A power distribution network partitioning method and system based on a graph neural network and reinforcement learning

Abstract

This application provides a distribution network partitioning method and system based on graph neural networks and reinforcement learning. The method includes: constructing a topology graph model containing node feature vectors and edge feature vectors according to the topology and electrical parameters of the distribution network; hierarchically dividing the topology graph model to generate several sub-topology graph models with different voltage levels; aggregating the neighborhood information of each node feature vector in each sub-topology graph model through a preset graph neural network to generate node embedding vectors of each node in each sub-topology graph model; inputting the node embedding vectors of each sub-topology graph model into a preset partitioning agent, so that the partitioning agent partitions each sub-topology graph model according to a preset order under corresponding hierarchical constraints to generate corresponding sub-partitioning schemes; and combining the sub-partitioning schemes to obtain the partitioning scheme of the distribution network, thereby improving the efficiency and rationality of distribution network partitioning.

Description

Technical Field

[0001] This application relates to the field of automatic distribution network zoning technology, and in particular to a distribution network zoning method and system based on graph neural networks and reinforcement learning. Background Technology

[0002] As the final link connecting end users in the power system, the efficiency and reliability of the distribution network directly affect social production and people's livelihood. With the large-scale integration of distributed power sources and the increasing diversification of load types, the traditional radial distribution network is gradually evolving into a complex system of interactive coordination between "source-grid-load-storage". To cope with this complexity, improve fault isolation speed, optimize load balancing, and enhance operational economy, scientific and rational hierarchical and zoning planning of the distribution network has become a key technical requirement. Scientific zoning can effectively reduce the scope of fault impact, optimize power flow distribution, and provide clear physical boundaries and management units for the flexible consumption and coordinated control of distributed energy resources.

[0003] Currently, distribution network zoning methods mainly rely on traditional clustering algorithms (such as K-means) and heuristic methods based on human experience. Traditional clustering algorithms typically divide nodes based solely on physical distance or limited static electrical parameters, making it difficult to capture the dynamic and complex electrical coupling relationships between nodes. This results in loose electrical connections within the zoning results and uneven power exchange between zoning areas, affecting actual operational performance. While heuristic methods can incorporate some operational rules, their constraints are mostly manually set, making them poorly adaptable to dynamic operating conditions such as load fluctuations and network reconfiguration. They also suffer from delayed scheme updates and difficulty in achieving real-time optimization. Furthermore, most existing methods focus on "planar" zoning within a single voltage level, lacking a comprehensive consideration of the zoning coordination relationships between different voltage levels (e.g., 10kV medium-voltage level and 0.4kV low-voltage level). This leads to mismatches between higher-level and lower-level zoning boundaries, affecting cross-level fault isolation and load transfer efficiency. Meanwhile, distribution network zoning must also strictly meet practical engineering constraints such as relay protection setting matching and power supply reliability indicators. Traditional methods are often computationally cumbersome and difficult to obtain the global optimal solution when dealing with such complex optimization problems with multiple objectives and constraints. Therefore, there is an urgent need for a new intelligent zoning method that can deeply integrate power grid topology and real-time electrical characteristics, automatically adapt to dynamic changes, and achieve cross-level collaborative optimization. Summary of the Invention

[0004] To address the aforementioned technical problems, this application provides a distribution network partitioning method and system based on graph neural networks and reinforcement learning, which improves the efficiency and rationality of distribution network partitioning.

[0005] In a first aspect, embodiments of this application provide a power distribution network zoning method based on graph neural networks and reinforcement learning, including: Based on the topology and electrical parameters of the distribution network, a topological graph model containing node feature vectors and edge feature vectors is constructed. Based on the preset voltage level division conditions, the topology model is hierarchically divided to generate several sub-topology models with different voltage levels. The neighborhood information of the feature vectors of each node in each of the sub-topology graph models is aggregated by a preset graph neural network, thereby generating the node embedding vector of each node in each of the sub-topology graph models. The embedding vectors of each node of each of the sub-topology graph models are input into a preset partitioning agent, so that the partitioning agent partitions each of the sub-topology graph models according to a preset order under the corresponding hierarchical constraints, and generates each corresponding sub-partitioning scheme. The hierarchical constraints of any current sub-topology graph model are constructed based on the sub-partitioning scheme of the previous sub-topology graph model. The partitioning agent is constructed based on a reinforcement learning model and trained based on historical operation data of different distribution networks. By combining the various sub-partition schemes, the partition scheme of the power distribution network is obtained.

[0006] This application provides a distribution network partitioning method based on graph neural networks and reinforcement learning. Through steps such as constructing a topology graph model, hierarchical partitioning, node embedding vector generation, and partitioning agent decision-making, it achieves scientific and efficient distribution network partitioning. First, this embodiment constructs a topology graph model containing node feature vectors and edge feature vectors by deeply integrating the power grid's topology and real-time electrical parameters, thus comprehensively reflecting the physical connections and electrical characteristics of the distribution network. Second, the topology graph is hierarchically partitioned using preset voltage level partitioning conditions, generating sub-topology graph models for different voltage levels. This effectively solves the problem of insufficient cross-level partitioning coordination in traditional methods, making the boundaries between higher-level and lower-level partitions more aligned and improving the efficiency of fault coordination isolation and load transfer. Next, the graph neural network aggregates the neighborhood information of nodes in each sub-topology graph to generate node embedding vectors, which can deeply capture the dynamic and complex electrical coupling relationships between nodes, avoiding the loose internal electrical connections caused by traditional clustering algorithms relying solely on static parameters. Finally, by using a partitioning agent trained based on reinforcement learning, each sub-topology is partitioned sequentially in descending order of voltage level. The sub-partitioning scheme is generated by combining hierarchical constraints. While ensuring that each layer performs partitioning independently, the partitioning results of high voltage levels are used to constrain the partitioning process of secondary sub-topologies. This achieves collaborative partitioning of each layer of the distribution network, significantly improving the rationality, adaptability, and efficiency of distribution network partitioning.

[0007] Furthermore, the step of constructing a topology graph model containing node feature vectors and edge feature vectors based on the topology and electrical parameters of the distribution network includes: Based on the normalized value of apparent power of each node, voltage level information, node equipment type and load rate in the electrical parameters, a feature vector of each corresponding node is constructed. Based on the normalized values ​​of line resistance, line reactance, line length, connection relationship identifier, and line load rate of each node in the electrical parameters, the corresponding edge feature vectors are constructed. The topology graph model is constructed by combining the topology of the power distribution network, the feature vectors of each node, and the feature vectors of each edge.

[0008] This application describes in detail the specific steps for constructing a topology graph model, including methods for constructing node feature vectors and edge feature vectors. The node feature vectors encompass key electrical parameters such as the normalized value of apparent power, voltage level information, node equipment type, and load factor. The edge feature vectors include information such as line resistance, reactance, normalized length, connection relationship identifiers, and line load factor. This multi-dimensional feature construction method enables the topology graph model to not only reflect the physical structure of the power grid but also accurately capture the electrical state of nodes and the operating status of lines. Furthermore, the topology graph model constructed by combining node and edge feature vectors provides a high-quality data foundation for subsequent hierarchical partitioning, information embedding, and zoning decisions, ensuring accurate representation of the dynamic characteristics of the power grid during the zoning process. Compared to traditional methods that rely solely on limited static parameters, the model constructed by this method is more comprehensive and adaptable, better able to handle complex scenarios such as distributed power source integration and diversified load types, thus providing a reliable basis for generating zoning schemes and improving the rationality of the zoning results.

[0009] In one possible implementation, the step of hierarchically dividing the topology model according to preset voltage level division conditions to generate several sub-topology models with different voltage levels includes: Based on the voltage level classification conditions and the voltage level information in the feature vectors of each node, hierarchical labels are added to each node in the distribution network. Based on the node device type in the feature vector of each node, a number of transformer nodes are determined and voltage conversion node labels are added to each transformer node. The hierarchical affiliation of each node feature vector is determined based on the hierarchical labels, and the hierarchical boundary of the topology graph model is determined based on the voltage conversion node labels. Then, several sub-topology graph models with different voltage levels are generated based on the hierarchical affiliation and the hierarchical boundary.

[0010] This application provides a specific method for hierarchical partitioning. By adding hierarchical labels to nodes, identifying transformer nodes, and adding voltage conversion node labels, the hierarchical boundaries are determined, generating sub-topology models for different voltage levels. This achieves hierarchical processing of the distribution network and effectively distinguishes network sections of different voltage levels, such as high voltage, medium voltage, and low voltage. Through hierarchical partitioning, this embodiment can comprehensively consider cross-level collaborative relationships, avoiding the mismatch problem between high-level and low-level partitioning boundaries in traditional "planar" partitioning methods. The introduction of hierarchical labels and voltage conversion node labels enables the partitioning process to accurately identify key equipment such as transformers, ensuring that the partitioning scheme meets the voltage conversion and energy transmission requirements of actual engineering projects. The generated sub-topology models for different voltage levels provide a clear input structure for subsequent partitioning agents, enabling partitioning decisions to be made in descending order of voltage level, ensuring the hierarchy and coherence of the partitioning scheme. This hierarchical partitioning method significantly improves the speed and accuracy of fault isolation, optimizes load balancing, and provides clear physical boundaries for the collaborative control of distributed energy resources, improving the rationality of subsequent partitioning results.

[0011] In one possible implementation, for any of the sub-topology graph models, the step of aggregating the neighborhood information of the feature vectors of each node in each sub-topology graph model through a preset graph neural network to generate the node embedding vector of each node in each sub-topology graph model includes: The sub-topology graph model is input into the graph neural network, so that the graph neural network updates the feature vectors of each node in the sub-topology graph model through several graph convolutional network layers, and finally generates the embedding vectors of each node in the sub-topology graph model. In any of the iteration update processes, the first node embedding vectors output by the previous graph convolutional layer are input to the current graph convolutional layer, so that the current graph convolutional layer aggregates the neighborhood information of each first node embedding vector according to its own weight matrix and the current edge feature vectors of the current sub-topology graph model, and then updates each first node embedding vector based on the neighborhood information to generate each corresponding second node embedding vector and input it to the next graph convolutional layer.

[0012] This application provides a method for iteratively updating node feature vectors. For any current sub-topology graph model, a graph neural network can aggregate node neighborhood information and gradually fuse local and global features in the current sub-topology graph model through multiple iterations, thereby generating a current node embedding vector rich in contextual information. This process deeply captures the complex electrical coupling relationships between nodes, avoiding the limitations of traditional methods that rely solely on physical distance or simple electrical parameters. The generated node embedding vector serves as input to the partitioning agent, providing a high-dimensional, abstract feature representation for partitioning decisions, enabling the agent to scientifically partition based on the deep electrical relationships between nodes. Furthermore, the combination of the weight matrix of the graph convolutional layer and the edge feature vector ensures the accuracy and efficiency of information aggregation, improving the model's adaptability to changes in power grid topology. By generating node embedding vectors, this method significantly enhances the understanding of the dynamic characteristics of the power grid during partitioning, laying a solid foundation for generating partitioning schemes with tight internal electrical connections and balanced power exchange between partitions.

[0013] In one possible implementation, for any of the sub-topology graph models, the step of inputting the embedding vectors of each node of each sub-topology graph model into a preset partitioning agent, so that the partitioning agent partitions each sub-topology graph model under corresponding hierarchical constraints, and generates corresponding sub-partitioning schemes, includes: Based on the embedding vectors of each node in the sub-topology graph model and the initial partitioning progress parameters, construct the initial input vector; The initial input vector is input to the partitioning agent, so that the partitioning agent makes partitioning decisions on each node in the sub-topology graph model in sequence according to the preset policy network and the current level constraints, and finally generates the sub-partitioning scheme of the sub-topology graph model. In any partitioning decision process, the current state vector is obtained through the partitioning agent; based on the current state vector and the current level constraints, the partitioning action of the current node is generated through the policy network. The partitioning action includes assigning the current node to an existing partition, creating a new partition and assigning it to the current node, or dividing the current node and an adjacent node into a partition boundary; the current partitioning progress parameter in the current state vector is updated according to the partitioning action of the current node, thereby obtaining an updated state vector, and the updated state vector is sent to the next partitioning decision.

[0014] This application describes the specific decision-making process of a partitioning agent, including steps such as constructing an initial input vector, generating partitioning actions through a policy network, and updating the state vector. Based on a reinforcement learning model, the partitioning agent can dynamically select partitioning actions according to the current state vector and hierarchical constraints, such as assigning nodes to existing partitions, creating new partitions, or setting partition boundaries. This decision-making mechanism makes the partitioning process highly flexible and adaptable, capable of responding in real-time to changes in the power grid's operating state. By sequentially making partitioning decisions for nodes, the agent can gradually construct partitioning schemes and update partitioning progress parameters based on historical decision records, ensuring the continuity and completeness of the partitioning process. The initial input vector, combining node embedding vectors and partitioning progress parameters, provides the agent with comprehensive state information, enabling it to comprehensively consider electrical characteristics and partitioning progress for optimization decisions. The diversity of partitioning actions ensures the richness of partitioning schemes, meeting the needs of different operating scenarios. Overall, the introduction of the partitioning agent endows the partitioning method with self-learning and adaptive capabilities, automatically optimizing partitioning schemes and improving the efficiency and rationality of distribution network partitioning.

[0015] Furthermore, the step of updating the current partitioning progress parameter in the current state vector based on the partitioning action of the current node, thereby obtaining the updated state vector, includes: Based on the partitioning action of the current node and the historical decision records of the agent, determine the number of nodes that have been partitioned and the number of partitions that have been formed. The current partitioning progress parameter is updated based on the number of nodes that have been partitioned and the number of partitions that have been formed, thereby obtaining the updated state vector.

[0016] This application further refines the update mechanism for partitioning progress parameters. By determining the number of nodes that have been partitioned and the number of partitions formed through partitioning actions and historical decision records, the state vector is updated. This ensures that the partitioning agent can track the partitioning progress in real time during the decision-making process, avoiding duplicate or missed partitions and guaranteeing the integrity of the partitioning scheme. Updating the number of completed nodes and the number of partitions provides the agent with clear progress feedback, enabling it to adjust subsequent decision strategies based on the current partitioning state and optimize the partitioning order and boundary settings. This dynamic update mechanism enhances the controllability and predictability of the partitioning process, allowing the partitioning scheme to gradually approach the optimal solution. Simultaneously, the introduction of progress parameters provides a basis for calculating the reward function, helping the agent better evaluate the partitioning effect during training and accelerating the learning process. By finely managing partitioning progress, this method improves the efficiency and accuracy of partitioning decisions, providing strong support for generating scientific and reasonable partitioning schemes.

[0017] In one possible implementation, the construction of the partitioned agent based on a reinforcement learning model and training it using historical operating data from different distribution networks includes: An initial agent is constructed based on the reinforcement learning model, and the initial agent includes an initial policy network and an initial value network. Several training data sets are constructed based on the historical operation data of the different distribution networks. The training data sets include distribution network topology maps, corresponding node embedding vectors, manually optimized optimal partition action sequences, and reward values ​​corresponding to the optimal partition action sequences. Based on the aforementioned training data and a preset reward function, the internal parameters of the initial agent are updated several times using gradient descent to obtain the final partitioned agent.

[0018] This application provides a method for constructing and training a partitioned intelligent agent, including constructing an initial agent based on a reinforcement learning model, preparing training data, and updating parameters using gradient descent. The initial agent consists of an initial policy network and an initial value network, which are responsible for generating partitioned actions and evaluating state values, respectively, forming a complete reinforcement learning framework. The training data comes from historical operating data of different distribution networks, including topology graphs, node embedding vectors, manually optimized optimal partitioned action sequences, and corresponding reward values, ensuring the diversity and realism of the training process. Gradient descent is used to update the internal parameters of the agent in multiple rounds, enabling it to gradually learn the optimal partitioning strategy and adapt to the operating characteristics of different power grids. Overall, the training process of the partitioned intelligent agent significantly improves the intelligence level of the method and further enhances the efficiency and rationality of distribution network partitioning.

[0019] Furthermore, when iteratively updating the internal parameters of the current agent using gradient descent based on the aforementioned training data and a preset reward function to obtain the updated agent for the current round, the step of updating the internal parameters of the initial agent for several rounds using gradient descent based on the aforementioned training data and a preset reward function to obtain the final partitioned agent includes: A preset number of current training data are randomly selected from the plurality of training data; Construct corresponding initial state vectors based on the distribution network topology diagram of each current training data and the corresponding node embedding vectors. Each of the initial state vectors is input to the current agent, so that the current agent generates a corresponding decision action sequence and a state vector sequence respectively; Based on each decision action sequence, each state vector sequence, the reward function, the optimal partition action sequence in each of the current training data, and the corresponding reward value, the parameters of the current policy network and the current value network of the current agent are updated using the gradient descent method to obtain the updated agent for the current round.

[0020] This application further describes the specific training steps of the partitioning agent, including randomly sampling training data, constructing an initial state vector, generating a decision action sequence, and updating network parameters using gradient descent. Randomly sampling training data ensures the randomness and comprehensiveness of the training process, avoiding the impact of data bias on the agent's performance. The initial state vector is constructed based on the topology graph and node embedding vectors, providing the agent with an accurate input environment. The generation of the decision action sequence and state vector sequence simulates the entire partitioning decision-making process, allowing the agent to accumulate experience during training. By comparing the reward function, the optimal partitioning action sequence, and the reward value, gradient descent updates the parameters of the policy network and value network, gradually optimizing the agent's decision-making ability. This iterative training mechanism enables the agent to continuously improve its partitioning strategy, approaching the optimal solution optimized manually. Through systematic training steps, the partitioning agent ultimately possesses powerful partitioning capabilities, providing high-quality partitioning schemes for the distribution network.

[0021] Furthermore, the reward function is constructed from the electrical connection tightness index within the partition, the power exchange balance index between partitions, the constraint satisfaction reward, and the node partitioning completion reward.

[0022] This application's embodiments define the composition of the reward function, including an intra-partition electrical connectivity index, an inter-partition power exchange balance index, a constraint satisfaction reward, and a node partitioning completion reward. This multi-index reward function ensures that the partitioning agent can comprehensively consider multiple objectives during training, such as electrical performance, balance, constraint satisfaction, and partitioning integrity. The intra-partition electrical connectivity index encourages the agent to generate internally tightly connected partitions, improving fault isolation efficiency; the inter-partition power exchange balance index promotes power balance between partitions, optimizing power flow distribution; the constraint satisfaction reward ensures that the partitioning scheme meets engineering requirements such as relay protection and power supply reliability; and the node partitioning completion reward incentivizes the agent to complete the partitioning of all nodes, ensuring the integrity of the scheme. This comprehensive reward mechanism enables the agent to learn a partitioning strategy that considers multiple objectives, further improving the rationality of distribution network partitioning.

[0023] Secondly, embodiments of this application provide a power distribution network partitioning system based on graph neural networks and reinforcement learning, including a topology graph construction module, a hierarchy partitioning module, an information embedding module, a partitioning module, and a scheme merging module; The topology graph construction module is used to construct a topology graph model containing node feature vectors and edge feature vectors based on the topology and electrical parameters of the power distribution network. The hierarchical division module is used to perform hierarchical division of the topology model according to preset voltage level division conditions, and generate several sub-topology models with different voltage levels. The information embedding module is used to aggregate the neighborhood information of the feature vectors of each node in each of the sub-topology graph models through a preset graph neural network, thereby generating the node embedding vector of each node in each of the sub-topology graph models. The partitioning module is used to input the embedding vectors of each node of each of the sub-topology graph models into a preset partitioning agent, so that the partitioning agent partitions each of the sub-topology graph models according to a preset order under the corresponding hierarchical constraints, and generates each corresponding sub-partitioning scheme. The hierarchical constraints of any current sub-topology graph model are constructed based on the sub-partitioning scheme of the previous sub-topology graph model. The partitioning agent is constructed based on a reinforcement learning model and trained based on historical operation data of different distribution networks. The scheme merging module is used to combine the various sub-partition schemes to obtain the partition scheme of the distribution network. Attached Figure Description

[0024] Figure 1 A flowchart illustrating a power distribution network zoning method based on graph neural networks and reinforcement learning, provided for an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a power distribution network zoning system based on graph neural networks and reinforcement learning, provided as an embodiment of this application. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0026] It should be noted that the step numbers in this document are only for the convenience of explaining the specific embodiments and are not intended to limit the order in which the steps are performed. In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0027] Example 1: like Figure 1 As shown, Embodiment 1 provides a distribution network partitioning method based on graph neural networks and reinforcement learning, including steps S1-S5: Step S1: Based on the topology and electrical parameters of the distribution network, construct a topology graph model containing node feature vectors and edge feature vectors; Step S2: According to the preset voltage level division conditions, the topology model is divided into levels to generate several sub-topology models with different voltage levels. Step S3: Aggregate the neighborhood information of the feature vectors of each node in each of the sub-topology graph models through a preset graph neural network, and then generate the node embedding vector of each node in each of the sub-topology graph models. Step S4: Input the embedding vectors of each node of each of the sub-topology graph models into a preset partitioning agent, so that the partitioning agent partitions each of the sub-topology graph models according to a preset order under the corresponding hierarchical constraints, and generates each corresponding sub-partitioning scheme. The hierarchical constraints of any current sub-topology graph model are constructed based on the sub-partitioning scheme of the previous sub-topology graph model. The partitioning agent is constructed based on a reinforcement learning model and trained based on the historical operation data of different distribution networks. Step S5: Combine the various sub-partition schemes to obtain the partition scheme of the power distribution network.

[0028] This application provides a distribution network partitioning method based on graph neural networks and reinforcement learning. Through steps such as constructing a topology graph model, hierarchical partitioning, node embedding vector generation, and partitioning agent decision-making, it achieves scientific and efficient distribution network partitioning. First, this embodiment constructs a topology graph model containing node feature vectors and edge feature vectors by deeply integrating the power grid's topology and real-time electrical parameters, thus comprehensively reflecting the physical connections and electrical characteristics of the distribution network. Second, the topology graph is hierarchically partitioned using preset voltage level partitioning conditions, generating sub-topology graph models for different voltage levels. This effectively solves the problem of insufficient cross-level partitioning coordination in traditional methods, making the boundaries between higher-level and lower-level partitions more aligned and improving the efficiency of fault coordination isolation and load transfer. Next, the graph neural network aggregates the neighborhood information of nodes in each sub-topology graph to generate node embedding vectors, which can deeply capture the dynamic and complex electrical coupling relationships between nodes, avoiding the loose internal electrical connections caused by traditional clustering algorithms relying solely on static parameters. Finally, by using a partitioning agent trained based on reinforcement learning, each sub-topology is partitioned sequentially in descending order of voltage level. The sub-partitioning scheme is generated by combining hierarchical constraints. While ensuring that each layer performs partitioning independently, the partitioning results of high voltage levels are used to constrain the partitioning process of secondary sub-topologies. This achieves collaborative partitioning of each layer of the distribution network, significantly improving the rationality, adaptability, and efficiency of distribution network partitioning.

[0029] Furthermore, in step S1, constructing a topology graph model containing node feature vectors and edge feature vectors based on the topology and electrical parameters of the distribution network includes: Based on the normalized value of apparent power of each node, voltage level information, node equipment type and load rate in the electrical parameters, a feature vector of each corresponding node is constructed. Based on the normalized values ​​of line resistance, line reactance, line length, connection relationship identifier, and line load rate of each node in the electrical parameters, the corresponding edge feature vectors are constructed. The topology graph model is constructed by combining the topology of the power distribution network, the feature vectors of each node, and the feature vectors of each edge.

[0030] This application describes in detail the specific steps for constructing a topology graph model, including methods for constructing node feature vectors and edge feature vectors. The node feature vectors encompass key electrical parameters such as the normalized value of apparent power, voltage level information, node equipment type, and load factor. The edge feature vectors include information such as line resistance, reactance, normalized length, connection relationship identifiers, and line load factor. This multi-dimensional feature construction method enables the topology graph model to not only reflect the physical structure of the power grid but also accurately capture the electrical state of nodes and the operating status of lines. Furthermore, the topology graph model constructed by combining node and edge feature vectors provides a high-quality data foundation for subsequent hierarchical partitioning, information embedding, and zoning decisions, ensuring accurate representation of the dynamic characteristics of the power grid during the zoning process. Compared to traditional methods that rely solely on limited static parameters, the model constructed by this method is more comprehensive and adaptable, better able to handle complex scenarios such as distributed power source integration and diversified load types, thus providing a reliable basis for generating zoning schemes and improving the rationality of the zoning results.

[0031] In a preferred embodiment, the process of constructing the topology model of the distribution network is as follows: Based on the topology and electrical parameters of the distribution network, the physical parameters of the distribution network are transformed into node and edge features of a graph structure using feature mapping rules. The mathematical expression for this is: ; In the formula: The representative node feature set corresponds to various physical nodes in the distribution network, including bus nodes, transformer nodes, load access nodes, etc. This represents the set of edge features, corresponding to the lines connecting various nodes in the distribution network.

[0032] Among them, the node feature vector is used to accurately describe the core electrical attributes of each power grid node. Its node feature vector The expression is: ; In the formula: For nodes The normalized value of the apparent power of the load is calculated using the following formula: , For nodes The real-time apparent power of the load can be collected in real time through the distribution network monitoring system. The maximum apparent power value of the load at all nodes in the distribution network is determined by traversing the historical and real-time load data of all nodes. For voltage level encoding values, a one-hot encoding method is used to transform non-numerical voltage level information into numerical features that can be processed by a graph neural network. For node device type characteristics, if node For transformer nodes, this characteristic is the normalized value of the transformer capacity, calculated using the following formula: , This is the rated capacity of the transformer. This represents the maximum rated capacity of all transformers in the distribution network; if the node For load nodes, this feature is a load type code, represented using one-hot encoding. For nodes The load factor. The calculation formula is: , For nodes Real-time active power, For nodes The rated active power.

[0033] The edge feature vector is used to describe the key characteristics of the path connecting two nodes. and nodes The line, its edge feature vector The expression is: ; In the formula: This is the normalized value of the line resistance. , For connecting nodes and nodes The line resistance value; This represents the maximum resistance value of all lines in the distribution network. To normalize the line reactance, , For connecting nodes and nodes The line reactance value, This represents the maximum reactance value of all lines in the distribution network. This is the normalized value for the line length. , For connecting nodes and nodes The physical length of the line, This represents the maximum length of all lines in the distribution network. As a connection identifier, when a node and nodes When connected directly by a line When node and nodes When there is no direct line connection between them . For line load rate, , For flow through connected nodes and nodes The real-time current of the line, This is the rated current of the line.

[0034] Furthermore, to avoid interference from differences in numerical magnitude between different types of features during subsequent model training, all feature values ​​in the feature vector need to be standardized, mapping them to the [0,1] interval. Standardization uses the min-max standardization method, calculated as follows: In the formula: These are the original eigenvalues. This is the minimum value of this characteristic among all corresponding entities in the distribution network. This is the maximum value of this feature across all corresponding entities in the distribution network.

[0035] Through the above transformation process, the nodes, lines, and related electrical parameters of the distribution network are completely converted into the node and edge features of the graph structure. This transformation not only preserves the topological information of the distribution network but also covers key electrical operating parameters, enabling the graph structure to comprehensively reflect the physical characteristics and operating status of the distribution network.

[0036] In one possible implementation, in step S2, the step of hierarchically dividing the topology model according to preset voltage level division conditions to generate several sub-topology models with different voltage levels includes: Based on the voltage level classification conditions and the voltage level information in the feature vectors of each node, hierarchical labels are added to each node in the distribution network. Based on the node device type in the feature vector of each node, a number of transformer nodes are determined and voltage conversion node labels are added to each transformer node. The hierarchical affiliation of each node feature vector is determined based on the hierarchical labels, and the hierarchical boundary of the topology graph model is determined based on the voltage conversion node labels. Then, several sub-topology graph models with different voltage levels are generated based on the hierarchical affiliation and the hierarchical boundary.

[0037] This application provides a specific method for hierarchical partitioning. By adding hierarchical labels to nodes, identifying transformer nodes, and adding voltage conversion node labels, the hierarchical boundaries are determined, generating sub-topology models for different voltage levels. This achieves hierarchical processing of the distribution network and effectively distinguishes network sections of different voltage levels, such as high voltage, medium voltage, and low voltage. Through hierarchical partitioning, this embodiment can comprehensively consider cross-level collaborative relationships, avoiding the mismatch problem between high-level and low-level partitioning boundaries in traditional "planar" partitioning methods. The introduction of hierarchical labels and voltage conversion node labels enables the partitioning process to accurately identify key equipment such as transformers, ensuring that the partitioning scheme meets the voltage conversion and energy transmission requirements of actual engineering projects. The generated sub-topology models for different voltage levels provide a clear input structure for subsequent partitioning agents, enabling partitioning decisions to be made in descending order of voltage level, ensuring the hierarchy and coherence of the partitioning scheme. This hierarchical partitioning method significantly improves the speed and accuracy of fault isolation, optimizes load balancing, and provides clear physical boundaries for the collaborative control of distributed energy resources, improving the rationality of subsequent partitioning results.

[0038] In a preferred embodiment, the specific steps for generating several sub-topology models with different voltage levels based on the constructed topology model and electrical parameters are as follows: 1. Clarify the basis for hierarchical division and the rules for defining the levels. Based on the voltage level coding value of the mid-node feature vector as the sole quantification criterion, and combined with the distribution network control hierarchy, the distribution network is clearly divided into a 10kV high-voltage level and a 0.4kV medium-voltage level, and the coverage area and boundary nodes of each level are defined: 10kV high voltage layer: covering 10kV busbar nodes, 10kV line connection nodes, 10kV / 0.4kV transformer high voltage side nodes, and 10kV level distributed power supply / load access nodes; 0.4kV medium voltage layer: covering 0.4kV busbar nodes, 0.4kV line connection nodes, 10kV / 0.4kV transformer low voltage side nodes, and 0.4kV level load / distributed power supply access nodes; Hierarchical Boundary Hub: The 10kV / 0.4kV transformer is defined as a voltage conversion node, serving as a natural boundary between the two levels. Its high-voltage side belongs to the 10kV level, and its low-voltage side belongs to the 0.4kV level, making it the core node for inter-level coordination.

[0039] 2. Hierarchical Node Identification and Labeling Based on Topology Graph Model. Based on the distribution network topology graph model already constructed in S1, accurate identification and labeling of all node levels are achieved: Extract the voltage level encoding value from the feature vectors of all nodes, and automatically match the voltage level attribute of the nodes by the model. Add a level label to each node, labeled as "10kV level" or "0.4kV level". Add an additional "cross-level hub node" label to voltage conversion nodes to record the node identifier and transformer capacity normalization value corresponding to its high-voltage / low-voltage side; The edge features are hierarchically associated and labeled. Based on the hierarchical attributes of the nodes at both ends of the line, they are marked as "10kV line", "0.4kV line" or "cross-level transformer connection line". Only transformer lines are retained as valid cross-level lines.

[0040] 3. Hierarchical decomposition and subgraph construction of the topology graph model. Based on the node hierarchy annotation results, the original topology graph model is logically decomposed into two independent hierarchical topology subgraphs. , The subgraph fully preserves the node features, edge features, and electrical parameters of the original model, only limiting the hierarchical affiliation of nodes and lines; at the same time, a hierarchical association mapping table is established to record the correspondence between voltage conversion nodes in the two subgraphs and the electrical parameters of cross-level lines, ensuring that the electrical associations between levels can be accurately traced.

[0041] 4. Quantitative Embedding of Hierarchical Collaborative Constraints. The collaborative logic of "high-level constraints on low-level constraints" is transformed into quantifiable constraint parameters and embedded into the global features of the two-level topological subgraphs as prerequisite hard constraints for subsequent partitioning steps, as detailed below: To ensure that lower-level nodes are completely contained within their corresponding higher-level nodes, the spatial inclusion constraint formula is as follows: ; In the formula: Let i be the set of nodes in the i-th 10kV partition; This is the set of nodes corresponding to the 0.4kV partition.

[0042] To meet the 2-3 times allocation requirement, the formula for the quantity allocation constraint of each layer's partitions is as follows: ; In the formula: This represents the number of the i-th 10kV partitions; This corresponds to the number of 0.4kV partitions.

[0043] Hierarchical power exchange constraint formula: ; ; In the formula: This represents the power exchange value between levels; This represents the power transfer value at the voltage conversion node; This represents the power exchange value between partitions within the k-th level; This represents the total load power of the k-th level.

[0044] To ensure alignment between the lower-level partition boundaries and the voltage transformation nodes, the boundary alignment constraint formula is as follows: ; In the formula: Let i be the set of boundary nodes for the i-th 0.4kV partition; This refers to the set of voltage conversion nodes within the corresponding 10kV zone.

[0045] 5. Validation of hierarchical results and output of hierarchical subgraphs.

[0046] The results of the hierarchical operation are fully validated. The core validation metrics include: node hierarchical labeling accuracy ≥ 100%, no missing features in the hierarchical subgraphs, all hierarchical constraints have been quantized and embedded, and no omissions in the hierarchical association mapping table. After the validation is passed, standardized 10kV and 0.4kV layer topology subgraphs are output. Subsequently, GNN feature extraction and DRL model partitioning operations will be performed independently on the two hierarchical subgraphs, following the order of 10kV layer partitioning first and then 0.4kV layer partitioning.

[0047] In one possible implementation, in step S3, for any of the sub-topology graph models, the aggregation of neighborhood information of the feature vectors of each node in each sub-topology graph model through a preset graph neural network, thereby generating the node embedding vector of each node in each sub-topology graph model, includes: The sub-topology graph model is input into the graph neural network, so that the graph neural network updates the feature vectors of each node in the sub-topology graph model through several graph convolutional network layers, and finally generates the embedding vectors of each node in the sub-topology graph model. In any of the iteration update processes, the first node embedding vectors output by the previous graph convolutional layer are input to the current graph convolutional layer, so that the current graph convolutional layer aggregates the neighborhood information of each first node embedding vector according to its own weight matrix and the current edge feature vectors of the current sub-topology graph model, and then updates each first node embedding vector based on the neighborhood information to generate each corresponding second node embedding vector and input it to the next graph convolutional layer.

[0048] This application provides a method for iteratively updating node feature vectors. For any current sub-topology graph model, a graph neural network can aggregate node neighborhood information and gradually fuse local and global features in the current sub-topology graph model through multiple iterations, thereby generating a current node embedding vector rich in contextual information. This process deeply captures the complex electrical coupling relationships between nodes, avoiding the limitations of traditional methods that rely solely on physical distance or simple electrical parameters. The generated node embedding vector serves as input to the partitioning agent, providing a high-dimensional, abstract feature representation for partitioning decisions, enabling the agent to scientifically partition based on the deep electrical relationships between nodes. Furthermore, the combination of the weight matrix of the graph convolutional layer and the edge feature vector ensures the accuracy and efficiency of information aggregation, improving the model's adaptability to changes in power grid topology. By generating node embedding vectors, this method significantly enhances the understanding of the dynamic characteristics of the power grid during partitioning, laying a solid foundation for generating partitioning schemes with tight internal electrical connections and balanced power exchange between partitions.

[0049] In a preferred embodiment, a GNN (Graph Neural Network) is used to extract features and learn representations of the topological graph. Multi-layer graph convolution operations are used to capture electrical connections and topological dependencies between nodes, generating node embedding vectors that contain global topological information and local electrical characteristics. The specific steps are as follows: 1. Construct a basic graph convolutional layer structure, using 3-5 layers of graph convolutional layers connected in series to form a feature extraction network. Each graph convolutional layer perceives node connection relationships based on the topological adjacency matrix of the distribution network topology graph model. The topological adjacency matrix A is n×n in dimension, where n is the total number of nodes in the distribution network. When node i and node j have a line connection... ,otherwise The number of layers in the graph convolutional layer is determined by the number of nodes: L=3 when the number of nodes is ≤50, L=4 when 50 < the number of nodes is ≤200, and L=5 when the number of nodes is >200.

[0050] 2. Determine the feature update rule for the graph convolutional layer. For the k-th graph convolutional layer, the feature vector of node i... It is calculated using the following formula: ; In the formula: The electrical association weights are obtained by normalizing the inverse of the line impedance. , The line impedance magnitude between nodes i and j is calculated based on the corresponding edge feature vectors. for A trainable weight matrix of dimension d k ₋1 represents the dimension of the node feature vector output by the (k-1)th graph convolutional layer, d kLet be the dimension of the node feature vector output by the k-th graph convolutional layer; for Bias vector of dimension; Using the LeakyReLU activation function This avoids information loss due to feature values ​​returning to zero during feature extraction in the distribution network.

[0051] Regarding the weight matrix The training process is as follows: (1) Alignment of training objectives: Align the weight matrix The training is tied to the "optimization objective of hierarchical zoning of distribution networks," and gradient updates are performed through backpropagation of the reward function of deep reinforcement learning (DRL). (2) Iterative update: Initialize the weight matrix of all graph convolutional layers. The distribution network topology, node features, and edge features are input into a graph convolutional layer to obtain node embedding vectors. These vectors are then input into the DRL agent to output a hierarchical partitioning scheme. The reward value corresponding to this partitioning scheme is calculated. The gradient of the reward value is backpropagated to the graph convolutional layer via backpropagation to update the weight matrix. Repeat step (2) until the reward value converges.

[0052] Regarding the bias vector It is related to the weight matrix The learnable parameters obtained through the end-to-end training process are optimized synchronously. The specific process is as follows: (1) Initialization: bias vector The initial value is zero-initialized, and its dimension is (2) Iterative update: In the backpropagation update of the graph convolutional layer, the bias vector With weight matrix Sharing the same gradient backpropagation path—when the reward value corresponding to the partitioning scheme output by the deep reinforcement learning (DRL) agent is backpropagated, the bias vector... It will adjust its own values ​​based on the "deviation between the feature output and the optimization target", and finally adjust them according to the weight matrix. This allows the feature output of the graph convolutional layer to better meet the task requirements of hierarchical and partitioned distribution networks.

[0053] 3. Perform multi-layer feature aggregation. The first layer uses the node feature vector generated in step S1. As initial input Local electrical characteristics are extracted through first-order neighborhood aggregation, and the output features are generated. The first layer focuses on the node's own load, voltage level, and impedance characteristics of adjacent lines to meet the needs of local node electrical characteristic analysis in distribution networks; the second layer is based on... Second-order neighborhood aggregation is performed to capture indirect connections in the topology, including the power transfer path characteristics between nodes connected by two lines in the distribution network; third-level and above high-order neighborhood aggregation breaks through local topology limitations and obtains topology and electrical association information in a wider range of the distribution network.

[0054] 4. Generate node embedding vectors, the feature vector output by the Lth layer (i.e., the last graph convolutional layer). L2 normalization This forms the final node embedding vector. This vector has two feature dimensions in the distribution network scenario. It retains the local micro characteristics of the node itself, such as load fluctuation rate and voltage deviation rate, while incorporating global macro characteristics such as the shortest path length from the node to the power source and the load distribution entropy of the feeder.

[0055] In one possible implementation, in step S4, for any of the sub-topology graph models, the embedding vectors of each node of each sub-topology graph model are respectively input to a preset partitioning agent, so that the partitioning agent partitions each sub-topology graph model under corresponding hierarchical constraints, generating corresponding sub-partitioning schemes, including: Based on the embedding vectors of each node in the sub-topology graph model and the initial partitioning progress parameters, construct the initial input vector; The initial input vector is input to the partitioning agent, so that the partitioning agent makes partitioning decisions on each node in the sub-topology graph model in sequence according to the preset policy network and the current level constraints, and finally generates the sub-partitioning scheme of the sub-topology graph model. In any partitioning decision process, the current state vector is obtained through the partitioning agent; based on the current state vector and the current level constraints, the partitioning action of the current node is generated through the policy network. The partitioning action includes assigning the current node to an existing partition, creating a new partition and assigning it to the current node, or dividing the current node and an adjacent node into a partition boundary; the current partitioning progress parameter in the current state vector is updated according to the partitioning action of the current node, thereby obtaining an updated state vector, and the updated state vector is sent to the next partitioning decision.

[0056] This application describes the specific decision-making process of a partitioning agent, including steps such as constructing an initial input vector, generating partitioning actions through a policy network, and updating the state vector. Based on a reinforcement learning model, the partitioning agent can dynamically select partitioning actions according to the current state vector and hierarchical constraints, such as assigning nodes to existing partitions, creating new partitions, or setting partition boundaries. This decision-making mechanism makes the partitioning process highly flexible and adaptable, capable of responding in real-time to changes in the power grid's operating state. By sequentially making partitioning decisions for nodes, the agent can gradually construct partitioning schemes and update partitioning progress parameters based on historical decision records, ensuring the continuity and completeness of the partitioning process. The initial input vector, combining node embedding vectors and partitioning progress parameters, provides the agent with comprehensive state information, enabling it to comprehensively consider electrical characteristics and partitioning progress for optimization decisions. The diversity of partitioning actions ensures the richness of partitioning schemes, meeting the needs of different operating scenarios. Overall, the introduction of the partitioning agent endows the partitioning method with self-learning and adaptive capabilities, automatically optimizing partitioning schemes and improving the efficiency and rationality of distribution network partitioning.

[0057] Furthermore, the step of updating the current partitioning progress parameter in the current state vector based on the partitioning action of the current node, thereby obtaining the updated state vector, includes: Based on the partitioning action of the current node and the historical decision records of the agent, determine the number of nodes that have been partitioned and the number of partitions that have been formed. The current partitioning progress parameter is updated based on the number of nodes that have been partitioned and the number of partitions that have been formed, thereby obtaining the updated state vector.

[0058] This application further refines the update mechanism for partitioning progress parameters. By determining the number of nodes that have been partitioned and the number of partitions formed through partitioning actions and historical decision records, the state vector is updated. This ensures that the partitioning agent can track the partitioning progress in real time during the decision-making process, avoiding duplicate or missed partitions and guaranteeing the integrity of the partitioning scheme. Updating the number of completed nodes and the number of partitions provides the agent with clear progress feedback, enabling it to adjust subsequent decision strategies based on the current partitioning state and optimize the partitioning order and boundary settings. This dynamic update mechanism enhances the controllability and predictability of the partitioning process, allowing the partitioning scheme to gradually approach the optimal solution. Simultaneously, the introduction of progress parameters provides a basis for calculating the reward function, helping the agent better evaluate the partitioning effect during training and accelerating the learning process. By finely managing partitioning progress, this method improves the efficiency and accuracy of partitioning decisions, providing strong support for generating scientific and reasonable partitioning schemes.

[0059] In a preferred embodiment, the partitioning intelligence uses the node embedding vector output by the GNN as the core state input during the decision-making process, and combines it with the partitioning progress information such as the current percentage of partitioned nodes and the current number of partitions to construct the current state vector as follows: ; In the formula: Let be the embedding vector of node i. This represents the partitioning progress parameter at step t (value range [0,1]). The action space defines the partition boundary division action as "assign node i to partition k" or "set nodes i and j as partition boundaries", and the action set is as follows: ; In the formula: The maximum number of zones is preset to 5-15 based on the scale of the distribution network.

[0060] The partitioning progress parameter θt is a comprehensive progress indicator calculated by normalization based on the proportion of partitioned nodes at step t and the current number of partitions. The specific formula is as follows: ; In the formula: Let be the number of nodes that have been partitioned in step t. This represents the total number of nodes in the distribution network. This indicates the "completion degree of node partitioning"; Let be the number of partitions formed in step t. This is the preset maximum number of partitions. This indicates the "completeness of the number of partitions"; , The weighting coefficients are ω1+ω2=1, usually ω1=0.6 and ω2=0.4, which can be adjusted according to actual needs. They are used to balance the progress ratio of node partitioning and the number of partitions.

[0061] Specifically, at the t-th step (t=1,2,...,n) decision-making, the partitioned agent bases its decision on the current state. Select Action (Assign node t to the target partition); after the action is executed, update the partition state and proceed to step t+1 for decision-making, until all nodes are partitioned; The partitioned agent checks constraints in real time during the decision-making process. If a constraint is not met, the current sequence is terminated and the solution is marked as invalid. In addition to the current-level constraints, the constraints also include relay protection constraints or power supply reliability constraints. The relay protection constraint requires the maximum short-circuit current within the partition to be: [Specific constraint details would be needed here]. ; In the formula: The maximum value of the three-phase short-circuit current of all nodes within the partition is calculated according to the "Code for Calculation of Short-Circuit Current of Power System" (DL / T 5429-2009). Specifically, based on the maximum operating mode of the system, the short-circuit current of each node is calculated through the node impedance matrix, and the maximum value is taken. The protection device's operating threshold is determined based on the protection device model parameters.

[0062] The power supply reliability constraint is applied to power distribution network scenarios and must meet the following requirement: the average power supply reliability rate of load points within the zone is: ; In the formula: The preset reliability threshold is set to a value no lower than 0.999.

[0063] For a partitioning scheme that satisfies the constraints under the current operating conditions, the scheme must include: The sub-partitioning scheme ultimately generated by the partitioning agent includes: a partitioning table that specifies the partition to which each node belongs (e.g., nodes 1-10 belong to partition 1, and nodes 11-20 belong to partition 2); and a boundary line list that marks the connecting lines between partitions (e.g., lines 5-6 and 15-16 are partition boundaries).

[0064] In one possible implementation, the construction of the partitioned agent based on a reinforcement learning model and training it using historical operating data from different distribution networks includes: An initial agent is constructed based on the reinforcement learning model, and the initial agent includes an initial policy network and an initial value network. Several training data sets are constructed based on the historical operation data of the different distribution networks. The training data sets include distribution network topology maps, corresponding node embedding vectors, manually optimized optimal partition action sequences, and reward values ​​corresponding to the optimal partition action sequences. Based on the aforementioned training data and a preset reward function, the internal parameters of the initial agent are updated several times using gradient descent to obtain the final partitioned agent.

[0065] This application provides a method for constructing and training a partitioned intelligent agent, including constructing an initial agent based on a reinforcement learning model, preparing training data, and updating parameters using gradient descent. The initial agent consists of an initial policy network and an initial value network, which are responsible for generating partitioned actions and evaluating state values, respectively, forming a complete reinforcement learning framework. The training data comes from historical operating data of different distribution networks, including topology graphs, node embedding vectors, manually optimized optimal partitioned action sequences, and corresponding reward values, ensuring the diversity and realism of the training process. Gradient descent is used to update the internal parameters of the agent in multiple rounds, enabling it to gradually learn the optimal partitioning strategy and adapt to the operating characteristics of different power grids. Overall, the training process of the partitioned intelligent agent significantly improves the intelligence level of the method and further enhances the efficiency and rationality of distribution network partitioning.

[0066] Furthermore, when iteratively updating the internal parameters of the current agent using gradient descent based on the aforementioned training data and a preset reward function to obtain the updated agent for the current round, the step of updating the internal parameters of the initial agent for several rounds using gradient descent based on the aforementioned training data and a preset reward function to obtain the final partitioned agent includes: A preset number of current training data are randomly selected from the plurality of training data; Construct corresponding initial state vectors based on the distribution network topology diagram of each current training data and the corresponding node embedding vectors. Each of the initial state vectors is input to the current agent, so that the current agent generates a corresponding decision action sequence and a state vector sequence respectively; Based on each decision action sequence, each state vector sequence, the reward function, the optimal partition action sequence in each of the current training data, and the corresponding reward value, the parameters of the current policy network and the current value network of the current agent are updated using the gradient descent method to obtain the updated agent for the current round.

[0067] This application further describes the specific training steps of the partitioning agent, including randomly sampling training data, constructing an initial state vector, generating a decision action sequence, and updating network parameters using gradient descent. Randomly sampling training data ensures the randomness and comprehensiveness of the training process, avoiding the impact of data bias on the agent's performance. The initial state vector is constructed based on the topology graph and node embedding vectors, providing the agent with an accurate input environment. The generation of the decision action sequence and state vector sequence simulates the entire partitioning decision-making process, allowing the agent to accumulate experience during training. By comparing the reward function, the optimal partitioning action sequence, and the reward value, gradient descent updates the parameters of the policy network and value network, gradually optimizing the agent's decision-making ability. This iterative training mechanism enables the agent to continuously improve its partitioning strategy, approaching the optimal solution optimized manually. Through systematic training steps, the partitioning agent ultimately possesses powerful partitioning capabilities, providing high-quality partitioning schemes for the distribution network.

[0068] Furthermore, the reward function is constructed from the electrical connection tightness index within the partition, the power exchange balance index between partitions, the constraint satisfaction reward, and the node partitioning completion reward.

[0069] This application's embodiments define the composition of the reward function, including an intra-partition electrical connectivity index, an inter-partition power exchange balance index, a constraint satisfaction reward, and a node partitioning completion reward. This multi-index reward function ensures that the partitioning agent can comprehensively consider multiple objectives during training, such as electrical performance, balance, constraint satisfaction, and partitioning integrity. The intra-partition electrical connectivity index encourages the agent to generate internally tightly connected partitions, improving fault isolation efficiency; the inter-partition power exchange balance index promotes power balance between partitions, optimizing power flow distribution; the constraint satisfaction reward ensures that the partitioning scheme meets engineering requirements such as relay protection and power supply reliability; and the node partitioning completion reward incentivizes the agent to complete the partitioning of all nodes, ensuring the integrity of the scheme. This comprehensive reward mechanism enables the agent to learn a partitioning strategy that considers multiple objectives, further improving the rationality of distribution network partitioning.

[0070] In a preferred embodiment, with the optimization objectives of tight electrical connections within the partition, balanced power exchange between partitions, and meeting relay protection and power supply reliability constraints, a DRL model is constructed to transform the distribution network partitioning task into a sequential decision problem. The node embeddings output by the GNN are used as state inputs, and the partition boundary division action is used as the decision output. The agent is guided to learn the optimal partitioning strategy through a reward function, thereby obtaining the partitioning agent. The specific training process is as follows: 1. Construct an initial partitioning agent. The initial partitioning agent adopts an Actor-Critic architecture. The Actor network is a 3-layer fully connected neural network. The input dimension is the state vector dimension, and the output dimension is the action space size, outputting the action probability distribution. The Critic network is a two-layer fully connected neural network that outputs state-action value. In a scenario where distribution network nodes process tasks sequentially, the partitioned agent decomposes the partitioned task into n decision steps (where n is the total number of distribution network nodes).

[0071] 2. Determine the optimization objectives of the intelligent agent, including tight electrical connections within the partition, balanced power exchange between partitions, and meeting the constraints of relay protection and power supply reliability. The tightness of electrical connections within the partition is measured by the average electrical distance between nodes within the partition, and the calculation formula is as follows: ; In the formula: M is the total number of node pairs within the partition. N is the total number of nodes in the partition; Let k be the electrical distance between the k-th pair of nodes; This represents the average electrical distance between all node pairs within the partition.

[0072] Inter-regional power exchange balance is assessed using the standard deviation of inter-regional tie-line power, and the specific calculation formula is as follows: ; In the formula: The total number of connecting lines, Let k be the power value of the k-th tie line. This represents the average power of the tie line. The relay protection constraints and power supply reliability constraints have already been given previously and will not be repeated here.

[0073] 3. Design a reward function for the agent's training process based on the optimization objective. The reward function integrates multi-dimensional indicators to guide the agent in learning the optimal strategy in the power distribution network scenario. The calculation formula is as follows: ; In the formula: Weighting coefficients ( ); To ensure the reward is met, the value is 1 if all constraints are met, otherwise -0.5. To account for the distribution network loss after zoning, To minimize potential network loss, an additional terminal reward is added when all nodes are partitioned (t=n). (M is the number of tie lines) (Maximum number of tie lines).

[0074] 4. Agent Training Process: An offline experience playback mechanism is adopted to record each decision... Data is stored in the experience pool; during each training session, a batch of data is randomly sampled from the experience pool, and the network parameters are updated using the loss function: Actor network loss. Critic network loss (Q' is the target Critic network output); iterative training is performed under typical distribution network conditions (such as peak load and off-peak load) until the partitioning scheme output by the agent satisfies the optimization objective for several consecutive rounds, where, This is the KL divergence penalty coefficient (value 0.1), used to avoid excessively large parameter update magnitudes.

[0075] Specifically, for typical operating conditions of the distribution network (such as peak load on weekdays, off-peak load on holidays, and seasonal load fluctuations), at least 1000 sets of historical operating data will be collected. Each set of data will include distribution network topology, node load, line power, and fault simulation data. The data will be divided into training and validation sets in a 7:3 ratio, with the training set used for model parameter updates and the validation set used for evaluating the training effect. The training sample format is defined as follows: ; In the formula: This is the distribution network topology diagram for the m-th data set. The node embedding matrix output by the GNN. This is the optimal partitioning action sequence optimized manually. This is the reward value for the corresponding action sequence.

[0076] Then, the offline training process is executed: the DRL model parameters are initialized, including the Actor network weights. Critic network weights The training batch size was set to 64, and the number of iterations was set to 500. In each training round, a batch of samples was randomly selected from the training set, and the parameters were updated using gradient descent. The average reward was calculated every 50 rounds using the validation set. When 3 consecutive rounds When the improvement is less than 1%, stop training and save the current model parameters (denoted as ). , In peak load scenarios, the load balance of the trained model should reach 0.85 or higher (balance = 1 - load standard deviation / average load).

[0077] The training process of a DRL model consists of the following 6 specific steps: (1) Sample sampling: 64 samples were randomly selected from the training set, and each sample contained a distribution network topology map G. m Node embedding matrix z mOptimal action sequence A m * Corresponding reward value R m ; (2) State and action input: embed the node of each sample into the matrix z. m Combined with the initial partitioning progress parameter θ0, it forms the state vector s t And extract the action a∈A from the sample. m ∗ ; (3) Forward computation of the Actor network: s t Input an Actor network, output the action probability distribution π(a|st), and record the current probability distribution π. old ; (4) Forward computation of the Critic network: s t The action 'a' is input into the Critic network, and the output state-action value Q(s,a) is output. (5) Loss calculation and gradient update: Calculate the losses of the Actor and Critic networks based on the aforementioned loss function. Then, using the gradient descent algorithm, backpropagate the gradient of the loss to the Actor and Critic networks to update the weights W. a W c ; (6) End of training: After the parameter update of this batch of samples is completed, the next batch of samples will be trained until all batches in this round are processed.

[0078] Furthermore, embodiments of this application can also perform online fine-tuning of deployed regional agents. During the operation of the distribution network, the operating status of the distribution network is monitored in real time, and online fine-tuning is triggered when any of the following conditions are met: When load fluctuates: the node load deviates from the historical same period by more than 20%, that is: ; In the formula: For the current load, This is the historical load for the same period.

[0079] Network reconfiguration: Changes in the state of line switches lead to changes in the topology, with the adjacency matrix changing by more than 10%, i.e.: ; Partition failure: The reward value of the current partitioning scheme is less than 70% of the optimal value during offline training, i.e.: ; When the above triggering conditions are met, an online fine-tuning mechanism is implemented: real-time data is collected within 30 minutes after the fine-tuning is triggered, a small batch of samples (10-20 groups) is generated, and the model is updated by freezing the pre-trained layers (fine-tuning only the output layer); the fine-tuning objective is to minimize the real-time reward bias. ; In the formula: This serves as the actual reward for real-time actions. Fine-tuning iterations should not exceed 10 rounds, ensuring parameter updates are completed within 1 minute. In network reconstruction scenarios, the time for the model to output the partitioning scheme after fine-tuning should be controlled within 20 seconds.

[0080] Furthermore, the method also includes a step to verify the effectiveness of the hierarchical partitioning scheme, specifically: Select typical operating scenarios of the distribution network, including normal operation, load change, single line fault, and multiple line fault, and apply the zoning scheme obtained in step S5 to the scenario simulation. Calculate and verify the following indicators: fault isolation time, the time from the occurrence of a fault to the determination of the isolation boundary; network loss reduction rate, the ratio of the difference between the network loss after partitioning and the network loss before partitioning to the network loss before partitioning; and relay protection coordination success rate, the ratio of the number of partitions that meet the protection action sequence requirements to the total number of partitions. The hierarchical partitioning scheme is deemed effective when the fault isolation time is ≤5 seconds, the network loss reduction rate is ≥8%, and the relay protection coordination success rate is ≥95% in all scenarios; otherwise, the reward function weight coefficients of the DRL model are readjusted and a new partitioning agent is relearned and generated.

[0081] In summary, the embodiments of this application have the following advantages over existing technologies: By capturing deep electrical connections and topological dependencies between nodes through multi-layer graph convolution operations of GNNs, the generated node embedding vectors simultaneously integrate global topological information and local electrical characteristics, solving the problem of loose electrical connections within partitions caused by traditional clustering algorithms relying solely on a single distance index, thus making the partitioning results more closely match the physical characteristics of the distribution network. Leveraging the sequential decision-making capability and offline training-online fine-tuning mechanism of the DRL model, it can quickly respond to dynamic operating conditions such as load fluctuations and network reconfiguration, and can output partitioning schemes adapted to new operating conditions in a short time. Furthermore, this embodiment innovatively designs a hierarchical logic of "high-level constraints on low-level," clearly defining the partitioning inclusion relationship and quantity ratio between the 10kV and 0.4kV levels, solving the problem of insufficient hierarchical coordination caused by the traditional method of "partitioning without hierarchical layering," and providing support for cross-level fault isolation. By quantifying constraints such as relay protection and power supply reliability into the reward function, the output scheme naturally meets engineering specifications, significantly improving the rationality, adaptability, and efficiency of distribution network partitioning.

[0082] Example 2: like Figure 2As shown, Embodiment 2 provides a power distribution network partitioning system based on graph neural networks and reinforcement learning, including a topology graph construction module 10, a hierarchy partitioning module 20, an information embedding module 30, a partitioning module 40, and a scheme merging module 50. The topology graph construction module 10 is used to construct a topology graph model containing node feature vectors and edge feature vectors based on the topology and electrical parameters of the power distribution network. The hierarchical division module 20 is used to perform hierarchical division of the topology model according to preset voltage level division conditions, and generate several sub-topology models with different voltage levels. The information embedding module 30 is used to aggregate the neighborhood information of the feature vectors of each node in each of the sub-topology graph models through a preset graph neural network, thereby generating the node embedding vector of each node in each of the sub-topology graph models. The partitioning module 40 is used to input the embedding vectors of each node of each of the sub-topology graph models into a preset partitioning agent, so that the partitioning agent partitions each of the sub-topology graph models according to a preset order under the corresponding hierarchical constraints, and generates each corresponding sub-partitioning scheme. The hierarchical constraints of any current sub-topology graph model are constructed based on the sub-partitioning scheme of the previous sub-topology graph model. The partitioning agent is constructed based on a reinforcement learning model and trained based on historical operation data of different distribution networks. The scheme merging module 50 is used to combine the various sub-partition schemes to obtain the partition scheme of the power distribution network.

[0083] Furthermore, the topology graph construction module 10 constructs a topology graph model containing node feature vectors and edge feature vectors based on the topology and electrical parameters of the distribution network, including: Based on the normalized value of apparent power of each node, voltage level information, node equipment type and load rate in the electrical parameters, a feature vector of each corresponding node is constructed. Based on the normalized values ​​of line resistance, line reactance, line length, connection relationship identifier, and line load rate of each node in the electrical parameters, the corresponding edge feature vectors are constructed. The topology graph model is constructed by combining the topology of the power distribution network, the feature vectors of each node, and the feature vectors of each edge.

[0084] Furthermore, the hierarchical division module 20 performs hierarchical division of the topology model according to preset voltage level division conditions, generating several sub-topology models with different voltage levels, including: Based on the voltage level classification conditions and the voltage level information in the feature vectors of each node, hierarchical labels are added to each node in the distribution network. Based on the node device type in the feature vector of each node, a number of transformer nodes are determined and voltage conversion node labels are added to each transformer node. The hierarchical affiliation of each node feature vector is determined based on the hierarchical labels, and the hierarchical boundary of the topology graph model is determined based on the voltage conversion node labels. Then, several sub-topology graph models with different voltage levels are generated based on the hierarchical affiliation and the hierarchical boundary.

[0085] In one possible implementation, for any of the sub-topology graph models, the information embedding module 30 aggregates the neighborhood information of the feature vectors of each node in each of the sub-topology graph models through a preset graph neural network, thereby generating the node embedding vector of each node in each of the sub-topology graph models, including: The sub-topology graph model is input into the graph neural network, so that the graph neural network updates the feature vectors of each node in the sub-topology graph model through several graph convolutional network layers, and finally generates the embedding vectors of each node in the sub-topology graph model. In any of the iteration update processes, the first node embedding vectors output by the previous graph convolutional layer are input to the current graph convolutional layer, so that the current graph convolutional layer aggregates the neighborhood information of each first node embedding vector according to its own weight matrix and the current edge feature vectors of the current sub-topology graph model, and then updates each first node embedding vector based on the neighborhood information to generate each corresponding second node embedding vector and input it to the next graph convolutional layer.

[0086] In one possible implementation, for any of the sub-topology graph models, the partitioning module 40 inputs the embedding vectors of each node of each sub-topology graph model into a preset partitioning agent, so that the partitioning agent partitions each sub-topology graph model under corresponding hierarchical constraints, generating corresponding sub-partitioning schemes, including: Based on the embedding vectors of each node in the sub-topology graph model and the initial partitioning progress parameters, construct the initial input vector; The initial input vector is input to the partitioning agent, so that the partitioning agent makes partitioning decisions on each node in the sub-topology graph model in sequence according to the preset policy network and the current level constraints, and finally generates the sub-partitioning scheme of the sub-topology graph model. In any partitioning decision process, the current state vector is obtained through the partitioning agent; based on the current state vector and the current level constraints, the partitioning action of the current node is generated through the policy network. The partitioning action includes assigning the current node to an existing partition, creating a new partition and assigning it to the current node, or dividing the current node and an adjacent node into a partition boundary; the current partitioning progress parameter in the current state vector is updated according to the partitioning action of the current node, thereby obtaining an updated state vector, and the updated state vector is sent to the next partitioning decision.

[0087] Furthermore, the step of updating the current partitioning progress parameter in the current state vector based on the partitioning action of the current node, thereby obtaining the updated state vector, includes: Based on the partitioning action of the current node and the historical decision records of the agent, determine the number of nodes that have been partitioned and the number of partitions that have been formed. The current partitioning progress parameter is updated based on the number of nodes that have been partitioned and the number of partitions that have been formed, thereby obtaining the updated state vector.

[0088] In one possible implementation, the construction of the partitioned agent based on a reinforcement learning model and training it using historical operating data from different distribution networks includes: An initial agent is constructed based on the reinforcement learning model, and the initial agent includes an initial policy network and an initial value network. Several training data sets are constructed based on the historical operation data of the different distribution networks. The training data sets include distribution network topology maps, corresponding node embedding vectors, manually optimized optimal partition action sequences, and reward values ​​corresponding to the optimal partition action sequences. Based on the aforementioned training data and a preset reward function, the internal parameters of the initial agent are updated several times using gradient descent to obtain the final partitioned agent.

[0089] Furthermore, when iteratively updating the internal parameters of the current agent using gradient descent based on the aforementioned training data and a preset reward function to obtain the updated agent for the current round, the step of updating the internal parameters of the initial agent for several rounds using gradient descent based on the aforementioned training data and a preset reward function to obtain the final partitioned agent includes: A preset number of current training data are randomly selected from the plurality of training data; Construct corresponding initial state vectors based on the distribution network topology diagram of each current training data and the corresponding node embedding vectors. Each of the initial state vectors is input to the current agent, so that the current agent generates a corresponding decision action sequence and a state vector sequence respectively; Based on each decision action sequence, each state vector sequence, the reward function, the optimal partition action sequence in each of the current training data, and the corresponding reward value, the parameters of the current policy network and the current value network of the current agent are updated using the gradient descent method to obtain the updated agent for the current round.

[0090] Furthermore, the reward function is constructed from the electrical connection tightness index within the partition, the power exchange balance index between partitions, the constraint satisfaction reward, and the node partitioning completion reward.

[0091] This application provides a power distribution network zoning system based on graph neural networks and reinforcement learning. Through steps such as constructing a topology graph model, hierarchical partitioning, node embedding vector generation, and zoning agent decision-making, it achieves scientific and efficient zoning of the power distribution network. First, this embodiment constructs a topology graph model containing node feature vectors and edge feature vectors by deeply integrating the power grid's topology and real-time electrical parameters, thus comprehensively reflecting the physical connections and electrical characteristics of the power distribution network. Second, the topology graph is hierarchically partitioned using preset voltage level division conditions, generating sub-topology graph models for different voltage levels. This effectively solves the problem of insufficient cross-level zoning coordination in traditional methods, making the boundaries between higher-level and lower-level zoning more closely matched, and improving the efficiency of fault coordination isolation and load transfer. Next, the graph neural network aggregates the neighborhood information of nodes in each sub-topology graph to generate node embedding vectors. This deeply captures the dynamic and complex electrical coupling relationships between nodes, avoiding the loose internal electrical connections caused by traditional clustering algorithms relying solely on static parameters. Finally, by using a partitioning agent trained based on reinforcement learning, each sub-topology is partitioned sequentially in descending order of voltage level. The sub-partitioning scheme is generated by combining hierarchical constraints. While ensuring that each layer performs partitioning independently, the partitioning results of high voltage levels are used to constrain the partitioning process of secondary sub-topologies. This achieves collaborative partitioning of each layer of the distribution network, significantly improving the rationality, adaptability, and efficiency of distribution network partitioning.

[0092] For a more detailed explanation of the working principle and procedures of this embodiment, please refer to the relevant description in Embodiment 1.

[0093] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of this application. It should be understood that the above descriptions are merely specific embodiments of this application and are not intended to limit the scope of protection of this application. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application for those skilled in the art.

Claims

1. A distribution network zoning method based on graph neural networks and reinforcement learning, characterized in that, include: Based on the topology and electrical parameters of the distribution network, a topological graph model containing node feature vectors and edge feature vectors is constructed. Based on the preset voltage level division conditions, the topology model is hierarchically divided to generate several sub-topology models with different voltage levels. The neighborhood information of the feature vectors of each node in each of the sub-topology graph models is aggregated by a preset graph neural network, thereby generating the node embedding vector of each node in each of the sub-topology graph models. The embedding vectors of each node of each of the sub-topology graph models are input into a preset partitioning agent, so that the partitioning agent partitions each of the sub-topology graph models according to a preset order under the corresponding hierarchical constraints, and generates each corresponding sub-partitioning scheme. The hierarchical constraints of any current sub-topology graph model are constructed based on the sub-partitioning scheme of the previous sub-topology graph model. The partitioning agent is constructed based on a reinforcement learning model and trained based on historical operation data of different distribution networks. By combining the various sub-partition schemes, the partition scheme of the power distribution network is obtained.

2. The distribution network partitioning method based on graph neural networks and reinforcement learning as described in claim 1, characterized in that, The construction of a topology graph model containing node feature vectors and edge feature vectors based on the topology and electrical parameters of the distribution network includes: Based on the normalized value of apparent power of each node, voltage level information, node equipment type and load rate in the electrical parameters, a feature vector of each corresponding node is constructed. Based on the normalized values ​​of line resistance, line reactance, line length, connection relationship identifier, and line load rate of each node in the electrical parameters, the corresponding edge feature vectors are constructed. The topology graph model is constructed by combining the topology of the power distribution network, the feature vectors of each node, and the feature vectors of each edge.

3. The distribution network partitioning method based on graph neural networks and reinforcement learning as described in claim 1, characterized in that, The step of hierarchically dividing the topology model according to preset voltage level division conditions to generate several sub-topology models with different voltage levels includes: Based on the voltage level classification conditions and the voltage level information in the feature vectors of each node, hierarchical labels are added to each node in the distribution network. Based on the node device type in the feature vector of each node, a number of transformer nodes are determined and voltage conversion node labels are added to each transformer node. The hierarchical affiliation of each node feature vector is determined based on the hierarchical labels, and the hierarchical boundary of the topology graph model is determined based on the voltage conversion node labels. Then, several sub-topology graph models with different voltage levels are generated based on the hierarchical affiliation and the hierarchical boundary.

4. The distribution network partitioning method based on graph neural networks and reinforcement learning as described in claim 1, characterized in that, For any of the sub-topology graph models, the step of aggregating the neighborhood information of the feature vectors of each node in each sub-topology graph model through a preset graph neural network, and then generating the node embedding vector of each node in each sub-topology graph model, includes: The sub-topology graph model is input into the graph neural network, so that the graph neural network updates the feature vectors of each node in the sub-topology graph model through several graph convolutional network layers, and finally generates the embedding vectors of each node in the sub-topology graph model. In any of the iteration update processes, the first node embedding vectors output by the previous graph convolutional layer are input to the current graph convolutional layer, so that the current graph convolutional layer aggregates the neighborhood information of each first node embedding vector according to its own weight matrix and the current edge feature vectors of the current sub-topology graph model, and then updates each first node embedding vector based on the neighborhood information to generate each corresponding second node embedding vector and input it to the next graph convolutional layer.

5. The distribution network partitioning method based on graph neural networks and reinforcement learning as described in claim 1, characterized in that, For any of the sub-topology graph models, the step of inputting the embedding vectors of each node of each sub-topology graph model into a preset partitioning agent, so that the partitioning agent partitions each sub-topology graph model under corresponding hierarchical constraints, and generates corresponding sub-partitioning schemes, includes: Based on the embedding vectors of each node in the sub-topology graph model and the initial partitioning progress parameters, construct the initial input vector; The initial input vector is input to the partitioning agent, so that the partitioning agent makes partitioning decisions on each node in the sub-topology graph model in sequence according to the preset policy network and the current level constraints, and finally generates the sub-partitioning scheme of the sub-topology graph model. In any partitioning decision process, the current state vector is obtained through the partitioning agent; based on the current state vector and the current level constraints, the partitioning action of the current node is generated through the policy network. The partitioning action includes assigning the current node to an existing partition, creating a new partition and assigning it to the current node, or dividing the current node and an adjacent node into a partition boundary; the current partitioning progress parameter in the current state vector is updated according to the partitioning action of the current node, thereby obtaining an updated state vector, and the updated state vector is sent to the next partitioning decision.

6. The distribution network partitioning method based on graph neural networks and reinforcement learning as described in claim 5, characterized in that, The step of updating the current partitioning progress parameter in the current state vector based on the partitioning action of the current node, thereby obtaining the updated state vector, includes: Based on the partitioning action of the current node and the historical decision records of the agent, determine the number of nodes that have been partitioned and the number of partitions that have been formed. The current partitioning progress parameter is updated based on the number of nodes that have been partitioned and the number of partitions that have been formed, thereby obtaining the updated state vector.

7. The distribution network partitioning method based on graph neural networks and reinforcement learning as described in claim 1, characterized in that, The method of constructing the regional agent based on a reinforcement learning model and training it using historical operating data from different power distribution networks includes: An initial agent is constructed based on the reinforcement learning model, and the initial agent includes an initial policy network and an initial value network. Several training data sets are constructed based on the historical operation data of the different distribution networks. The training data sets include distribution network topology maps, corresponding node embedding vectors, manually optimized optimal partition action sequences, and reward values ​​corresponding to the optimal partition action sequences. Based on the aforementioned training data and a preset reward function, the internal parameters of the initial agent are updated several times using gradient descent to obtain the final partitioned agent.

8. The distribution network partitioning method based on graph neural networks and reinforcement learning as described in claim 7, characterized in that, When the internal parameters of the current agent are iteratively updated using gradient descent based on the aforementioned training data and a preset reward function to obtain the updated agent for the current round, the step of updating the internal parameters of the initial agent for several rounds using gradient descent based on the aforementioned training data and a preset reward function to obtain the final partitioned agent includes: A preset number of current training data are randomly selected from the plurality of training data; Construct corresponding initial state vectors based on the distribution network topology diagram of each current training data and the corresponding node embedding vectors. Each of the initial state vectors is input to the current agent, so that the current agent generates a corresponding decision action sequence and a state vector sequence respectively; Based on each decision action sequence, each state vector sequence, the reward function, the optimal partition action sequence in each of the current training data, and the corresponding reward value, the parameters of the current policy network and the current value network of the current agent are updated using the gradient descent method to obtain the updated agent for the current round.

9. The distribution network partitioning method based on graph neural networks and reinforcement learning as described in claim 7, characterized in that, The reward function is constructed from the electrical connection tightness index within the partition, the power exchange balance index between partitions, the constraint satisfaction reward, and the node partitioning completion reward.

10. A power distribution network zoning system based on graph neural networks and reinforcement learning, characterized in that, It includes a topology graph construction module, a hierarchy partitioning module, an information embedding module, a partitioning module, and a scheme merging module; The topology graph construction module is used to construct a topology graph model containing node feature vectors and edge feature vectors based on the topology and electrical parameters of the power distribution network. The hierarchical division module is used to perform hierarchical division of the topology model according to preset voltage level division conditions, and generate several sub-topology models with different voltage levels. The information embedding module is used to aggregate the neighborhood information of the feature vectors of each node in each of the sub-topology graph models through a preset graph neural network, thereby generating the node embedding vector of each node in each of the sub-topology graph models. The partitioning module is used to input the embedding vectors of each node of each of the sub-topology graph models into a preset partitioning agent, so that the partitioning agent partitions each of the sub-topology graph models according to a preset order under the corresponding hierarchical constraints, and generates each corresponding sub-partitioning scheme. The hierarchical constraints of any current sub-topology graph model are constructed based on the sub-partitioning scheme of the previous sub-topology graph model. The partitioning agent is constructed based on a reinforcement learning model and trained based on historical operation data of different distribution networks. The scheme merging module is used to combine the various sub-partition schemes to obtain the partition scheme of the distribution network.

Images