Active support-based distribution network distributed resource multi-relationship graph clustering method

Abstract

This invention provides a multi-relationship graph clustering method for distributed resources in power distribution networks based on active support, belonging to the field of artificial intelligence modeling and optimization technology for power systems. This invention constructs active support capability relationship views, electrical coupling relationship views, and node attribute relationship views. It utilizes a multi-relationship graph neural network to jointly model and learn multi-view information, achieving feature representation and structural fusion of distributed resources in a unified embedding space. By introducing a soft clustering module and a soft modular optimization objective function, and combining view consistency constraints, active support capability constraints, and cluster equilibrium constraints, it achieves joint optimization of the clustering process and physical characteristic constraints, thereby realizing adaptive clustering based on active support characteristics. This invention can form a distributed resource aggregate model with physical consistency and can be used to evaluate the inertia support capability, frequency support capability, and voltage support capability of the aggregate, providing a theoretical basis for improving the active support capability and scheduling optimization of power distribution networks.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence modeling and optimization technology for power systems, and particularly relates to a multi-relationship graph clustering method for distributed resources in distribution networks based on active support. Background Technology

[0002] With the large-scale integration of distributed resources such as distributed photovoltaics, energy storage, electric vehicles, and controllable loads into distribution networks, power grid operation exhibits significant decentralization and heterogeneity. A large number of distributed resources differ significantly in capacity, control strategies, and dynamic response characteristics, making traditional resource modeling methods based on equivalent modeling or static classification unable to accurately reflect their dynamic operating characteristics and active support capabilities. Existing research methods mainly suffer from the following shortcomings: (1) Modeling based on a single feature: Existing clustering methods are mostly based on electrical distance or capacity parameters for division, ignoring the differences in control strategies and dynamic support capabilities of distributed resources, resulting in insufficient physical interpretability of clustering results.

[0003] (2) Lack of multi-source information fusion capability: Multi-dimensional information such as electrical topology relationship, control strategy characteristics and active support capability are usually modeled independently, failing to achieve joint learning in a unified expression space.

[0004] (3) Difficult to characterize dynamic operating characteristics: Traditional static indicators cannot effectively describe the dynamic response capability of distributed resources under frequency and voltage disturbances, resulting in large deviations in aggregate modeling.

[0005] (4) Clustering process is disconnected from physical constraints: Existing machine learning clustering methods are mostly data-driven and lack physical constraints of power systems (such as inertia support, reactive power regulation capability, etc.), making them difficult to apply to engineering scheduling scenarios. Therefore, there is an urgent need for a multi-source information modeling method that can integrate electrical relationships, control characteristics and active support capabilities to achieve adaptive clustering and aggregate equivalent modeling of distributed resources. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, this invention provides a clustering method for distributed resource multi-relationship graphs in distribution networks based on active support, thereby solving problems such as single-feature modeling, lack of multi-source information fusion, and insufficient physical constraints in existing technologies.

[0007] To achieve the above objectives, the technical solution adopted by this invention is: a clustering method for multiple relationship graphs of distributed resources in a distribution network based on active support, comprising the following steps: S1. Collect distribution network topology information and distributed resource operation parameters, and construct a distributed resource node information model that includes resource access location, capacity configuration, control strategy and operation status; S2. Generate different resource configurations and operating conditions based on the node information model, and construct an offline database of typical operating scenarios; obtain the dynamic response characteristics of resources through offline electromagnetic transient simulation, establish the mapping relationship between operating status and active support capability indicators, and generate training datasets based on the mapping relationship; wherein, each training data includes randomly generated steady-state operating data and dynamic response data generated according to the mapping relationship; S3. Based on the node information model and active support capability indicators, construct the active support capability relationship diagram G respectively. 1 Electrical coupling diagram G 2 and node attribute relationship graph G 3 These are then integrated to form a multi-relationship graph structure; each training data point corresponds to an active support capability relationship graph G. 1 Electrical coupling diagram G 2 and node attribute relationship graph G 3 ; S4. Using the multi-relation graph structure as the input of the multi-relation graph neural network model, the features of the multi-relation graph structure are extracted by utilizing the multi-relation graph neural network model to obtain the comprehensive representation after multi-relation fusion embedding; S5. Optimize the joint loss function of the multi-relationship graph neural network model. When the joint loss function converges or reaches the maximum number of iterations, output the final clustering result based on the comprehensive representation to achieve adaptive clustering of distributed resources.

[0008] The beneficial effects of this invention are as follows: By constructing an active support capability relationship view, an electrical coupling relationship view, and a node attribute relationship view, and utilizing a multi-relationship graph neural network model to jointly model and learn multi-view information, this invention achieves feature representation and structural fusion of distributed resources in a unified embedding space. Furthermore, by introducing a soft clustering module and a soft modularization optimization objective function, and combining view... Figure 1 By incorporating consistency constraints, active support capability constraints, and cluster equilibrium constraints, this invention achieves joint optimization of the clustering process and physical characteristic constraints, thereby realizing adaptive clustering based on active support characteristics. Ultimately, this invention can form a distributed resource aggregate model with physical consistency and be used to evaluate the aggregate's inertia support capability, frequency support capability, and voltage support capability, providing a theoretical basis for improving the active support capability and scheduling optimization of distribution networks.

[0009] Further, S1 includes: Collect power distribution network topology information and distributed resource operation parameters; Based on the collected data, the distribution network is represented as a topology network consisting of a set of nodes and line connections, and the characteristics of heterogeneous resource nodes are constructed according to the distributed resource type, resource access location, capacity configuration and control strategy. Based on the characteristics of heterogeneous resource nodes, the resource attribute characteristics, control attribute characteristics, and operational status attribute characteristics of each node are combined to form a distributed resource node information model, the expression of which is: ] Where n represents the number of nodes in the distribution network, and the information vector of the i-th node is represented as:

[0010] in, , and These represent the resource attributes, control attributes, and running status attributes of the i-th node, respectively.

[0011] The beneficial effects of the above-mentioned further solutions are: by establishing a node information model for heterogeneous distributed resources, resource attributes, control characteristics and operating status are uniformly represented, realizing the feature fusion and differentiated description of different types of distributed resources, and providing reliable data support for the subsequent construction of active support capability relationship graphs and multi-relationship graph neural network clustering.

[0012] Furthermore, the construction of the offline database includes: Different resource configurations and operating conditions are generated based on the node information model; Based on the generated data, identify the operational variables that affect proactive support capabilities. Based on runtime variables Randomly generate typical operating combinations; Simulations were conducted under various typical disturbance conditions, and the active support capability indices corresponding to these conditions were calculated based on the simulation results. Based on typical operating combinations and corresponding active support capability indicators, an offline database with a one-to-one correspondence between steady-state operating condition characteristics and active support capability indicators is established as the basis for generating active support capability indicators under new operating conditions, thus completing the construction of the offline database.

[0013] Furthermore, the generation of the training dataset includes: The dynamic response characteristics of resources are obtained through offline electromagnetic transient simulation; Based on the dynamic response characteristics of resources, for newly generated stable power flow conditions, the corresponding running state variables are extracted. ; Based on extracted running state variables The system retrieves the closest running state from the offline database, where the running state similarity is... The expression is:

[0014] in, This represents the m-th runtime variable in the offline database. Based on the search results, the k samples with the lowest similarity were selected, and the active support capability index was calculated. Among them, the active support capability indicator Used to construct an active support capability relationship diagram in S3; Establish operational status and proactive support capability indicators The mapping relationship between them is established, and a training dataset is generated based on the mapping relationship; where, when the running state variable... When the similarity of the running state of the samples with the original samples in the offline database exceeds a set threshold, a dynamic simulation is performed again, and the simulated samples are added to the offline database to dynamically update the offline sample library.

[0015] The beneficial effects of the above-mentioned further scheme are: by establishing a mapping relationship between operating state variables and active support capability indicators, the active support capability indicators are generated rapidly, avoiding a large number of repetitive dynamic simulation calculations and improving the efficiency of training sample construction; at the same time, the sample coverage is enhanced through the offline database dynamic update mechanism.

[0016] Furthermore, the active support capability relationship diagram G 1 The inertia support capability index, frequency support capability index, and voltage support capability index are expressed as follows:

[0017]

[0018]

[0019]

[0020] in, Represents a node The active support capability feature vector, This represents the virtual inertia coefficient of node i. This represents the change in active power during the disturbance period. Indicates the rate of change of frequency. and All represent weighting coefficients. Indicator representing inertia support capability. Indicators representing frequency support capabilities This indicates the voltage support capability index. This represents the active frequency droop coefficient of node i. and Let represent the adjustable active power limit and the current active power of node i, respectively. Represents the frequency response time constant. , and All represent the indicator weight coefficients. This represents the reactive voltage droop factor. and These represent the adjustable reactive power limit and the current reactive power of node i, respectively. Indicates node voltage sensitivity. , and All of these represent the indicator weighting coefficients.

[0021] The beneficial effects of the above-mentioned further scheme are as follows: By constructing a multi-dimensional active support capability index that includes inertia support capability, frequency support capability, and voltage support capability, a comprehensive characterization of the active support characteristics of distributed resources under different control methods and operating states is achieved. Specifically, the inertia support index, combining virtual inertia parameters and disturbance response characteristics, reflects the resource's ability to rapidly support system frequency changes; the frequency support index, comprehensively considering active power margin, frequency control characteristics, and response speed, accurately describes the resource's ability to participate in frequency regulation; and the voltage support index, integrating reactive power regulation capability, voltage control characteristics, and node voltage sensitivity, reflects the resource's ability to regulate local voltage fluctuations. Constructing an active support capability feature space through these multi-dimensional indices improves the distinguishability of functional characteristics of different types of distributed resources, providing effective physical feature basis for subsequent active support capability relationship diagram construction and resource adaptive clustering.

[0022] Furthermore, the electrical coupling relationship view G 2 The electrical coupling strength between nodes is quantified by a weighted sum of active-voltage coupling distance and reactive-voltage coupling distance.

[0023] in, Represents a node With nodes The combined electrical distance between them The active-voltage coupling distance is represented by the node. When injecting unit active power, the node The amount of voltage change. The reactive-voltage coupling distance is represented by the node. node when injecting unit reactive power The amount of voltage change. and All represent weighting coefficients, satisfying ; Node attribute relationship graph G 3 The characteristic index expression is as follows:

[0024]

[0025] in, Represents a node and nodes Attribute similarity distance between them and All represent weighting coefficients, satisfying , Indicates the distance attribute of the control strategy. Represents a node and nodes Hardware attribute similarity, Indicates the number of hardware attribute parameters. and Representing nodes respectively and nodes The kth hardware attribute parameter.

[0026] The beneficial effects of the above-mentioned further scheme are: by constructing an electrical coupling relationship view, the electrical influence relationship between different distributed resource nodes can be effectively reflected, improving the perception ability of clustering results on the physical connection characteristics of the distribution network. At the same time, by constructing a node attribute relationship view, the differences in resource hardware attributes and control strategies are integrated, enabling the subsequent clustering process to simultaneously consider resource functional similarity, electrical coupling characteristics, and node attribute correlation relationships.

[0027] Furthermore, the expression for the multi-relationship graph structure is as follows:

[0028]

[0029]

[0030] in, This represents an undirected weighted graph of the r-th view. Represents a set of nodes. Indicates the first A set of edges for each view. Indicates the first The adjacency weight matrix of each view, its elements Represents a node With nodes Similarity under the r-th view ;

[0031] in, Represents a node With nodes Similarity under the active support capability view and Representing nodes respectively The active support capability feature vector of node j, This represents the Gaussian kernel width parameter;

[0032] in, Represents a node With nodes Similarity in the electrical coupling relationship view Represents a node With nodes The combined electrical distance between them This represents the Gaussian kernel width parameter;

[0033] in, Represents a node With nodes Similarity in the node attribute relationship view Represents a node With nodes Attribute similarity distance between them This represents the Gaussian kernel width parameter.

[0034] The beneficial effects of the above-mentioned further scheme are: by using the Gaussian kernel function to convert the relationships between nodes under different views into graph structure weights, the ability to fuse multi-source heterogeneous information is improved, providing effective structural feature support for subsequent multi-relationship graph neural network clustering, and improving the accuracy and interpretability of clustering results.

[0035] Furthermore, the multi-relationship graph neural network model includes an input layer, a multi-relationship feature extraction layer, a cross-relationship attention fusion layer, and an output layer; The input layer is used to receive the distributed resource node information matrix and the corresponding active support capability relationship diagram, electrical coupling relationship diagram and node attribute relationship diagram to form multi-relationship diagram input data; The multi-relationship feature extraction layer comprises multiple relation graph convolutional networks. Based on the multi-relation graph input data, it performs node feature propagation and aggregation for different relation graphs, extracting node feature representations under different relation views. The calculation process is as follows:

[0036] in, This represents the set of neighboring nodes of node i in the r-th view. and All represent the r-th view at the th time.l The trainable parameters of the layer, This represents the feature representation of neighbor node j in the previous layer. Represents a nonlinear activation function; The cross-relationship attention fusion layer is used to represent node features based on each relationship view. The system adaptively learns the importance of different relation views and fuses them according to the feature weights of each relation to obtain a unified node representation. :

[0037] Where R represents the number of views, Indicates that node i is at the 1st rank. l The attention weight of the layer to the r-th view. This indicates that node i, under the r-th relation view, has passed through the [missing information - likely a specific point or context]. l Node feature representation obtained after layer relation graph convolutional network; The output layer is used to represent unified nodes. The model is mapped to a low-dimensional feature space, and the probability distribution of nodes belonging to different categories is calculated through a soft clustering layer; the Student-t distribution is used to define the soft clustering probability.

[0038] in, Represents a node The probability of belonging to class k. This represents the low-dimensional embedding representation of node i learned through a multi-relation graph neural network model. This represents the center vector of the k-th cluster. Let P represent the center vector of the j-th cluster, where j represents the cluster index, and P represents the initial total number of clusters. This represents the degree of freedom parameter.

[0039] The beneficial effects of the above-mentioned further scheme are as follows: by constructing a multi-relation graph neural network containing a multi-relation feature extraction layer, a cross-relation attention fusion layer, and a soft clustering output layer, joint learning and adaptive fusion of information from different relation views are achieved; feature representations of resource nodes under different relations are extracted by relation-specific graph convolution, and the contribution of each relation is dynamically adjusted through an attention mechanism, thereby improving the utilization capability of heterogeneous relation information; and the joint optimization of node representation learning and clustering process is achieved by combining soft clustering mapping, avoiding the problem of traditional clustering methods relying on manual feature selection.

[0040] Furthermore, the joint optimization objective function of the multi-relationship graph neural network model is expressed as follows:

[0041]

[0042]

[0043]

[0044]

[0045]

[0046] in, This represents the joint optimization objective function of a multi-relationship graph neural network model. This represents a modularity index for soft clustering. Information entropy constraint The weight, Indicates vision Figure 1 Consistency constraints The weight, Indicates active support capability constraints The weight, This represents the total number of edges in the graph. Represents the elements of the adjacency matrix. and Representing nodes respectively The degree of node j, The total number of clusters is represented by , and k represents the index of the summation for each cluster. and Let represent the membership degrees of node i and node j to the k-th cluster category, respectively. and Representing nodes respectively In the The representation of the first view and the s-th view. Indicates the first The set of all nodes in each cluster category Represents a node The active support capability feature vector, Indicates the first The average vector of all nodes in each cluster category.

[0047] The beneficial effects of the above-mentioned further solutions are: the soft clustering modularity index can enhance the internal connectivity of similar resources, and the information entropy constraint improves the discriminativeness of cluster assignment. Figure 1 Consistency constraints ensure the consistency of node representations across different relational views, while proactive support capability constraints enhance the physical functional consistency of clustering results. By optimizing network parameters and cluster centers through multi-objective joint constraints, distributed resources with similar proactive support characteristics are automatically grouped into the same category. Attached Figure Description

[0048] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0049] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0050] Example like Figure 1 As shown, this invention provides a clustering method for multiple relationship graphs of distributed resources in a distribution network based on active support, and its implementation method is as follows: S1. Collect distribution network topology information and distributed resource operation parameters, and construct a distributed resource node information model that includes resource access location, capacity configuration, control strategy, and operating status. The implementation method is as follows: Collect power distribution network topology information and distributed resource operation parameters; Based on the collected data, the distribution network is represented as a topology network consisting of a set of nodes and line connections, and the characteristics of heterogeneous resource nodes are constructed according to the distributed resource type, resource access location, capacity configuration and control strategy. Based on the characteristics of heterogeneous resource nodes, the resource attribute characteristics, control attribute characteristics, and operating status attribute characteristics of each node are combined to form a distributed resource node information model. S2. Generate various operating scenarios based on the node information model and establish an offline scenario database. Establish a mapping relationship between resource operating status and active support capability indicators through electromagnetic transient simulation, and generate training datasets in batches. Each training dataset includes randomly generated steady-state operating data and dynamic response data generated according to the mapping relationship. The construction of an offline database includes: Different resource configurations and operating conditions are generated based on the node information model; Based on the generated data, identify the operational variables that affect proactive support capabilities. Based on runtime variables Randomly generate typical operating combinations; Simulations were conducted under various typical disturbance conditions, and the active support capability indices corresponding to these conditions were calculated based on the simulation results. Based on typical operating combinations and corresponding active support capability indicators, an offline database with a one-to-one correspondence between steady-state operating condition characteristics and active support capability indicators is established as the basis for generating active support capability indicators under new operating conditions, thus completing the construction of the offline database. Generate a training dataset, including: The dynamic response characteristics of resources are obtained through offline electromagnetic transient simulation; Based on the dynamic response characteristics of resources, for newly generated stable power flow conditions, the corresponding running state variables are extracted. ; Based on extracted running state variables Retrieve the closest running status from the offline database; Based on the search results, the k samples with the lowest similarity were selected, and the active support capability index was calculated. Among them, the active support capability indicator Used to construct an active support capability relationship diagram in S3; Establish operational status and proactive support capability indicators The mapping relationship between them is established, and a training dataset is generated based on the mapping relationship; where, when the running state variable... When the similarity of the running state of the original samples in the offline database exceeds the set threshold, a dynamic simulation is performed again, and the simulated samples are added to the offline database to dynamically update the offline sample library. S3. Based on the node information model and active support capability indicators, construct the active support capability relationship diagram G respectively. 1 Electrical coupling diagram G 2 and node attribute relationship graph G 3 These are then integrated to form a multi-relationship graph structure; each training data point corresponds to an active support capability relationship graph G. 1 Electrical coupling diagram G 2 and node attribute relationship graph G 3 ; S4. Using the multi-relationship graph structure as the input of the multi-relationship graph neural network model, the features of the multi-relationship graph structure are extracted by the multi-relationship graph neural network model to obtain the comprehensive representation after multi-relationship fusion embedding; the multi-relationship graph neural network model includes an input layer, a multi-relationship feature extraction layer, a cross-relationship attention fusion layer, and an output layer; The input layer is used to receive the distributed resource node information matrix and the corresponding active support capability relationship diagram, electrical coupling relationship diagram and node attribute relationship diagram to form multi-relationship diagram input data; The multi-relationship feature extraction layer comprises multiple relation graph convolutional networks. Based on the multi-relation graph input data, it performs node feature propagation and aggregation for different relation graphs, extracting node feature representations under different relation views. ; The cross-relationship attention fusion layer is used to represent node features based on each relationship view. The algorithm adaptively learns the importance of different relation views and fuses them according to the feature weights of each relation to obtain a comprehensive representation after multi-relation fusion embedding. ; The output layer is used to embed the comprehensive representation after multiple relations are fused. Map to a low-dimensional feature space and calculate the probability distribution of nodes belonging to different categories through a soft clustering layer; S5. Optimize the joint loss function of the multi-relationship graph neural network model. When the joint loss function converges or reaches the maximum number of iterations, output the final clustering result based on the comprehensive representation to achieve adaptive clustering of distributed resources.

[0051] In this embodiment, each data point in the training dataset contains randomly generated steady-state operating data and dynamic response data generated according to the mapping relationship. For each randomly generated steady-state operating scenario, corresponding dynamic response data is generated according to the mapping relationship. Each "steady-state data + dynamic response data" data point uses the multi-relationship graph construction method in S3 (the feature indicators under each of the three views are calculated separately) to form three relationship graph structures (one data point corresponds to three relationship graphs: active support capability relationship graph G). 1 Electrical coupling diagram G 2 and node attribute relationship graph G 3 The three relationship graphs are input into the multi-relationship graph neural network model to obtain the fused embedding representation of the three relationship graphs, which is a comprehensive expression of "containing multiple relationships" for each data point. S5 optimizes multiple objectives simultaneously through a joint loss function: 1) Optimize the contribution weights of the three multi-relationship graphs and adaptively select effective information; 2) Enhance the consistency of node representations learned by the three types of views; 3) Enhance the similar active support capabilities within the same cluster and emphasize functional partitioning; 4) Optimize the modularity between groups and enhance the differences between different clusters. The training objects and updates of S5 in each round are: 1) Relationship graph neural network parameters; 2) Cross-relationship fusion parameters; 3) Cluster centers. The effect is reflected in the update of the "comprehensive expression" of nodes in S4. Through multiple rounds of learning and training, a comprehensive expression form that is easy to actively support functional partitioning is finally obtained, and the category to which each node belongs is determined according to the probability of the maximum membership degree of the node.

[0052] In this embodiment, a distributed resource node information model is constructed as follows: The IEEE 33-node system, which is widely used in distribution network research, is selected as the basic distribution network model, and a multi-scenario node information model is constructed in combination with the characteristics of distributed resource access in actual distribution networks: The distribution network is represented as a topological network composed of node sets and line connection relationships, and heterogeneous resource node characteristics are constructed according to the distributed resource type, resource access location, capacity configuration and control strategy, and the nodes are uniformly modeled.

[0053] First, the distribution network topology is established, including nodes, line impedance, load parameters, and distribution line topology. Then, various distributed resource access scenarios are constructed based on different penetration rates, including typical operating scenarios with low penetration (10%–20%), medium penetration (20%–50%), and high penetration (above 50%). Distributed resource capacity and access locations are randomly configured to simulate the operating characteristics of different development stages of the actual distribution network. For each operating scenario, typical daily load curves, photovoltaic output curves, and energy storage operating status are further considered. MATPOWER (Power System Analysis Toolkit) is used to generate steady-state power flow data for different operating sections in batches, forming a node-based operating database.

[0054] Based on the characteristics of heterogeneous resource nodes, the resource attribute characteristics, control attribute characteristics, and operational status attribute characteristics of each node are combined to establish a distributed resource node information model. : ]

[0055] in, Represents a node The complete feature information consists of resource attribute sub-vectors, control attribute sub-vectors, and operating status sub-vectors, which serve as the unified input for subsequent multi-relationship views. Here, n represents the number of distribution network nodes. This represents the vector of the nth node. Indicates resource attributes, Indicates control attributes, This indicates the running status attribute.

[0056] Resource attributes are defined as follows:

[0057] in, The code indicates the resource type, distinguishing distributed resources such as photovoltaic (0), integrated photovoltaic and energy storage (1), and adjustable load (2). Indicates the apparent capacity of the resource. , , and These represent the upper and lower limits of active power output and reactive power output, respectively. and These represent the upward and downward ramp rates of active power, respectively. It represents the resource response delay constant, characterizing the time lag between the issuance of a command and the power response.

[0058] The control attribute is defined as follows:

[0059] Where C represents the control strategy code, distinguishing PQ control (00001), GFL droop control (00010), GFM droop control (00100), GFM matching control (01000), and GFM virtual synchronizer control (10000). and These are the active power-frequency droop coefficient and the reactive power-voltage droop coefficient, respectively. This represents the virtual inertia coefficient, characterizing the inverter's inertia support capability. and These represent the proportional and integral coefficients of the current control inner loop, respectively. This represents the virtual impedance value, used to characterize the output impedance characteristics of the inverter.

[0060] The running status attribute is defined as follows:

[0061] in, and These represent the current active power output and reactive power output, respectively. Indicates node voltage. This indicates the state of charge of energy storage (non-energy storage resources are set to zero). It represents the availability of resources and is used to reflect the adjustability of new energy resources such as photovoltaic and wind power to weather conditions.

[0062] In this embodiment, a dynamic index mapping database and simulation data generation method are constructed as follows: MATPOWER has steady-state power flow calculation capabilities and can output steady-state operating sections such as active power, reactive power, and voltage of nodes. However, it cannot simulate dynamic transient processes such as inverter grid-connected control, fault disturbances, and frequency and voltage fluctuations, and cannot directly extract dynamic support indicators such as inertia, frequency regulation, and voltage regulation of distributed resources. While directly relying on Simulink (a graphical simulation tool) to generate samples in batches for electromagnetic transient simulation can completely reproduce the dynamic response process, time-domain simulation calculations are time-consuming and computationally expensive, making it difficult to support the construction of large-scale datasets required for deep learning multi-view clustering models. To address the shortcomings of these two types of tools, this invention proposes a steady-state-dynamic index mapping database scheme. A mapping relationship between steady-state operating states and dynamic active support indicators is established using a small number of typical dynamic simulation samples, enabling rapid generation of dynamic indicators. The entire database establishment process includes an offline database establishment stage and an online calling stage.

[0063] 1) Offline database setup Based on Simulink, a refined electromagnetic transient simulation model of various types of distributed resources such as photovoltaics, energy storage, and controllable loads was built. Time-domain simulations were carried out through typical disturbance conditions. A complete set of dynamic quantitative indicators of inertia support, frequency regulation, and voltage support of distributed resources under each condition were extracted. An offline mapping database with a one-to-one correspondence between steady-state operating condition characteristics and dynamic support indicators was established.

[0064] First, identify the main operational variables that affect the active support capability:

[0065] in, , and These represent the node's active power output, reactive power output, and voltage, respectively; SOC represents the energy storage state of charge. C represents the availability of new energy resources, C represents the control method, and R represents the resource type.

[0066] Secondly, randomize typical operating combinations. :

[0067] Where M represents the number of offline simulation samples, and for each running state, the Simulink dynamic model is called to apply four typical disturbance conditions.

[0068] Load surges by 10%; photovoltaic active power output drops sharply by 20%; a three-phase short-circuit fault lasts for 100ms; system frequency drops by 0.5Hz instantaneously. Calculate the corresponding dynamic indicators:

[0069] in, , and These represent the inertia support capability index, frequency support capability index, and voltage support capability index, respectively. A database sample can be established based on this. : .

[0070] in, Indicates sample The corresponding active support capability indicator vector.

[0071] 2) Dynamic indicator mapping mechanism During the large-scale sample batch generation phase, leveraging the efficient steady-state solution advantages of MATPOWER, a massive number of power flow sections under multiple operating conditions in the distribution network are generated in batches. During modeling, the distributed resources of each node are uniformly connected to the network using an equivalent PQ power injection model, abandoning the traditional synchronous generator node modeling method. Compared to the generator model, the PQ injection model does not require configuring unit dynamic parameters or control constraint modules, resulting in faster power flow solution speeds, more flexible adjustments to operating parameters such as output, load, and topology, and greater controllability of data output.

[0072] To avoid MATPOWER having to perform dynamic simulations again for each generated operating condition, this invention establishes a dynamic index mapping mechanism.

[0073] After MATPOWER generates a new steady-state power flow condition, the corresponding operating state variables are extracted. It can retrieve the most similar running state from the offline database. Define the running state similarity. for:

[0074] Select The sample with the lowest similarity is used to calculate a dynamic index using a similarity-weighted interpolation method. : ) Weight Calculated using the following formula:

[0075] To improve the coverage of the offline database, a dynamic update mechanism is introduced, which updates when new operating conditions arise. The similarity between the samples and the original samples in the database all exceeded the set threshold. When needed, Simulink is invoked to perform a dynamic simulation again, and the simulation samples are added to the offline sample library to update the samples.

[0076] In this embodiment, an active support capability relationship view, an electrical coupling relationship view, and a node attribute relationship view are constructed as follows: 1) Active support capability relationship view Inertia Support Capability Indicators: Traditional synchronous generators provide inertia support to the system through the rotational kinetic energy of the rotor, suppressing rapid frequency changes. For distributed resources controlled by grid-connected inverters, their inertia support capability mainly depends on the control strategy and operating state. This invention uses the node virtual inertia coefficient and inertia dynamic response to quantify the inertia support capability of node-based distributed resources, defined as:

[0077] in, This represents the virtual inertia coefficient of node i. This represents the change in active power during the disturbance period. Indicates the rate of change of frequency. and All represent weighting coefficients. This indicator reflects the amount of kinetic energy that distributed resources on a node can release or absorb when the system frequency changes; a larger value indicates a stronger inertia support capability.

[0078] Frequency support capability index: Frequency support capability refers to the ability of distributed resources to participate in frequency regulation by adjusting active power output when the system frequency deviates from the rated value. It is defined as follows:

[0079] in, This represents the active frequency droop coefficient of node i. and Let represent the adjustable active power limit and the current active power of node i, respectively. Represents the frequency response time constant. , and All of these represent the indicator weighting coefficients.

[0080] Voltage support capability index: Voltage support capability refers to the ability of distributed resources to maintain node voltage stability by adjusting reactive power output, and is defined as follows:

[0081] in, This represents the reactive voltage droop factor. and These represent the adjustable reactive power limit and the current reactive power of node i, respectively. Indicates node voltage sensitivity. , and All of these represent the indicator weighting coefficients.

[0082] 2) Electrical coupling relationship view The electrical coupling between nodes in a power distribution network directly affects the control effect of distributed resource clusters. Nodes with closer electrical distances have a higher degree of coupling in electrical quantities such as voltage and power flow, making them more suitable for being grouped into the same cluster for unified control. This invention uses a weighted sum of active-voltage coupling distance and reactive-voltage coupling distance to quantify the electrical coupling strength between nodes:

[0083] in, Represents a node With nodes The combined electrical distance between them The active-voltage coupling distance is represented by the node. When injecting unit active power, the node The amount of voltage change. The reactive-voltage coupling distance is represented by the node. node when injecting unit reactive power The amount of voltage change. and All represent weighting coefficients, satisfying It can be adjusted according to the actual operating characteristics of the distribution network.

[0084] The active-voltage coupling distance and reactive-voltage coupling distance can be calculated using the Jacobian matrix of the distribution network:

[0085] The sensitivity matrix can be obtained by inverting the Jacobian matrix:

[0086] in, Matrix elements That is, a node When injecting unit active power, the node The voltage change, i.e. ; Matrix elements That is, a node node when injecting unit reactive power The voltage change, i.e. .

[0087] 3) Node attribute relationship view Node characteristics reflect the hardware features and control strategies of distributed resources, and are intrinsic factors determining the control capabilities of distributed resources. This invention mainly considers the following two aspects of node attributes.

[0088] Hardware attributes: Hardware attributes include parameters such as the rated capacity, ramp rate, response time, and adjustable capacity limits of the distributed resources. These parameters directly determine the adjustability and response speed of the distributed resources. For nodes... With nodes The similarity of their hardware attributes can be calculated using Euclidean distance:

[0089] in, and Representing nodes respectively and nodes The k-th hardware attribute parameter, where m represents the number of hardware attribute parameters.

[0090] Control strategy attributes: Different grid-connected control strategies determine the external characteristics and active support capabilities of distributed resources. This invention considers five typical distributed photovoltaic control strategies: 1. Pure grid-following GFL photovoltaic: It adopts maximum power point tracking (MPPT) control and current power dual-loop control, and does not have active support capability.

[0091] 2. GFL Photovoltaic + Droop Control: Based on traditional GFL control, active power-frequency droop and reactive power-voltage droop control are added, providing basic frequency and voltage support capabilities.

[0092] 3. GFM Photovoltaic: It adopts a matching control strategy and can autonomously establish grid voltage and frequency, making it suitable for weak grids and islanded operation scenarios.

[0093] 4. GFM Photovoltaic + Energy Storage Hybrid System Droop Control: It combines the power generation capacity of photovoltaics with the rapid adjustment capability of energy storage, and has flexible active and reactive power adjustment capabilities.

[0094] 5. GFM Photovoltaic + Energy Storage Hybrid System VSG Control: Simulates the operating characteristics of a synchronous generator, providing virtual inertia and damping, and possessing strong active support capabilities.

[0095] For two nodes with the same control strategy, further compare their control loop PI control parameters, including the proportional coefficient. and integral coefficient The similarity of control parameters is also calculated using Euclidean distance. For two nodes with different control strategies, the similarity of their control strategy attributes is set to 0.

[0096] Combining hardware attributes and control strategy attributes, the node and nodes Attribute similarity distance between for:

[0097] in, Indicates the distance attribute of the control strategy. and All represent weighting coefficients, satisfying .

[0098] In this embodiment, the construction of the multi-relationship view is as follows: the features of the three relationship views mentioned above are transformed into three different connection graphs, forming a multi-view graph structure data. For a distribution network containing N distributed resource nodes, three undirected weighted graphs are constructed. , and These correspond to the active support capability relationship, electrical coupling relationship, and node attribute relationship, respectively. Let r represent the undirected weighted graph of the r-th view, consisting of the node set, edge relationships, and adjacency weight matrix. Represents a set of nodes. Indicates the first A set of edges for each view. Indicates the first The adjacency weight matrix of each view, its elements Represents a node With nodes In the Similarities from different perspectives.

[0099] 1) The method for constructing the adjacency weight matrix is ​​as follows: 1. Active Support Capability View: Cosine similarity is calculated based on the feature vectors of node inertia, frequency, and voltage support capabilities.

[0100] in, Represents a node With nodes Similarity under the active support capability view , Representing nodes respectively The active support capability feature vector of node j, Represents a node The active support capability feature vector, Represents a node Inertia support capability index This represents the frequency support capability index of node i. This represents the voltage support capability index of node i. This represents the Gaussian kernel width parameter, used to control the rate at which node similarity decreases with feature distance in the corresponding view. The smaller the value, the more sensitive the view is to differences in node features; only nodes with highly similar features have high similarity weights. The larger the value, the smoother the view, and nodes that are far apart still retain a certain degree of similarity.

[0101] 2. Electrical Coupling View: Constructing a similarity matrix based on comprehensive electrical distance:

[0102] in, Represents a node With nodes Similarity in the electrical coupling relationship view Represents a node With nodes The combined electrical distance between them This represents the Gaussian kernel width parameter.

[0103] 3. Node Attribute View: Construct a similarity matrix based on the node attribute similarity distance.

[0104] in, Represents a node With nodes Similarity in the node attribute relationship view Represents a node With nodes Attribute similarity distance between them This represents the Gaussian kernel width parameter.

[0105] In this embodiment, a multi-view clustering deep learning model is constructed and trained, specifically as follows: This invention constructs a Multi-Relational Graph Neural Network (MR-GNN) model to fuse multi-view graph structural information and learn low-dimensional representations of nodes. The MR-GNN model can simultaneously process graph data of multiple relation types, and adaptively learns the importance weights of different views through an attention mechanism, achieving deep fusion of multi-view information.

[0106] 1) Graph convolution feature extraction and cross-view attention fusion The Multi-Relationship Graph Neural Network (MR-GNN) model includes an input layer, a multi-relationship feature extraction layer, a cross-relationship attention fusion layer, and an output layer. The specific network structures of the multi-relationship feature extraction layer and the cross-relationship attention fusion layer are implemented by stacking multiple layers of network units. Each network unit contains multiple relationship-specific graph convolutional sub-layers and a cross-attention fusion sub-layer.

[0107] For the lLayer network unit, node In the The representation under each view is updated as follows:

[0108] in, This represents the set of neighboring nodes of node i in the r-th view. and All represent the r-th view at the th time. l The trainable parameters of the layer, This represents the feature representation of neighbor node j in the previous layer. This represents a non-linear activation function.

[0109] Then, the node representations from different views are fused through a cross-relationship attention fusion layer:

[0110] Where R represents the number of views, and in this invention R=3; Represents a node In the l Layer to the first The attention weight of each view is calculated as follows:

[0111] in, Represents the attention weight vector. This represents a vector concatenation operation. Let i be the feature representation of node i in the previous layer. This indicates that node i, under the r-th relation view, has passed through the [missing information - likely a specific point or context]. l Node feature representations obtained after a layer relation graph convolutional network. This represents the feature representation of node i in another view, and k is the relational view index used to traverse all views.

[0112] go through l The propagation and fusion of the multi-layer graph neural network model MR-GNN ultimately yields a low-dimensional representation matrix of the nodes. Where d represents the dimension of the low-dimensional embedding vector corresponding to the node. This represents the total number of nodes.

[0113] 2) Construct a soft clustering mapping layer To achieve end-to-end clustering optimization, this invention adds a soft clustering layer to the output layer of the Multi-Relational Graph Neural Network (MR-GNN) model, mapping node representations to probability distributions of belonging to each cluster, where K represents the initial number of clusters. The initial set of cluster centers is: The soft clustering layer uses the Student-t distribution to calculate the similarity between nodes and cluster centers, i.e., nodes. belong Probability of class:

[0114] in, Represents a node The probability of belonging to class k. This represents the low-dimensional embedding representation of node i obtained through a multi-relation graph neural network. This represents the center vector of the k-th cluster. Let P represent the center vector of the j-th cluster, where j represents the cluster index, and P represents the initial total number of clusters. The parameter represents the degrees of freedom of the student-t (t-distribution) distribution.

[0115] 3) Design a multi-objective constrained loss function To construct an end-to-end joint optimization objective function, a distributed resource adaptive clustering and grouping based on a multi-relationship graph neural network is achieved through the synergistic effect of multiple structural constraints and physical prior constraints. The overall optimization objective function is defined as follows:

[0116] Among them, each loss function jointly constrains the model learning process from multiple perspectives, such as graph structure consistency, information entropy constraint, physical characteristic constraint, and cluster distribution balance.

[0117] Modular Evaluation Metric for Soft Clustering: Traditional graph clustering methods typically require pre-specifying the number of clusters and are prone to getting trapped in local optima. This invention uses a modular evaluation metric for clustering as an unsupervised optimization objective to achieve end-to-end adaptive clustering without pre-specifying the number of clusters. The modularity metric is used to measure the quality of graph partitioning and is defined as:

[0118] Where A represents the adjacency matrix of the graph, and m represents the total number of edges in the graph. Represents a node The degree, For nodes The cluster to which it belongs, For indicator functions, when The value is 1 if it is true, and 0 otherwise.

[0119] The range of values ​​for the modularity index Q is: A higher value indicates a better clustering result, meaning that nodes within the same cluster are more tightly connected, while connections between different clusters are sparser. Therefore, a soft clustering modularity index can be calculated based on the clustering assignment probability matrix.

[0120] The optimization objective is to maximize the soft modularity metric. .

[0121] Information entropy constraint: Without constraints, soft clustering can easily lead to nodes being distributed approximately evenly across multiple categories, thus reducing class discriminative power and the clarity of the cluster structure. To avoid overly ambiguous cluster assignments that result in all nodes being evenly distributed across multiple categories, an entropy regularization term is added:

[0122] Improve the separability and stability of clustering results.

[0123] See Figure 1 Consistency Constraint: To ensure the consistency of node representations learned under different relation views in the latent feature space, a consistency constraint is introduced:

[0124] Active Support Capability Constraint: In power system scenarios, the clustering of distributed resources not only relies on graph structure similarity but also must meet the requirement of physical functional consistency. That is, resources of the same type should have similar inertia support capabilities, frequency regulation capabilities, and voltage support capabilities. Therefore, this constraint achieves convergence constraint on the internal structure of the cluster by minimizing the distance from nodes within a cluster to its cluster center. This ensures that the same cluster exhibits a high degree of physical consistency, thereby enhancing the interpretability and usability of the clustering results in engineering applications.

[0125]

[0126]

[0127] in, This represents the joint optimization objective function of a multi-relationship graph neural network model. This represents a modularity index for soft clustering. Information entropy constraint The weight, Indicates vision Figure 1 Consistency constraints The weight, Indicates active support capability constraints The weight, This represents the total number of edges in the graph. Represents the elements of the adjacency matrix. and Representing nodes respectively The degree of node j, The total number of clusters is represented by , and k represents the index of the summation for each cluster. and Let represent the membership degrees of node i and node j to the k-th cluster category, respectively. and Representing nodes respectively In the The representation of the first view and the s-th view. Indicates the first The set of all nodes in each cluster category Represents a node The active support capability feature vector, Indicates the first The average vector of all nodes in each cluster category.

[0128] 4) Parameter optimization and model training process The backpropagation algorithm is used to optimize the overall loss function, and the weight parameters of the multi-relationship graph neural network model MR-GNN are updated sequentially. Attention vector and cluster center During training, similar cluster centers are dynamically merged, and clusters that are too small are removed, ultimately achieving an adaptive determination of the number of clusters. Training stops when the loss function converges or the maximum number of iterations is reached, and the final clustering results are output.

Claims

1. A clustering method for distributed resource multi-relationship graphs in distribution networks based on active support, characterized in that, Includes the following steps: S1. Collect distribution network topology information and distributed resource operation parameters, and construct a distributed resource node information model that includes resource access location, capacity configuration, control strategy and operation status; S2. Generate different resource configurations and operating conditions based on the node information model, and construct an offline database of typical operating scenarios; obtain the dynamic response characteristics of resources through offline electromagnetic transient simulation, establish the mapping relationship between operating status and active support capability indicators, and generate training datasets based on the mapping relationship; wherein, each training data includes randomly generated steady-state operating data and dynamic response data generated according to the mapping relationship; S3. Based on the node information model and active support capability indicators, construct the active support capability relationship diagram G respectively. 1 Electrical coupling diagram G 2 and node attribute relationship graph G 3 These are then integrated to form a multi-relationship graph structure; each training data point corresponds to an active support capability relationship graph G. 1 Electrical coupling diagram G 2 and node attribute relationship graph G 3 ; S4. Using the multi-relation graph structure as the input of the multi-relation graph neural network model, the features of the multi-relation graph structure are extracted by utilizing the multi-relation graph neural network model to obtain the comprehensive representation after multi-relation fusion embedding; S5. Optimize the joint loss function of the multi-relationship graph neural network model. When the joint loss function converges or reaches the maximum number of iterations, output the final clustering result based on the comprehensive representation to achieve adaptive clustering of distributed resources.

2. The active support-based distributed resource multi-relationship graph clustering method for distribution networks according to claim 1, characterized in that, S1 includes: Collect power distribution network topology information and distributed resource operation parameters; Based on the collected data, the distribution network is represented as a topology network consisting of a set of nodes and line connections, and the characteristics of heterogeneous resource nodes are constructed according to the distributed resource type, resource access location, capacity configuration and control strategy. Based on the characteristics of heterogeneous resource nodes, the resource attribute characteristics, control attribute characteristics, and operational status attribute characteristics of each node are combined to form a distributed resource node information model, the expression of which is: ] Where n represents the number of nodes in the distribution network, and the information vector of the i-th node is represented as: in, , and These represent the resource attributes, control attributes, and running status attributes of the i-th node, respectively.

3. The active support-based distributed resource multi-relationship graph clustering method for distribution networks according to claim 1, characterized in that, The construction of the offline database includes: Different resource configurations and operating conditions are generated based on the node information model; Based on the generated data, identify the operational variables that affect proactive support capabilities. Based on runtime variables Randomly generate typical operating combinations; Simulations were conducted under various typical disturbance conditions, and the active support capability indices corresponding to these conditions were calculated based on the simulation results. Based on typical operating combinations and corresponding active support capability indicators, an offline database with a one-to-one correspondence between steady-state operating condition characteristics and active support capability indicators is established as the basis for generating active support capability indicators under new operating conditions, thus completing the construction of the offline database.

4. The active support-based distributed resource multi-relationship graph clustering method for distribution networks according to claim 1, characterized in that, The generated training dataset includes: The dynamic response characteristics of resources are obtained through offline electromagnetic transient simulation; Based on the dynamic response characteristics of resources, for newly generated stable power flow conditions, the corresponding running state variables are extracted. ; Based on extracted running state variables The system retrieves the closest running state from the offline database, where the running state similarity is... The expression is: in, This represents the m-th runtime variable in the offline database. Based on the search results, the k samples with the lowest similarity were selected, and the active support capability index was calculated. Among them, the active support capability indicator Used to construct an active support capability relationship diagram in S3; Establish operational status and proactive support capability indicators The mapping relationship between them is established, and a training dataset is generated based on the mapping relationship; where, when the running state variable... When the similarity of the running state of the samples with the original samples in the offline database exceeds a set threshold, a dynamic simulation is performed again, and the simulated samples are added to the offline database to dynamically update the offline sample library.

5. The active support-based distributed resource multi-relationship graph clustering method for distribution networks according to claim 1, characterized in that, The active support capability relationship diagram G 1 The inertia support capability index, frequency support capability index, and voltage support capability index are expressed as follows: in, Represents a node The active support capability feature vector, This represents the virtual inertia coefficient of node i. This represents the change in active power during the disturbance period. Indicates the rate of change of frequency. and All represent weighting coefficients. Indicator representing inertia support capability. Indicators representing frequency support capabilities This indicates the voltage support capability indicator. This represents the active frequency droop coefficient of node i. and Let represent the adjustable active power limit and the current active power of node i, respectively. Represents the frequency response time constant. , and All represent the indicator weight coefficients. This represents the reactive voltage droop factor. and These represent the adjustable reactive power limit and the current reactive power of node i, respectively. Indicates node voltage sensitivity. , and All of these represent the indicator weighting coefficients.

6. The active support-based distributed resource multi-relationship graph clustering method for distribution networks according to claim 1, characterized in that, Electrical Coupling Relationship View G 2 The electrical coupling strength between nodes is quantified by a weighted sum of active-voltage coupling distance and reactive-voltage coupling distance. in, Represents a node With nodes The combined electrical distance between them The active-voltage coupling distance is represented by the node. When injecting unit active power, the node The amount of voltage change. The reactive-voltage coupling distance is represented by the node. node when injecting unit reactive power The amount of voltage change. and All represent weighting coefficients, satisfying ; Node attribute relationship graph G 3 The characteristic index expression is as follows: in, Represents a node and nodes Attribute similarity distance between them and All represent weighting coefficients, satisfying , Indicates the distance attribute of the control strategy. Represents a node and nodes Hardware attribute similarity, Indicates the number of hardware attribute parameters. and Representing nodes respectively and nodes The kth hardware attribute parameter.

7. The active support-based distributed resource multi-relationship graph clustering method for distribution networks according to claim 1, characterized in that, The expression for the multi-relationship graph structure is as follows: in, This represents an undirected weighted graph of the r-th view. Represents a set of nodes. Indicates the first A set of edges for each view. Indicates the first The adjacency weight matrix of each view, its elements Represents a node With nodes Similarity under the r-th view ; in, Represents a node With nodes Similarity under the active support capability view and Representing nodes respectively The active support capability feature vector of node j, This represents the Gaussian kernel width parameter; in, Represents a node With nodes Similarity in the electrical coupling relationship view Represents a node With nodes The combined electrical distance between them This represents the Gaussian kernel width parameter; in, Represents a node With nodes Similarity in the node attribute relationship view Represents a node With nodes Attribute similarity distance between them This represents the Gaussian kernel width parameter.

8. The active support-based distributed resource multi-relationship graph clustering method for distribution networks according to claim 1, characterized in that, The multi-relationship graph neural network model includes an input layer, a multi-relationship feature extraction layer, a cross-relationship attention fusion layer, and an output layer. The input layer is used to receive the distributed resource node information matrix and the corresponding active support capability relationship diagram, electrical coupling relationship diagram and node attribute relationship diagram to form multi-relationship diagram input data; The multi-relationship feature extraction layer comprises multiple relation graph convolutional networks. Based on the multi-relation graph input data, it performs node feature propagation and aggregation for different relation graphs, extracting node feature representations under different relation views. The calculation process is as follows: in, This represents the set of neighboring nodes of node i in the r-th view. and All represent the r-th view at the th time. l The trainable parameters of the layer, This represents the feature representation of neighbor node j in the previous layer. Represents a non-linear activation function; The cross-relationship attention fusion layer is used to represent node features based on each relationship view. The algorithm adaptively learns the importance of different relation views and fuses them according to the feature weights of each relation to obtain a comprehensive representation after multi-relation fusion embedding. : Where R represents the number of views, Indicates that node i is at the 1st rank. l The attention weight of the layer to the r-th view. This indicates that node i, under the r-th relation view, has passed through the [missing information - likely a specific point or context]. l Node feature representation obtained after layer relation graph convolutional network; The output layer is used to embed the comprehensive representation after multiple relations are fused. The model is mapped to a low-dimensional feature space, and the probability distribution of nodes belonging to different categories is calculated through a soft clustering layer; the Student-t distribution is used to define the soft clustering probability. in, Represents a node The probability of belonging to class k. This represents the low-dimensional embedding representation of node i learned through a multi-relation graph neural network model. This represents the center vector of the k-th cluster. Let P represent the center vector of the j-th cluster, where j represents the cluster index, and P represents the initial total number of clusters. This represents the degree of freedom parameter.

9. The active support-based distributed resource multi-relationship graph clustering method for distribution networks according to claim 1, characterized in that, The joint optimization objective function of the multi-relationship graph neural network model is expressed as follows: in, This represents the joint optimization objective function of a multi-relationship graph neural network model. This represents a modularity index for soft clustering. Information entropy constraint The weight, Indicates view consistency constraints The weight, Indicates active support capability constraints The weight, This represents the total number of edges in the graph. Represents the elements of the adjacency matrix. and Representing nodes respectively The degree of node j, The total number of clusters is represented by , and k represents the index of the summation for each cluster. and Let represent the membership degrees of node i and node j to the k-th cluster category, respectively. and Representing nodes respectively In the The representation of the first view and the s-th view. Indicates the first The set of all nodes in each cluster category Represents a node The active support capability feature vector, Indicates the first The average vector of all nodes in each cluster category.

Images