A fast identification method of power topology based on deep graph matching

Abstract

The present invention relates to the technical field of power topology identification and discloses a method for rapid power topology identification based on deep graph matching, comprising the following steps: Step S1, multi-source heterogeneous data fusion and adaptive noise reduction, using an adaptive threshold noise reduction algorithm; Step S2, dynamic spatiotemporal graph modeling; Step S3, attention-driven graph matching, using a dual-channel graph neural network; and Step S4, incremental topology library update. For situations where real-time data from low-voltage distribution networks generates significant noise due to equipment aging and acquisition errors, an adaptive threshold noise reduction algorithm is used to accurately filter out high-frequency noise components based on a threshold, and combined with the spatiotemporal kriging interpolation method to intelligently repair missing data, ensuring data purity and integrity, and providing a data foundation for subsequent topology identification.

Description

Technical Field

[0001] The present invention relates to the technical field of power topology recognition, and in particular to a method for rapid power topology recognition based on depth graph matching. Background Art

[0002] The power topology rapid identification method based on deep graph matching is mainly used to quickly and accurately identify the topology structure in the power system. When a fault occurs in the power system, rapid and accurate identification of the topology structure helps to quickly locate the fault location. By comparing it with the normal topology structure, abnormal points can be discovered in time, thereby guiding maintenance personnel to quickly reach the fault site for repairs, reducing the power outage time and scope, and improving power supply reliability.

[0003] After searching, the Chinese patent number CN120067700A discloses a method for verifying the distribution of topological similarity data of a power system, which includes the following steps: collecting the operating data of the power system; calculating the similarity between topological states; screening out multiple historical topological states that are closest to the current topological state; performing fast approximate similarity query to quickly match similar topological states; verifying the distribution of power system operating data under different topological states, analyzing the impact of topological changes on system operating data, identifying abnormal distribution or inconsistent operating states, and introducing a streaming data processing framework. The present invention utilizes a combination of different similarity measurement methods.

[0004] However, in the above technical solution, there is significant noise in the real-time data of the low-voltage distribution network, such as equipment aging and collection errors, and the multi-source data formats are heterogeneous, and the network nodes and connection relationships are growing exponentially. The traditional algorithm has high computational complexity and it is difficult to meet the real-time response requirements in fault scenarios. Although the above patent introduces approximate similarity query, the parallel processing capability of large-scale topologies is limited, resulting in network deployment delays. In addition, the power network load fluctuates and the topology switches frequently, and the existing model relies on historical data training. It has insufficient generalization ability for sudden structural changes and requires repeated training to adapt to new scenarios. Based on this, the present invention designs a method for rapid identification of power topology based on deep graph matching to solve the above problems. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for rapid identification of power topology based on deep graph matching, which solves the problem of real-time response in fault scenarios in the background technology.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] A method for rapid identification of power topology based on deep graph matching includes the following steps:

[0008] Step S1: Multi-source heterogeneous data fusion and adaptive noise reduction, synchronously collecting high-precision real-time data streams of voltage amplitude U, current phase angle θ, active power P, and reactive power Q of each node in the power system, with a sampling frequency of not less than 1kHz;

[0009] Adopting an adaptive threshold noise reduction algorithm, based on the threshold λ = 0.6745·median(|Dhighfrequency|), we can accurately filter out high-frequency noise components caused by equipment aging to ensure data purity.

[0010] Step S2: Dynamic spatiotemporal graph modeling. Using the electrical connection relationships between power equipment as edges, a binary adjacency matrix A0 is initialized and constructed. A matrix element A0(i,j) of 1 indicates that node i is electrically connected to node j, and 0 indicates that it is not connected. A three-layer spatiotemporal graph convolutional network is constructed, with a reasonable configuration of the number of convolution kernels K=3 and the hidden layer dimension dh=128. The network is used to deeply fuse the power system dynamic data within the sliding time window Tw=15 minutes to achieve real-time dynamic update of the adjacency matrix.

[0011] Step S3: attention-driven graph matching, using a dual-channel graph neural network to perform deep feature extraction on the currently constructed topology graph Gc and the historical topology graph library {Gh};

[0012] Step S4, incremental topology library update, injects the newly identified topology structure Gc into the historical topology library based on the ball tree structure index in an orderly manner, ensuring that the neighbor search complexity is maintained at O(logM) to achieve efficient data organization and retrieval.

[0013] Preferably, in step S1, the spatiotemporal Kriging interpolation method is used, the spatial correlation scale L=1.2km is set, and the missing data is intelligently repaired in combination with the spatiotemporal correlation to ensure data integrity and time series continuity; the preprocessed multi-dimensional data is uniformly encoded into a 128-dimensional feature vector Fv, and a node feature matrix X of a scale of N×128 is constructed, where N represents the total number of nodes in the power system.

[0014] Preferably, in step S2, in order to enhance the training effect of the spatiotemporal graph convolutional network, a transfer learning strategy is introduced, and the current network is initialized using model parameters pre-trained on a large-scale general power topology dataset to reduce the number of samples required for training and speed up convergence.

[0015] Preferably, in step S3, in order to improve computing efficiency, a block parallel computing strategy is adopted to evenly divide the complex topology graph into multiple relatively independent subgraphs. With the help of zero-copy RDMA communication technology, and the communication delay is strictly controlled to ≤2ms, asynchronous parallel processing is realized in a GPU cluster environment to accelerate the graph matching process; through scientific calculation of the node similarity matrix S, the abnormal power nodes are accurately marked according to the preset threshold max(Si) less than 0.85, and the geographic fault coordinates (xi,yi) of the abnormal points are further output.

[0016] Preferably, in step S4, whenever the number of newly added topologies reaches Nnew=100, the online knowledge distillation mechanism is automatically triggered to encourage the student model to quickly learn new features while retaining key knowledge and continuously optimize model parameters.

[0017] Preferably, in the noise reduction process of step S1, an improved sliding window CUSUM algorithm is further introduced to monitor current mutation events in real time, and the event extraction sensitivity coefficient β=0.35 is accurately set; the power data is combined with the isolation forest algorithm to detect and filter anomalies, and the anomaly score threshold τ=0.65 is strictly limited.

[0018] Preferably, in step S3, the two metrics of cosine similarity and Euclidean distance are combined to construct a comprehensive similarity metric function to match the current topology map with the maps in the historical topology map library; the dropout technology is used to randomly discard some neurons with a certain probability in the hidden layer of the graph neural network, which prevents overfitting while enabling the model to perform graph matching more stably when facing different scenarios and data changes.

[0019] Preferably, in step S4, the historical topology library is compressed and streamlined regularly, and the usage frequency of each topology structure in the past T days is counted. When the usage frequency is lower than ω=0.01, it is removed from the topology library and stored in the cold backup area to optimize the storage space and maintain the efficiency and timeliness of the topology library.

[0020] Preferably, in step S4, the topology database needs to be backed up regularly, with the backup cycle set to once a week, and the backup data is stored in a remote distributed storage system.

[0021] Preferably, during the entire power topology identification process, a real-time monitoring and logging system is established to monitor and record each link such as data collection, data processing, model training, graph matching, and topology update in real time; the system can promptly detect abnormal conditions during system operation and record abnormal information in the log file.

[0022] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0023] 1. For the situation where the real-time data of the low-voltage distribution network is significantly noisy due to equipment aging and acquisition errors, the present invention adopts an adaptive threshold noise reduction algorithm to accurately filter out high-frequency noise components based on the threshold, and combines the spatiotemporal Kriging interpolation method to intelligently repair missing data, ensuring data purity and integrity, and providing a data foundation for subsequent topology identification.

[0024] 2. The present invention builds a spatiotemporal graph convolutional network, deeply integrates the dynamic data of the power system within the sliding time window, dynamically updates the adjacency matrix in real time, and can capture the dynamic changes of the power system topology structure in real time; uses a dual-channel graph neural network to perform deep feature extraction on the current topology map and the historical topology map library respectively, and adopts a block parallel computing strategy to accelerate the graph matching process, meet the real-time response requirements in fault scenarios, and effectively avoid network delay problems.

[0025] 3. The present invention automatically triggers the online knowledge distillation mechanism whenever the number of newly added topologies reaches a certain value, prompting the student model to quickly learn new features while retaining key knowledge and continuously optimize model parameters. This solves the problem that existing models rely on historical data training, have insufficient generalization capabilities for sudden structural changes, and require repeated training. The historical topology library can be dynamically updated and improved, improving the model's adaptability and accurate description capabilities to changes in power system topology. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a diagram of the overall system workflow and monitoring architecture of the present invention;

[0027] Figure 2 This is a decomposition diagram of the data noise reduction and restoration technology of the present invention;

[0028] Figure 3 It is a workflow diagram of the present invention;

[0029] Figure 4 It is a dual-channel workflow diagram of the present invention. DETAILED DESCRIPTION

[0030] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0031] Example 1;

[0032] See also Figure 1-Figure 4 In an embodiment of the present invention, a method for rapid identification of power topology based on depth map matching includes the following steps:

[0033] Step S1: Multi-source heterogeneous data fusion and adaptive noise reduction, synchronously collecting high-precision real-time data streams of voltage amplitude U, current phase angle θ, active power P, and reactive power Q of each node in the power system, with a sampling frequency of not less than 1kHz;

[0034] Adopting an adaptive threshold noise reduction algorithm, based on the threshold λ = 0.6745·median(|Dhighfrequency|), we can accurately filter out high-frequency noise components caused by equipment aging to ensure data purity.

[0035] Step S2: Dynamic spatiotemporal graph modeling. Using the electrical connection relationships between power equipment as edges, a binary adjacency matrix A0 is initialized and constructed. A matrix element A0(i,j) of 1 indicates that node i is electrically connected to node j, and 0 indicates that it is not connected. A three-layer spatiotemporal graph convolutional network is constructed, with a reasonable configuration of the number of convolution kernels K=3 and the hidden layer dimension dh=128. The network is used to deeply fuse the power system dynamic data within the sliding time window Tw=15 minutes to achieve real-time dynamic update of the adjacency matrix.

[0036] Step S3: attention-driven graph matching, using a dual-channel graph neural network to perform deep feature extraction on the currently constructed topology graph Gc and the historical topology graph library {Gh};

[0037] Step S4, incremental topology library update, injects the newly identified topology structure Gc into the historical topology library based on the ball tree structure index in an orderly manner, ensuring that the neighbor search complexity is maintained at O(logM) to achieve efficient data organization and retrieval.

[0038] In step S1, the spatiotemporal kriging interpolation method is used, the spatial correlation scale L is set to 1.2 km, and the missing data is intelligently repaired in combination with the spatiotemporal correlation to ensure data integrity and time series continuity; the preprocessed multi-dimensional data is uniformly encoded into a 128-dimensional feature vector Fv, and a node feature matrix X of a scale of N×128 is constructed, where N represents the total number of nodes in the power system.

[0039] In step S2, in order to enhance the training effect of the spatiotemporal graph convolutional network, a transfer learning strategy is introduced to initialize the current network using the model parameters pre-trained on a large-scale general power topology dataset, in order to reduce the number of samples required for training and accelerate the convergence speed.

[0040] In step S3, to improve computing efficiency, a block parallel computing strategy is adopted to evenly divide the complex topology graph into multiple relatively independent subgraphs. With the help of zero-copy RDMA communication technology, and the communication delay is strictly controlled to ≤2ms, asynchronous parallel processing is realized in a GPU cluster environment to accelerate the graph matching process; through scientific calculation of the node similarity matrix S, the abnormal power nodes are accurately marked according to the preset threshold max(Si) less than 0.85, and the geographic fault coordinates (xi,yi) of the abnormal points are further output.

[0041] In step S4, whenever the number of new topologies reaches Nnew=100, the online knowledge distillation mechanism is automatically triggered to encourage the student model to quickly learn new features while retaining key knowledge and continuously optimize model parameters.

[0042] The working principle of this embodiment of the present invention is as follows: In step S1, focusing on the working principles of multi-source heterogeneous data fusion and adaptive noise reduction, when the power system is operating, each node generates four key data types: voltage amplitude U, current phase angle θ, active power P, and reactive power Q. These data are synchronously collected in the form of high-precision real-time data streams, with a sampling frequency strictly set to no less than 1kHz, aiming to accurately capture the ever-changing dynamic characteristics of the power system. Given that factors such as equipment aging can easily introduce high-frequency noise, this invention utilizes an adaptive threshold noise reduction algorithm. Its core principle is to use a threshold λ = 0.6745·median(|Dhigh|) to accurately identify and filter out high-frequency noise components. Here, D high frequency represents the signal after filtering through a specific high-frequency band, and median represents the median operation, which achieves accurate quantification and effective filtering of noise, thereby ensuring data purity. At the same time, to address possible missing data during the data collection process, the spatiotemporal kriging interpolation method is introduced, and the spatial correlation scale is set to L = 1.2km. Combined with the spatiotemporal correlation, the missing data is intelligently repaired to ensure data integrity and time series continuity, and to restore the true operating status of the power system to the greatest extent possible.

[0043] Entering step S2, the core lies in the working principle of dynamic spatiotemporal graph modeling. First, a binary adjacency matrix A0 is constructed based on the electrical connection relationship between power equipment. The matrix element A0(i,j) is 1 or 0, which intuitively indicates whether node i and node j are electrically connected. This matrix preliminarily constructs the topological framework of the power system. On this basis, a three-layer spatiotemporal graph convolutional network is built, in which the number of convolution kernels K is determined to be 3, and the hidden layer dimension is set to dh=128. This configuration aims to achieve deep fusion and efficient processing of dynamic data of the power system. By setting the sliding time window Tw=15min, the dynamic data within a period of time is included in the analysis scope. The powerful feature extraction capability of the spatiotemporal graph convolutional network is used to capture the dynamic changes of the topological structure of the power system in real time, and the adjacency matrix is ​​dynamically updated accordingly, so that the topological model always fits the actual operating conditions.

[0044] In step S3, the key principle lies in attention-driven graph matching. A dual-channel graph neural network architecture is employed, with each channel performing deep feature extraction on both the currently constructed topology graph Gc and the historical topology graph library {Gh}. This dual-channel design enables the network to concurrently and focus on mining features from two different sources of topology graphs, fully extracting deep-level feature information about nodes, edges, and the overall topological structure, providing accurate feature representations for subsequent matching operations. To improve computational efficiency, a block-based parallel computing strategy is employed, evenly partitioning the complex power topology graph into multiple relatively independent subgraphs based on specific rules. In a GPU cluster environment, zero-copy RDMA (Remote Direct Memory Access) communication technology is used to achieve efficient data transmission and exchange between subgraphs, while strictly controlling communication latency to ≤2ms. This ensures the efficiency of asynchronous parallel processing and significantly accelerates the graph matching process. In addition, by scientifically calculating the node similarity matrix S and based on the judgment rule that the preset threshold max(Si) is less than 0.85, abnormal power nodes can be accurately marked. This not only quickly locates abnormal parts in the power system, but also further outputs the geographic fault coordinates (xi, yi) of the abnormal points, providing key positioning information for fault inspection and system maintenance.

[0045] As of step S4, its working principle revolves around the incremental topology library update. Whenever the number of newly identified topological structures Gc reaches a certain number, that is, the number of newly added topologies reaches Nnew=100, the online knowledge distillation mechanism will be automatically triggered. The core principle of this mechanism is to enable the student model to quickly learn new features on the basis of retaining the key knowledge of the existing model. By constructing an effective inheritance and fusion path between new and old knowledge, and continuously optimizing the model parameters, the historical topology library can be dynamically updated and continuously improved, always maintaining efficient adaptability and accurate description capabilities to changes in the power system topology. At the same time, the data organization method based on the ball tree structure index ensures that the complexity of the neighbor search is maintained at O(logM), achieving efficient data organization and retrieval.

[0046] Example 2;

[0047] See also Figure 1-Figure 4 In the embodiment of the present invention, in the noise reduction process of step S1, an improved sliding window CUSUM algorithm is further introduced to monitor current mutation events in real time, and the event extraction sensitivity coefficient β is accurately set to 0.35; the isolation forest algorithm is combined to detect and filter anomalies in power data, and the anomaly score threshold τ is strictly limited to 0.65.

[0048] In step S3, a comprehensive similarity measurement function is constructed by combining the cosine similarity and Euclidean distance measurements to match the current topology map with the maps in the historical topology map library. The dropout technology is used to randomly discard some neurons in the hidden layer of the graph neural network with a certain probability, which prevents overfitting while enabling the model to perform graph matching more stably when facing different scenarios and data changes.

[0049] In step S4, the historical topology library is compressed and streamlined regularly, and the usage frequency of each topology structure in the past T days is counted. When the usage frequency is lower than ω=0.01, it is removed from the topology library and stored in the cold backup area to optimize storage space and maintain the efficiency and timeliness of the topology library.

[0050] In step S4, the topology database needs to be backed up regularly. The backup cycle is set to once a week, and the backup data is stored in a distributed storage system at a different location.

[0051] During the entire power topology identification process, a real-time monitoring and logging system is established to monitor and record each link, including data collection, data processing, model training, graph matching, and topology update. The system can promptly detect abnormal situations during system operation and record abnormal information in log files.

[0052] The working principle of this embodiment of the present invention is as follows: During the noise reduction process in step S1, an improved sliding window CUSUM algorithm is introduced. This algorithm sets the event extraction sensitivity coefficient β to 0.35 to monitor current mutation events in the power system in real time. Its core principle is to perform a cumulative sum test on the current data within a sliding window based on statistical process control theory. When a sudden change occurs in the current data, the algorithm can quickly detect when the cumulative sum exceeds the set threshold, thereby accurately identifying the current mutation event.

[0053] The Isolation Forest algorithm is used to detect and filter anomalies in power data. This algorithm is combined with the Isolation Forest algorithm to detect and filter anomalies in power data, with a strict anomaly score threshold of τ = 0.65. Based on the principles of random forests, the Isolation Forest algorithm constructs multiple isolation trees to measure the degree of anomaly in data points. For each data point in the power data, the algorithm calculates its anomaly score. If the score exceeds the threshold of τ = 0.65, the data is considered anomaly and filtered.

[0054] A comprehensive similarity metric function is constructed by combining cosine similarity and Euclidean distance to match the current topology map with the maps in the historical topology map library. Specifically, cosine similarity is used to measure the directional similarity between two topology maps in the feature vector space. The ratio of their dot product and module length is calculated to reflect the angular relationship between the topology maps. Euclidean distance measures the straight-line distance between two topology maps in the feature vector space, reflecting their numerical difference.

[0055] The historical topology library is compressed and streamlined regularly. The usage frequency of each topology structure in the past T days is counted. When the usage frequency falls below ω = 0.01, it is removed from the topology library and stored in the cold backup area. This process is based on an analysis of the frequency of topology usage. Infrequently used topologies have low reference value in current power system operations. Removing them optimizes storage space. These topologies are also stored in the cold backup area for emergency use.

[0056] Throughout the entire power topology identification process, a real-time monitoring and logging system is established. This system monitors and records every step of the process, including data collection, data processing, model training, graph matching, and topology updates. Monitoring points are set up at key stages to collect information such as system operating status, data flow, model training progress, and performance indicators, and record this information in log files. If an anomaly occurs during system operation, the real-time monitoring and logging system promptly detects it, issues an alert, and records the anomaly in detail in the log files.

[0057] Example 3;

[0058] See also Figure 1-Figure 4 This embodiment provides a method for rapid identification of power topology based on depth map matching, and the specific steps are as follows:

[0059] An adaptive threshold denoising algorithm was used, using a threshold of λ = 0.6745 × median(|Dhigh|), where Dhigh represents the signal filtered through a specific high-frequency band and median represents the median value. This algorithm accurately removes high-frequency noise components caused by equipment aging, ensuring data purity. A spatiotemporal kriging interpolation method was introduced, with a spatial correlation scale of L = 1.2 km. Missing data was intelligently repaired based on spatiotemporal correlations to ensure data integrity and temporal continuity. The preprocessed multidimensional data was uniformly encoded into a 128-dimensional feature vector Fv, and a node feature matrix X of size N × 128 was constructed, where N represents the total number of nodes in the power system.

[0060] At the same time, an improved sliding window CUSUM algorithm is further introduced into the noise reduction process to monitor current mutation events in real time, and the event extraction sensitivity coefficient β is accurately set to 0.35. The isolation forest algorithm is combined to detect and filter anomalies in power data, and the anomaly score threshold τ is strictly limited to 0.65.

[0061] Taking the electrical connection relationship between power equipment as the edge, the binary adjacency matrix A0 is initialized and constructed. The matrix element A0 (i, j) is 1, which means that node i and node j are electrically connected, and 0 means they are not connected. The topological framework of the power system is preliminarily constructed.

[0062] Using a block-based parallel computing strategy, the complex power topology graph is evenly divided into eight relatively independent subgraphs. Using zero-copy RDMA communication technology, with communication latency strictly controlled to ≤2ms, asynchronous parallel processing is implemented on a GPU cluster environment, accelerating the graph matching process. By scientifically calculating the node similarity matrix S, anomalous power nodes are accurately marked based on a preset threshold max(Si) less than 0.85, and the geographic fault coordinates (xi, yi) of the anomaly points are further output.

[0063] Whenever a newly identified topological structure Gc is injected into the historical topological library based on the ball tree structure index in an orderly manner and the number of newly added topologies Nnew=100 is reached, the online knowledge distillation mechanism is automatically triggered to enable the student model to quickly learn new features while retaining key knowledge and continuously optimize the model parameters.

[0064] The historical topology library is compressed and streamlined regularly, and the usage frequency of each topology structure in the past T = 30 days is counted. When the usage frequency is lower than ω = 0.01, it is removed from the topology library and stored in the cold backup area to optimize storage space and maintain the efficiency and timeliness of the topology library. At the same time, the topology library is backed up regularly, with the backup cycle set to once a week, and the backup data is stored in a distributed storage system in a different location.

[0065] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and alterations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for rapid identification of power topology based on deep graph matching, characterized in that: The following steps are involved: Step S1: Multi-source heterogeneous data fusion and adaptive noise reduction, synchronously collecting high-precision real-time data streams of voltage amplitude U, current phase angle θ, active power P, and reactive power Q of each node in the power system, with a sampling frequency of not less than 1kHz; Adopting an adaptive threshold noise reduction algorithm, based on the threshold λ = 0.6745·median(|Dhighfrequency|), we can accurately filter out high-frequency noise components caused by equipment aging to ensure data purity. Step S2: Dynamic spatiotemporal graph modeling. Using the electrical connection relationships between power equipment as edges, a binary adjacency matrix A0 is initialized and constructed. A matrix element A0(i,j) of 1 indicates that node i is electrically connected to node j, and 0 indicates that it is not connected. A three-layer spatiotemporal graph convolutional network is constructed, with a reasonable configuration of the number of convolution kernels K=3 and the hidden layer dimension dh=128. The network is used to deeply fuse the power system dynamic data within the sliding time window Tw=15 minutes to achieve real-time dynamic update of the adjacency matrix. Step S3: attention-driven graph matching, using a dual-channel graph neural network to perform deep feature extraction on the currently constructed topology graph Gc and the historical topology graph library {Gh}; Step S4, incremental topology library update, the newly identified topology structure Gc is sequentially injected into the historical topology library based on the ball tree structure index, ensuring that the neighbor search complexity is maintained at O(logM), achieving efficient data organization and retrieval; In step S3, to improve computational efficiency, a block-parallel computing strategy is adopted to evenly divide the complex topology graph into multiple relatively independent subgraphs. With the help of zero-copy RDMA communication technology, and with communication latency strictly controlled to ≤2ms, asynchronous parallel processing is implemented in a GPU cluster environment to accelerate the graph matching process. By scientifically calculating the node similarity matrix S, abnormal power nodes are accurately marked based on a preset threshold max(Si) less than 0.85, and the geographic fault coordinates (xi, yi) of the abnormal points are further output. In step S3, a comprehensive similarity measurement function is constructed by combining the cosine similarity and Euclidean distance measurement methods to match the current topology map with the maps in the historical topology map library; the dropout technology is used to randomly discard some neurons in the hidden layer of the graph neural network with a certain probability, which prevents overfitting while enabling the model to perform graph matching more stably when facing different scenarios and data changes.

2. The method for rapid identification of power topology based on depth map matching according to claim 1, characterized in that: In step S1, the spatiotemporal kriging interpolation method is used, the spatial correlation scale L is set to 1.2 km, and the missing data is intelligently repaired in combination with the spatiotemporal correlation to ensure data integrity and time series continuity; the preprocessed multi-dimensional data is uniformly encoded into a 128-dimensional feature vector Fv, and a node feature matrix X of a scale of N×128 is constructed, where N represents the total number of nodes in the power system.

3. The method for rapid identification of power topology based on depth map matching according to claim 1, characterized in that: In step S2, in order to enhance the training effect of the spatiotemporal graph convolutional network, a transfer learning strategy is introduced to initialize the current network using the model parameters pre-trained on a large-scale general power topology dataset, in order to reduce the number of samples required for training and accelerate the convergence speed.

4. The method for rapid identification of power topology based on depth map matching according to claim 1, characterized in that: In step S4, whenever the number of new topologies reaches Nnew=100, the online knowledge distillation mechanism is automatically triggered to encourage the student model to quickly learn new features while retaining key knowledge and continuously optimize model parameters.

5. The method for rapid identification of power topology based on depth map matching according to claim 1, characterized in that: In the noise reduction process of step S1, an improved sliding window CUSUM algorithm is further introduced to monitor current mutation events in real time, and the event extraction sensitivity coefficient β is accurately set to 0.

35. The isolation forest algorithm is combined to detect and filter anomalies in power data, and the anomaly score threshold τ is strictly limited to 0.

65.

6. The method for rapid identification of power topology based on depth map matching according to claim 1, characterized in that: In step S4, the historical topology library is compressed and streamlined regularly, and the usage frequency of each topology structure in the past T days is counted. When the usage frequency is lower than ω=0.01, it is removed from the topology library and stored in the cold backup area to optimize storage space and maintain the efficiency and timeliness of the topology library.

7. The method for rapid identification of power topology based on depth map matching according to claim 1, characterized in that: In step S4, the topology database needs to be backed up regularly. The backup cycle is set to once a week, and the backup data is stored in a distributed storage system at a different location.

8. The method for rapid identification of power topology based on depth map matching according to claim 1, characterized in that: During the entire power topology identification process, a real-time monitoring and logging system is established to monitor and record each link of data collection, data processing, model training, graph matching, and topology update in real time; the system can promptly detect abnormal situations during system operation and record abnormal information in log files.

Images