A power distribution network topology intelligent identification method based on a deep convolutional network

Abstract

The application discloses a power distribution network topology intelligent identification method based on a deep convolutional network, which comprises the following steps: collecting electrical measurement data of multiple nodes in a power distribution network; inputting the electrical measurement data into a pre-trained deep convolutional network model for topology identification; and the construction method of the deep convolutional network model comprises the following steps: in the attention mechanism of the network model, a physical confidence gating factor is introduced to modulate the attention weight between any two nodes, so as to inhibit false association generated based on accidental similarity of data; a virtual impedance jitter index is obtained by calculating the discrete degree of a dynamic virtual impedance in a time window; and the virtual impedance jitter index is input into a pre-set nonlinear gating function to obtain the physical confidence gating factor. According to the scheme of the application, the accuracy and robustness of power distribution network topology identification are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of power systems and their automation technology, and in particular to a method for intelligent identification of distribution network topology based on deep convolutional networks. Background Technology

[0002] As the final link connecting users in the power system, the accurate identification of the distribution network's topology is fundamental for state estimation, fault location, and load transfer. With the integration of distributed power sources and the expansion of the distribution network, network operation modes are becoming more varied, and topology structures are frequently switching.

[0003] Existing topology identification methods mainly suffer from the following problems: 1. Reliance on complete measurements: Traditional methods rely on remote signaling / telemetry data from SCADA systems. When data is missing or contains errors, the identification accuracy decreases.

[0004] 2. Low computational efficiency: Mathematical optimization-based methods face combinatorial explosion problems when the number of nodes increases, making it difficult to meet real-time requirements.

[0005] 3. Scarcity of sample data: Machine learning-based methods require a large number of training samples, but in actual operation, sample data for faults or specific topologies are scarce and singular, resulting in insufficient generalization ability of the model.

[0006] 4. Existing technologies have significant limitations when processing actual measurement data. Distribution network measurement devices (such as PMUs and smart meters) often experience time-asynchronous transmission of voltage and power flow data due to communication delays, clock drift, and other factors, frequently accompanied by electromagnetic noise interference. In traditional attention mechanisms, weight calculation relies solely on the similarity of feature vectors in the numerical space (such as dot product or cosine similarity). When data exhibits temporal misalignment, two physically unconnected or weakly correlated nodes may exhibit accidental numerical synchronization (spurious correlation), causing the attention mechanism to incorrectly assign high weights and introduce noise features as topological criteria, resulting in a significant decrease in model recognition accuracy in noisy environments.

[0007] Therefore, the urgent problem to be solved is the low accuracy of topology identification under conditions of incomplete measurement data, high noise interference, and imbalanced samples. Summary of the Invention

[0008] To address the technical problems of low topology identification accuracy under conditions of incomplete measurement data, high noise interference, and imbalanced samples, this invention provides a distribution network topology intelligent identification method based on deep convolutional networks, which effectively improves the accuracy and robustness of distribution network topology identification.

[0009] To address this, the present invention employs the following technical solution: a distribution network topology intelligent identification method based on deep convolutional networks, comprising: collecting electrical measurement data of multiple nodes in the distribution network, including node voltage and injected power; inputting the electrical measurement data into a pre-trained deep convolutional network model for topology identification; the method for constructing the deep convolutional network model comprises: introducing a physical confidence gating factor in the attention mechanism of the network model to modulate the attention weights between any two nodes, so as to suppress false associations based on accidental data similarity; wherein, the method for calculating the physical confidence gating factor comprises: obtaining the ratio of the voltage difference between two nodes within a preset time window to the injected power of one of the nodes, to characterize the dynamic virtual impedance between the two nodes; calculating the dispersion of the dynamic virtual impedance within the time window to obtain the virtual impedance jitter index; and inputting the virtual impedance jitter index into a preset nonlinear gating function to obtain the physical confidence gating factor, wherein the value of the physical confidence gating factor is negatively correlated with the value of the virtual impedance jitter index.

[0010] Preferably, the method for constructing the deep convolutional network model further includes: performing data augmentation on the original training sample set to expand the training data; the data augmentation method includes at least one of the following: performing data augmentation based on generative adversarial networks to generate augmented data similar to the distribution of real data; performing data augmentation based on continuous power flow calculation to generate supplementary samples by simulating electrical data under different load levels and operating modes.

[0011] Preferably, the enhancement based on generative adversarial networks includes: constructing a generative adversarial network model containing a generator and a discriminator; the generator receiving random noise and outputting simulated electrical samples; the discriminator being responsible for distinguishing between real electrical samples and simulated electrical samples; and training the generator and discriminator through a game between them, enabling the generator to produce augmented data that approximates the real distribution.

[0012] Preferably, the enhancement based on continuous power flow calculation includes: when the number of training samples is insufficient, using continuous power flow calculation to simulate electrical data under different load levels and operating modes to generate supplementary samples to cover more operating scenarios.

[0013] Preferably, the network structure of the deep convolutional network model includes, in sequence: an input layer for receiving an electrical feature matrix, at least one convolutional layer for extracting local electrical features, a pooling layer for compressing feature dimensions, and a fully connected layer for mapping features to topological categories. The convolutional layer uses convolution kernels to perform convolution operations on the input matrix to extract local electrical features of the power distribution network, and introduces a nonlinear activation function to increase the model's fitting ability; after the convolutional layer, an attention mechanism is introduced to assign different weights to the electrical features of different nodes. The pooling layer downsamples the local electrical features extracted by convolution, compressing the feature dimension, reducing computational complexity, and improving the model's generalization ability. The fully connected layer flattens out the multidimensional features and maps the features to the sample label space through the fully connected structure.

[0014] Preferably, the nonlinear gating function is a variant of the Sigmoid function. When the virtual impedance jitter index is lower than a preset threshold, the value of the physical confidence gating factor approaches 1; when the virtual impedance jitter index is higher than the preset threshold, the value of the physical confidence gating factor approaches 0.

[0015] Preferably, the formula for calculating the physical confidence gating factor is:

[0016] In the formula, represents the physical confidence gating factor, a dimensionless coefficient with a value range of (0,1). This represents the impedance jitter tolerance threshold derived from historical operational data statistics. This represents the attenuation sensitivity coefficient, which controls the steepness of the nonlinear gating function.

[0017] Preferably, the formula for calculating attention weights is:

[0018] In the formula, This represents the normalized attention weights ultimately used for feature aggregation. and Let i and j represent the high-dimensional feature vectors of nodes i and j after feature transformation, respectively. The weight matrix is ​​a learnable matrix. The parameter vector representing the attention mechanism. This represents the activation function. This represents a vector concatenation operation. Represents a set of nodes.

[0019] Preferably, the formula for calculating the degree of dispersion of the physical consistency of electrical connections between nodes is as follows:

[0020] In the formula, This indicates the degree of dispersion in the physical consistency of the electrical connection between nodes i and j. and Let represent the voltage vectors of node i and node j at time t, respectively. and Let these represent the active power injection and reactive power injection at node i at time t, respectively. This represents a small constant used to prevent the denominator from being zero. This represents the average value of the virtual impedance magnitude calculated within the time window T.

[0021] Preferably, the electrical measurement data comes from a SCADA system, an AMI system, or a PMU device; the construction steps of the distribution network topology intelligent identification method also include cleaning and semantic analysis of the collected unstructured text data.

[0022] The beneficial effects of this invention are as follows: By introducing a virtual impedance jitter index and a physical confidence gating factor, this invention adds physical constraints to a purely data-driven neural network. The model can automatically identify and suppress spurious correlations caused by asynchronous measurement data or electromagnetic interference, ensuring that topology identification results are not misled by low-quality data, and exhibiting extremely high robustness in real-world distribution network scenarios with low signal-to-noise ratios. By verifying the physical consistency of voltage-power changes, it distinguishes between numerical similarity and physical connectivity from a mechanistic perspective, significantly reducing the topology misidentification rate in complex networks (such as ring networks and multi-branch structures). Attached Figure Description

[0023] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. In the drawings, several embodiments of the invention are illustrated by way of example and not limitation, and like or corresponding reference numerals denote like or corresponding parts, wherein: Figure 1 This is a flowchart illustrating a power distribution network topology intelligent identification method based on deep convolutional networks according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating a method for calculating the physical confidence gating factor according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the working mechanism of a generative adversarial network according to an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0026] Figure 1 This is a flowchart illustrating a distribution network topology intelligent identification method 100 based on a deep convolutional network according to an embodiment of the present invention.

[0027] like Figure 1 As shown, in step S101, electrical measurement data of multiple nodes in the distribution network are collected, including node voltage and injected power. In some embodiments, the electrical measurement data comes from a SCADA system, an AMI system, or a PMU device. The construction steps of the above-described distribution network topology intelligent identification method also include cleaning and semantic analysis of the collected unstructured text data.

[0028] In step S102, electrical measurement data is input into a pre-trained deep convolutional network model for topology identification.

[0029] Figure 2 This is a flowchart illustrating a method 200 for calculating the physical confidence gating factor according to an embodiment of the present invention.

[0030] In the attention mechanism of the network model, a physical confidence gating factor is introduced to modulate the attention weights between any two nodes in order to suppress false associations based on accidental data similarity.

[0031] The calculation method for the physical confidence gating factor is as follows: Figure 2 As shown, in step S201, the ratio of the voltage difference between the two nodes within a preset time window to the injected power of one of the nodes is obtained to characterize the dynamic virtual impedance between the two nodes.

[0032] In step S202, the dispersion of the dynamic virtual impedance within the time window is calculated to obtain the virtual impedance jitter index. In some embodiments, this virtual impedance jitter index can be characterized by the dispersion of the physical consistency of the electrical connections between nodes. The formula for calculating the dispersion of the physical consistency of the electrical connections between nodes is as follows:

[0033] In the formula, This indicates the degree of dispersion in the physical consistency of the electrical connection between nodes i and j. and Let represent the voltage vectors of node i and node j at time t, respectively. and Let these represent the active power injection and reactive power injection at node i at time t, respectively. This represents a small constant used to prevent the denominator from being zero. This represents the average value of the virtual impedance magnitude calculated within the time window T.

[0034] In step S203, the virtual impedance jitter index is input into a preset nonlinear gating function to obtain a physical confidence gating factor, wherein the value of the gating factor is negatively correlated with the value of the jitter index. In some embodiments, the nonlinear gating function is a variant of the Sigmoid function. When the virtual impedance jitter index is lower than a preset threshold, the value of the physical confidence gating factor approaches 1; when the virtual impedance jitter index is higher than the preset threshold, the value of the physical confidence gating factor approaches 0.

[0035] The formula for calculating the physical confidence gating factor is:

[0036] In the formula, represents the physical confidence gating factor, a dimensionless coefficient with a value range of (0,1). This represents the impedance jitter tolerance threshold derived from historical operational data statistics. This represents the attenuation sensitivity coefficient, which controls the steepness of the gating function.

[0037] Attention weights are adjusted based on the aforementioned physical confidence gating factor, and the formula for calculating these attention weights is as follows:

[0038] In the formula, This represents the normalized attention weights ultimately used for feature aggregation. and Let i and j represent the high-dimensional feature vectors of nodes i and j after feature transformation, respectively. The weight matrix is ​​a learnable matrix. The parameter vector representing the attention mechanism. This represents the activation function. This represents a vector concatenation operation. Represents a set of nodes.

[0039] In some embodiments, the method for constructing a deep convolutional network model further includes data augmentation of the original training sample set to expand the training data; the data augmentation method includes at least one of the following: data augmentation based on generative adversarial networks to generate augmented data similar to the distribution of real data; data augmentation based on continuous power flow calculation to generate supplementary samples by simulating electrical data under different load levels and operating modes.

[0040] When using a generative adversarial network (GAN)-based augmentation approach, a GAN model is first constructed, comprising a generator and a discriminator. The generator receives random noise and outputs simulated electrical samples, while the discriminator distinguishes between real and simulated electrical samples. Then, through game-based training between the generator and discriminator, the generator produces augmented data that approximates the true distribution.

[0041] When using an enhancement method based on continuous power flow calculation, even when the number of training samples is insufficient, continuous power flow calculation can be used to simulate electrical data under different load levels and operating modes to generate supplementary samples to cover more operating scenarios.

[0042] The network structure of the aforementioned deep convolutional network model includes, in sequence: an input layer for receiving the electrical feature matrix, at least one convolutional layer for extracting local electrical features, a pooling layer for compressing the feature dimension, and a fully connected layer for mapping features to topological categories.

[0043] The present invention will now be described in detail with reference to specific embodiments. This embodiment mainly includes the following steps.

[0044] Step 1: Construct a sample mapping relationship between topology status and electrical data. Collect historical operating data of the distribution network from the SCADA or AMI system. Extract line labels (topology categories) for both on / off states in the historical operating status, as well as the corresponding node voltage amplitude, active power, reactive power, and other electrical characteristic data for that state. Normalize the collected electrical measurement data to construct an original sample set of "topology label - electrical characteristic".

[0045] In addition to basic SCADA / AMI data, high-frequency data from PMUs (Phasor Measurement Units) can be integrated to capture more subtle dynamic features. Structured rules and semantic analysis techniques are introduced during processing to clean unstructured text (such as scheduling instructions) and eliminate format heterogeneity issues caused by differences in data sources.

[0046] Step Two: Sample Data Augmentation Based on GAN or Power Flow Calculation. To address the issue of insufficient data volume for certain topological categories in the original sample set, data augmentation is performed using any one or a combination of the following methods to construct an augmented training set: 1. Enhancements based on Generative Adversarial Networks (GANs): such as Figure 3 As shown, an adversarial network model is constructed, comprising a generator G and a discriminator D. The generator receives random noise and outputs simulated electrical samples; the discriminator is responsible for distinguishing real samples from generated samples. Through game-like training between the two, the generator produces augmented data that approximates the true distribution, addressing the problems of small sample size and data imbalance.

[0047] 2. Enhancement based on continuous power flow calculation: When the number of training samples is insufficient, continuous power flow calculation is used to simulate electrical data under different load levels and operating modes to generate supplementary samples to cover more operating scenarios.

[0048] Step 3: Construct a topology classification network based on convolutional layers, pooling layers, and fully connected layers. Build a deep convolutional neural network (CNN) model to explore the nonlinear relationship between electrical features and topology. The network structure specifically includes: 1. Input layer: Receives the electrical feature matrix (such as the node voltage-power matrix) processed in step two.

[0049] 2. Convolutional Layer: This layer uses convolutional kernels to perform convolution operations on the input matrix, extracting local electrical features of the distribution network, and introduces a nonlinear activation function to enhance the model's fitting ability. In this embodiment, an attention mechanism is introduced after the convolutional layer to assign different weights to the electrical features of different nodes. This allows the model to focus on key nodes (such as nodes near tie switches), quickly capture topological abrupt changes, and improve identification accuracy.

[0050] 3. Pooling Layer: Performs downsampling operations (such as max pooling or average pooling) on ​​convolutional features to compress feature dimensions, reduce computational complexity, and improve the model's generalization ability.

[0051] 4. Fully Connected Layer: Flattens out multidimensional features and maps them to the sample label space through a fully connected structure. The sample label space is a set of all possible topological categories.

[0052] 5. Output layer: Outputs the probability of the current distribution network topology category.

[0053] To avoid situations where two physically unconnected or weakly correlated nodes might exhibit accidental numerical synchronization (spurious correlation), leading to the attention mechanism incorrectly assigning high weights, this embodiment constructs a virtual impedance jitter index based on physical laws when introducing the attention mechanism, and uses it to adjust the weights.

[0054] According to Ohm's law, if there is a real physical connection between two nodes i and j and their data is synchronized, the ratio of their voltage difference to the current (or power) flowing through them (i.e., impedance) should remain relatively stable for a short period of time. Conversely, if the data is not synchronized or the similarity is merely a coincidence caused by noise, this ratio will fluctuate drastically.

[0055] Based on this physical logic, a virtual impedance jitter index is constructed. This index does not depend on the actual line current (because the topology is unknown), but rather uses node-injected power as an excitation source for approximate deduction.

[0056] The formula is as follows:

[0057] In the formula, This indicates the degree of dispersion in the physical consistency of the electrical connection between nodes i and j. and Let represent the voltage vectors of node i and node j at time t, respectively. and Let these represent the active power injection and reactive power injection at node i at time t, respectively. This represents a small constant used to prevent the denominator from being zero. This represents the average value of the virtual impedance magnitude calculated within the time window T.

[0058] This formula defines the instantaneous ratio of voltage drop to power as dynamic virtual impedance. When The smaller the value, the more stable the voltage and power fluctuations at nodes i and j are within the time window, indicating good data synchronization as they conform to the laws of physical transmission. An increase in the value indicates that the voltage difference and power change between the two nodes are not synchronized, suggesting the presence of phase drift or random noise, a low probability of physical connection, or extremely poor data quality. This step transforms the raw electrical data into physical stability characteristics.

[0059] After obtaining the aforementioned index representing the degree of dispersion of the physical consistency of electrical connections between nodes, it cannot be directly used for attention calculation because the jitter index is an unbounded physical quantity and needs to be transformed into a non-linear suppression probability. Considering the scene characteristics, when the jitter exceeds a certain threshold, the weight of that edge should be drastically reduced to shield against noise interference.

[0060] Based on this, the present invention sets a physical confidence gating factor to softly truncate the "spurious correlations" generated by noise and asynchronous data.

[0061] The formula for calculating the physical confidence gating factor is as follows:

[0062] In the formula, represents the physical confidence gating factor, a dimensionless coefficient with a value range of (0,1). This represents the impedance jitter tolerance threshold derived from historical operational data statistics. This represents the attenuation sensitivity coefficient, which controls the steepness of the gating function.

[0063] The above formula is a variant of the Sigmoid function. When the virtual impedance jitter... Below the threshold hour, A value close to 1 indicates that the data relationship on that side has a high degree of physical confidence and should be retained; when jitter... When the threshold is exceeded, It rapidly approaches 0. This process maps the discreteness of the physical consistency of electrical connections between nodes to probability weights, effectively truncating the "spurious correlations" caused by noise and asynchronous data.

[0064] Introducing an attention mechanism, the corresponding attention weights are represented as follows:

[0065] In the formula, This represents the normalized attention weights ultimately used for feature aggregation. and Let i and j represent the high-dimensional feature vectors of nodes i and j after feature transformation, respectively. The weight matrix is ​​a learnable matrix. The parameter vector representing the attention mechanism. This represents the activation function. This indicates a vector concatenation operation.

[0066] In this formula, the original feature similarity score (LeakyReLU term) is reduced by a physical confidence gating factor. Direct modulation. Even if the data from two nodes are numerically very similar (i.e., the original scores are high), if their physical consistency is poor (i.e., ... (Approaching 0), the final exponential term is also forced to be lowered. This ensures that the final weights not only reflect the similarity of data patterns but also strictly adhere to physical and electrical constraints. Through this improved network layer-by-layer aggregation feature, the final output is the on / off status of distribution network lines classified by the fully connected layer.

[0067] Embedding physical laws into the generation process of attention weights gives the intermediate layer weights of the neural network a clear physical meaning (i.e., the physical reliability of the connection), which not only improves the recognition accuracy, but also provides a side reference for maintenance personnel to troubleshoot measurement equipment faults.

[0068] Step 4: Offline training of the model.

[0069] The enhanced training set obtained in step two is used to train the CNN model constructed in step three offline. The network weights are adjusted using the backpropagation algorithm, enabling the model to accurately classify the corresponding topology from the input electrical data.

[0070] Step 5: Online real-time topology prediction.

[0071] Real-time cross-sectional measurement data (such as voltage amplitude and injected power) of the distribution network are collected, normalized, and then input into a trained CNN model. The model outputs the predicted topology category, realizing automated and intelligent identification of the distribution network topology.

[0072] While this specification has shown and described numerous embodiments of the invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in the practice of this invention.

Claims

1. A method for intelligent identification of power distribution network topology based on deep convolutional networks, characterized in that, include: Collect electrical measurement data from multiple nodes in the distribution network, including node voltage and injected power; The electrical measurement data is input into a pre-trained deep convolutional network model for topology identification; The method for constructing the deep convolutional network model includes: introducing a physical confidence gating factor in the attention mechanism of the network model to modulate the attention weights between any two nodes, so as to suppress false associations based on accidental data similarity; wherein, the calculation method of the physical confidence gating factor includes: The ratio of the voltage difference between two nodes within a preset time window to the injected power of one of the nodes is obtained to characterize the dynamic virtual impedance between the two nodes. The virtual impedance jitter index is obtained by calculating the dispersion of the dynamic virtual impedance within a time window; and The virtual impedance jitter index is input into a preset nonlinear gating function to obtain the physical confidence gating factor, wherein the value of the physical confidence gating factor is negatively correlated with the value of the virtual impedance jitter index; The virtual impedance jitter index is characterized by the dispersion of the physical consistency of the electrical connections between nodes. The formula for calculating the dispersion of the physical consistency of the electrical connections between nodes is as follows: In the formula, This indicates the degree of dispersion in the physical consistency of the electrical connection between nodes i and j. and Let represent the voltage vectors of node i and node j at time t, respectively. and Let these represent the active power injection and reactive power injection at node i at time t, respectively. This represents a small constant used to prevent the denominator from being zero. This represents the average value of the virtual impedance magnitude calculated within the time window T.

2. The intelligent identification method for power distribution network topology based on deep convolutional networks according to claim 1, characterized in that, The method for constructing the deep convolutional network model also includes: Data augmentation is performed on the original training sample set to expand the training data; the data augmentation method includes at least one of the following: Data augmentation is performed based on generative adversarial networks to generate augmented data that is similar to the distribution of real data. Data augmentation is performed based on continuous power flow calculations, generating supplementary samples by simulating electrical data under different load levels and operating modes.

3. The intelligent identification method for power distribution network topology based on deep convolutional networks according to claim 2, characterized in that, The enhancements based on generative adversarial networks include: A generative adversarial network model is constructed, which includes a generator and a discriminator. The generator receives random noise and outputs simulated electrical samples. The discriminator is responsible for distinguishing between real electrical samples and simulated electrical samples. By training the generator and discriminator through a game-like process, the generator can produce augmented data that approximates the true distribution.

4. The intelligent identification method for power distribution network topology based on deep convolutional networks according to claim 2, characterized in that, The enhancements based on continuous power flow calculation include: When the training sample size is insufficient, continuous power flow calculation is used to simulate electrical data under different load levels and operating modes to generate supplementary samples to cover more operating scenarios.

5. The intelligent identification method for power distribution network topology based on deep convolutional networks according to claim 2, characterized in that, The network structure of the deep convolutional network model includes, in sequence: an input layer for receiving the electrical feature matrix, at least one convolutional layer for extracting local electrical features, a pooling layer for compressing the feature dimension, and a fully connected layer for mapping features to topological categories. The convolutional layer uses convolution kernels to perform convolution operations on the input matrix to extract local electrical features of the power distribution network, and introduces a nonlinear activation function to increase the model's fitting ability; after the convolutional layer, an attention mechanism is introduced to assign different weights to the electrical features of different nodes. The pooling layer downsamples the local electrical features extracted by convolution, compressing the feature dimension, reducing computational complexity, and improving the model's generalization ability. The fully connected layer flattens out the multidimensional features and maps the features to the sample label space through the fully connected structure.

6. The intelligent identification method for power distribution network topology based on deep convolutional networks according to claim 1, characterized in that, The nonlinear gating function is a variant of the Sigmoid function. When the virtual impedance jitter index is lower than a preset threshold, the value of the physical confidence gating factor approaches 1; when the virtual impedance jitter index is higher than the preset threshold, the value of the physical confidence gating factor approaches 0.

7. The intelligent identification method for power distribution network topology based on deep convolutional networks according to claim 1, characterized in that, The formula for calculating the physical confidence gating factor is: In the formula, represents the physical confidence gating factor, a dimensionless coefficient with a value range of (0,1). This represents the impedance jitter tolerance threshold derived from historical operational data statistics. This represents the attenuation sensitivity coefficient, which controls the steepness of the nonlinear gating function. This indicates the degree of dispersion in the physical consistency of the electrical connection between nodes i and j.

8. The intelligent identification method for distribution network topology based on deep convolutional networks according to claim 7, characterized in that, The formula for calculating attention weights is: In the formula, This represents the normalized attention weights ultimately used for feature aggregation. and Let i and j represent the high-dimensional feature vectors of nodes i and j after feature transformation, respectively. The weight matrix is ​​a learnable matrix. The parameter vector representing the attention mechanism. This represents the activation function. This represents a vector concatenation operation. Represents a set of nodes.

9. The intelligent identification method for power distribution network topology based on deep convolutional networks according to claim 1, characterized in that, The electrical measurement data comes from SCADA systems, AMI systems, or PMU devices; the construction steps of the distribution network topology intelligent identification method also include cleaning and semantic analysis of the collected unstructured text data.

Images