence and the second current sampling sequence, respectively. Global average pooling is performed on the feature matrices of the first node and the feature matrices of the second node respectively to obtain the corresponding global feature vectors of the sequence. Cosine similarity is used as a similarity metric for global features of sequences to calculate the similarity of graph convolution sequences.

9. The active distribution network longitudinal protection method according to claim 8, characterized in that, The step of performing global average pooling on the feature matrices of the first node and the feature matrices of the second node includes: Based on the first node feature matrix and the second node feature matrix, obtain all row vectors of the node feature matrix respectively; Sum the row vectors of each row in the node feature matrix; Divide the summation result by the total number of rows in the node feature matrix to obtain the global feature vector of the sequence; The expression for the global feature vector of the sequence is: In the formula, f m z is the global feature vector of the sequence, and m = 1 or 2; m,k is the k-th row vector of the node feature matrix; n is the total number of rows in the node feature matrix.

10. The active distribution network longitudinal protection method according to claim 8, characterized in that, The method of using cosine similarity as a similarity metric for global sequence features, and calculating the similarity of graph convolutional sequences, includes: Calculate the dot product of the sequential global eigenvectors of the first node's feature matrix and the sequential global eigenvectors of the second node's feature matrix; Calculate the L2 norm of the sequential global eigenvectors of the first node feature matrix and the sequential global eigenvectors of the second node feature matrix, respectively. Divide the dot product by the product of the two L2 norms to obtain the graph convolution sequence similarity; The expression for the similarity of the graph convolutional sequences is: In the formula, S is the graph convolution sequence similarity; f1 is the sequence global feature vector of the first node feature matrix; and f2 is the sequence global feature vector of the second node feature matrix.

11. The active distribution network longitudinal protection method according to claim 10, characterized in that, The range of the similarity value of the graph convolution sequence is [-1, 1]; The fault identification criteria inside and outside the zone include determining the dynamic characteristics of the first current sampling sequence and the second current sampling sequence based on the similarity of graph convolution sequences, specifically including: When the similarity value of the graph convolution sequence is greater than or equal to the preset threshold, the system is judged to be in normal operation or to have experienced an external short circuit fault. When the similarity value of the graph convolution sequence is less than a preset threshold, it is determined that a short circuit fault has occurred in the system.

12. An active distribution network longitudinal protection system, characterized in that, The active power distribution network longitudinal protection system includes: The acquisition module is started to calculate the phase current mutation in real time. When the phase current mutation exceeds the load current of a preset multiple, the three-phase current sampling data after a preset time at the installation location of the protection device at this end is obtained as the first current sampling sequence and sent to the protection device at the other end. The data conversion module is used to acquire the three-phase current sampling data after a preset time at the installation location of the protection device at the other end as the second current sampling sequence, and convert the first current sampling sequence and the second current sampling sequence into topology data. The feature calculation module is used to perform convolution operations on the transformed topological structure data to obtain the node feature matrix, and then perform global pooling on the node feature matrix to calculate the similarity of the graph convolution sequence. The fault identification module is used to identify faults inside and outside the zone based on fault identification criteria. If the similarity of the graph convolution sequence is less than a preset threshold, it is judged as a short circuit fault inside the zone and a trip command is sent; otherwise, it is judged as normal operation or a short circuit fault outside the zone and the protection device resets.

13. An active distribution network longitudinal protection system according to claim 12, characterized in that, When converting data into topological data, the data conversion module includes mapping the first current sampling sequence and the second current sampling sequence into undirected graphs, respectively. Specifically, this includes: constructing a node set by using the values ​​of the current sampling points as nodes in the graph based on the first and second current sampling sequences; constructing an edge set by determining the connection relationships between nodes through the nearest neighbor rule; establishing a connection between each node and its adjacent sampling points in the nearest neighbor rule; calculating the edge weights between each node in the edge set; obtaining the edge weights by normalizing the sampling point time difference; establishing an adjacency matrix based on the weight values ​​to represent the connection relationships of the graph; and constructing undirected graphs of the first and second current sampling sequences according to the node set, the edge set, and the adjacency matrix.

14. An active distribution network longitudinal protection system according to claim 12, characterized in that, The feature calculation module, when performing global pooling on the node feature matrices to calculate the similarity of graph convolution sequences, includes: obtaining a first node feature matrix and a second node feature matrix by performing graph convolution operations on the first current sampling sequence and the second current sampling sequence, respectively; performing global average pooling on the first node feature matrix and the second node feature matrix to obtain the corresponding global feature vectors of the sequence; and using cosine similarity as a similarity metric for the global features of the sequence to calculate the similarity of the graph convolution sequences.

15. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 11.

16. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 11.

Images