Quick recovery method and system based on regional distribution network fault reconstruction

Abstract

The invention discloses a quick recovery method and system based on regional power distribution network fault reconstruction, and belongs to the technical field of digital simulation and power distribution automation of a power system. The method comprises the following steps: collecting system operation data of a power distribution network after a fault in real time, constructing a topological graph of the power distribution network, extracting an island sub-graph taking a link switch as a source-target node, further constructing a sub-graph feature information matrix, inputting a pre-trained sub-graph neural network model, obtaining a connectivity prediction result of each potential connection link, and predicting the connectivity of each potential connection link. According to the method, a dynamic reconstruction control instruction is generated and executed so as to realize rapid recovery of the power distribution network, and the system comprises a sensing acquisition module, a communication module, an edge calculation module and a control execution module. According to the method, physical constraints and local graph features are fused, the calculation complexity and the recovery time are remarkably reduced while the difference of all the sub-graphs is guaranteed, and rapid and reliable recovery after the power distribution network breaks down is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system digital simulation, in particular to a fast recovery method and system based on regional distribution network fault reconstruction. BACKGROUND

[0002] In recent years, extreme weather events such as hurricanes, heavy rains, and droughts have occurred frequently, significantly increasing the risk of large-scale cascading failures in power grids. Such large-scale blackouts caused by interdependent fault chains not only cause serious economic losses, but also pose a major threat to public safety. It is worth noting that under the background of increasing penetration of renewable energy and increasing interconnection of power grids, system uncertainty has intensified, further amplifying the risk of cascading failures. Under this dual challenge, dynamic reconstruction technology of distribution network has become a key means to improve the resilience of power grid, as it can optimize the topology structure by adjusting the switch state in real time.

[0003] Distribution network reconstruction is a complex high-dimensional combinatorial optimization problem, the core of which is to optimize the network topology by adjusting the switch state to ensure the efficiency and reliability of power transmission. Traditional solutions mainly rely on heuristic algorithms and optimization algorithms, but these model-based solving methods have two limitations: (1) As the network structure becomes more complex and controllable resources become more diverse, the reconstruction problem grows exponentially, making it difficult to meet the rapid solving needs of practical applications; (2) Existing researches mostly use assumptions and model simplifications, making it difficult to achieve maximum connectivity of power sources and loads under multi-source coordination, and different simplification strategies and mathematical processing methods have a significant impact on the solution results and efficiency. In contrast, neural network models can improve the solution speed by 20-100 times in solving network reconstruction timing decisions by avoiding explicit modeling of complex physical constraints in power systems. In recent years, deep neural networks, convolutional neural networks, and graph neural networks have been widely used in power system operation control links such as fault detection, optimal power flow, and state estimation. In particular, graph neural networks effectively utilize the inherent network topology characteristics of power systems, thereby ensuring the integrity and accuracy of the topology structure. This has achieved certain research in distribution network fault reconstruction.

[0004] However, the graph neural network tends to flatten the power network topology representation, and can only capture the overall pattern, making it difficult to accurately identify and distinguish local key structural differences. Especially in large-scale distribution networks with thousands of nodes, relying only on single features of nodes or edges for whole graph reasoning makes it difficult to accurately analyze differentiated topology information. In addition, when the system is split into multiple islands due to faults, the same node may assume different functional roles in different islands. The homogeneous node embedding of traditional graph neural networks cannot capture this dynamic semantic information with topology, which may lead to connection prediction errors. In addition, in actual engineering applications, existing fast recovery systems still mostly use centralized decision-making architecture, relying on global optimization and remote control of the main station system, which has the problem of inflexible algorithm deployment. At the same time, traditional systems are difficult to effectively integrate with neural network models, especially in terms of processing heterogeneous representations and real-time inference of graph neural networks, which lack adaptive hardware and software support.

[0005] In view of the above problems, there is an urgent need for a fast recovery method and system based on regional power distribution network fault reconstruction to solve the above problems existing in traditional methods. SUMMARY

[0006] The purpose of the present application is to provide a fast recovery method and system based on regional power distribution network fault reconstruction, which significantly reduces the computational load and time cost by fusing physical constraints and key local features, and ensures the difference of subgraphs corresponding to different links.

[0007] To achieve the above purpose, the technical scheme adopted by the present application is as follows: A fast recovery method based on regional power distribution network fault reconstruction, comprising: Step 1: Real-time acquisition of system operation data of the power distribution network after the fault; Step 2: Based on the system operation data, a power distribution network topology graph is constructed, and a link switch is extracted as a source-target node for the topology structure of the power distribution network to obtain an island subgraph; Step 3: Feature processing is performed on the island subgraph to construct a subgraph feature information matrix; Step 4: The subgraph feature information matrix is input into a pre-trained subgraph neural network model to obtain the connectivity prediction results of each potential connection link; Step 5: Based on the connectivity prediction results, a dynamic reconstruction control instruction is generated, and the control instruction is executed to realize the recovery of the power distribution network.

[0008] Further, in step 2, the topology structure of the power distribution network is extracted as a source-target node with a link switch to obtain an island subgraph, specifically: The two endpoints of the link switch are taken as the source node and the target node; Determine the k-hop neighbor range of the source node and the target node, obtain all neighbor nodes in the range, and collectively constitute an island subgraph.

[0009] Further, the determination method of the hop number k is: (1) In the formula, DG is the DG regulation capability factor, R is the line impedance mean value, p is the island topology density parameter, is the k-hop subgraph node set, and Ɛ is the fluctuation sensitivity parameter, is the island load proportion weight. is the island internal node voltage threshold, is the load fluctuation rate, is the number of island internal buses.

[0010] Further, in step 3, the island subgraph is processed for features, and a subgraph feature information matrix is constructed, specifically: Calculate the feature distance of each neighbor node in the k-hop range with respect to the source node and the target node; Based on the feature distance, each neighbor node in the k-hop range is assigned a discrete label; According to the level determined by the discrete label, the feature vectors of the neighbor nodes are weighted and aggregated, and the aggregation result is spliced with the features and label embedding vectors of the source node or the target node itself to form a subgraph feature information matrix.

[0011] Further, the feature distance of each neighbor node in the k-hop range with respect to the source node and the target node is calculated, specifically: (2) In the formula, is the feature distance, V, p are the voltage and distributed power penetration rate of the node respectively, is the weight coefficient. The subscript u, , are the source node, the target node, and any neighbor node in the target link respectively, is the reference voltage, is the maximum penetration rate of the distributed power in the island.

[0012] Further, based on the feature distance, each neighbor node in the k-hop range is assigned a discrete label, specifically: (3) In the formula, is the node assigned a discrete label, is the system rated parameter, which is used to fix the label hierarchical granularity, The reference penetration rate of the distributed power supply in the island.

[0013] Further, according to the level determined by the discrete label, the feature vectors of the neighbor nodes are weighted and aggregated, and the aggregation result is spliced with the features of the source node or the target node itself and the label embedding vector to form a subgraph feature information matrix, specifically: The feature vectors of the remaining neighbor nodes in the island subgraph are accumulated into the target node, and the feature vectors of the neighbor nodes are weighted and aggregated according to the level determined by the discrete label, which is: (4) In the formula, is the feature vector of the target node u after aggregation, is the feature vector of the neighbor node ω, , is the active power, reactive power and voltage of the neighbor node . is the maximum label level number, is the weight function, , is the decay coefficient, which is 0.5; is the k-hop neighbor set of the source node u; Finally, the subgraph feature information matrix is obtained, which is: (5) (6) In the formula, is the label embedding vector, d e is the embedding dimension, which is 16 or 32; is one-hot encoding; is the embedding matrix, which is a trainable parameter matrix for mapping discrete labels to low-dimensional continuous variables. is the feature vector of node u.

[0014] Further, in step 4, the subgraph feature information matrix is input into the pre-trained subgraph neural network model to obtain the connectivity prediction result of each potential connection link, specifically: The subgraph feature information matrix is power encoded and node feature encoded, and the noise influence caused by the difference in feature scale is suppressed through LayerNorm; Based on the graph convolutional neural network model, the encoded subgraph feature information matrix is updated for feature information, the updating process is connected through residual connection, and the batch normalization processing is performed on the residual connection structure; The updated result is calculated for connectivity prediction to obtain the connectivity prediction result.

[0015] The application further provides a fast recovery system based on regional power distribution network fault reconstruction, which is used for realizing the fast recovery method based on regional power distribution network fault reconstruction. The sensing and collecting module is used for collecting system operation data of the power distribution network after the fault in real time. The communication module is used for transmitting the system operation data and control instructions. The edge computing module comprises a memory and a processor, the memory stores a computer program, and the processor realizes the steps in claims 1-8 when executing the program. The control execution module is used for executing switch operation based on the connectivity prediction result.

[0016] Further, the sensing and collecting module comprises at least one of a phasor measurement unit, a feeder terminal unit, a distribution terminal unit and a smart meter arranged at a power distribution network node. The edge computing module is specifically an edge computing gateway arranged at a power distribution substation or a key node, or a station server arranged at a power distribution automation master station or a substation; the edge computing gateway or the station server is internally provided with an artificial intelligence acceleration chip. The control execution module comprises remotely controllable automatic sectionalizing switches, tie switches and distributed power source grid-connected switches. In summary, the application has at least one of the following beneficial technical effects: 1. Significantly improving the reconstruction accuracy and reliability: by introducing an island subgraph extraction mechanism, the application can accurately capture the heterogeneous topological characteristics between each island of the power distribution network after the fault, avoiding the connection prediction errors caused by the isomorphic node embedding of the traditional graph neural network. Combined with the feature aggregation method fused with physical rules, the accuracy of connectivity judgment and dynamic reconstruction is significantly improved.

[0017] 2. Greatly improving the reconstruction decision speed: the complex global network reconstruction problem is decomposed into a local link prediction problem centered on the link switch, greatly reducing the problem solving dimension. Combined with efficient feature engineering and a lightweight subgraph neural network model, the fast decision of milliseconds to seconds is realized, meeting the stringent real-time requirements of fault recovery.

[0018] 3. Effectively reducing the consumption of computing resources: compared with the traditional graph neural network method of processing the entire power distribution network graph, the local subgraph extraction and calculation strategy used in the application significantly reduces the calculation load and memory occupation in the model training and inference process, so that the algorithm is easy to deploy on edge computing devices or resource-limited industrial controllers.

[0019] 4. Enhanced Model Generalization and Adaptability: The method proposed in this invention does not rely on specific network topologies or fixed fault scenarios. By dynamically determining the k-hop range and adaptive feature aggregation, the model can adaptively cope with distribution networks of different sizes and structures, and maintain stable performance under various complex operating conditions such as fault evolution, branch switching, and distributed power output fluctuations.

[0020] 5. Deep integration of physical constraints to ensure the feasibility of the scheme: In the process of feature extraction and aggregation, Kirchhoff's laws, voltage constraints and other core physical rules of power systems are innovatively integrated to ensure that the model's learning and decision-making process conforms to the basic laws of power grid operation, thereby ensuring the physical feasibility and security of the generated reconfiguration scheme. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the algorithm flow of the present invention; Figure 2 This is a schematic diagram of the IEEE 33-node system.

[0022] Figure 3 This is a schematic diagram of the system architecture of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0024] like Figure 1 and Figure 2 As shown, this invention provides a fast recovery method based on regional distribution network fault reconfiguration (SDGCN), comprising: Step 1: Collect real-time system operation data of the distribution network after a fault; Step 2: Based on system operation data, simplify the actual line distribution wiring diagram into a distribution network topology diagram, and extract the distribution network topology with link switches as the source-target nodes to obtain the island sub-diagram; Step 3: Perform feature processing on the isolated subgraphs and construct a subgraph feature information matrix; Step 4: Input the subgraph feature information matrix into the pre-trained subgraph neural network model to obtain the connectivity prediction results of each potential connection link; Step 5: Generate dynamic reconfiguration control commands based on connectivity prediction results, and execute the control commands to restore the distribution network.

[0025] In step 2, the topology of the power distribution network is extracted with the link switch as the source-target node to obtain an island subgraph, specifically: An undirected topology graph G=(V, E) of the post-fault regional power distribution network is obtained, where V represents a bus node set, and E represents a power line set; For the post-fault power distribution network, all neighbor nodes within k hops with the line link switch as the source and target node are extracted as an island subgraph node, and the determination method of k is as follows: (1) In the formula, is a DG regulation capability factor, is a decay coefficient, R is the average line impedance, p is an island topology density parameter, is a k-hop subgraph node set, and Ɛ is a fluctuation sensitivity parameter, is an island load proportion weight. is an island internal node voltage threshold, is a load fluctuation rate, is the number of buses in the island.

[0026] In step 3, the island subgraph is processed to construct a subgraph feature information matrix, specifically: For the two endpoints (u, v) of the link switch, the feature distance of each neighbor node within k hops relative to the source node and the target node is calculated, specifically: (2) In the formula, is a feature distance, V, p are the voltage and distributed power penetration rate of the node, respectively, is a weight coefficient. Subscripts u, , are the source node, target node, and any neighbor node in the target link, respectively, is a reference voltage, is the maximum penetration rate of distributed power in the island.

[0027] Based on the feature distance, each neighbor node within k hops is assigned a discrete label, specifically: (3) In the formula, is a node assigned discrete label, is a system rated parameter for fixing the label hierarchical granularity, is a reference penetration rate of distributed power in the island.

[0028] According to the level determined by the discrete label, the feature vectors of the neighbor nodes are weighted and aggregated, and the aggregation result is spliced with the features of the source node or the target node itself and the label embedding vector to form a subgraph feature information matrix, specifically: The feature vectors of the remaining neighbor nodes in the isolated subgraph are accumulated into the target node, and the feature vectors of the neighbor nodes are weighted and aggregated according to the level determined by the discrete label, which is: (4) In the formula, is the aggregated feature vector of the target node u, is the feature vector of the neighbor node ω, , is the active power, reactive power and voltage of the neighbor node . is the maximum label level, is the weight function, , is the decay coefficient, which is 0.5; is the k-hop neighbor set of the source node u; Finally, the subgraph feature information matrix is obtained, which is: (5) (6) In the formula, is the label embedding vector, d e is the embedding dimension, which is 16 or 32; is one-hot encoding; is the embedding matrix, which is a trainable parameter matrix for mapping discrete labels to low-dimensional continuous variables. is the feature vector of node u.

[0029] In step 4, the subgraph feature information matrix is input into the pre-trained subgraph neural network model to obtain the connectivity prediction result of each potential connection link, specifically: The subgraph feature information matrix, i.e. the feature information matrix of the target node u, is power encoded and node feature encoded to enhance feature separability, and LayerNorm is used to suppress the noise influence caused by feature scale difference, the calculation method is: (7) (8) (9) In the formula, are the active power and reactive power eigenvalues of the target node and its neighbor nodes, respectively; LayerNorm is used to normalize the power features, stabilize the training process, and prevent gradient explosion / vanishing caused by excessively large or small power; is the original attribute vector of the target node u, W is the weight matrix, and b is the bias term.

[0030] The graph convolutional neural network model is used to update the feature information using the update process of the graph convolutional network. The update mechanism of the graph convolutional network is: (10) (11) wherein, and are the node feature matrices of the lth and l+1th layers, respectively, N is the number of link switch endpoints after subgraph extraction, and d is the dimension of the aggregated target node feature information. is the adjacency matrix of the graph, is the identity matrix that ensures that the node can aggregate its own features; is the adjacency matrix with a self-loop, and D is the degree matrix of ; is the learnable weight matrix of the lth layer; is the activation function; Residual connection is added to the update process, and the specific method is: (12) wherein, is the initial input feature; is the residual projection matrix of the lth layer, which is used to linearly map the initial feature to the current spatial layer; l Batch normalization is performed on the residual connection structure, and the specific method is: (13) (14) (15) wherein, , m is the number of small batches, is the feature vector of the ith sample at the lth layer; is the mean, and denotes the variance; is a very small constant to prevent the denominator from being zero; are the scaling factor and the offset factor, respectively, which are learnable parameters, and the initial values are usually 1 and 0, used to restore the expression ability of the network; The connectivity prediction calculation is performed on the training results, and the prediction probability is: (16) (17) wherein, are the first and second layer weight matrices respectively, are the first and second layer bias vectors respectively, is a Sigmoid function, is the true label.

[0031] In combination with the IEEE33 node system shown in FIG. 1, the performance of the algorithm of the present application is verified. Figure 2 Nodes 1-33 are node numbers, node 1 is connected to the external power grid, nodes 7, 11, 12, 21, 25, 27, and 33 are connected to distributed power sources, the dashed line represents the line tie switch, and the diagonal line represents the line fault. The reference voltage of the entire experiment is 12.66KV, and the total load of the network is 6.431+j2.821MVA. By introducing random faults into the network, 21 different fault topologies are generated. Each topology contains 300 sets of scene data, and all samples are divided according to the ratio of 10% for verification, 80% for training, and 10% for evaluation. During the training process, the number of cycles is set to 300, and the batch size is 512. Figure 2 In the comparative example analysis, the algorithm (SDGCN) proposed in the present application is compared with the conventional graph convolution network (GCN) without subgraph extraction and the current widely used "subgraph-embedding-prediction" algorithm (SEAL) in terms of performance. The performance of each model in fault network processing and calculation tasks is analyzed from multiple evaluation indicators, including F1 score, training loss, fault network information capture ability (ICC), and precision. The specific results are shown in Table 1.

[0032] Table 1 Performance evaluation of three algorithms

[0033] As can be seen from the table, the algorithm SDGCN proposed in the present application is superior to the other two conventional algorithms in terms of precision, network information capture ability, and training loss.

[0034] In order to enable the method described in the present application to be applied in actual distribution networks, the present application also provides a fast recovery system based on regional distribution network fault reconstruction, as shown in FIG. 2, for implementing the fast recovery method based on regional distribution network fault reconstruction described above. The architecture maps the algorithm steps to specific hardware modules, realizing full-process automation from data acquisition to control execution.

[0035] Figure 3 In order to enable the method described in the present application to be applied in actual distribution networks, the present application also provides a fast recovery system based on regional distribution network fault reconstruction, as shown in FIG. 2, for implementing the fast recovery method based on regional distribution network fault reconstruction described above. The architecture maps the algorithm steps to specific hardware modules, realizing full-process automation from data acquisition to control execution. ​

[0036] The hardware system includes the following four levels: 1. Sensing and acquisition layer Hardware composition: This layer mainly includes intelligent sensing devices deployed at key nodes of the power distribution network, such as busbars, switch stations, and distributed power grid connection points, such as synchronized phasor measurement units (PMUs), feeder terminal units (FTUs), distribution terminal units (DTUs), smart meters, and remote terminal units (RTUs) of SCADA systems.

[0037] Function correspondence (step 1): These devices are responsible for real-time acquisition of post-fault system operation data, including node voltage V, current I, active power P, reactive power Q, real-time on-off state of switches, and real-time output of distributed power sources (DGs), which are used to calculate parameters such as penetration rate p. The analog signals collected are converted to digital signals on-site, forming standardized digital data.

[0038] 2. Communication network layer Hardware composition: This layer is composed of wired and wireless communication devices, including industrial Ethernet switches, fiber ring network devices, power line carrier (PLC) communication modules, and 5G / 4G wireless private network communication terminals.

[0039] Function: Provide high-reliability, low-latency bidirectional data channels. In the upstream direction, it is responsible for aggregating and transmitting the massive operation data collected by the sensing layer to the computing layer; in the downstream direction, it is responsible for accurately and quickly distributing the control instructions generated by the computing layer to each execution device. This layer meets the communication standards of the power system, ensuring that the transmission delay of critical data is within milliseconds.

[0040] 3. Edge / station-end computing layer Hardware composition: This layer includes edge computing gateways deployed at substations or important distribution nodes, as well as station-end monitoring servers deployed at power distribution automation master stations or sub-stations.

[0041] Edge computing gateway: Uses a high-performance embedded hardware platform, equipped with a multi-core processor, and sufficient memory and solid-state storage (eMMC / SSD).

[0042] Function and algorithm step correspondence: Topology construction and subgraph extraction (corresponding to step 2): The computing unit (edge gateway or station-end server) receives global data sent through the communication network, dynamically constructs and maintains the undirected topology graph G= (V, E) of the power distribution network in memory. According to formula (1), dynamically calculate the subgraph extraction range k for each potential connection link (i.e., the link between the two nodes), and run the k-hop neighbor search algorithm centered on the source node and target node to extract the corresponding island subgraph data structure from the global graph.

[0043] Feature processing and matrix construction (corresponding to step 3): for each extracted island subgraph, the calculation unit calls its mathematical calculation library to calculate the feature distance of each neighbor node according to formula (2), assigns a discrete label to it according to formula (3), and finally completes feature weighted aggregation and splicing to generate a standard subgraph feature information matrix M according to formulas (4)-(6) u Neural network model inference (corresponding to step 4): The pre-trained subgraph neural network model (SDGCN) parameter file (.pt) is stored in the storage medium of the calculation unit. When making fault recovery decisions, the model is loaded into the memory. The forward inference calculation of the model is performed by the GPU, which encapsulates a series of operations such as feature encoding, graph convolution, residual connection, normalization and classifier calculation described in formulas (7)-(17), and finally outputs the connectivity prediction probability of each potential connection link This layer supports online model updating.

[0044] 4. Control execution layer Hardware composition: This layer is composed of various remotely operated switch devices, mainly including automatic sectionalizing switches, tie switches and grid-connected switches of distributed power sources. These switches have built-in electric operating mechanisms and intelligent controllers, and have "remote signaling, remote measurement and remote control" functions.

[0045] Function correspondence (step 5): Receive the dynamic reconstruction control instruction set issued from the calculation layer. The instructions are issued to the controllers of the target switches in the form of standard power protocols through the communication network. After the controller parses the instructions, it drives the operating mechanism to perform "closing" or "opening" operation, thereby changing the network topology and realizing the restoration of the distribution network. To increase safety, before performing critical operations, the switch controller can integrate simple local logic verification (such as voltage difference and phase angle check).

[0046] Beneficial effects of the cooperative work of the hardware system: Through the close combination of the above hierarchical hardware architecture and the algorithm of the present application, the following engineering advantages are further realized: Distributed computing, reducing central pressure: deploying high-complexity graph neural network inference tasks on the edge side close to the data source greatly reduces the amount of data that needs to be uploaded to the main station, reducing the demand for main station computing resources and communication bandwidth, and improving the overall throughput and robustness of the system.

[0047] Software and hardware decoupling and efficient cooperation: the algorithm design takes into account the characteristics of embedded platforms (such as fixed-dimension matrix operations), which are easy to optimize in parallel using processor vector instruction sets, fully utilizing hardware computing power and achieving millisecond-level inference.

[0048] Easy to integrate and smooth evolution: the architecture is highly compatible with existing distribution automation systems. It can be implemented by adding edge computing gateways in existing substations or ring network cabinets, and upgrading the master station software in a modular manner, effectively protecting the user's existing investment.

[0049] High reliability and autonomous capability: the edge computing node has certain local storage and computing capability. Even in the case of communication interruption with the master station, it can still make autonomous recovery decisions for the key link using the locally cached model and recent data, greatly enhancing the survivability of the distribution network.

[0050] Embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0051] The present application is described with reference to flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions that are executed by the processor of the computer or other programmable data processing apparatus generate an apparatus that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in a flow or multiple flows and / or blocks Figure 1 The functions specified in a flow or multiple flows and / or blocks

[0052] These computer program instructions can also be stored in a computer-readable memory that can direct the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction apparatus, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in a flow or multiple flows and / or blocks Figure 1 The functions specified in a flow or multiple flows and / or blocks

[0053] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, so that a series of operation steps are performed on the computer or other programmable data processing apparatus to produce a computer-implemented process, so that the instructions executed on the computer or other programmable data processing apparatus provide a process for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in a flow or multiple flows and / or blocks Figure 1steps of the functions specified in the block or blocks.

[0054] The content not described in detail in the specification of the present application belongs to the prior art known to the person skilled in the art. It is indicated herein that the above description helps the person skilled in the art to understand the present application, but does not limit the protection scope of the present application. Any implementation of equivalent replacement, modification, improvement and deletion of the above description without departing from the essential content of the present application falls within the protection scope of the present application.

Claims

1. A fast restoration method based on regional distribution network fault reconfiguration, characterized in that, The method comprises the following steps: Step 1: Real-time acquisition of system operation data of the power distribution network after failure; Step 2: Based on the system operation data, a power distribution network topology graph is constructed, and a link switch is extracted as a source-target node to obtain an island subgraph; Step 3: Feature processing is performed on the island subgraph to construct a subgraph feature information matrix; Step 4: The subgraph feature information matrix is input into a pre-trained subgraph neural network model to obtain connectivity prediction results of each potential connection link; Step 5: Based on the connectivity prediction results, a dynamic reconstruction control instruction is generated, and the control instruction is executed to realize the restoration of the power distribution network.

2. The method of claim 1, wherein, In step 2, the topology structure of the power distribution network is extracted as a link switch as a source-target node to obtain an island subgraph, specifically: The two endpoints of the link switch are taken as the source node and the target node; Determine the k-hop neighbor range of the source node and the target node, and obtain all neighbor nodes in the range to form an island subgraph.

3. The method of claim 2, wherein, The determination method of the number of hops k is: （1） In the formula, is the DG regulation capability factor, is the attenuation coefficient, R is the average line impedance, In step 3, the feature processing is performed on the island subgraph to construct a subgraph feature information matrix, specifically: is the island topology density parameter, is the k-hop subgraph node set, and Ɛ is the volatility sensitivity parameter, is the island load proportion weight, is the island internal node voltage threshold, is the load fluctuation rate, is the number of island internal buses.

4. The method of claim 3, wherein, Calculate the feature distance of each neighbor node in the k-hop range relative to the source node and the target node; Based on the feature distance, a discrete label is assigned to each neighbor node in the k-hop range; According to the level determined by the discrete label, the feature vectors of the neighbor nodes are weighted and aggregated, and the aggregation result is spliced with the feature and label embedding vector of the source node or the target node itself to form a subgraph feature information matrix. The feature distance of each neighbor node in the k-hop range relative to the source node and the target node is calculated, specifically:

5. The method of claim 4, wherein, Based on the feature distance, a discrete label is assigned to each neighbor node in the k-hop range, specifically: （2） wherein, is the characteristic distance, V, According to the level determined by the discrete label, the feature vectors of the neighbor nodes are weighted and aggregated, and the aggregation result is spliced with the feature and label embedding vector of the source node or the target node itself to form a subgraph feature information matrix, specifically: are the voltage of the node and the penetration of the distributed generation, respectively, are the weight coefficients, different subscripts u, , are the source node, the target node, any neighbor node in the target link, respectively, is the reference voltage, is the maximum penetration of the distributed generation in the island.

6. The method of claim 5, wherein, The feature vectors of the remaining neighbor nodes in the island subgraph are accumulated into the target node, and the feature vectors of the neighbor nodes are weighted and aggregated according to the level determined by the discrete label, which is: （3） wherein, is assigned to the node, u, , are the source node, the target node, any neighbor node in the target link, respectively, is the system rated parameter, used to fix the label hierarchical granularity, is the reference penetration of distributed generation in the island.

7. The method of claim 6, wherein, Finally, the subgraph feature information matrix is obtained, which is: In step 4, the subgraph feature information matrix is input into a pre-trained subgraph neural network model to obtain connectivity prediction results of each potential connection link, specifically: （4） In the formula, are the feature vectors of the target node u after aggregation, respectively, is the feature vector of the neighbor node ω, , are the active power, reactive power and voltage of the neighbor node , is the maximum label level number, is the weight function, , is the attenuation coefficient, taken as 0.5; is the k-hop neighbor set of the source node u; Power coding and node feature coding are performed on the subgraph feature information matrix, and LayerNorm is used to suppress the noise influence caused by the difference in feature scale; （5） （6） wherein, is the label embedding vector, d e is the embedding dimension, taking values 16 or 32; is one-hot encoding; is the embedding matrix, a trainable parameter matrix that maps discrete labels to low-dimensional continuous variables, is the feature vector of node u.

8. The method of claim 7, wherein, Based on the graph convolutional neural network model, the feature information of the coded subgraph feature information matrix is updated, the update process is connected by residual connection, and the residual connection structure is batch normalized; The update result is used for connectivity prediction calculation to obtain the connectivity prediction result. The method comprises the following steps: A sensing acquisition module is used to real-time acquisition of system operation data of the power distribution network after failure; 9. A system for fast restoration based on regional power distribution network fault reconfiguration, for implementing the method of any one of claims 1-8, characterized in that, A communication module is used to transmit the system operation data and control instructions; An edge computing module includes a memory and a processor, the memory stores a computer program, and the processor executes the program to realize the steps of claims 1-8. ​ ​ The control execution module is configured to perform a switching operation based on the connectivity prediction result.

10. The fast restoration system based on regional power distribution network fault reconfiguration of claim 9, wherein, The sensing acquisition module includes at least one of a synchrophasor measurement unit, a feeder terminal unit, a distribution terminal unit and a smart meter deployed at a distribution network node. The edge computing module is specifically an edge computing gateway deployed at a distribution substation or a key node, or a station server deployed at a distribution automation master station or a substation; the edge computing gateway or the station server is built-in with an artificial intelligence acceleration chip. The control execution module includes a remotely controllable automatic sectionalizing switch, a tie switch and a distributed power source grid-connected switch.

Images