High-proportion photovoltaic regional power distribution network state sensing method, system and device based on heterogeneous graph neural network and medium

Abstract

The invention relates to the technical field of intelligent power grid operation monitoring and state recognition, and discloses a method, a system and equipment for sensing the state of a high-proportion photovoltaic regional power distribution network based on a heterogeneous graph neural network and a medium. The limitation of a traditional method in processing complex network topology, multi-source heterogeneous data and disturbance driving behaviors is broken through. In the state sensing process, from acquisition of multi-source data, a series of preprocessing, construction of a reasonable graph structure, extraction of node embedding vectors and output of state labels and abnormal scores are performed by using a preset state recognition model, and all links are closely connected and have innovativeness. And key nodes or regions with potential risks or abnormal states in the power distribution network are identified through the abnormal score function, and state online sensing and risk early warning of the power distribution network containing the photovoltaic region are realized.

Description

Technical Field

[0001] This invention relates to the field of smart grid operation monitoring and status identification technology, and in particular to a method, system, device and medium for status sensing of high-proportion photovoltaic regional distribution networks based on heterogeneous graph neural networks. Background Technology

[0002] With the increasing penetration of new energy sources in regional power distribution systems, especially the large-scale integration of distributed power sources such as photovoltaics, the dynamic characteristics of power distribution network operation are becoming increasingly complex. In this complex state, the node states of the power distribution network may exhibit stronger time-varying characteristics and greater sensitivity to disturbances. This leads to frequent occurrences of voltage fluctuations, power flow reversals, and short-term disturbances, further increasing the demand for comprehensive perception of the power distribution network's operational status and early warning of anomalies. Against this backdrop, how to construct a state perception method with generalization capabilities and strong robustness has become a key issue in ensuring the safe and stable operation of regional power systems.

[0003] Currently, mainstream methods for power distribution network status awareness rely heavily on centralized measurement systems and rule-driven approaches. For example, measurement data can be collected through SCADA systems, PMU devices, or smart meters, and combined with traditional algorithms such as threshold judgment and statistical analysis for reasonable operational status identification. However, these methods have significant limitations when dealing with complex network topologies, multi-source heterogeneous data, and disturbance-driven behaviors. For instance, the uncertainty and drastic fluctuations in photovoltaic output make it difficult for existing static feature extraction models to adapt to the requirements of real-world dynamic scenarios. Furthermore, power distribution networks typically contain numerous low-voltage branches, distribution area users, and distributed equipment on both the power supply and load sides. The operational status of these devices is significantly affected by the coupling of spatial correlation and temporal evolution, and current traditional methods cannot effectively model the nonlinear relationships between multidimensional features.

[0004] With advancements in technology, graph neural networks have demonstrated significant advantages in processing structured data in recent years, and some research has attempted to apply them to power system topology modeling and state classification. However, most current work still focuses on conventional distribution network topologies or single data source inputs, failing to systematically integrate disturbance source characteristics, multi-source heterogeneous node attributes, and edge attribute features, making it difficult to meet the high requirements for comprehensive state perception of real-world regional distribution networks.

[0005] Therefore, there is an urgent need for a new methodological framework that integrates photovoltaic disturbance perception, multi-source heterogeneous data fusion, and graph neural network modeling capabilities. This framework can be used to improve the perception accuracy and anomaly identification capabilities of complex operating states of regional power distribution networks, and further provide technical support for refined power grid management and flexible dispatching. Summary of the Invention

[0006] In view of the aforementioned existing problems, the present invention is proposed.

[0007] Therefore, this invention provides a method, system, device, and medium for high-proportion photovoltaic regional distribution network status perception based on heterogeneous graph neural networks. It can solve the shortcomings of existing distribution network status perception methods in handling complex network topologies, multi-source heterogeneous data, and disturbance-driven behaviors, as well as the problem that existing graph neural network applications cannot systematically integrate multiple characteristics and features to meet the high requirements of comprehensive status perception of actual regional distribution networks.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for state perception of a high-proportion photovoltaic regional distribution network based on a heterogeneous graph neural network, comprising: Acquire multi-source data from various nodes of the distribution network within the target area, and perform a first preprocessing on the multi-source data; The first preprocessing is used to process the multi-source data into several node time-series feature vectors; Establish a graph structure based on various nodes of the distribution network within the target area. The graph structure also includes multiple edge relationships such as flexible electrical connections, temporal evolution, and geographical boundaries. The various types of nodes include distribution transformer nodes, user nodes, photovoltaic nodes, and feeder switch nodes; The graph structure is subjected to a second preprocessing, which is used to extract node embedding vectors through an edge-enhanced graph attention mechanism. A preset state recognition model is provided, wherein the input of the state recognition model is the node embedding vector, and the state recognition model is used to output the state label and anomaly score of the corresponding node.

[0009] As a preferred embodiment of the high-proportion photovoltaic regional distribution network state sensing method based on heterogeneous graph neural networks described in this invention, it further includes: Based on the output of the state recognition model and the electrical centrality index of the corresponding node in the graph structure, the importance of the corresponding node is calculated. Solve and sort each node of the distribution network within the target area; Output the operational status perception results of the target key nodes and the corresponding anomaly level.

[0010] As a preferred embodiment of the high-proportion photovoltaic regional distribution network state perception method based on heterogeneous graph neural network described in this invention, the graph structure further includes mapping the feeder electrical connection relationship, the temporal dependency relationship in the historical observation sequence, and the physical geographical location information between nodes to electrical connection edges, temporal change edges, and geographical location edges in the graph, respectively.

[0011] As a preferred embodiment of the high-proportion photovoltaic regional distribution network state perception method based on heterogeneous graph neural networks described in this invention, the preset state recognition model includes: The structure of the preset state recognition model includes a state classification and anomaly detection joint module, which is a fusion of the state classification module and the anomaly detection module. The state classification module adopts a multi-class classifier structure, mapping the node embedding vector to a state label, which includes a normal operation label, a fluctuation abnormal label, and a continuous abnormal label. The anomaly detection module is based on the Euclidean distance between the current node embedding vector and the historical mean embedding vector.

[0012] This preferred solution can efficiently and accurately identify and assess the node status of distribution networks in areas with a high proportion of photovoltaic power. By combining state classification and anomaly detection modules, state classification and anomaly detection are organically integrated, avoiding the waste of computational resources and increased time costs caused by performing state classification and anomaly detection separately in traditional methods.

[0013] As a preferred embodiment of the high-proportion photovoltaic regional distribution network status perception method based on heterogeneous graph neural network described in this invention, the status classification and anomaly detection joint module combines status labels and anomaly scores for the status perception result output, and combines photovoltaic disturbance time windows for anomaly tracing and dynamic screening.

[0014] As a preferred embodiment of the high-proportion photovoltaic regional distribution network status perception method based on heterogeneous graph neural networks described in this invention, the output of the operational status perception results of key target nodes and the corresponding anomaly levels includes: Set an upper threshold for anomaly scoring, and compare the importance of all nodes with the upper threshold for anomaly scoring; All nodes whose importance exceeds the threshold of the anomaly score are marked as high-risk nodes; Extract all high-risk nodes from the state matrix to form a subset.

[0015] As a preferred embodiment of the high-proportion photovoltaic regional distribution network state perception method based on heterogeneous graph neural network described in this invention, it further includes: if photovoltaic disturbance time periods co-occur, then the edge weights of these nodes are enhanced in the graph structure to form a high-risk region clustering map.

[0016] Secondly, the present invention provides a high-proportion photovoltaic regional distribution network state sensing system based on heterogeneous graph neural networks, comprising: The first preprocessing module is used to acquire multi-source data from various nodes of the distribution network within the target area and to perform a first preprocessing on the multi-source data. The first preprocessing is used to process the multi-source data into several node time-series feature vectors; The graph structure acquisition module is used to establish a graph structure based on various nodes of the distribution network within the target area. The graph structure also includes multiple edge relationships such as flexible electrical connections, temporal evolution, and geographical boundaries. The various types of nodes include distribution transformer nodes, user nodes, photovoltaic nodes, and feeder switch nodes; The second preprocessing module is used to perform a second preprocessing on the graph structure. The second preprocessing is used to extract node embedding vectors through an edge-enhanced graph attention mechanism. The recognition module is used to preset a state recognition model. The input of the state recognition model is the node embedding vector. The state recognition model is used to output the state label and anomaly score of the corresponding node.

[0017] Thirdly, the present invention provides an electronic device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described above.

[0018] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0019] Compared with existing technologies, the beneficial effects of this invention are that it proposes a state perception method for high-proportion photovoltaic (PV) regional distribution networks based on heterogeneous graph neural networks. This method effectively integrates PV disturbance perception and multi-source heterogeneous data by establishing a heterogeneous graph neural network, overcoming the limitations of traditional methods in handling complex network topologies, multi-source heterogeneous data, and disturbance-driven behaviors. In the state perception process, starting from acquiring multi-source data, a series of preprocessing steps are performed to construct a reasonable graph structure, extract node embedding vectors, and then a preset state recognition model is used to output state labels and anomaly scores. Each step is closely linked and innovative. The anomaly scoring function identifies key nodes or areas in the distribution network that have potential risks or state anomalies, achieving online state perception and risk warning for PV-containing regional distribution networks. This invention also has advantages such as strong multi-source data fusion capability, high structural adaptability, and high anomaly location accuracy. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 The present invention provides a flowchart of a method for sensing the state of a high-proportion photovoltaic regional distribution network based on a heterogeneous graph neural network, as an embodiment of the present invention.

[0022] Figure 2 This is an internal structure diagram of an electronic device for a high-proportion photovoltaic regional distribution network state sensing method based on heterogeneous graph neural networks, provided as an embodiment of the present invention. Detailed Implementation

[0023] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0024] It should be noted in advance that the system mentioned in the embodiments as the subject of real-time operation refers to any system configured with this method.

[0025] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides a method for sensing the state of a high-proportion photovoltaic regional distribution network based on a heterogeneous graph neural network, including: This invention provides a method that can effectively solve the problems mentioned above. The following will describe in detail how to implement the high-proportion photovoltaic regional distribution network state perception method based on heterogeneous graph neural network in conjunction with several embodiments. Figure 1 A flowchart illustrating a method for state perception of a high-proportion photovoltaic regional distribution network based on heterogeneous graph neural networks is shown, including: S101, acquire multi-source data from various nodes of the distribution network within the target area, and perform the first preprocessing on the multi-source data; In an embodiment of the present invention, the first preprocessing is used to process multi-source data into several node time-series feature vectors; In an optional implementation, the raw multi-source data collected from distribution transformers, user load nodes, distributed photovoltaic nodes and feeder switches can be uniformly cleaned and standardized. For example, sliding window normalization can be performed on heterogeneous characteristics such as active power, reactive power, voltage amplitude, voltage phase angle, photovoltaic output, total irradiance, temperature and wind speed to eliminate dimensional differences and suppress noise. In an optional implementation, time alignment and temporal completion can be performed on data with different sampling frequencies. For example, linear interpolation or forward padding strategies can be used to complete missing values ​​to ensure that each node has complete features under the same timestamp. Furthermore, the temporal similarity between nodes can be calculated based on the dynamic time warping algorithm. For example, the DTW distance can be calculated for the feature sequence of each node within a set time window to generate a coupling strength matrix. Valid associations can be screened according to a preset similarity threshold to construct temporal edges that reflect state synchronization.

[0026] Among them, the node temporal feature vector refers to the temporal sequence formed after the above processing, which is organized with a unified time granularity and contains multi-dimensional electrical and environmental features, and serves as the basic input for subsequent graph structure modeling and embedding learning.

[0027] It should be noted that S101 acquires multi-source data from various nodes of the distribution network within the target area and performs the first preprocessing, providing unified, clean, and time-aligned node feature inputs for subsequent steps.

[0028] S102, Establish a graph structure based on various nodes of the distribution network within the target area. The graph structure also includes flexible electrical connections, temporal evolution, and multiple edge relationships of geographical boundaries. In this embodiment of the invention, the various nodes include distribution transformer nodes, user nodes, photovoltaic nodes, and feeder switch nodes; In an embodiment of the present invention, the graph structure further includes mapping the feeder electrical connection relationship, the temporal dependency relationship in the historical observation sequence, and the physical geographical location information between nodes to electrical connection edges, temporal change edges, and geographical location edges in the graph, respectively.

[0029] In an optional implementation, distribution transformer nodes, user nodes, photovoltaic nodes, and feeder switch nodes can be used as vertices in the diagram, and electrical connection edges can be established according to the actual feeder topology. For example, electrical connection edges can be established between transformers and switches, and between switches and user loads that are directly connected by cables or busbars, to reflect the actual power supply path. In one optional implementation, time-series evolution edges are constructed based on historical observation data. For example, the state characteristics of the same node within a continuous time window are self-connected to form time-series edges to capture its dynamic evolution trend. Geographical critical edges are established by combining the latitude and longitude coordinates of the nodes. For example, geographical adjacency edges are established between adjacent photovoltaic nodes or user nodes that are geographically less than a set threshold but are not electrically directly connected, in order to characterize the cooperative change relationship caused by meteorological disturbances or local load fluctuations.

[0030] Among them, the multiple edge relationships consisting of soft electrical connections, temporal evolution, and geographical criticality refer to three types of structured connections that respectively represent physical power supply dependence, temporal dynamic correlation, and spatial proximity influence in a unified heterogeneous graph, which together support the modeling of the complex coupling characteristics of the distribution network under high proportion of photovoltaic access.

[0031] It should be noted that S102 establishes a graph structure that integrates multiple edge relationships, including electrical connections, temporal evolution, and geographical boundaries, providing structured topological support for the edge-enhanced graph attention mechanism in S103.

[0032] S103, Perform a second preprocessing on the graph structure. The second preprocessing is used to extract node embedding vectors through the edge-enhanced graph attention mechanism. In an optional implementation, a second preprocessing is performed on the graph structure to extract node embedding vectors through an edge-enhanced graph attention mechanism. For example, first, the neighbor set of each node in the graph is determined, and then the attention weights are calculated based on the composite attribute vector of the node feature vector and the edge. For any node i, its neighbor set is denoted as Ni, and its edge-enhanced attention weights can be calculated by the linear transformation matrix W and the attention weight vector a. Then, these weights are applied to update the node embedding representation. For example, the nonlinear activation function ReLU is used to update the embedding vector of node i in the (l+1)th layer of the graph neural network to ensure that the influence of different types of edges on the node state can be captured. In an optional implementation, to further enhance the model's ability to model complex spatiotemporal dynamic relationships, a stacked structure can be introduced for multi-round information propagation. In each round of propagation, attention weights and node embeddings are updated to ultimately obtain stable node vectors for state classification. In this process, the node embedding vectors are extracted by combining the node's own attributes and the features of its adjacent edges using an edge-enhanced graph attention mechanism. This not only considers different types of edges such as electrical connection edges, temporal change edges, and geographical location edges, but also enables the model to more accurately distinguish the influence of different types of neighbors on the node state by explicitly introducing edge features and edge types.

[0033] Among them, the edge-enhanced graph attention mechanism essentially incorporates the semantic information of edges into the calculation of attention weights, enabling the model to adaptively learn the degree of influence of different types of adjacency information on node states, thereby improving the representational power of node embedding vectors and the prediction accuracy of distribution network state change trends.

[0034] It should be noted that S103 extracts node embedding vectors through an edge-enhanced graph attention mechanism, effectively integrating node attributes and multi-class edge semantic information, providing highly discriminative feature representations for the state recognition model of S104.

[0035] S104, Preset state recognition model. The input of the state recognition model is the node embedding vector. The state recognition model is used to output the state label and anomaly score of the corresponding node.

[0036] In an embodiment of the present invention, the preset state recognition model includes: The pre-defined state recognition model has a structure that includes a state classification and anomaly detection joint module, which is a fusion of the state classification module and the anomaly detection module. The status classification module adopts a multi-class classifier structure, mapping the node embedding vector to status labels, which include normal operation labels, fluctuation abnormality labels, and continuous abnormality labels. The anomaly detection module is based on the Euclidean distance between the current node embedding vector and the historical mean embedding vector.

[0037] In this embodiment of the invention, the state classification and anomaly detection joint module combines state labels and anomaly scores for the state perception result output, and combines them with the photovoltaic disturbance time window for anomaly tracing and dynamic screening.

[0038] In an embodiment of the present invention, the importance of a corresponding node is calculated based on the output of the state recognition model and the electrical centrality index of the corresponding node in the graph structure. Solve and sort each node of the distribution network within the target area; Output the operational status perception results of the target key nodes and the corresponding anomaly level.

[0039] In one optional implementation, node importance can be calculated by jointly using the anomaly score output by the state recognition model and the electrical centrality index of the node in the graph structure. For example, for each node i, its anomaly score is weighted and fused with electrical centrality indices such as betweenness centrality or current betweenness to obtain a comprehensive importance value. Subsequently, the importance values ​​of all distribution transformer nodes, user nodes, photovoltaic nodes, and feeder switch nodes in the target area are uniformly solved and sorted in descending order. For example, a fast sorting algorithm is used to sort all network nodes from high to low importance to ensure that key nodes are identified first. Finally, the nodes with the highest ranking and anomaly scores exceeding a set threshold are output as target key nodes, and their operating status labels and corresponding anomaly levels are marked. For example, nodes with high importance, a status label of "continuous anomaly", and an anomaly score greater than the threshold θ are marked as first-level high-risk nodes, forming a structured perception result for the scheduling system to call.

[0040] Among them, node importance refers to a comprehensive evaluation index that integrates the degree of operational anomalies and the criticality of topological electrical systems, and is used to quantify the priority of nodes in distribution network status perception and risk warning.

[0041] In this embodiment of the invention, the output of the operational status perception results of the target key node and the corresponding anomaly level includes: Set an upper threshold for anomaly scoring and compare the importance of all nodes with the upper threshold for anomaly scoring; All nodes whose importance exceeds the threshold of the anomaly score are marked as high-risk nodes; Extract all high-risk nodes from the state matrix to form a subset.

[0042] In an optional implementation, an anomaly scoring threshold θ can be set, and the importance values ​​of all nodes can be compared with this threshold one by one. For example, the importance value of each node i in the state-aware output matrix can be compared with this threshold. Determine whether it satisfies >θ; Mark all nodes that meet this condition as high-risk nodes. For example, if the importance of a photovoltaic node is 0.85 and θ is set to 0.8, then this node is explicitly identified as high-risk. Subsequently, these high-risk nodes are extracted from the state matrix containing the status labels and anomaly scores of all nodes in the network to form an independent high-risk node subset. For example, a submatrix containing only node indexes, operating status labels, and anomaly scores is constructed for subsequent visualization or scheduling alarms. The high-risk node subset refers to the set of nodes whose importance exceeds the threshold of the anomaly score, used to focus on key anomaly areas and support accurate risk warnings.

[0043] In an embodiment of the present invention, if the photovoltaic disturbance time period is co-occurring, the edge weights of these nodes are enhanced in the graph structure to form a high-risk region clustering graph.

[0044] In one optional implementation, it can be determined whether the time of anomaly at a high-risk node coincides with the period of drastic fluctuation in photovoltaic output. For example, if the peak of an anomaly score for a user node occurs during the period corresponding to a sudden drop in irradiance or the start-up and shutdown of a photovoltaic inverter, then the node is considered to co-occur with the photovoltaic disturbance period. Subsequently, the edge weights between the node and its neighboring nodes are enhanced in the heterogeneous graph structure. For example, the weights of electrical connection edges or geographical adjacency edges are multiplied by a preset amplification factor α (α>1) to strengthen their influence in graph propagation. Finally, community detection or clustering algorithms are performed based on the enhanced edge weights. For example, the Louvain method is used to partition the graph structure, aggregating closely connected high-risk nodes into local clusters to form a high-risk area clustering map. Here, the high-risk area clustering map refers to the subgraph structure generated by edge weight enhancement after considering the co-occurrence of photovoltaic disturbance time, which reflects the spatial aggregation characteristics of abnormal nodes and is used to intuitively locate vulnerable areas of the distribution network significantly affected by photovoltaic disturbances.

[0045] Example 2: Based on the above examples, a specific implementation of a high-proportion photovoltaic regional distribution network state perception method based on heterogeneous graph neural networks can be designed as follows: The core objective of collecting multi-source, heterogeneous electrical operation data from regional distribution networks is to provide a data foundation with spatiotemporal representativeness and physical significance for subsequent state sensing. Considering the differences in operating characteristics and measurement methods among different types of nodes in regional distribution networks, a unified data collection and processing framework needs to be designed.

[0046] Specifically, the raw data sequence consisting of the following features is first collected from each node: in, Let represent the multi-source feature vector of the i-th node at time t. Active power Reactive power Voltage amplitude, The voltage phase angle, Contribute to photovoltaic nodes, For full irradiance, For temperature, This refers to wind speed.

[0047] Furthermore, due to the inconsistent sampling frequencies, missing measurements, and noise in the aforementioned data sources, sliding window normalization, time-series completion, and noise suppression algorithms are required to improve data quality; this constitutes the first preprocessing step. in, Let be the feature mean and standard deviation of node i, respectively.

[0048] Furthermore, the next step is to measure the correlation between nodes to reflect the synchronicity and dynamic correlation between their state changes. The Dynamic Time Warping (DTW) algorithm is introduced for temporal similarity calculation: in, It is a dynamic time warping algorithm. For node i in the time window The feature sequence within. Calculate the similarity for all node pairs and form a coupling strength matrix. : in, The similarity threshold is set. Based on this, the distribution transformer is used as the starting node. The initial graph structure is generated by combining the feeder connection relationship, voltage level, physical geographical proximity and topological connectivity. The temporal dynamic edges are established according to the coupling strength matrix. Finally, a heterogeneous graph structure with three types of edge attributes is obtained, which provides input for the next step of graph neural network state modeling.

[0049] Where the threshold It can be set based on experience to control the sparsity of graph edges.

[0050] Furthermore, based on this, considering the physical connections, power supply cascading relationships, and topology of the distribution network, the edges are further divided into three categories: Physical connection: such as the direct connection between transformer, switch and load.

[0051] Temporal edge: such as a self-connection of data from the same node under different time windows.

[0052] Geographic adjacency edge: such as voltage covariance between adjacent geographic coordinate nodes.

[0053] It should be noted that the above steps complete the entire process from raw data acquisition to graph structure construction, providing structural input for subsequent graph neural network modeling.

[0054] Furthermore, after completing the fusion expression of multi-source electrical characteristic data of the regional distribution network, in order to further establish a state-aware model with structural information and adjacency semantics, this invention proposes a graph structure modeling method for the actual topology and operating state of the distribution network, and combines graph neural network technology to dynamically model and embed the spatial topology and temporal evolution characteristics of each node.

[0055] Furthermore, this method first uses key nodes such as distribution transformers, user load nodes, distributed photovoltaic nodes, and feeder switchgear as vertices in the graph, and maps them to edges in the graph structure through the physical connection relationships of feeder cables or busbars. On this basis, a multi-layer graph structure with heterogeneous characteristics is constructed to fully reflect the structural diversity and information flow transmission channels in the distribution network.

[0056] Furthermore, to fully reflect the functional differences and connection semantics of different types of equipment in the distribution network, this method constructs three core edge types, namely: Electrical connection edge: indicates the actual physical connection relationship between nodes, such as feeder connection, busbar parallel connection, etc.

[0057] Time series edges: connect the representations of the same node within adjacent time windows to capture the trend of its features evolving over time.

[0058] Geographic edge: Connects geographically adjacent but electrically unconnected nodes, reflecting geographic spatial correlation.

[0059] Furthermore, the entire heterogeneous graph can be modeled as a triple: in, Represents a set of nodes. Let each represent a set of edges consisting of electrical edges, temporal edges, and geographical edges, respectively. This represents the time evolution of the graph.

[0060] Furthermore, after constructing the graph structure, a reasonable mechanism needs to be designed to extract adjacency information strongly correlated with the evolution of node states from the graph. This invention introduces an edge-enhanced graph attention mechanism to guide the model in learning the contribution of each type of edge to node representation learning, i.e., the second preprocessing.

[0061] Furthermore, for any node in the graph Its set of neighboring nodes is denoted as The formula for calculating the edge-enhanced attention weights is as follows: in, For node feature vectors, A composite attribute vector representing an edge, including connection type identifier and association features. It is a linear transformation matrix. For attention weight vectors, Let i be the set of neighbors of node i.

[0062] It should be noted that the above attention mechanism essentially incorporates the semantic information of edges into the calculation of attention weights, enabling the model to adaptively learn the degree of influence of different types of adjacency information on the node state, thereby improving the representational power of the node embedding vector.

[0063] Furthermore, after calculating the attention weights, the node representation will be updated using the following formula: in, It is a non-linear activation function. Let be the updated embedding vector of node i in the (l+1)th layer of the graph neural network.

[0064] Furthermore, to enhance the model's ability to model complex spatiotemporal dynamic relationships, a stacked structure can be introduced for multi-round information propagation, with attention weights updated in each round. and node embedding Finally, a stable node vector is obtained for state classification.

[0065] It should be noted that, through the above method, this invention establishes a graph neural network modeling method oriented towards the structural characteristics of power distribution networks, which significantly improves the model's ability to express topological connectivity, electrical correlation, and spatiotemporal dynamic evolution, providing a high-quality input vector foundation for subsequent state recognition and anomaly detection.

[0066] Furthermore, after constructing the graph structure, an edge-enhanced graph attention mechanism is used to extract the embedding vector representation of each node, combining node features with the features of its adjacent edges. Based on the above attention mechanism, the updated embedding of the node is obtained: in, For nodes The final embedded representation.

[0067] Furthermore, embedding nodes into vectors The input is fed into a state classifier, which uses a fully connected neural network (MLP) to identify the running state. Let the output of the classifier be: in, , For classifier parameters, Predict the probability vector for the state of a node.

[0068] The predicted label is: Common state labels include: normal operation, fluctuating abnormality, and persistent abnormality. In this invention, the state classification module employs a multi-class classifier structure, embedding nodes into vectors. Mapped to status labels These correspond to normal operation, abnormal fluctuations, and continuous abnormality, respectively.

[0069] Furthermore, in addition to the state classifier, to further enhance the model's ability to identify unknown anomalies, an anomaly detection module based on embedding distribution deviation is designed. This module uses the mean of the embedding vectors of the training sample nodes. As the center of the normal samples, the anomaly score for each test node is defined as: Rating value This value is used to characterize the degree of deviation of the node's current operating state. The larger the value, the more obvious the difference between its embedded features and the normal state.

[0070] Set threshold ,like If so, the node is determined to be an abnormal node.

[0071] The union module will use status labels With abnormal scoring It is used in conjunction with the state perception result output, and can be combined with the photovoltaic disturbance time window for anomaly tracing and dynamic screening, so as to realize the real-time identification of node status and potential fault warning in the regional distribution network.

[0072] Furthermore, based on the state classification results of each node... With abnormal score Construct a state-aware output matrix Combined with the anomalous distribution map, and the time window caused by photovoltaic disturbance. Dynamic screening and key markers are implemented.

[0073] The state-aware output matrix is ​​defined as follows: in, This represents the set of indices for all nodes in a regional distribution network. Let i be the label representing the running status of node i. Assign an abnormal score to it.

[0074] Furthermore, to achieve more visualized and operable state perception results, this invention constructs a node risk level map based on the output matrix: Set an upper threshold for abnormal ratings All satisfied Nodes are marked as high-risk nodes. All high-risk nodes are extracted from the state matrix to form a subset. If photovoltaic disturbances coexist during certain time periods, the edge weights of these nodes are enhanced in the graph structure to form a high-risk area clustering map. This map can be output in tabular or graph form for subsequent use by the scheduling system.

[0075] Example 3, referring to Figure 2 This embodiment also provides a high-proportion photovoltaic regional distribution network status sensing system based on heterogeneous graph neural networks, including: The first preprocessing module is used to acquire multi-source data from various nodes of the distribution network within the target area and to perform first preprocessing on the multi-source data. The first preprocessing step is used to process multi-source data into several node time-series feature vectors. The graph structure acquisition module is used to establish a graph structure based on various nodes of the distribution network within the target area. The graph structure also includes multiple edge relationships such as flexible electrical connections, temporal evolution, and geographical boundaries. Various types of nodes include distribution transformer nodes, user nodes, photovoltaic nodes, and feeder switch nodes; The second preprocessing module is used to perform a second preprocessing on the graph structure. The second preprocessing is used to extract node embedding vectors through the edge-enhanced graph attention mechanism. The recognition module is used to preset the state recognition model. The input of the state recognition model is the node embedding vector. The state recognition model is used to output the state label and anomaly score of the corresponding node.

[0076] The above-mentioned unit modules can be embedded in the processor of the electronic device in hardware form or independent of it, or they can be stored in the memory of the electronic device in software form, so that the processor can call and execute the corresponding operations of the above modules.

[0077] This embodiment also provides an electronic device, which can be a terminal, and its internal structure diagram can be as follows: Figure 2As shown, the electronic device includes a processor, memory, communication interface, display screen, and input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage medium. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a method for sensing the state of a high-proportion photovoltaic regional distribution network based on a heterogeneous graph neural network. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the device's casing, or an external keyboard, touchpad, or mouse.

[0078] This embodiment also provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, it performs the following steps: Acquire multi-source data from various nodes of the distribution network within the target area, and perform the first preprocessing on the multi-source data; The first preprocessing step is used to process multi-source data into several node time-series feature vectors. Establish a graph structure based on various nodes of the distribution network within the target area. The graph structure also includes multiple edge relationships such as flexible electrical connections, temporal evolution, and geographical boundaries. Various types of nodes include distribution transformer nodes, user nodes, photovoltaic nodes, and feeder switch nodes; A second preprocessing is performed on the graph structure, which is used to extract node embedding vectors through the edge-enhanced graph attention mechanism. A preset state recognition model is used. The input of the state recognition model is the node embedding vector. The state recognition model is used to output the state label and anomaly score of the corresponding node.

[0079] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0080] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0081] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for state perception of a high-proportion photovoltaic regional distribution network based on heterogeneous graph neural networks, characterized in that, include: Acquire multi-source data from various nodes of the distribution network within the target area, and perform a first preprocessing on the multi-source data; The first preprocessing is used to process the multi-source data into several node time-series feature vectors; Establish a graph structure based on various nodes of the distribution network within the target area. The graph structure also includes multiple edge relationships such as flexible electrical connections, temporal evolution, and geographical boundaries. The various types of nodes include distribution transformer nodes, user nodes, photovoltaic nodes, and feeder switch nodes; The graph structure is subjected to a second preprocessing, which is used to extract node embedding vectors through an edge-enhanced graph attention mechanism. A preset state recognition model is provided, wherein the input of the state recognition model is the node embedding vector, and the state recognition model is used to output the state label and anomaly score of the corresponding node.

2. The method for sensing the state of a high-proportion photovoltaic regional distribution network based on a heterogeneous graph neural network as described in claim 1, characterized in that, Also includes: Based on the output of the state recognition model and the electrical centrality index of the corresponding node in the graph structure, the importance of the corresponding node is calculated. Solve and sort each node of the distribution network within the target area; Output the operational status perception results of the target key nodes and the corresponding anomaly level.

3. The method for sensing the state of a high-proportion photovoltaic regional distribution network based on a heterogeneous graph neural network as described in claim 2, characterized in that, The graph structure further includes mapping the feeder electrical connection relationship, the temporal dependency relationship in the historical observation sequence, and the physical geographical location information between nodes to electrical connection edges, temporal change edges, and geographical location edges in the graph, respectively.

4. The method for sensing the state of a high-proportion photovoltaic regional distribution network based on a heterogeneous graph neural network as described in claim 3, characterized in that, The preset state recognition model includes: The structure of the preset state recognition model includes a state classification and anomaly detection joint module, which is a fusion of the state classification module and the anomaly detection module. The state classification module adopts a multi-class classifier structure, mapping the node embedding vector to a state label, which includes a normal operation label, a fluctuation abnormal label, and a continuous abnormal label. The anomaly detection module is based on the Euclidean distance between the current node embedding vector and the historical mean embedding vector.

5. The method for state perception of a high-proportion photovoltaic regional distribution network based on a heterogeneous graph neural network as described in claim 4, characterized in that, The state classification and anomaly detection combined module uses state labels and anomaly scores in combination for the state perception result output, and combines photovoltaic disturbance time windows for anomaly tracing and dynamic screening.

6. The method for state perception of a high-proportion photovoltaic regional distribution network based on heterogeneous graph neural networks as described in claim 5, characterized in that, The output target key node's operational status perception results and corresponding anomaly levels include: Set an upper threshold for anomaly scoring, and compare the importance of all nodes with the upper threshold for anomaly scoring; All nodes whose importance exceeds the threshold of the anomaly score are marked as high-risk nodes; Extract all high-risk nodes from the state matrix to form a subset.

7. The method for state perception of a high-proportion photovoltaic regional distribution network based on heterogeneous graph neural networks as described in claim 6, characterized in that, Also includes: If the photovoltaic disturbance time period is co-occurring, the edge weights of these nodes in the graph structure are enhanced to form a clustering map of high-risk areas.

8. A state sensing system for a high-proportion photovoltaic regional distribution network based on a heterogeneous graph neural network, employing the method described in any one of claims 1 to 7, characterized in that, include: The first preprocessing module is used to acquire multi-source data from various nodes of the distribution network within the target area and to perform a first preprocessing on the multi-source data. The first preprocessing is used to process the multi-source data into several node time-series feature vectors; The graph structure acquisition module is used to establish a graph structure based on various nodes of the distribution network within the target area. The graph structure also includes multiple edge relationships such as flexible electrical connections, temporal evolution, and geographical boundaries. The various types of nodes include distribution transformer nodes, user nodes, photovoltaic nodes, and feeder switch nodes; The second preprocessing module is used to perform a second preprocessing on the graph structure. The second preprocessing is used to extract node embedding vectors through an edge-enhanced graph attention mechanism. The recognition module is used to preset a state recognition model. The input of the state recognition model is the node embedding vector. The state recognition model is used to output the state label and anomaly score of the corresponding node.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method for sensing the state of a high-proportion photovoltaic regional distribution network based on heterogeneous graph neural networks as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for sensing the state of a high-proportion photovoltaic regional distribution network based on heterogeneous graph neural networks as described in any one of claims 1 to 7.

Images