A low-voltage substation topology identification method using a graph diffusion model and its system, equipment, and medium

Abstract

The present invention relates to the field of power distribution network technology, and more specifically, to a low-voltage substation topology identification method, system, device, and medium using a graph diffusion model. The method provided by the present invention mainly includes establishing a topology identification model, inputting target time series features into the topology identification model, implicitly identifying pairs of nodes with synchronized fluctuations through the topology identification model, outputting node features that integrate neighbor information, predicting the connection probability between each node based on the node features, and outputting a new substation topology structure. This solution is different from traditional methods that rely on specific substation rules or require a lot of manual adjustments. The graph neural network trained by this solution is designed to learn general "electrical feature-topology structure" mapping rules and denoising logic. The model is not limited to the specific structure of the training data and can be effectively generalized and applied to new substations with large differences in topology structures, solving the pain points of poor adaptability and weak scalability of traditional methods.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network technology, and in particular to a low-voltage substation topology identification method, system, equipment, and medium using a graph diffusion model. Background Art

[0002] As a key link connecting the power system and end users, the real-time accuracy of the topology of the distribution network (i.e., the connection relationship between components) is the basis for ensuring the safe and efficient operation of the power grid and realizing advanced applications (such as fault location, load transfer, and new energy consumption). However, the distribution network itself has the characteristics of complex structure, large scale, changeable operation mode (frequent switch operation), and often incomplete measurement information, making accurate grasp of its real-time operating topology a long-standing industry challenge. Traditional topology acquisition methods, such as relying on communication to upload switch status information, face the problems of high cost, unreliable communication or incomplete coverage; while the signal injection method requires additional equipment and may interfere with the power grid.

[0003] In recent years, data-driven approaches using existing electrical quantities (such as voltage and current) for analysis, including the application of neural networks (NNs) or generative adversarial networks (GANs), have shown promise. However, these approaches often suffer from limitations such as insufficient generalization to specific topologies, the requirement for large amounts of labeled data, difficulty in training, and high data quality requirements. Therefore, the industry urgently needs a new intelligent topology analysis method that relies solely on readily available conventional electrical quantity data, is low-cost, and can automatically adapt to various complex topological changes. Summary of the Invention

[0004] The purpose of the present invention is to provide a low voltage area topology identification method and system, equipment and medium using a graph diffusion model to solve the problems in the above-mentioned prior art.

[0005] The present invention is achieved through the following technical solutions:

[0006] In a first aspect, the present invention provides a method for identifying low-voltage substation topology using a graph diffusion model, comprising:

[0007] Step 1: Obtain historical data from several different distribution substations, normalize the historical data, divide the historical data into sliding windows, establish a time series modeling module, and extract the time series features of each sliding window through the time series modeling module;

[0008] Step 2: The extracted time series features are used as the initial features of the nodes and input into the graph neural network. The graph neural network is trained based on the initial features to obtain a topology recognition model.

[0009] Step 3: Get the target data of several different distribution stations, process the target data as in step 1, and get the target time series features.

[0010] Step 4: Input the target time series features into the topology recognition model, implicitly identify the node pairs with synchronous fluctuations through the topology recognition model, output the node features that integrate the neighbor information, predict the connection probability between each node based on the node features, and output the new substation topology structure.

[0011] Preferably, obtaining the historical data or target data includes obtaining substation infrastructure data and obtaining electrical quantity data;

[0012] The acquisition of the substation infrastructure data includes acquiring the actual topological relationship diagram of each substation, and converting the actual topological relationship diagram into a mathematical diagram. The representation of the mathematical diagram is G=(V,E), where G is a relationship diagram, V is a node combination, E is an edge set, and E=0 or 1. When E is equal to 0, there is no power connection between the two nodes, and when E is equal to 1, there is a power connection between the two nodes.

[0013] Preferably, the normalization process includes:

[0014]

[0015] Where, is the normalized data, For the input data, is the minimum value in the data set, is the maximum value in the data set.

[0016] Preferably, the performing sliding window division on the historical data includes:

[0017] Set the time series for the normalized data and set the window parameters, which include window size, sliding step size and prediction range;

[0018] Determine the data point of the starting index of the first window and the starting index of the last window, and set the number of valid window samples. Through a loop, the variable starts from the window of the starting index and is incremented by the step size. The number of divisions is increased until it reaches the maximum number of valid window samples, and then the division is ended;

[0019] The divided data is generated into data groups and then output.

[0020] Preferably, the point of determining the starting index of the last window includes:

[0021]

[0022] in, The sequence value of the point with the starting index of the last window, The sequence value of the data point with the starting index of the first window, is the window size, is the prediction range, is the offset, The total number of data.

[0023] Preferably, determining the number of valid window samples includes:

[0024]

[0025] Where, is the number of valid window samples.

[0026] Preferably, the training of the graph neural network based on the initial features includes:

[0027] Get the initial features, add noise multiple times in a cycle, generate noisy topology samples, set the number of iterations, and when the number of iterations is reached, store the current noisy topology samples and noise information as training targets;

[0028] The training objective provides training data about noisy topological samples and noise information, and the training data is input into the graph neural network for training while calculating the loss. The parameters of the graph neural network are updated by the calculated loss, and the topology recognition model is output based on the updated parameters of the graph neural network and the training data.

[0029] In a second aspect, the present invention further provides a low-voltage substation topology identification system using a graph diffusion model, comprising:

[0030] The model building module is configured to obtain historical data from several different distribution substations, normalize the historical data, divide the historical data into sliding windows, establish a time series modeling module, and extract time series features of each sliding window through the time series modeling module; the extracted time series features are used as initial features of the nodes and input into the graph neural network, and the graph neural network is trained based on the initial features to obtain a topology recognition model;

[0031] The recognition module is configured to obtain the target data of several different distribution station areas, and obtain the target time series characteristics after processing the target data as in step 1.

[0032] The target time series features are input into the topology recognition model, and the node pairs with fluctuation synchronization are implicitly identified through the topology recognition model. The node features that integrate neighbor information are output, and the connection probability between each node is predicted based on the node features, and the new substation topology structure is output.

[0033] In a third aspect, the present invention also provides an electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the above-mentioned low-voltage substation topology identification method using a graph diffusion model.

[0034] In a fourth aspect, the present invention further provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned low-voltage substation topology identification method using a graph diffusion model.

[0035] The technical solution of the present invention has at least the following advantages and beneficial effects:

[0036] The method provided by the present invention mainly includes establishing a topology recognition model, inputting the target time series features into the topology recognition model, implicitly identifying the node pairs with fluctuation synchronization through the topology recognition model, outputting the node features that integrate the neighbor information, predicting the connection probability between each node based on the node features, and outputting the new substation topology structure. This solution is different from the traditional method that relies on specific substation rules or requires a lot of manual adjustments. The graph neural network trained by this solution is designed to learn the general "electrical feature-topology structure" mapping rules and denoising logic. Therefore, the model is not limited to the specific structure of the training data, and can be effectively generalized and applied to new substations with large differences in topology structures. It solves the pain points of poor adaptability and weak scalability of traditional methods, and greatly reduces the cost and difficulty of deploying the model in different regions. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0038] Figure 1 Schematic diagram of the control flow of the topology recognition model of the present invention;

[0039] Figure 2 It is a schematic diagram of the station area structure of the present invention. DETAILED DESCRIPTION

[0040] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0041] The division of modules in this application is a logical division. In actual application, there may be other division methods. For example, multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0042] Independently described modules or submodules may or may not be physically separate; they may be implemented in software or hardware. Some modules or submodules may be implemented in software, with the processor invoking the software to implement the functionality of these modules or submodules, while other modules or submodules may be implemented in hardware, such as hardware circuits. Furthermore, some or all of the modules may be selected based on actual needs to achieve the objectives of the present application.

[0043] Please refer to Figure 1 The present invention provides a low-voltage substation topology identification method using a graph diffusion model, comprising:

[0044] Step 1: Obtain historical data from several different distribution substations, normalize the historical data, divide the historical data into sliding windows, establish a time series modeling module, and extract the time series features of each sliding window through the time series modeling module;

[0045] like Figure 2 The figure shows a typical residential distribution substation structure. The incoming power line enters the transformer directly, then passes through a multi-layered distribution network to the meter layer (meter box layer). This is what we often call a one-meter-per-household system. The meter and transformer (concentrator) can communicate, so the correspondence between each meter and transformer is known. However, the relationship between each intermediate layer and the meter is unknown and often confusing. Therefore, this relationship between the intermediate layer and the meter is a key topology issue for many power supply companies. If necessary, detection equipment can be installed at other layers to identify the corresponding topology.

[0046] Secondly, a sliding window technique is used to process the time series data of electrical quantities at each node, such as voltage, current, or power. The historical data is divided into multiple time windows, each processing only the values ​​from the previous period. A time series modeling module (such as a one-dimensional convolutional neural network or a recurrent neural network) is used to extract features within each window. These features reflect the local dynamic characteristics of the electrical quantity, such as fluctuation amplitude and trend, and provide input for the subsequent graph neural network. Unlike traditional methods that process the entire historical data at once, this segmented processing approach can flexibly adapt to changes at different time scales, ensuring targeted and effective feature extraction.

[0047] Step 2: The extracted time series features are used as the initial features of the nodes and input into the graph neural network. The graph neural network is trained based on the initial features to obtain a topology recognition model.

[0048] The extracted time series features are used as the initial features of the nodes and input into the graph neural network. GNN automatically learns the relationship between node features through a multi-layer message passing mechanism. For example, when using the graph attention network (GAT), the model can dynamically assign weights to neighboring nodes and implicitly focus on nodes with similar electrical quantity fluctuations without manually designing similarity calculation rules. This method enables the model to capture node pairs that have experienced electrical quantity fluctuations at the same time multiple times in history, and then infer possible connection relationships. The end-to-end learning characteristics of GNN ensure its robustness and adaptability in complex data scenarios.

[0049] Step 3: Get the target data of several different distribution stations, process the target data as in step 1, and get the target time series features.

[0050] Step 4: Input the target time series features into the topology recognition model, implicitly identify the node pairs with synchronous fluctuations through the topology recognition model, output the node features that integrate the neighbor information, predict the connection probability between each node based on the node features, and output the new substation topology structure.

[0051] Regarding implicit recognition, it is specifically about automatically learning and capturing the inherent laws of dynamic relationships between nodes (such as fluctuation synchronization) through deep learning models, rather than relying on preset rules or explicit feature definitions.

[0052] Models (such as neural networks (GNNs) or attention mechanisms) automatically extract deep features from raw data (such as node voltage fluctuation time series), eliminating the need for manually designed synchronization criteria. For example, fluctuation synchronization between pairs of nodes may manifest as similarity in voltage curves or phase correlation. The model autonomously discovers these correlation patterns through training.

[0053] Dynamic Relationship Modeling: Implicit recognition does not explicitly label "synchronized node pairs." Instead, it uses nonlinear transformations in a neural network to encode the dynamic interactions between nodes as vector distances or attention weights in a latent space. If the latent vectors of two nodes are highly similar or the attention weights are significant, they are considered a synchronized pair.

[0054] Based on the features output by the GNN, the model uses a simple fully connected layer to predict the connection probability between each pair of nodes and generate a complete substation topology. The advantage of this design lies in its combination of sliding windows and the automatic feature learning capabilities of GNNs, avoiding the tedious manual feature engineering required by traditional methods while efficiently processing both time series and graph-structured data in distribution network substations.

[0055] Leveraging the powerful graph structure learning capabilities of graph neural networks (GNNs), we can deeply explore and understand the inherent and unique network connectivity patterns of different distribution substations. More importantly, GNNs can learn and model the complex physical mapping relationship between topological structures and electrical quantities (such as voltage, current, and power).

[0056] The core computational mechanisms of GNNs (such as message passing) are based on local operations between nodes and their neighbors, and their parameters (weights) are shared across nodes throughout the entire graph. This property makes GNN models naturally capable of handling topological graphs of varying scale and complexity. Compared to many traditional methods that require fixed input dimensions or whose computational complexity grows dramatically with scale, trained GNN models can be directly applied to topological recognition tasks ranging from small and simple areas to large and complex areas without structural modification, demonstrating excellent inherent scalability.

[0057] Through the core training mechanism of "generating noisy samples to training GNN denoising" (similar to the gradual denoising idea of ​​the diffusion model), the model is explicitly trained to cope with various topological noises, such as incorrect switch states, sensor data disturbances, and data loss caused by communication interruptions.

[0058] Specifically, the model is explicitly trained to cope with various topological noises by converting the denoising process into an optimizable parameter task through controllable noise injection and structured learning.

[0059] During the training phase, noise is actively injected into the original graph structure (such as randomly adding / deleting edges, flipping edge types) to generate noisy samples.

[0060] For example:

[0061] Random edge removal: simulates missing connections in real scenarios (e.g., PTDNet parameterizes edge retention probability through Bernoulli distribution);

[0062] Adversarial noise: Generate noise edges that are contrary to the downstream task objectives (such as intentionally connecting nodes with large label differences);

[0063] The noise range can cover structural noise (edge ​​missing / redundancy), feature noise (node ​​attribute perturbation) and label noise (incorrect labeling).

[0064] Noise level controllability: By adjusting the noise injection intensity (such as Gaussian noise variance and edge perturbation ratio), the model can be adapted to different noise scenarios (such as blind denoising tasks). For example, the DnCNN-B model is trained within the noise level range of [0,55], enabling a single model to process unknown noise.

[0065] Secondly, each GNN layer introduces a parameterized mask to dynamically filter out invalid edge information. For example, PTDNet designs a Bernoulli distribution parameter πuvl for each layer, learns edge importance through gradient optimization, suppresses the propagation of noisy edges, and adds regularization terms (such as structure reconstruction loss) to the loss function to force the model to focus on high-confidence neighbors.

[0066] Residual learning and noise separation: the model does not directly predict the clean graph structure, but learns the noise residual (such as v=y−x, y is the noisy input and x is the clean target)

[0067] Advantages: Convert denoising into an additive task to avoid overfitting the model to noise patterns (e.g., DnCNN achieves blind denoising through residual mapping)

[0068] Dynamic topology adjustment capabilities, through neighbor comparison loss, pseudo-label enhancement and other technologies, correct incorrect topological dependencies caused by noise (for example, CR-GNN uses prediction consistency to screen reliable nodes).

[0069] Simultaneous explicit training features: parameterization and end-to-end optimization, guiding the denoising direction through the loss function, joint loss design: combining reconstruction loss (such as MSE) and topology regularization loss (such as neighbor smoothing constraint); adaptive threshold: setting noise filtering thresholds (such as alpha / beta thresholds) and dynamically adjusting denoising strength (see Cycle-GAN-Unet); end-to-end optimization framework, the entire process from noisy input → intermediate representation → clean output is differentiable, realizing closed-loop training of noise injection and denoising.

[0070] This makes the model naturally resistant to the inevitable measurement noise and data uncertainty in actual operation, and can recover the most likely true topological state from ambiguous and incomplete information, which is significantly better than traditional methods that are sensitive to noise.

[0071] Currently, there are very few existing datasets for substations that provide complete topology and time series data. This solution uses a diffusion approach to generate training datasets, allowing the model to learn gradually, reducing the learning difficulty. Furthermore, it doesn't require a large amount of data to support it. Furthermore, this data generation method is similar to real-world situations, where the initial topology may be known, but over time, illegal construction may occur.

[0072] Unlike traditional methods that rely on specific substation rules or require extensive manual tuning, the GNN trained in this solution is designed to learn universal "electrical feature-topology" mapping rules and denoising logic. As a result, the model is not limited to the specific structure of the training data and can be effectively generalized and applied to new substations with widely varying topologies. This addresses the pain points of traditional methods, such as poor adaptability and scalability, and significantly reduces the cost and difficulty of deploying the model in different regions.

[0073] Specifically, obtaining the historical data or target data includes obtaining substation infrastructure data and obtaining electrical quantity data;

[0074] Substation infrastructure data: Collect the actual topology diagram for each substation and convert it into a standard mathematical graph representation, G=(V,E), where V represents the node set, including node types (such as transformers, switchgear, branch boxes, smart meters, etc.) and their attribute information; E represents the edge set, where E(i,j)=1 indicates a direct electrical connection between nodes i and j, and E(i,j)=0 indicates no connection. Basic information such as equipment ratings and installation locations is also recorded to establish a complete substation static topology database, providing standard training labels for the model. The dataset should cover typical topologies to ensure that the model can learn diverse topological patterns.

[0075] Electrical quantity data: Obtain historical time series data on electrical quantities such as voltage, current, active / reactive power, and power factor for all nodes within each substation in the dataset, with a sampling frequency of at least 15 minutes per point and a coverage period of at least one month to ensure that different electricity consumption patterns, such as weekdays and weekends, peak and off-peak periods, are captured. Comprehensive preprocessing is performed on the raw measurement data, including outlier detection and processing (based on statistical thresholds or density clustering methods), unified timestamp alignment, missing value filling (using time series interpolation techniques), and data standardization. Data enhancement techniques such as sliding window segmentation, load scaling, and moderate noise injection are used to expand the dataset size and increase its diversity, simulating electrical behavior characteristics under different operating conditions. The dataset should pay special attention to the correlation between electrical quantities between nodes, which is the key basis for subsequent topology inference.

[0076] Specifically, the normalization process includes:

[0077]

[0078] Where, is the normalized data, For the input data, is the minimum value in the data set, is the maximum value in the data set.

[0079] In an exemplary embodiment of the present invention, the sliding window partitioning of the historical data includes:

[0080] Set the time series for the normalized data and set the window parameters, which include window size, sliding step size and prediction range;

[0081] Determine the data point of the starting index of the first window and the starting index of the last window, and set the number of valid window samples. Through a loop, the variable starts from the window of the starting index and is incremented by the step size. The number of divisions is increased until it reaches the maximum number of valid window samples, and then the division is ended;

[0082] The divided data is generated into data groups and then output.

[0083] Specifically, determining the starting index of the last window includes:

[0084]

[0085] in, The sequence value of the point with the starting index of the last window, The sequence value of the data point with the starting index of the first window, is the window size, is the prediction range, is the offset, The total number of data.

[0086] Secondly, determining the number of valid window samples includes:

[0087]

[0088] Where, is the number of valid window samples.

[0089] In one exemplary embodiment of the present invention, the model training employs a diffusion model-based training method, designed to simulate complex real-world scenarios and recover the true distribution network topology from noisy data. The entire training process consists of two key stages: noise addition and denoising learning. By gradually introducing noise and learning the inverse recovery process, the model is able to adapt to the uncertainties of real-world scenarios. Training the graph neural network based on initial features includes:

[0090] Get the initial features, add noise multiple times in a cycle, generate noisy topology samples, set the number of iterations, and when the number of iterations is reached, store the current noisy topology samples and noise information as training targets;

[0091] The training objective provides training data about noisy topological samples and noise information, and the training data is input into the graph neural network for training while calculating the loss. The parameters of the graph neural network are updated by the calculated loss, and the topology recognition model is output based on the updated parameters of the graph neural network and the training data.

[0092] During the noise addition process, the model uses the actual topology as a starting point and gradually introduces noise to simulate various realistic scenarios that may arise during distribution network substation operation. For example, by randomly adding or removing edges, temporary line disconnections or additions can be reflected; by applying perturbations to node characteristics (such as voltage and current), measurement device errors or environmental interference can be simulated; and by partially masking data, data loss can be reproduced. The introduction of these noises is not arbitrary but rather constrained by underlying physical boundary conditions. Specifically, noise addition is required to ensure that the number of node connections does not exceed the upper limit of the ports of transformers or equipment in the power system, maintain overall substation connectivity, avoid isolated nodes, and ensure that electrical quantities (such as voltage and current) remain within reasonable ranges, avoiding unrealistic values. This design ensures that the training data truly reflects the operating characteristics of the distribution network.

[0093] During the denoising learning process, the model, through multiple iterations, gradually recovers the true connectivity from the noised topology. The model uses a graph neural network (GNN) as its core denoising module. Leveraging the GNN's message-passing mechanism, it analyzes the inherent correlations between electrical characteristics between nodes and predicts the true topological connectivity. In each training iteration, the model receives the current noise state and diffusion step length information. Based on these inputs, it predicts the connection probability between nodes, calculates the loss based on the difference with the true topology, and optimizes the model parameters using a backpropagation algorithm. After multiple rounds of training, the model effectively reduces the impact of noise and ultimately generates a highly accurate topological structure.

[0094] This training method simulates real-world scenarios and imposes physical constraints during the noise addition phase, ensuring that the model not only learns theoretical topological features but also adapts to the complex conditions of actual distribution networks. This approach significantly improves the model's robustness and accuracy, enabling it to excel in the task of identifying distribution network substation topology, fully meeting the practical needs of intelligent management.

[0095] In a second aspect, the present invention further provides a low-voltage substation topology identification system using a graph diffusion model, comprising:

[0096] The model building module is configured to obtain historical data from several different distribution substations, normalize the historical data, divide the historical data into sliding windows, establish a time series modeling module, and extract time series features of each sliding window through the time series modeling module; the extracted time series features are used as initial features of the nodes and input into the graph neural network, and the graph neural network is trained based on the initial features to obtain a topology recognition model;

[0097] The recognition module is configured to obtain the target data of several different distribution station areas, and obtain the target time series characteristics after processing the target data as in step 1.

[0098] The target time series features are input into the topology recognition model, and the node pairs with fluctuation synchronization are implicitly identified through the topology recognition model. The node features that integrate neighbor information are output, and the connection probability between each node is predicted based on the node features, and the new substation topology structure is output.

[0099] The process of actual use of the present invention is as follows:

[0100] During the new substation topology identification phase, the present invention utilizes a trained temporal graph neural network (GNN) model to automatically infer the topology of the new substation based on its time-series data of electrical quantities (such as voltage and current). This model requires no prior topological information and adapts to unknown connections. The process involves first collecting and normalizing the time-series data of each node. This data is then divided into multiple time periods using a sliding window technique, and a time-series modeling module is used to extract local features and capture the dynamic characteristics of the electrical quantities. These features are then input into the GNN, which automatically learns inter-node correlations using a message-passing mechanism, implicitly identifies pairs of nodes with synchronized fluctuations, and outputs node features that incorporate neighbor information. Finally, based on these features, the connection probability between each pair of nodes is predicted to generate the topology of the new substation. This process is efficient, accurate, and can quickly adapt to unknown scenarios, providing strong technical support for dynamic distribution network management and the integration of new substations.

[0101] In addition, the functional units in the various embodiments of the present invention may be integrated into a single processing unit, each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0102] If the integrated unit is implemented as a software functional unit and sold or used as a standalone product, it can be stored on a computer-readable storage medium. This computer software product, stored on a storage medium, includes instructions for causing a computer device (which may be a personal computer, server, or network device, etc.) to perform all or part of the steps of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0103] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A low voltage substation topology identification method using a graph diffusion model, characterized in that: include: Step 1: Obtain historical data from several different distribution substations, normalize the historical data, divide the historical data into sliding windows, establish a time series modeling module, and extract the time series features of each sliding window through the time series modeling module; Step 2: The extracted time series features are used as the initial features of the nodes and input into the graph neural network. The graph neural network is trained based on the initial features to obtain a topology recognition model. Step 3: Get the target data of several different distribution stations, process the target data as in step 1, and get the target time series features. Step 4: Input the target time series features into the topology recognition model, implicitly identify the node pairs with synchronous fluctuations through the topology recognition model, output the node features that integrate the neighbor information, predict the connection probability between each node based on the node features, and output the new substation topology structure.

2. The method for identifying low-voltage substation topology using a graph diffusion model according to claim 1, characterized in that: Acquiring the historical data or target data includes acquiring substation infrastructure data and acquiring electrical quantity data; The acquisition of the substation infrastructure data includes acquiring the actual topological relationship diagram of each substation, and converting the actual topological relationship diagram into a mathematical diagram. The representation of the mathematical diagram is G=(V,E), where G is a relationship diagram, V is a node combination, E is an edge set, and E=0 or 1. When E is equal to 0, there is no power connection between the two nodes, and when E is equal to 1, there is a power connection between the two nodes.

3. The method for identifying low-voltage substation topology using a graph diffusion model according to claim 2, characterized in that: The normalization process includes: Where, is the normalized data, For the input data, is the minimum value in the data set, is the maximum value in the data set.

4. The method for identifying low-voltage substation topology using a graph diffusion model according to claim 3, characterized in that: The performing sliding window division on the historical data includes: Set the time series for the normalized data and set the window parameters, which include window size, sliding step size and prediction range; Determine the data point of the starting index of the first window and the starting index of the last window, and set the number of valid window samples. Through a loop, the variable starts from the window of the starting index and is incremented by the step size. The number of divisions is increased until it reaches the maximum number of valid window samples, and then the division is ended; The divided data is generated into data groups and then output.

5. The method for identifying low-voltage substation topology using a graph diffusion model according to claim 4, characterized in that: The points for determining the starting index of the last window include: in, The sequence value of the point with the starting index of the last window, The sequence value of the data point with the starting index of the first window, is the window size, is the prediction range, is the offset, The total number of data.

6. The method for identifying low-voltage substation topology using a graph diffusion model according to claim 5, characterized in that: Determining the number of valid window samples includes: Where, is the number of valid window samples.

7. The method for identifying low-voltage substation topology using a graph diffusion model according to claim 5, characterized in that: The training of the graph neural network based on the initial features includes: Get the initial features, add noise multiple times in a cycle, generate noisy topology samples, set the number of iterations, and when the number of iterations is reached, store the current noisy topology samples and noise information as training targets; The training objective provides training data about noisy topological samples and noise information, and the training data is input into the graph neural network for training while calculating the loss. The parameters of the graph neural network are updated by the calculated loss, and the topology recognition model is output based on the updated parameters of the graph neural network and the training data.

8. A low voltage substation topology identification system using a graph diffusion model, characterized in that: include: a model building module configured to obtain historical data of several different distribution substations, normalize the historical data, divide the historical data into sliding windows, establish a time series modeling module, and extract time series features of each sliding window through the time series modeling module; The extracted time series features are used as the initial features of the nodes and input into the graph neural network. The graph neural network is trained based on the initial features to obtain a topology recognition model. The recognition module is configured to obtain the target data of several different distribution station areas, and obtain the target time series characteristics after processing the target data as in step 1. The target time series features are input into the topology recognition model, and the node pairs with fluctuation synchronization are implicitly identified through the topology recognition model. The node features that integrate neighbor information are output, and the connection probability between each node is predicted based on the node features, and the new substation topology structure is output.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method for identifying low-voltage substation topology using a graph diffusion model according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for identifying low-voltage substation topology using a graph diffusion model according to any one of claims 1 to 7 is implemented.

Images