Power simulation data checking method and system based on artificial intelligence and graph database

Abstract

This invention relates to the field of power system verification technology, and discloses a method and system for verifying power simulation data based on artificial intelligence and graph databases. The method acquires the dataset to be verified output by the power simulation system, constructs a graph structure based on power network topology information, maps the data to the graph structure, and stores it in a graph database; it generates node embedding vectors through embedding representation learning to calculate anomaly metrics; it performs joint judgment by combining topological constraint rules and physical law constraint rules and generates a verification report; and it queries anomaly propagation links and dynamically adjusts weight coefficients. This invention improves the accuracy and efficiency of power simulation data verification and reduces the risk of anomaly propagation.

Description

Technical Field

[0001] This invention relates to the field of power system verification technology, and in particular to a method and system for verifying power simulation data based on artificial intelligence and graph databases. Background Technology

[0002] As power systems continue to expand in scale and increase in complexity, power simulation systems play a crucial role in power grid planning, dispatching, and security analysis. The accuracy of power simulation data directly affects the reliability and security of power system decisions. Traditional power simulation data verification mainly relies on manual inspection and static rule comparison, which is insufficient to meet the demands of efficient verification of massive amounts of data. In recent years, the development of artificial intelligence technology and graph databases has provided new technical means and solutions for power simulation data verification.

[0003] In power systems, complex relationships exist between network topology, node parameters, and time-series data. These relationships are constrained by both physical laws and topological structures. Traditional relational databases struggle to efficiently represent and query these complex relationships, while graph databases, with their inherent relational representation and fast path lookup capabilities, provide a more suitable technological foundation for the storage and analysis of power data. Meanwhile, artificial intelligence, particularly graph neural network technology, has demonstrated significant potential in complex network anomaly detection. By learning the topological and temporal characteristics of nodes, it can discover anomaly patterns that are difficult to identify using traditional methods.

[0004] However, existing power simulation data verification technologies have the following shortcomings: Traditional verification methods lack sufficient consideration of the overall power network topology and inter-node relationships, often employing isolated point-to-point verification strategies, making it difficult to detect complex anomalies caused by the coordinated action of multiple components. Existing technologies fail to effectively integrate topological constraint rules and physical law constraint rules for joint verification, resulting in verification results that are either too strict, ignoring reasonable deviations in actual operation, or too lenient, failing to detect hidden data anomalies. Existing anomaly detection methods typically lack interpretability and anomaly propagation analysis capabilities, making it difficult to trace the root cause of anomalies and assess their potential impact. This results in a lack of targeted and forward-looking anomaly handling, reducing the safety and reliability of power system operation. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for verifying power simulation data based on artificial intelligence and graph database, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of the present invention provides a method for verifying power simulation data based on artificial intelligence and graph databases, comprising: Obtain the dataset to be verified output by the power simulation system. The dataset to be verified includes power network topology information, node operating parameters and time-series measurement data. A graph structure is constructed based on the power network topology information, and the node operating parameters and time-series measurement data are mapped to the attribute fields of the corresponding nodes and edges in the graph structure; The graph structure is stored in a graph database, and predefined topological constraint rules and physical law constraint rules are extracted using the query language of the graph database. Embedding representation learning is performed on the graph structure. By aggregating the topological features of the neighboring nodes and the temporal dynamic features of each node in the graph structure, embedding vectors of each node are generated. Based on the distribution of the embedding vectors in the feature space, the anomaly metric of the dataset to be verified is calculated. The anomaly metric is jointly determined with the topological constraint rules and physical law constraint rules to generate a verification report that includes the anomaly location, anomaly type, and violated constraint items. The upstream associated nodes and downstream affected nodes of the abnormal location are queried using the path query language of the graph database to obtain the abnormal propagation link. The weight coefficients of different topological regions in the aggregation process are then adjusted based on the historical verification records of the nodes in the abnormal propagation link.

[0007] A graph structure is constructed based on the power network topology information, and the node operating parameters and time-series measurement data are mapped to the attribute fields of the corresponding nodes and edges in the graph structure, including: The device identifiers in the power network topology information are parsed to obtain device type labels and device level information. The node type attributes are determined based on the device type labels, and node level coordinates are assigned in the graph structure based on the device level information. Based on the connection relationship between device identifiers in the power network topology information, a directed edge or an undirected edge is created, and the starting node identifier and ending node identifier of the directed edge or the two end node identifiers of the undirected edge are recorded in the edge connection relationship table. Establish a bidirectional index mapping table between device identifiers and graph structure node identifiers; Perform time-domain analysis on the node operating parameters to determine the steady-state and dynamic characteristic quantities in the node operating parameters; The steady-state characteristic quantity is written into the static attribute field of the corresponding node identifier through the forward query path of the bidirectional index mapping table, and the dynamic characteristic quantity and the time series measurement data are written into the time-varying attribute field of the corresponding node identifier.

[0008] The graph structure is stored in a graph database, and predefined topological constraint rules and physical law constraint rules are extracted using the query language of the graph database, including: The nodes and edges in the graph structure are serialized according to the storage mode of the graph database, and a multi-dimensional index structure is established for the nodes during the serialization process. The multi-dimensional index structure includes a type dimension index grouped according to the node type attribute and a hierarchical dimension index layered according to the node hierarchical coordinates. By traversing the type dimension index using the query language of the graph database, the set of neighbor node types that each node type is allowed to connect to and the connection quantity constraints are determined for each node type, thus obtaining the topology constraint rules; The connection table of the edges is traversed using the query language of the graph database. For each directed edge, the starting node identifier and the ending node identifier recorded in the connection table of the edges are read. The starting node identifier is determined as the starting point of energy flow, and the ending node identifier is determined as the ending point of energy flow. The node storage locations corresponding to the starting and ending points of the energy flow are determined by the multi-dimensional index structure, the node operation parameters are extracted from the node storage locations, and the physical law constraint rules are generated based on the physical correlation between the node operation parameters.

[0009] Embedding representation learning is performed on the graph structure. This involves generating embedding vectors for each node by aggregating the topological features of its neighboring nodes and temporal dynamic features. Based on the distribution of these embedding vectors in the feature space, anomaly metrics for the dataset to be verified are calculated, including: The set of neighboring nodes of each node is determined from the graph structure, and the topological influence coefficient of each neighboring node on the node is determined by the connection relationship of the edges in the set of neighboring nodes. Extract the time-series measurement data from the time-varying attribute field of each node as time-series dynamic features, and weight and aggregate the time-series dynamic features with the node features of the neighboring nodes according to the topological influence coefficient to generate the aggregated feature vector of each node. The aggregated feature vector is subjected to a nonlinear transformation through iterative updates, and the transformation results after multiple iterations are used as the embedding vectors of each node. In a predefined feature space, the distance metric between the embedding vectors of each node is calculated, and the distance metric between the embedding vector of each node and the center embedding vector of the node type to which each node belongs is used as the node-level anomaly metric. Based on the distribution characteristics of the node-level anomaly metrics in the graph structure, the anomaly metrics of the dataset to be verified are determined.

[0010] From the graph structure, determine the set of neighboring nodes for each node, and for each neighboring node in the set of neighboring nodes, determine the topological influence coefficient of each neighboring node on the node through the connection relationship of the edges, including: Based on the edge connection information of each node in the graph structure, the identifiers of neighboring nodes directly connected to the node are determined based on the edge connection information, and the nodes corresponding to the neighboring node identifiers are formed into a set of neighboring nodes. For each neighbor node in the set of neighbor nodes, the directional information of the edge connecting the node and the neighbor node is determined by the edge connection table; The energy flow direction of the edge is determined based on the directional information. When the energy flow direction is from the neighbor node to the node, the neighbor node is marked as an input neighbor node. When the energy flow direction is from the node to the neighbor node, the neighbor node is marked as an output neighbor node. The transmission capacity strength value of the edge is calculated based on the edge weight information of the edge connecting the node to each neighboring node. The transmission capability strength value is associated with the labeling type of the neighboring nodes. A positive topology influence coefficient is calculated for neighboring nodes labeled as input neighboring nodes, and a negative topology influence coefficient is calculated for neighboring nodes labeled as output neighboring nodes.

[0011] The upstream and downstream affected nodes of the anomaly location are queried using the path query language of the graph database to obtain the anomaly propagation path. Based on the historical verification records of the nodes in the anomaly propagation path, the weight coefficients of different topological regions during the aggregation process are adjusted, including: Use the node identifier corresponding to the abnormal location as the starting node for path query; By using the path query language of the graph database, a reverse traversal is performed along the directional information of the edges to extract all predecessor nodes whose energy flow direction points to the node corresponding to the abnormal position, which are then used as upstream associated nodes. By using the path query language of the graph database to perform a forward traversal along the directional information of the edges, all successor nodes reached from the node corresponding to the abnormal position along the energy flow direction of the edge are extracted as downstream affected nodes; The upstream associated nodes, the nodes corresponding to the abnormal locations, and the downstream affected nodes are arranged into an abnormal propagation link according to the energy flow sequence. Based on the historical verification records of each node in the anomaly propagation link, the anomaly tendency metric of each node is calculated, the anomaly propagation link is divided into topological regions containing multiple nodes, and the weight coefficients of different topological regions are adjusted during the aggregation process based on the anomaly tendency metric of nodes in each topological region.

[0012] Based on the historical verification records of each node in the anomaly propagation link, an anomaly tendency metric is calculated for each node. The anomaly propagation link is then divided into topological regions containing multiple nodes. The weighting coefficients of different topological regions during the aggregation process are adjusted based on the anomaly tendency metric of nodes within each topological region, including: Based on the historical verification records of each node in the abnormal propagation link, the verification timestamp and verification status identifier are determined, and the time interval between each failed verification status and the current time is calculated according to the verification timestamp. The time interval is subjected to time decay processing to obtain time decay weights. The number of times each node is marked as failing the verification in the historical verification record is weighted and summed with the corresponding time decay weights. Based on the weighted summation result and the number of historical verification records, the abnormal tendency measure value of each node is calculated. The abnormal propagation link is divided into multiple topological regions based on the connection relationship and directionality information of the edges between nodes in the abnormal propagation link. Calculate the connectivity of each node within each topological region, and aggregate the anomaly tendency metric values ​​of each node based on the connectivity of each node to obtain the regional anomaly tendency metric value of the topological region; The anomaly tendency metric of the region is converted into a weight adjustment coefficient, and the initial weight coefficient corresponding to the topological region is adjusted based on the weight adjustment coefficient.

[0013] A second aspect of the present invention provides a power simulation data verification system based on artificial intelligence and graph databases, comprising: The first unit is used to acquire the dataset to be verified output by the power simulation system. The dataset to be verified includes power network topology information, node operating parameters and time series measurement data. The second unit is used to construct a graph structure based on the power network topology information, and to map the node operating parameters and time-series measurement data to the attribute fields of the corresponding nodes and edges in the graph structure. The third unit is used to store the graph structure in a graph database and extract predefined topological constraint rules and physical law constraint rules through the query language of the graph database. The fourth unit is used to perform embedding representation learning on the graph structure. It generates embedding vectors for each node by aggregating the topological features of the neighboring nodes and the temporal dynamic features of each node in the graph structure, and calculates the anomaly metric of the dataset to be verified based on the distribution of the embedding vectors in the feature space. The fifth unit is used to jointly determine the anomaly metric value with the topological constraint rules and physical law constraint rules, and generate a verification report containing the anomaly location, anomaly type and violation of constraint items; The sixth unit is used to query the upstream associated nodes and downstream affected nodes of the abnormal location through the path query language of the graph database to obtain the abnormal propagation link, and adjust the weight coefficients of different topological regions in the aggregation process based on the historical verification records of the nodes in the abnormal propagation link.

[0014] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0016] This invention maps power network topology information, node operating parameters, and time-series measurement data into a graph structure, achieving an intuitive expression and efficient storage of complex relationships in power systems, overcoming the limitations of traditional relational databases in handling highly interconnected data. This invention uses a graph database to store and extract constraint rules, enabling complex topological and physical law constraints to be expressed and queried in a structured manner, improving the management efficiency and flexibility of verification rules. This invention introduces a graph embedding representation learning method, which can simultaneously consider the topological characteristics and temporal dynamic characteristics of nodes, capturing complex patterns in power systems more comprehensively than traditional methods and improving the accuracy of anomaly detection. This invention achieves a fusion of data-driven and knowledge-driven verification by jointly judging anomaly metrics and constraint rules, enabling the discovery of both latent anomalies based on statistical distributions and explicit anomalies that violate clear rules, significantly improving the comprehensiveness of verification. This invention innovatively introduces an anomaly propagation link analysis mechanism, which can track the upstream and downstream impact range of anomalies and dynamically adjust aggregation weights based on historical verification records, giving the model adaptive learning capabilities and improving the early warning capability for cascading system faults. This invention organically combines graph computing, deep learning, and domain knowledge. Compared with traditional verification methods, it can more accurately locate anomalies, identify anomaly types, and provide specific explanations of constraint violations, thus providing a strong guarantee for the reliability of power system simulation results. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the power simulation data verification method based on artificial intelligence and graph database according to an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram illustrating the process of determining abnormal metric values ​​in a dataset to be verified according to an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0021] Figure 1 This is a flowchart illustrating the power simulation data verification method based on artificial intelligence and graph database, according to an embodiment of the present invention. Figure 1 As shown, the method includes: Obtain the dataset to be verified output by the power simulation system. The dataset to be verified includes power network topology information, node operating parameters and time-series measurement data. A graph structure is constructed based on the power network topology information, and the node operating parameters and time-series measurement data are mapped to the attribute fields of the corresponding nodes and edges in the graph structure; The graph structure is stored in a graph database, and predefined topological constraint rules and physical law constraint rules are extracted using the query language of the graph database. Embedding representation learning is performed on the graph structure. By aggregating the topological features of the neighboring nodes and the temporal dynamic features of each node in the graph structure, embedding vectors of each node are generated. Based on the distribution of the embedding vectors in the feature space, the anomaly metric of the dataset to be verified is calculated. The anomaly metric is jointly determined with the topological constraint rules and physical law constraint rules to generate a verification report that includes the anomaly location, anomaly type, and violated constraint items. The upstream associated nodes and downstream affected nodes of the abnormal location are queried using the path query language of the graph database to obtain the abnormal propagation link. The weight coefficients of different topological regions in the aggregation process are then adjusted based on the historical verification records of the nodes in the abnormal propagation link.

[0022] In one optional implementation, a graph structure is constructed based on the power network topology information, and the node operating parameters and time-series measurement data are mapped to the attribute fields of the corresponding nodes and edges in the graph structure, including: The device identifiers in the power network topology information are parsed to obtain device type labels and device level information. The node type attributes are determined based on the device type labels, and node level coordinates are assigned in the graph structure based on the device level information. Based on the connection relationship between device identifiers in the power network topology information, a directed edge or an undirected edge is created, and the starting node identifier and ending node identifier of the directed edge or the two end node identifiers of the undirected edge are recorded in the edge connection relationship table. Establish a bidirectional index mapping table between device identifiers and graph structure node identifiers; Perform time-domain analysis on the node operating parameters to determine the steady-state and dynamic characteristic quantities in the node operating parameters; The steady-state characteristic quantity is written into the static attribute field of the corresponding node identifier through the forward query path of the bidirectional index mapping table, and the dynamic characteristic quantity and the time series measurement data are written into the time-varying attribute field of the corresponding node identifier.

[0023] First, the device identifiers in the power network topology information are parsed. Device identifiers in a power network typically include a device type code and a serial number. For example, a transformer is represented by "TR001," where "TR" represents the transformer type and "001" is the serial number. Device type tags are extracted using regular expression matching or string splitting, such as parsing "TR001" as the type "transformer." Simultaneously, device hierarchy information, such as voltage level (e.g., 500kV, 220kV, 110kV) or regional division (e.g., main grid, distribution network), is extracted from the topology information. Based on the parsed device type tags, nodes are classified into different types such as generation nodes, substation nodes, and load nodes, and assigned corresponding node type attributes. Based on the device hierarchy information, hierarchical coordinates are assigned to each node in the graph structure, for example, using voltage level as the vertical axis and geographical location as the horizontal axis, constructing a multi-level graph structure representation.

[0024] Next, edges are created based on the connection relationships between devices in the power network topology information. Connections between devices in a power system have clear physical meanings, such as the direction of power flow or control relationships. For connections with clear directionality, such as lines where power flows from a generator to a load, directed edges are created; for bidirectional peer-to-peer connections, such as connections between buses, undirected edges are created. When creating an edge, a connection relationship table is recorded, showing the connection relationship between the edge's starting node identifier and ending node identifier (for directed edges) or both endpoint node identifiers (for undirected edges). This table can be stored using data structures such as adjacency lists or adjacency matrices. An example of an edge connection relationship representation is: {edge identifier: [starting node identifier, ending node identifier, edge type]}, where the edge type can be either "directed" or "undirected".

[0025] Next, a bidirectional index mapping table is established between device identifiers and graph node identifiers, creating a mapping relationship between the actual power system and the graph model. This allows for rapid query conversion between the two domains. The bidirectional index mapping table contains forward mappings (from device identifier to graph node identifier) ​​and reverse mappings (from graph node identifier to device identifier). For example, the mapping table can be designed as two dictionary structures: {device identifier: graph node identifier} and {graph node identifier: device identifier}. This design supports efficient bidirectional queries. When accessing actual device information from the graph model is needed, the reverse mapping can quickly locate the relevant information; when updating the parameters of the actual device in the graph model is needed, the forward mapping can quickly locate the corresponding node.

[0026] Time-domain analysis is performed on node operating parameters to distinguish between steady-state and dynamic characteristics. Node operating parameters typically include long-term stable parameters (steady-state characteristics) and parameters that change over time (dynamic characteristics). Steady-state characteristics include basic parameters such as rated capacity and rated voltage, which remain constant over a relatively long period. Dynamic characteristics include operating state parameters such as real-time power, voltage amplitude, and phase angle, which change continuously with the system's operating conditions. Time-domain analysis methods (such as sliding window analysis and Fourier transform) are used to process the parameter sequences, identifying steady-state and dynamic characteristics based on the frequency and amplitude of parameter changes. For example, parameters with variance less than a preset threshold can be identified as steady-state characteristics; otherwise, they are classified as dynamic characteristics.

[0027] Finally, the parameters are written to the corresponding graph structure attribute fields. Through the forward lookup path of the bidirectional index mapping table, steady-state characteristics are written to the static attribute fields of the corresponding nodes. For example, the transformer's rated capacity "500MVA" is written to the static attribute of node "n001". Simultaneously, dynamic characteristics and time-series measurement data are written to the time-varying attribute fields of the nodes. Data can be organized using timestamp indexes, such as in the form {timestamp: parameter value}. Each node in the graph structure can be designed with two main attribute containers: a static attribute dictionary and a time-varying attribute dictionary, storing steady-state and dynamic characteristics respectively. Time-varying attributes can be further organized into a time-series database format, supporting efficient time-series querying and analysis.

[0028] In practical applications, when the power grid undergoes topological changes, such as the addition of new equipment or alterations to connectivity, these changes can be reflected by updating the nodes and edges in the graph structure. When equipment parameters change, the corresponding node can be quickly located and the relevant attribute fields updated using a bidirectional index mapping table. This graph-based power network representation method provides a unified data foundation for subsequent network analysis, fault diagnosis, and optimized control.

[0029] In one optional implementation, the graph structure is stored in a graph database, and predefined topological constraint rules and physical law constraint rules are extracted using the query language of the graph database, including: The nodes and edges in the graph structure are serialized according to the storage mode of the graph database, and a multi-dimensional index structure is established for the nodes during the serialization process. The multi-dimensional index structure includes a type dimension index grouped according to the node type attribute and a hierarchical dimension index layered according to the node hierarchical coordinates. By traversing the type dimension index using the query language of the graph database, the set of neighbor node types that each node type is allowed to connect to and the connection quantity constraints are determined for each node type, thus obtaining the topology constraint rules; The connection table of the edges is traversed using the query language of the graph database. For each directed edge, the starting node identifier and the ending node identifier recorded in the connection table of the edges are read. The starting node identifier is determined as the starting point of energy flow, and the ending node identifier is determined as the ending point of energy flow. The node storage locations corresponding to the starting and ending points of the energy flow are determined by the multi-dimensional index structure, the node operation parameters are extracted from the node storage locations, and the physical law constraint rules are generated based on the physical correlation between the node operation parameters.

[0030] When storing a graph structure in a graph database, the first step is to serialize the nodes and edges. During this process, for each node, its identifier, type attribute, and level coordinates are encoded. For example, a node representing an electrical device can be serialized as: "Node:E001:Type=Transformer:Level=2:Position=(x,y)", where E001 is the node identifier, Transformer is the node type, Level=2 indicates the level, and Position indicates the coordinate position.

[0031] To improve query efficiency, a multi-dimensional index structure is established during serialization. The type dimension index is implemented using a hash table, with the key being the node type and the value being a set of nodes of the same type. For example, Type_Index{Transformer} ={E001, E008, E015} represents a set of nodes of type transformer. The level dimension index adopts a tree structure, organized according to the level in which the node is located. Each level node contains information about all nodes in that level. For example, Level_Index[2] = {E001, E005, E009} represents a set of nodes located at the second level.

[0032] The edge serialization process encodes the starting node identifier, ending node identifier, and edge attribute information. For example, "Edge:E001->E005:Type=PowerFlow:Capacity=100kW" represents an energy flow edge from E001 to E005, with the type being power flow and a capacity of 100kW. Simultaneously, an edge connection table is established to record the starting and ending point information of all edges.

[0033] After serialization, topology constraint rules are extracted using the query language of the graph database. First, each node type in the type dimension index is traversed, and the other types of nodes connected to that type are statistically analyzed. Taking the type "Transformer" as an example, the query language can return the types of nodes connected to the transformer node and the number of connections. By analyzing the query results, the allowed types and number of neighboring nodes for the transformer node are determined. For example, if it is found that the transformer must be connected to at least one power node and one or more load nodes, then the topology constraint rule is generated: "The transformer node must be connected to at least one power type node and one to ten load type nodes."

[0034] To extract the constraints of physical laws, the starting point and ending point of energy flow are first determined by traversing the edge connection table. For each edge record, such as "Edge:E001->E005", E001 is determined as the starting point of energy flow and E005 is determined as the ending point of energy flow.

[0035] A multi-dimensional index structure allows for rapid location of node storage locations. For example, to find node E001, the node identifier can be searched in the index to obtain detailed information about the node. Node operating parameters, such as voltage, current, and power, can be extracted from the node storage location.

[0036] Based on the extracted node operating parameters, the physical correlation between the starting and ending nodes is analyzed. Taking a power system as an example, if the starting point is a generator and the ending point is a transformer, the output power parameters of the generator and the input power parameters of the transformer are extracted. The law of conservation of energy is applied to generate the constraint rule: "The output power of the generator must be equal to the sum of the input power of the transformer and the line loss".

[0037] In practical applications, specific topological constraint rules and physical law constraint rules can be customized according to different graph structure types. For power system graphs, Kirchhoff's current law constraints can be extracted; for fluid network graphs, Bernoulli's equation constraints can be extracted.

[0038] The above methods achieve efficient storage of graph structures and automatic extraction of constraint rules, laying the foundation for subsequent graph structure verification and optimization. The entire process fully utilizes the query capabilities and indexing mechanisms of graph databases, ensuring the accuracy and efficiency of constraint rule extraction.

[0039] In one optional implementation, embedding representation learning is performed on the graph structure. This involves generating embedding vectors for each node by aggregating the topological features of its neighboring nodes and temporal dynamic features. Based on the distribution of these embedding vectors in the feature space, an anomaly metric for the dataset to be verified is calculated, including: The set of neighboring nodes of each node is determined from the graph structure, and the topological influence coefficient of each neighboring node on the node is determined by the connection relationship of the edges in the set of neighboring nodes. Extract the time-series measurement data from the time-varying attribute field of each node as time-series dynamic features, and weight and aggregate the time-series dynamic features with the node features of the neighboring nodes according to the topological influence coefficient to generate the aggregated feature vector of each node. The aggregated feature vector is subjected to a nonlinear transformation through iterative updates, and the transformation results after multiple iterations are used as the embedding vectors of each node. In a predefined feature space, the distance metric between the embedding vectors of each node is calculated, and the distance metric between the embedding vector of each node and the center embedding vector of the node type to which each node belongs is used as the node-level anomaly metric. Based on the distribution characteristics of the node-level anomaly metrics in the graph structure, the anomaly metrics of the dataset to be verified are determined.

[0040] In graph structures, embedding representation learning methods can be used to represent nodes in the graph, which can effectively capture the topological relationships and dynamic features between nodes, providing strong support for anomaly detection.

[0041] Figure 2 This is a schematic diagram illustrating the process of determining abnormal metric values ​​in a dataset to be verified, according to an embodiment of the present invention. Figure 2 As shown, in this embodiment, the set of neighboring nodes for each node is first determined from the graph structure. For each node v in the graph, its set of neighboring nodes N(v) is identified by traversing the edges connected to it. For each neighboring node u in set N(v), the topological influence coefficient of that neighboring node on node v is determined based on the edge connection relationship. The topological influence coefficient can be directly represented by the edge weight, or different calculation methods can be used depending on the specific application scenario. For example, in power network monitoring, the influence coefficient can be calculated based on the transmission capacity and distance of the power lines; in social network analysis, the influence coefficient can be determined based on the interaction frequency and density.

[0042] Specifically, the topological influence coefficient α(u,v) can be expressed as the influence strength of node u on node v. When calculating, the weight of the edge w(u,v) and the degree of node u d(u) are taken into account, so that the influence is proportional to the edge weight and inversely proportional to the degree of the node, thus avoiding the dominance of nodes with high height in the aggregation process.

[0043] Next, time-series measurement data from the time-varying attribute fields of each node are extracted as time-series dynamic features. For sensor nodes in industrial control systems, time-series data of parameters such as temperature, pressure, and flow rate can be extracted; for router nodes in communication networks, time series data of indicators such as data throughput, latency, and packet loss rate can be extracted. These time-series dynamic features are weighted and aggregated with the node features of neighboring nodes according to the aforementioned topology influence coefficient to generate an aggregated feature vector for each node.

[0044] The specific aggregation process is as follows: For node v, its aggregated feature vector h(v) consists of its own temporal dynamic feature f(v) and the weighted sum of the features of its neighboring nodes. The temporal dynamic features can be extracted by using a sliding window to extract the measurement values ​​of the most recent n time points, forming a feature vector of fixed dimensions. During aggregation, the feature vector of each neighboring node u is multiplied by its corresponding topological influence coefficient α(u,v) and then summed to finally obtain the aggregated feature vector of node v.

[0045] Then, the aggregated feature vectors are nonlinearly transformed through iterative updates. Each iteration includes two stages: a feature aggregation stage and a nonlinear transformation stage. In the feature aggregation stage, the feature vectors of all neighboring nodes from the previous iteration are collected and aggregated weighted according to the topological influence coefficient. In the nonlinear transformation stage, the aggregation results are nonlinearly mapped using an activation function (such as ReLU or tanh). After k iterations, the final embedding vectors of each node are obtained. This iterative approach can capture the structural information within the k-hop neighborhood, enabling the embedding vectors to contain a wider range of contextual features.

[0046] Taking the power network as an example, after two rounds of iterative updates, the embedding vector of substation node v not only includes its own time-series features such as voltage and current, but also integrates the information of directly connected transmission line nodes and the information of other indirectly connected substations, thus forming a more comprehensive node representation.

[0047] In a predefined feature space, the distance metric between the embedding vectors of each node is calculated. First, all nodes are grouped according to their type, such as grouping all substation nodes into one category and all transmission line nodes into another. A central embedding vector is calculated for each category, representing the typical characteristics of that category of nodes. Then, the distance between each node's embedding vector and the central vector of its category is calculated, using metrics such as Euclidean distance or cosine similarity. This distance value is the node-level anomaly metric, reflecting the degree to which a node deviates from the typical pattern.

[0048] Finally, based on the distribution characteristics of node-level anomaly metrics in the graph structure, the anomaly metrics of the dataset to be verified are determined. Statistical distribution characteristics of the anomaly metrics, such as mean, variance, and quantiles, can be calculated. Spatial distribution patterns of anomalous nodes, such as clustering and diffusion, can also be analyzed. For the entire dataset to be verified, its anomaly metrics can be comprehensively evaluated using indicators such as the proportion of anomalous nodes, the average anomaly degree of anomalous nodes, and the spatial distribution entropy of anomalous nodes.

[0049] For example, in communication network monitoring, if abnormal metrics of router nodes in a certain area are found to be generally high and clustered, it can be determined that the network in that area is under a distributed denial-of-service attack; if the abnormal nodes are scattered and the abnormality is low, it is due to normal fluctuations in the equipment. In this way, abnormal patterns in the graph structure can be effectively identified, providing support for network security monitoring, fault diagnosis, and other applications.

[0050] In one optional implementation, the set of neighboring nodes for each node is determined from the graph structure, and the topological influence coefficient of each neighboring node on the node is determined based on the edge connection relationship among the neighboring nodes in each set, including: Based on the edge connection information of each node in the graph structure, the identifiers of neighboring nodes directly connected to the node are determined based on the edge connection information, and the nodes corresponding to the neighboring node identifiers are formed into a set of neighboring nodes. For each neighbor node in the set of neighbor nodes, the directional information of the edge connecting the node and the neighbor node is determined by the edge connection table; The energy flow direction of the edge is determined based on the directional information. When the energy flow direction is from the neighbor node to the node, the neighbor node is marked as an input neighbor node. When the energy flow direction is from the node to the neighbor node, the neighbor node is marked as an output neighbor node. The transmission capacity strength value of the edge is calculated based on the edge weight information of the edge connecting the node to each neighboring node. The transmission capability strength value is associated with the labeling type of the neighboring nodes. A positive topology influence coefficient is calculated for neighboring nodes labeled as input neighboring nodes, and a negative topology influence coefficient is calculated for neighboring nodes labeled as output neighboring nodes.

[0051] In the process of graph structure data processing, it is first necessary to determine the set of neighboring nodes of each node in the graph structure, and then determine the topological influence coefficient of each neighboring node on the node based on the edge connection relationship of each neighboring node in the set of neighboring nodes.

[0052] Obtain graph structure data, which contains multiple nodes and edges connecting these nodes. Each node has a unique identifier, and each edge contains connection information, including direction and edge weight. The graph structure can be represented as G=(V,E), where V represents the set of nodes and E represents the set of edges.

[0053] Based on the edge connection information in the graph structure, for each node vi∈V, all neighboring nodes directly connected to that node are determined. This process is achieved by retrieving all edges in the edge set E that contain node vi. If there exists an edge eij connecting nodes vi and vj, then node vj is identified as a neighboring node of node vi, and its identifier is added to the neighboring node set N(vi) of node vi.

[0054] For example, for node v1, by searching the edge set, we find that edges e12, e13 and e14 connect node v1 to nodes v2, v3 and v4 respectively. Then the set of neighboring nodes of node v1 is N(v1)={v2,v3,v4}.

[0055] For each neighbor node in the neighbor node set, the directionality information of the edges connecting the current node and its neighbors is determined using the edge connection table. The directionality information of the edges is represented as pointing from the starting node to the ending node. In a directed graph, edge eij represents a directed edge from node vi to node vj; in an undirected graph, edge eij can be considered a bidirectional edge.

[0056] Based on the determined directional information, the direction of energy flow along the edges is determined. In the information or energy transfer model, the direction of the edge represents the direction of energy flow. When energy flows from neighbor node vj to the current node vi, neighbor node vj is marked as an input neighbor node; when energy flows from the current node vi to neighbor node vj, neighbor node vj is marked as an output neighbor node.

[0057] In practical applications, such as power grids, electricity flows from power plants to users, and power plant nodes are input neighbor nodes relative to user nodes; in social networks, if user A follows user B, information flows from B to A, and B is A's input neighbor node.

[0058] Based on the edge weights connecting the current node to its neighbors, the transmission capacity strength of each edge is calculated. Edge weights can represent the importance, transmission capacity, or association strength of an edge. The transmission capacity strength value can be obtained directly from the edge weights, or by transforming the edge weights using a specific function.

[0059] For example, for the edge eij connecting nodes vi and vj, its edge weight is wij. wij can be directly used as the transmission capacity strength value, or the function f(wij) can be applied to transform it, such as f(wij) = wij / (total edge weights), to obtain a standardized transmission capacity strength value.

[0060] The transmission capacity strength value is associated with the labeling type of neighboring nodes to calculate the topology influence coefficient. For a neighboring node vj labeled as an input neighbor, its topology influence coefficient on the current node vi is positive, which is represented as the positive topology influence coefficient PI(vj,vi); for a neighboring node vk labeled as an output neighbor, its topology influence coefficient on the current node vi is negative, which is represented as the negative topology influence coefficient NI(vk,vi).

[0061] The specific calculation method is as follows: For the input neighbor node vj, the positive topology influence coefficient PI(vj,vi) = Sji, where Sji represents the transmission capacity strength value of the edge from node vj to node vi; for the output neighbor node vk, the negative topology influence coefficient NI(vk,vi) = -Sik, where Sik represents the transmission capacity strength value of the edge from node vi to node vk.

[0062] The calculated topology influence coefficient can quantify the degree and nature of the influence of neighboring nodes on the current node in the network structure. The positive topology influence coefficient represents the enhancement or support effect of neighboring nodes on the current node, while the negative topology influence coefficient represents the weakening or depletion effect of neighboring nodes on the current node.

[0063] Ultimately, for each node in the graph structure, a set of neighboring nodes and their corresponding topological influence coefficients can be obtained, forming a node topological influence feature vector, which provides a basic feature representation for subsequent graph data analysis tasks.

[0064] In one optional implementation, the upstream associated nodes and downstream affected nodes of the abnormal location are queried using the path query language of the graph database to obtain the abnormal propagation path. Based on the historical verification records of the nodes in the abnormal propagation path, the weight coefficients of different topological regions during the aggregation process are adjusted, including: Use the node identifier corresponding to the abnormal location as the starting node for path query; By using the path query language of the graph database, a reverse traversal is performed along the directional information of the edges to extract all predecessor nodes whose energy flow direction points to the node corresponding to the abnormal position, which are then used as upstream associated nodes. By using the path query language of the graph database to perform a forward traversal along the directional information of the edges, all successor nodes reached from the node corresponding to the abnormal position along the energy flow direction of the edge are extracted as downstream affected nodes; The upstream associated nodes, the nodes corresponding to the abnormal locations, and the downstream affected nodes are arranged into an abnormal propagation link according to the energy flow sequence. Based on the historical verification records of each node in the anomaly propagation link, the anomaly tendency metric of each node is calculated, the anomaly propagation link is divided into topological regions containing multiple nodes, and the weight coefficients of different topological regions are adjusted during the aggregation process based on the anomaly tendency metric of nodes in each topological region.

[0065] First, the various equipment nodes of the power system and their connections are constructed into a directed graph structure and stored in a graph database. Each node represents a physical device, such as a substation, distribution unit, or terminal electrical equipment; edges represent the connections between devices and contain attribute information such as the direction of energy flow. When an abnormal event is detected, the following steps are performed to determine the abnormal propagation path and optimize its weights: When a device malfunctions, the node identifier corresponding to the location of the malfunction, such as "N1234", is obtained and used as the starting node for path querying. This node identifier uniquely corresponds to a device node in the graph database.

[0066] Next, a reverse traversal operation is performed using the path query language of the graph database. Specifically, the Cypher query language of the Neo4j graph database can be used to execute a query similar to the following: MATCH (n)-[r:CONNECTS_TO]->(target); WHERE target.nodeId = 'N1234' AND r.energyFlowDirection = 'TOWARDS_TARGET'; RETURN n; This query performs a reverse traversal along the directional information of the edges, extracting all predecessor nodes whose energy flow direction points to the node at the abnormal location as upstream associated nodes. For example, when an anomaly occurs at a substation, the upstream substation node and transmission line node connected to it will be identified.

[0067] Then, a forward traversal is performed, using similar query statements to extract all successor nodes reachable from the anomaly location as downstream affected nodes: MATCH (source)-[r:CONNECTS_TO]->(n); WHERE source.nodeId = 'N1234' AND r.energyFlowDirection = 'FROM_SOURCE'; RETURN n; This query identifies all downstream nodes affected by abnormal locations. For example, when a power distribution station malfunctions, it can affect the power supply to multiple residential communities or industrial power nodes.

[0068] The upstream associated nodes, abnormal location nodes, and downstream affected nodes obtained from the above query are combined according to the energy flow sequence to form a complete abnormal propagation link. This link can be represented as an ordered node sequence, such as: [upstream substation node, upstream line node, abnormal distribution station node, downstream line node, downstream user node].

[0069] For each node in the anomaly propagation chain, its verification records are extracted from the historical database, including factors such as the frequency of past anomalies, the time of the most recent anomaly, and the difficulty of remediation. Based on this data, an anomaly tendency metric is calculated for each node. The calculation method can combine the node's historical anomaly frequency, recent anomaly probability, and importance weight. Anomaly Probability Measure = Historical Anomaly Frequency Coefficient × Recent Anomaly Probability × Node Importance Weight Among them, the historical anomaly frequency coefficient reflects the frequency of anomalies occurring in the past; the recent anomaly probability takes into account the time decay factor, making the most recent anomalies have a higher impact; and the node importance weight is determined according to the node's criticality in the network.

[0070] Then, based on the topological relationships and energy flow direction between nodes, the anomaly propagation link is divided into multiple topological regions. A typical division method is to divide the link into an upstream region, an anomaly center region, and a downstream region. For longer propagation links, this can be further subdivided into multi-level upstream regions and multi-level downstream regions.

[0071] Based on the anomaly tendency measures of nodes within each topological region, a comprehensive anomaly tendency index for the region is calculated. For example, for multiple nodes within a region, a weighted average method can be used: Regional Anomaly Propensity Index = (Anomaly Propensity Measure of Node 1 × Node 1 Weight + ... + Anomaly Propensity Measure of Node n × Node n Weight) / Sum of Weights of All Nodes Finally, based on the anomaly tendency index of each topological region, the weight coefficients of different topological regions during the aggregation process are dynamically adjusted. For example, if the anomaly tendency index of the upstream region is high, the weight of the upstream region in subsequent anomaly analysis and early warning is increased; conversely, if the anomaly tendency of the downstream region is more significant, the weight of the downstream region is increased. The specific adjustment formula can be: Regional weight coefficient = Base weight × (1 + Regional anomaly tendency index adjustment factor) The adjustment factor is directly proportional to the regional anomaly tendency index. This dynamic adjustment mechanism enables anomaly analysis and early warning to more accurately reflect the actual situation, improving the accuracy of anomaly propagation path identification and the timeliness of early warning.

[0072] The above technical solutions enable accurate identification and dynamic weight adjustment of abnormal propagation links, providing effective technical support for power system anomaly handling.

[0073] In one optional implementation, an anomaly tendency metric is calculated for each node based on its historical verification records. The anomaly propagation link is then divided into topological regions containing multiple nodes. The weighting coefficients of different topological regions during the aggregation process are adjusted based on the anomaly tendency metric of nodes within each topological region, including: Based on the historical verification records of each node in the abnormal propagation link, the verification timestamp and verification status identifier are determined, and the time interval between each failed verification status and the current time is calculated according to the verification timestamp. The time interval is subjected to time decay processing to obtain time decay weights. The number of times each node is marked as failing the verification in the historical verification record is weighted and summed with the corresponding time decay weights. Based on the weighted summation result and the number of historical verification records, the abnormal tendency measure value of each node is calculated. The abnormal propagation link is divided into multiple topological regions based on the connection relationship and directionality information of the edges between nodes in the abnormal propagation link. Calculate the connectivity of each node within each topological region, and aggregate the anomaly tendency metric values ​​of each node based on the connectivity of each node to obtain the regional anomaly tendency metric value of the topological region; The anomaly tendency metric of the region is converted into a weight adjustment coefficient, and the initial weight coefficient corresponding to the topological region is adjusted based on the weight adjustment coefficient.

[0074] When calculating the anomaly tendency metric for each node in the anomaly propagation chain, the verification timestamp and verification status identifier are first determined based on historical verification records. Historical verification records typically contain information such as node ID, verification timestamp, and verification result (pass / fail). For each node, all its verification records are extracted, with particular attention paid to those marked "verification failed".

[0075] Calculate the time interval between each failed verification state and the current time based on the verification timestamp. Assume that a node A has three failed verification records with timestamps t1, t2, and t3, and the current time is t_now. Then the corresponding time intervals are Δt1=t_now-t1, Δt2=t_now-t2, and Δt3=t_now-t3.

[0076] The time interval is subjected to time decay processing, and the time decay weight is calculated using an exponential decay function: w_i = e^(-λΔt_i), where λ is the decay factor, controlling the decay rate. More recent records receive greater weight, while more distant records receive less weight, reflecting the characteristic that recent anomalies have a greater impact on the current state.

[0077] The number of times each node is marked as failing verification in the historical verification record is weighted and summed with the corresponding time decay weight. For node A, if there are n verification failure records, its abnormal weighted sum S_A is calculated as S_A = Σ(w_i), where i ranges from 1 to n.

[0078] The anomaly tendency measure of each node is calculated based on the weighted summation result and the total number of historical verification records, N_total. The anomaly tendency measure of node A is AD_A = S_A / N_total. This value ranges from 0 to 1, with a larger value indicating a greater tendency for the node to generate anomalies.

[0079] Next, based on the connectivity between nodes and the directionality of edges in the anomaly propagation path, the path is divided into multiple topological regions. First, a directed graph G=(V,E) is constructed, where V represents the set of nodes and E represents the set of directed edges. The graph G is then partitioned using either a strongly connected component algorithm or a community detection algorithm, resulting in a set of subgraphs {G1, G2, ..., G_k}, each representing a topological region.

[0080] When partitioning topological regions, considering the directionality of edges is crucial. For example, if there is a one-way connection from node A to B, and also a one-way connection from B to C, with no other connections, then A, B, and C are partitioned into the same topological region because anomalies can propagate along A→B→C. If a loop exists, such as A→B→C→A, then these three nodes must belong to the same strongly connected component, forming a single topological region.

[0081] Calculate the connectivity of each node within each topological region, including in-degree and out-degree. The in-degree of node v, in_degree(v), represents the number of edges pointing to that node, and the out-degree, out_degree(v), represents the number of edges originating from that node. The total connectivity of a node, degree(v), = in_degree(v) + out_degree(v).

[0082] The anomaly tendency metrics of each node are aggregated based on their connectivity to obtain the regional anomaly tendency metrics of the topological region. For the topological region G_r, its regional anomaly tendency metric RAD_r is calculated as a weighted average of the anomaly tendency metrics of each node. RAD_r = Σ(AD_v × degree(v)) / Σdegree(v), where v traverses all nodes in G_r.

[0083] Nodes with high connectivity play a crucial role in anomaly propagation and are therefore given higher weight. For example, if a topological region contains three nodes with connectivity of 2, 3, and 5, and anomaly tendency metrics of 0.3, 0.5, and 0.2, respectively, then the anomaly tendency metric for that region is (0.3×2 + 0.5×3 + 0.2×5) / (2+3+5) = 0.32.

[0084] The regional anomaly tendency metric is converted into weight adjustment coefficients, which are used to adjust the initial weight coefficients corresponding to the topological regions. The transformation function can be a linear or non-linear mapping. For example, the sigmoid function can be used for mapping: α_r = 1 + σ(RAD_r - θ), where σ is the sigmoid function and θ is the threshold parameter.

[0085] When the regional anomaly tendency metric RAD_r is greater than the threshold θ, α_r is greater than 1, increasing the weight of the topological region; when RAD_r is less than θ, α_r is less than 1, decreasing the weight of the region. Finally, the adjusted weight of the topological region G_r is W_r' = W_r × α_r, where W_r is the initial weight coefficient.

[0086] This method is also applicable to the field of network security. It can treat network devices as nodes and communication links as edges. By analyzing the historical security check records and topology of each device, more vulnerable network areas can be identified and given more attention in security protection strategies.

[0087] The power simulation data verification system based on artificial intelligence and graph database according to this invention includes: The first unit is used to acquire the dataset to be verified output by the power simulation system. The dataset to be verified includes power network topology information, node operating parameters and time series measurement data. The second unit is used to construct a graph structure based on the power network topology information, and to map the node operating parameters and time-series measurement data to the attribute fields of the corresponding nodes and edges in the graph structure. The third unit is used to store the graph structure in a graph database and extract predefined topological constraint rules and physical law constraint rules through the query language of the graph database. The fourth unit is used to perform embedding representation learning on the graph structure. It generates embedding vectors for each node by aggregating the topological features of the neighboring nodes and the temporal dynamic features of each node in the graph structure, and calculates the anomaly metric of the dataset to be verified based on the distribution of the embedding vectors in the feature space. The fifth unit is used to jointly determine the anomaly metric value with the topological constraint rules and physical law constraint rules, and generate a verification report containing the anomaly location, anomaly type and violation of constraint items; The sixth unit is used to query the upstream associated nodes and downstream affected nodes of the abnormal location through the path query language of the graph database to obtain the abnormal propagation link, and adjust the weight coefficients of different topological regions in the aggregation process based on the historical verification records of the nodes in the abnormal propagation link.

[0088] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0089] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0090] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A power simulation data verification method based on artificial intelligence and graph database, characterized in that, include: Obtain the dataset to be verified output by the power simulation system. The dataset to be verified includes power network topology information, node operating parameters and time-series measurement data. A graph structure is constructed based on the power network topology information, and the node operating parameters and time-series measurement data are mapped to the attribute fields of the corresponding nodes and edges in the graph structure; The graph structure is stored in a graph database, and predefined topological constraint rules and physical law constraint rules are extracted using the query language of the graph database. Embedding representation learning is performed on the graph structure. By aggregating the topological features of the neighboring nodes and the temporal dynamic features of each node in the graph structure, embedding vectors of each node are generated. Based on the distribution of the embedding vectors in the feature space, the anomaly metric of the dataset to be verified is calculated. The anomaly metric is jointly determined with the topological constraint rules and physical law constraint rules to generate a verification report that includes the anomaly location, anomaly type, and violated constraint items. The upstream associated nodes and downstream affected nodes of the abnormal location are queried using the path query language of the graph database to obtain the abnormal propagation link. The weight coefficients of different topological regions in the aggregation process are then adjusted based on the historical verification records of the nodes in the abnormal propagation link.

2. The method according to claim 1, characterized in that, A graph structure is constructed based on the power network topology information, and the node operating parameters and time-series measurement data are mapped to the attribute fields of the corresponding nodes and edges in the graph structure, including: The device identifiers in the power network topology information are parsed to obtain device type labels and device level information. The node type attributes are determined based on the device type labels, and node level coordinates are assigned in the graph structure based on the device level information. Based on the connection relationship between device identifiers in the power network topology information, a directed edge or an undirected edge is created, and the starting node identifier and ending node identifier of the directed edge or the two end node identifiers of the undirected edge are recorded in the edge connection relationship table. Establish a bidirectional index mapping table between device identifiers and graph structure node identifiers; Perform time-domain analysis on the node operating parameters to determine the steady-state and dynamic characteristic quantities in the node operating parameters; The steady-state characteristic quantity is written into the static attribute field of the corresponding node identifier through the forward query path of the bidirectional index mapping table, and the dynamic characteristic quantity and the time series measurement data are written into the time-varying attribute field of the corresponding node identifier.

3. The method according to claim 1, characterized in that, The graph structure is stored in a graph database, and predefined topological constraint rules and physical law constraint rules are extracted using the query language of the graph database, including: The nodes and edges in the graph structure are serialized according to the storage mode of the graph database, and a multi-dimensional index structure is established for the nodes during the serialization process. The multi-dimensional index structure includes a type dimension index grouped according to the node type attribute and a hierarchical dimension index layered according to the node hierarchical coordinates. By traversing the type dimension index using the query language of the graph database, the set of neighbor node types that each node type is allowed to connect to and the connection quantity constraints are determined for each node type, thus obtaining the topology constraint rules; The connection table of the edges is traversed using the query language of the graph database. For each directed edge, the starting node identifier and the ending node identifier recorded in the connection table of the edges are read. The starting node identifier is determined as the starting point of energy flow, and the ending node identifier is determined as the ending point of energy flow. The node storage locations corresponding to the starting and ending points of the energy flow are determined by the multi-dimensional index structure, the node operation parameters are extracted from the node storage locations, and the physical law constraint rules are generated based on the physical correlation between the node operation parameters.

4. The method according to claim 1, characterized in that, Embedding representation learning is performed on the graph structure. This involves generating embedding vectors for each node by aggregating the topological features of its neighboring nodes and temporal dynamic features. Based on the distribution of these embedding vectors in the feature space, anomaly metrics for the dataset to be verified are calculated, including: The set of neighboring nodes of each node is determined from the graph structure, and the topological influence coefficient of each neighboring node on the node is determined by the connection relationship of the edges in the set of neighboring nodes. Extract the time-series measurement data from the time-varying attribute field of each node as time-series dynamic features, and weight and aggregate the time-series dynamic features with the node features of the neighboring nodes according to the topological influence coefficient to generate the aggregated feature vector of each node. The aggregated feature vector is subjected to a nonlinear transformation through iterative updates, and the transformation results after multiple iterations are used as the embedding vectors of each node. In a predefined feature space, the distance metric between the embedding vectors of each node is calculated, and the distance metric between the embedding vector of each node and the center embedding vector of the node type to which each node belongs is used as the node-level anomaly metric. Based on the distribution characteristics of the node-level anomaly metrics in the graph structure, the anomaly metrics of the dataset to be verified are determined.

5. The method according to claim 4, characterized in that, From the graph structure, determine the set of neighboring nodes for each node, and for each neighboring node in the set of neighboring nodes, determine the topological influence coefficient of each neighboring node on the node through the connection relationship of the edges, including: Based on the edge connection information of each node in the graph structure, the identifiers of neighboring nodes directly connected to the node are determined based on the edge connection information, and the nodes corresponding to the neighboring node identifiers are formed into a set of neighboring nodes. For each neighbor node in the set of neighbor nodes, the directional information of the edge connecting the node and the neighbor node is determined by the edge connection table; The energy flow direction of the edge is determined based on the directional information. When the energy flow direction is from the neighbor node to the node, the neighbor node is marked as an input neighbor node. When the energy flow direction is from the node to the neighbor node, the neighbor node is marked as an output neighbor node. The transmission capacity strength value of the edge is calculated based on the edge weight information of the edge connecting the node to each neighboring node. The transmission capability strength value is associated with the labeling type of the neighboring nodes. A positive topology influence coefficient is calculated for neighboring nodes labeled as input neighboring nodes, and a negative topology influence coefficient is calculated for neighboring nodes labeled as output neighboring nodes.

6. The method according to claim 1, characterized in that, The upstream and downstream affected nodes of the anomaly location are queried using the path query language of the graph database to obtain the anomaly propagation path. Based on the historical verification records of the nodes in the anomaly propagation path, the weight coefficients of different topological regions during the aggregation process are adjusted, including: Use the node identifier corresponding to the abnormal location as the starting node for path query; By using the path query language of the graph database, a reverse traversal is performed along the directional information of the edges to extract all predecessor nodes whose energy flow direction points to the node corresponding to the abnormal position, which are then used as upstream associated nodes. By using the path query language of the graph database to perform a forward traversal along the directional information of the edges, all successor nodes reached from the node corresponding to the abnormal position along the energy flow direction of the edge are extracted as downstream affected nodes; The upstream associated nodes, the nodes corresponding to the abnormal locations, and the downstream affected nodes are arranged into an abnormal propagation link according to the energy flow sequence. Based on the historical verification records of each node in the anomaly propagation link, the anomaly tendency metric of each node is calculated, the anomaly propagation link is divided into topological regions containing multiple nodes, and the weight coefficients of different topological regions are adjusted during the aggregation process based on the anomaly tendency metric of nodes in each topological region.

7. The method according to claim 6, characterized in that, Based on the historical verification records of each node in the anomaly propagation link, an anomaly tendency metric is calculated for each node. The anomaly propagation link is then divided into topological regions containing multiple nodes. The weighting coefficients of different topological regions during the aggregation process are adjusted based on the anomaly tendency metric of nodes within each topological region, including: Based on the historical verification records of each node in the abnormal propagation link, the verification timestamp and verification status identifier are determined, and the time interval between each failed verification status and the current time is calculated according to the verification timestamp. The time interval is subjected to time decay processing to obtain time decay weights. The number of times each node is marked as failing the verification in the historical verification record is weighted and summed with the corresponding time decay weights. Based on the weighted summation result and the number of historical verification records, the abnormal tendency measure value of each node is calculated. The abnormal propagation link is divided into multiple topological regions based on the connection relationship and directionality information of the edges between nodes in the abnormal propagation link. Calculate the connectivity of each node within each topological region, and aggregate the anomaly tendency metric values ​​of each node based on the connectivity of each node to obtain the regional anomaly tendency metric value of the topological region; The anomaly tendency metric of the region is converted into a weight adjustment coefficient, and the initial weight coefficient corresponding to the topological region is adjusted based on the weight adjustment coefficient.

8. A power simulation data verification method system based on artificial intelligence and graph database, used to implement the method as described in any one of claims 1-7, characterized in that, include: The first unit is used to acquire the dataset to be verified output by the power simulation system. The dataset to be verified includes power network topology information, node operating parameters and time series measurement data. The second unit is used to construct a graph structure based on the power network topology information, and to map the node operating parameters and time-series measurement data to the attribute fields of the corresponding nodes and edges in the graph structure. The third unit is used to store the graph structure in a graph database and extract predefined topological constraint rules and physical law constraint rules through the query language of the graph database. The fourth unit is used to perform embedding representation learning on the graph structure. It generates embedding vectors for each node by aggregating the topological features of the neighboring nodes and the temporal dynamic features of each node in the graph structure, and calculates the anomaly metric of the dataset to be verified based on the distribution of the embedding vectors in the feature space. The fifth unit is used to jointly determine the anomaly metric value with the topological constraint rules and physical law constraint rules, and generate a verification report containing the anomaly location, anomaly type and violation of constraint items; The sixth unit is used to query the upstream associated nodes and downstream affected nodes of the abnormal location through the path query language of the graph database to obtain the abnormal propagation link, and adjust the weight coefficients of different topological regions in the aggregation process based on the historical verification records of the nodes in the abnormal propagation link.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images