Simulation result processing system and method for digital distribution network system

Abstract

The application discloses a simulation result processing system and method of a distribution network digital system, and relates to the technical field of distribution network digitalization.The system comprises a data module, a feature module, a causality module and a mode module.A fusion data packet is obtained based on simulation result data of a simulation model and distribution network data.The fusion data packet is input into a space-time graph convolution feature extractor to obtain a high-order feature matrix.The high-order feature matrix is input into a causality discovery engine to obtain a dynamic causal graph.The high-order feature matrix and the dynamic causal graph are used to obtain an operation mode.A mode evolution graph is obtained based on the operation mode and the dynamic causal graph.A knowledge base is obtained based on the mode evolution graph, which solves the problem that existing simulation result processing systems cannot automatically mine deep rules and thus cannot form reusable systematic knowledge.

Description

Technical Field

[0001] This invention relates to the field of digital distribution network technology, and more specifically, to a simulation result processing system and method for digital distribution network systems. Background Technology

[0002] With the deepening of the digital and intelligent transformation of power distribution networks, simulation systems based on technologies such as digital twins and cloud-edge collaboration are widely used in scenarios such as power grid planning, operation mode verification, and fault early warning, generating massive amounts of high-dimensional and multi-source simulation result data. Traditional result processing methods mainly rely on engineers manually analyzing charts and reports, which has the following significant drawbacks:

[0003] 1) Inefficient and lacking depth: When faced with trillions / petaflops of data, manual analysis is time-consuming and laborious, making it difficult to conduct comprehensive and in-depth mining. The analysis remains at the level of surface statistics and simple comparison.

[0004] 2) Weak pattern detection capability: It is difficult to automatically identify hidden operating patterns, correlation rules and evolution trends from complex data in order to provide early warning of potential risks;

[0005] 3) Difficulty in knowledge accumulation: The analysis results are mostly in the form of reports and charts, which are fragmented and isolated, and cannot form structured, interconnected and reusable systematic knowledge, making it difficult to support intelligent decision-making. Summary of the Invention

[0006] To address the problem that existing simulation result processing systems cannot automatically uncover deep-seated patterns, thus failing to form reusable systematic knowledge, this invention provides a simulation result processing system for a digital distribution network system. The system includes:

[0007] Data module: used to obtain fused data packets based on simulation results data from the simulation model and distribution network data;

[0008] Feature module: used to input the fused data package into the spatiotemporal graph convolutional feature extractor to obtain a high-order feature matrix;

[0009] Causal module: used to input the high-order feature matrix into the causal discovery engine to obtain a dynamic causal graph;

[0010] Mode module: used to obtain the operating mode based on the high-order feature matrix and the dynamic causal graph;

[0011] Knowledge module: used to obtain a pattern evolution graph based on the operating mode and the dynamic causal graph, and to obtain a knowledge base based on the pattern evolution graph.

[0012] This system constructs fused data packages from multiple sources, providing a high-quality, context-rich, and standardized data foundation. The spatiotemporal graph convolutional feature extractor automatically extracts feature matrices where each dimension is associated with explicit physical semantics, constructing high-order intelligent features with physical interpretability. Through a causal discovery engine, it deeply integrates with the physical constraints and temporal logic of the power grid, automatically identifying and quantifying the causal direction and strength between features, achieving a cognitive leap from data correlation to causal mechanisms. It identifies complex operating modes from massive, high-dimensional data, constructs pattern evolution maps between modes, reveals state transition patterns and risk paths, transforms tacit knowledge into explicit, structured digital assets, solves the problems of knowledge gaps and silos, and ensures the sustainable accumulation and inheritance of power grid operation knowledge. Thus, it automatically and intelligently transforms simulation data into actionable knowledge, enhancing the utilization value and decision support capabilities of distribution network simulation results.

[0013] Furthermore, the distribution network data includes distribution network topology data, equipment parameters, system logs, and external data from third parties. The fused data package includes timestamps, equipment entities, physical quantity values ​​and semantic metadata, associated events, and topology versions.

[0014] The data module is specifically used for:

[0015] A dynamic topology graph is obtained based on the topology data and the device parameters, and the first node of the dynamic topology graph is obtained.

[0016] The steady-state features of the second node in the simulation model are extracted to obtain the feature vector;

[0017] The first sub-topology graph of the simulation model and the second sub-topology graph of the dynamic topology graph are matched to obtain a matching result. Based on the matching result, the mapping relationship and the feature vector, the second node is mapped to the first node to obtain a feature-node topology graph.

[0018] Metadata and semantic event tags are obtained based on the distribution network data. A time-series data stream is obtained based on the semantic event tags and a first preset time window. The metadata includes physical quantity type, associated equipment, equipment type, functional role, and security boundary threshold.

[0019] A data-node topology graph is obtained based on the metadata, the time-series data stream, and the feature-node topology graph, and the fused data packet is obtained based on the data-node topology graph.

[0020] A dynamic topology graph containing node and branch connection relationships, as well as topology version numbers, is generated in real time. A feature-node topology graph is constructed, establishing a direct correspondence between the internal model numbers such as node and branch numbers in the simulation result file and the actual physical device numbers (such as the Chaoyang Line #12 tower switch). This ensures that each data point (such as voltage value) is accurately attached to a specific device entity under a specific version on the dynamic topology graph. A time-series data stream is obtained, and rich semantics and event context are injected into the attached data. This enables the real-time transformation of raw, heterogeneous, and multi-source data streams into spatiotemporally synchronized, topology-aligned, and semantically rich fused data packets, providing high-quality, directly computable input for upper-level intelligent analysis.

[0021] Furthermore, the feature module is specifically used for:

[0022] A first topology graph is constructed based on the topology version, the data-node topology graph, and the second preset time window. The third node and edge of the first topology graph are the connection relationships between the device entities and the device entities, respectively.

[0023] The first topological graph is input into the spatiotemporal graph convolutional feature extractor to obtain the high-order feature matrix at each time point.

[0024] Since the topology version may change over time, this system determines a stable topology for each time window: for each time window, a graph is constructed based on the topology version, and then a high-order feature matrix containing the topology and spatiotemporal context is generated through a spatiotemporal graph convolutional feature extractor.

[0025] Furthermore, the spatiotemporal graph convolutional feature extractor includes several spatiotemporal convolutional blocks, each of which includes a temporal convolutional layer and a spatial graph convolutional layer. The temporal convolutional layer includes several convolutional layers, and the spatial graph convolutional layer includes several graph convolutional layers.

[0026] A spatiotemporal graph convolutional feature extractor performs two operations simultaneously in the same layer: spatial convolution, which aggregates electrical quantities (such as voltage and current) of neighboring nodes along the edges of the topology graph to capture electrical coupling effects; and temporal convolution, which captures the dynamic evolution of each node along its own time axis to obtain a high-order feature vector for each device node at each time step. This vector not only encodes its own historical state but also its contextual state within the network structure. For example, it can characterize the transient response of node A when the bus voltage support is weak.

[0027] Furthermore, the causal module is specifically used for:

[0028] A third preset time window is set for different causal hypothesis types. The third preset time window, the higher-order feature matrix, and the constraints are input into the causal discovery engine to obtain the candidate causal graph at each time point.

[0029] Obtain the undirected edges of the candidate causal graph, and obtain undirected nodes based on the undirected edges;

[0030] Based on the system log, operation events are obtained; based on the operation events, electrical quantity change data is obtained; based on the electrical quantity change data, a first causal direction is obtained.

[0031] Obtain the information transfer metric between any two adjacent undirected nodes, and obtain the second causal direction based on the information transfer metric;

[0032] The prediction residual of the undirected node is obtained based on the physical equation, and the third causal direction is obtained based on the prediction residual;

[0033] The dynamic causal graph at each time point is obtained based on the candidate causal graph, the first causal direction, the second causal direction, and the third causal direction.

[0034] The time window is dynamically set according to different causal hypothesis types, and the time priority principle of cause must precede effect is strictly implemented to ensure that all events and data points are on the same time axis. The dynamic causal graph that determines the causal direction is obtained by using three physical information discrimination methods: intervention signal analysis, information flow time asymmetry analysis and physical model residual test.

[0035] Furthermore, the third preset time window, the higher-order feature matrix, and the constraint conditions are input into the causal discovery engine to obtain the candidate causal graph at each time point;

[0036] The constraints include:

[0037] Topological connectivity constraint: If any two device entities have no direct or indirect connection path in the electrical topology, then it is prohibited to establish a causal edge between the graph nodes corresponding to the two device entities.

[0038] Causal timing constraint: The causal event must occur no later than the result event.

[0039] Power flow direction constraint: In a radial network, first establish causal edges from upstream devices to downstream devices;

[0040] Control logic constraints: Preset automatic control logic relationships are added to the graph structure as known causal edges.

[0041] Injecting domain knowledge serves as both a hard constraint and a soft guide to ensure the reasonableness of the results.

[0042] Furthermore, the mode module is specifically used for:

[0043] Align the higher-order feature matrix and the dynamic causal graph at each time point;

[0044] The structural information of the dynamic causal graph is encoded to obtain a graph embedding vector, and the higher-order feature matrix and the graph embedding vector are fused to obtain several joint representation vectors;

[0045] Obtain the local density of the joint representation vector, obtain the joint representation vector corresponding to the maximum value of the local density to obtain the first prototype, and obtain several candidate prototypes based on the first prototype;

[0046] Based on the candidate prototype clustering of the joint representation vector, several initial clusters are obtained;

[0047] Obtain the center point of each initial cluster, and obtain several global clusters based on the center points;

[0048] The global clustering is clustered based on the preset distribution network operation hierarchy to obtain several regional clusters;

[0049] Several operating modes are obtained based on the region clustering.

[0050] By combining feature views and causal views, we ensure that patterns are not only statistically similar but also consistent in causal drivers. We employ a prototype learning algorithm to find a set of prototypes that can represent the entire data distribution in the feature-causal joint space. We introduce interpretable prototype learning and adaptive hierarchical clustering, and adopt different strategies for different levels of the distribution network (equipment level, feeder level, and regional level) to condense fine-grained operation patterns.

[0051] Furthermore, the specific steps for obtaining several candidate prototypes based on the first prototype include:

[0052] A1. Obtain the first distance between the first prototype and any of the joint representation vectors. Based on the first distance, obtain the joint representation vector that satisfies the iteration condition and whose local density is at its maximum value, and obtain the second prototype.

[0053] A2. Update the first prototype to the second prototype and return to A1;

[0054] A3. Iterate through A1 to A2 until the iteration condition is no longer met, then the iteration ends, and the candidate prototype is obtained based on the first prototype and the second prototype.

[0055] The iteration condition is:

[0056] The first distance is greater than a first threshold, and the local density is greater than a second threshold;

[0057] The specific steps for obtaining several global clusters based on the central points include:

[0058] B1. Based on the central points, cluster the joint representation vector to obtain several first clusters;

[0059] B2. Obtain the target center point of the first cluster, update the center point to the target center point, and return to B1;

[0060] B3. Iterate through B1 to B2 until the termination condition is met, then the iteration ends, and the global cluster is obtained based on the first cluster.

[0061] Furthermore, the knowledge module is specifically used for:

[0062] Based on the operating mode, a set of time points is obtained, the dynamic causal graph corresponding to the set of time points is obtained, several key causal graphs are obtained, and the key causal graphs are aligned.

[0063] Obtain the occurrence frequency of each edge in the key causal graph, obtain key edges based on the occurrence frequency and a preset frequency, obtain the core causal graph based on the key edges, obtain the causal intermediateness of each node in the core causal graph, and obtain the initial node, key node, and result node based on the causal intermediateness.

[0064] Several critical paths are obtained based on the aforementioned key nodes.

[0065] The causal description of the critical path is generated based on the feature-physical quantity mapping dictionary;

[0066] Based on the aforementioned operating mode, the core causal graph, and the causal narrative, several pattern-causal knowledge pairs are obtained.

[0067] If any two running modes share the same initial node, then the two running modes have a sharing relationship; if the initial node of any one running mode is the result node of another running mode, then the two running modes have a temporal relationship.

[0068] The pattern evolution map is obtained based on the pattern-causal knowledge pairs, the sharing relationship, and the temporal relationship.

[0069] Extract the common key causal structure of patterns from the causal graph and transform it into a causal chain described by physical quantities that engineers can understand. Provide a causal explanation for each pattern, and then encapsulate the pattern information and causal narrative into a structured knowledge pair object. Establish the relationship between knowledge pairs to form a networked knowledge system, and complete the sublimation from data to features, then to patterns, and finally to interpretable knowledge.

[0070] This invention also provides a method for processing simulation results of a digital distribution network system, the method comprising:

[0071] A fused data package is obtained based on simulation results data from the simulation model and distribution network data;

[0072] The fused data package is input into the spatiotemporal graph convolutional feature extractor to obtain a high-order feature matrix;

[0073] The high-order feature matrix is ​​input into the causal discovery engine to obtain a dynamic causal graph;

[0074] The operating mode is obtained based on the higher-order feature matrix and the dynamic causal graph;

[0075] Based on the operating mode and the dynamic causal graph, a pattern evolution map is obtained.

[0076] The principle and effect of this method are similar to those of this system, and therefore, no further details will be provided for this method.

[0077] One or more technical solutions provided by this invention have at least the following technical effects or advantages:

[0078] This system integrates data packets to provide a high-quality, context-rich, standardized data foundation; the spatiotemporal graph convolutional feature extractor constructs high-order intelligent features with physical interpretability; the causal discovery engine automatically identifies and quantifies the causal direction and strength between features, achieving a cognitive leap from data correlation to causal mechanisms; it identifies complex operating modes from massive, high-dimensional data, constructs pattern evolution maps between modes, reveals state transition laws and risk paths, transforms tacit knowledge into explicit, structured digital assets, solves the problems of knowledge gaps and silos, and ensures the sustainable accumulation and inheritance of power grid operation knowledge; thus, it automatically and intelligently transforms simulation data into operable knowledge to enhance the utilization value and decision support capabilities of distribution network simulation results. Attached Figure Description

[0079] The accompanying drawings, which are provided to further illustrate embodiments of the invention and constitute a part of this invention, are not intended to limit the scope of the invention.

[0080] Figure 1 This is a flowchart illustrating the simulation result processing system of the digital distribution network system in this invention. Detailed Implementation

[0081] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, where there is no conflict, the embodiments of the present invention and the features thereof can be combined with each other.

[0082] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0083] Example 1

[0084] refer to Figure 1 This embodiment provides a simulation result processing system for a digital distribution network system, the system comprising:

[0085] Data module: used to obtain fused data packets based on simulation results data from the simulation model and distribution network data;

[0086] Feature module: used to input the fused data package into the spatiotemporal graph convolutional feature extractor to obtain a high-order feature matrix;

[0087] Causal module: used to input the high-order feature matrix into the causal discovery engine to obtain a dynamic causal graph;

[0088] Mode module: used to obtain the operating mode based on the high-order feature matrix and the dynamic causal graph;

[0089] Knowledge module: used to obtain a pattern evolution graph based on the operating mode and the dynamic causal graph, and to obtain a knowledge base based on the pattern evolution graph.

[0090] The distribution network data includes distribution network topology data, equipment parameters, system logs, and external data from third parties (such as meteorological and geographic information). The fused data package includes timestamps (which can be obtained from the timestamps corresponding to the topology version number), equipment entities (which can be obtained from the equipment to which they belong), physical quantity values ​​and semantic metadata (which may include physical quantity values, physical quantity types, and physical quantity units), associated events (which can be obtained from time-series data streams and may include semantic event tags and their corresponding timestamps and equipment entities), and topology version.

[0091] The data module is specifically used for:

[0092] A dynamic topology map is obtained based on the topology data (such as topology data of the distribution network at different times and the topology change sequence) and the equipment parameters. This map not only includes the connection relationship of nodes and branches, but also the topology version number and the effective time window. For example, according to the topology changes of the power grid (such as the change of switch status), the topology is divided into different versions, and each version corresponds to a unique topology version number. For each timestamp, the topology version to which it belongs is determined.

[0093] Obtain the first node of the dynamic topology graph;

[0094] The steady-state characteristics of the second node in the simulation model (such as the reference voltage level, the number of connected branch types, and the typical range of injected power) are extracted to obtain the feature vector;

[0095] The first sub-topology graph of the simulation model and the second sub-topology graph of the dynamic topology graph are matched to obtain the matching result. Based on the SimRank algorithm, the local topology sub-graph of the simulation model and the real-time dynamic topology graph are matched for subgraph isomorphism or similarity. Similarity can also be measured from multiple perspectives, such as degree distribution, clustering coefficient, shortest path distance and subgraph matching graph.

[0096] Based on the matching results, mapping relationships, and feature vectors, the second node is mapped to the first node to obtain a feature-node topology graph. Alternatively, a lightweight neural network can be trained using historically verified mapping relationships (the correspondence between the internal number of the simulation model and the number of the real physical device) as a training set. For simulation results, based on their electrical quantities and topology context, the internal nodes are automatically mapped to the entity device nodes on the dynamic topology graph, so that each data point (such as voltage value) is precisely attached to a specific device entity under a specific version on the dynamic topology graph.

[0097] Based on the power distribution network data, metadata and semantic event tags are obtained. For example, using natural language processing technology, key information such as lightning strikes and insulation aging is extracted from the system logs, and these events are used as semantic event tags. Their timestamps and corresponding device names are obtained to associate and bind them with time-series data streams of specific time windows and specific devices.

[0098] Based on the semantic event tags and the first preset time window, a time-series data stream is obtained. For example, the text record: "14:30 Inspection found that the oil temperature of transformer #1 was too high" is taken as an event and associated with the load current and top oil temperature simulation data of the transformer before and after 14:30, providing cause-data pairs for subsequent analysis.

[0099] The metadata includes physical quantity type, associated device, device type, functional role, and security boundary threshold;

[0100] In this embodiment, the physical quantity type describes what physical quantity the data point represents. Examples include voltage, current, active power, reactive power, frequency, and temperature. For greater precision, finer-grained classifications can be included, such as voltage fundamental amplitude, voltage harmonic distortion rate, and current phase angle.

[0101] The device to which the physical quantity belongs specifies which specific physical device it pertains to. For example: transformer T1, line L12, busbar B3, and photovoltaic inverter PV_inv_5, etc.

[0102] The equipment type describes the category of the equipment, such as: transformer, overhead line, cable, circuit breaker, disconnector, photovoltaic inverter, wind turbine, and energy storage converter.

[0103] Functional roles describe the function of the device in the power grid, such as: critical interconnection node, photovoltaic access point, important load power supply node, backup power access point, and voltage support node.

[0104] The safety boundary threshold represents the safe range that a physical quantity is allowed under current operating conditions.

[0105] A data-node topology graph is obtained based on the metadata, the time-series data stream, and the feature-node topology graph. The fused data package is then obtained based on the data-node topology graph. All metadata, time-series data streams, and feature vectors are mapped to specific device entities under a specific version on the dynamic topology graph.

[0106] In this embodiment, before obtaining the dynamic topology graph, the data module further includes:

[0107] All incoming streaming and file data are assigned a unified system simulation clock and geographic / logical coordinate encoding to achieve spatiotemporal benchmark unification:

[0108] For data with inconsistent simulation step sizes, a resampling algorithm based on physical process interpolation is adopted. For example, when performing microsecond-level interpolation on electromechanical transient results, not only is mathematical interpolation (such as splines) used, but also the switching action event at that moment (from the event log in the system log) is combined to reduce the smooth transition that violates physical laws at the switching moment and ensure the physical authenticity of the data in the time domain.

[0109] Specifically, the feature module is used for:

[0110] A first topology graph is constructed based on the topology version, the data-node topology graph, and the second preset time window. For each time window, a graph is constructed based on the topology version and the data-node topology graph. The third node and edge of the first topology graph are the connection relationships between the device entity and the device entity, respectively.

[0111] The first topology graph is input into the spatiotemporal graph convolutional feature extractor to obtain the high-order feature matrix at each time point. In this embodiment, the high-order feature matrix includes several high-order feature vectors. Each high-order feature vector includes dimensional structure, feature source, feature type, feature organization method, and feature metadata (which may include feature name, topology version corresponding to the feature, device, device type, functional role, and security boundary threshold), physical meaning or mathematical definition of the feature, data source of the feature, and relationships between features (such as correlation and causality).

[0112] Dimensional structure: Assuming there are F device nodes and G time steps, and each time step extracts H features at each node, then the shape of the high-order feature matrix is ​​[F, G, H].

[0113] Feature types can include:

[0114] Transient features: Capture rapidly changing electromagnetic transient processes, such as voltage drops during faults and waveform characteristics of lightning overvoltages; Steady-state features: Describe the characteristics of the system during steady-state operation, such as voltage levels, load rates, and power factors; Topological features: Reflect the location and connectivity of nodes in the power grid topology, such as nodal degree, electrical distance, and node importance (eigenvector centrality); Propagation features: Describe the propagation characteristics of electrical quantities in the power grid, such as the direction, speed, and attenuation of fault disturbance propagation; Correlation features: Reflect the degree of correlation between electrical quantities between different nodes, such as voltage correlation and power transfer factor; Causal features: Based on the results of the causal discovery engine, encode the strength or direction of causal relationships as features.

[0115] Feature organization: The H dimensions of the feature matrix are not randomly arranged, but grouped according to certain physical meanings. For example, the first F1 dimensions are voltage-related features, then the F2 dimensions are current-related features, and then the power-related features.

[0116] The spatiotemporal graph convolutional feature extractor includes several spatiotemporal convolutional blocks. Each spatiotemporal convolutional block includes a temporal convolutional layer and a spatial graph convolutional layer. The temporal convolutional layer includes several convolutional layers (such as one-dimensional convolutions), and the spatial graph convolutional layer includes several graph convolutional layers (such as graph convolutional networks, GCNs).

[0117] Specifically, the causal module is used for:

[0118] A third preset time window is set for different causal hypothesis types. For example, for fast electromagnetic transient causality (such as overvoltage caused by lightning), the time interval between cause and effect is in the millisecond range; for slow electromechanical dynamic causality (such as transformer overload caused by load growth), the time window is widened to the minute range. This setting can be based on the power grid physical process knowledge base to make the search more accurate and efficient.

[0119] The third preset time window, the higher-order feature matrix, and the constraints are input into the causal discovery engine to obtain the candidate causal graph for each time point.

[0120] In this embodiment, the causal discovery engine can be a hybrid inference model formed by integrating multiple methods and power grid domain knowledge bases (such as physical laws (Kirchhoff's laws, power balance), equipment control logic, protection coordination principles, etc.). The multiple methods may include:

[0121] Structure learning based on conditional independence (such as the Peter-Clark algorithm and the FCI (Fast Causal Inference) algorithm): is used to initially discover conditional dependencies between variables;

[0122] Causal discovery based on time-series information (such as Granger causality, transitive entropy, and other algorithms): used to infer causality by utilizing the chronological order of time;

[0123] Inference based on structural causal models (such as Do-Calculus, counterfactual reasoning, etc.) is used to quantify causal effects and answer the question of what would happen if...

[0124] Machine learning-based causal representation learning (such as variational autoencoders (VAEs) and causal discovery techniques) is used to learn low-dimensional causal factors from high-dimensional features.

[0125] Obtain the undirected edges of the candidate causal graph, and obtain undirected nodes based on the undirected edges; determine the direction of the causal edges (XY) generated in the candidate causal graph that still have ambiguous directions;

[0126] Based on the system log, operation events are obtained, electrical quantity change data is obtained based on the operation events, and a first causal direction is obtained based on the electrical quantity change data. For example, if the system log records a clear capacitor bank switching operation, observe the changes in related electrical quantities before and after this operation. If the voltage of a certain node changes significantly after the switching while other factors remain unchanged, a strong causal chain of capacitor switching-node voltage can be established.

[0127] Obtain the information transfer metric between any two adjacent undirected nodes, and obtain the second causal direction based on the information transfer metric; such as calculating the information transfer function from X to Y and from Y to X (e.g., based on the existing Granger test) or transfer entropy (e.g., binning, kernel density estimation, and k-nearest neighbor algorithms). In physical systems, the information transfer from cause to effect usually has a specific time delay and decay pattern. By analyzing the asymmetry, the dominant causal direction can be inferred.

[0128] The prediction residuals of the undirected nodes are obtained based on the physical equations, and the third causal direction is obtained based on the prediction residuals. The prediction residuals are calculated using simplified physical equations (such as power flow equations). If the change of X can better explain the prediction residuals of Y, then X is tended to be the cause of Y.

[0129] The dynamic causal graph at each time point is obtained based on the candidate causal graph, the first causal direction, the second causal direction, and the third causal direction.

[0130] Specifically, the third preset time window, the higher-order feature matrix, and the constraints are input into the causal discovery engine to obtain the candidate causal graph at each time point; the constraints include:

[0131] Topological connectivity constraint: If any two device entities have no direct or indirect connection path in the electrical topology, then it is prohibited to establish a causal edge between the graph nodes corresponding to the two device entities.

[0132] Causal timing constraint: The occurrence of the causal event must be no later than the occurrence of the result event; for example, the load value at a future time cannot be the cause of the voltage value at a past time.

[0133] Power flow direction constraint: In a radial operating network, a causal edge is first established from the upstream device to the downstream device; in a normally operating radial distribution network, the power flow direction is from the power source to the load, which provides a strong prior for the causal direction (upstream is the cause, downstream is the effect).

[0134] Control logic constraints: Pre-defined automatic control logic relationships are added to the graph structure as known causal edges. For example, known automatic control logic (such as voltage-reactive power control) can be directly added to the graph as a known causal link.

[0135] Specifically, the mode module is used for:

[0136] Align the higher-order feature matrix and the dynamic causal graph at each time point;

[0137] The structural information of the dynamic causal graph (such as the adjacency matrix or causal edges between nodes) is encoded to obtain a graph embedding vector. The higher-order feature matrix and the graph embedding vector are fused to obtain several joint representation vectors. In this embodiment, the fusion method can be concatenation or weighted summation.

[0138] The prototype learning method is used to obtain the local density of the joint representation vector (which can be obtained by algorithms such as k-nearest neighbor density estimation, kernel density estimation and DBSCAN density), obtain the joint representation vector corresponding to the maximum value of the local density to obtain the first prototype, and obtain several candidate prototypes based on the first prototype.

[0139] Based on the candidate prototype clustering of the joint representation vector, several initial clusters are obtained;

[0140] Obtain the center point of each initial cluster, and obtain several global clusters based on the center points;

[0141] The global clustering is clustered based on the preset distribution network operation levels (such as overall operation mode, regional operation mode and equipment group operation mode) to obtain several regional clusters; for each global mode, its corresponding data points are extracted and clustered again to obtain a more granular regional mode.

[0142] Based on the region clustering, several operating modes are obtained. The generated joint representation vectors are clustered to group similar operating states into one category, thus forming an operating mode.

[0143] The specific steps for obtaining several candidate prototypes based on the first prototype include:

[0144] A1. Obtain the first distance between the first prototype and any of the joint representation vectors. Based on the first distance, obtain the joint representation vector that satisfies the iteration condition and whose local density is at its maximum value, and obtain the second prototype.

[0145] A2. Update the first prototype to the second prototype and return to A1;

[0146] A3. Iterate through A1 to A2 until the iteration condition is no longer met, then the iteration ends, and the candidate prototype is obtained based on the first prototype and the second prototype.

[0147] The iteration condition is:

[0148] The first distance is greater than a first threshold, and the local density is greater than a second threshold;

[0149] The specific steps for obtaining several global clusters based on the central points include:

[0150] B1. Based on the central points, cluster the joint representation vector to obtain several first clusters;

[0151] B2. Obtain the target center point of the first cluster, update the center point to the target center point, and return to B1;

[0152] B3. Iterate through B1 to B2 until the termination condition is met, at which point the iteration ends, and the global cluster is obtained based on the first cluster. In this embodiment, the termination condition can be reaching the required number of iterations or the target centroid no longer changing.

[0153] For example, calculate the local density of each data point and select the point with the highest density as the first prototype. Then, among the remaining points, select the point that is far enough away from the existing prototype (exceeding a threshold) and has the highest local density as the next prototype. Repeat this process until there are no more points that meet the criteria, thus obtaining several candidate prototypes.

[0154] Each data point is then assigned to the nearest candidate prototype, forming an initial cluster. Within each cluster, the center point of that cluster is reselected as the new prototype, and this process is repeated until the prototypes no longer change or the maximum number of iterations is reached.

[0155] Specifically, the knowledge module is used for:

[0156] Based on the operating mode, a set of time points is obtained, the dynamic causal graph corresponding to the set of time points is obtained, and several key causal graphs are obtained. For each mode, a set of all time points belonging to that mode is obtained, and each time point has its corresponding causal graph.

[0157] Align the key causal graphs; since the causal graphs at different time points may differ slightly, it is necessary to align all causal graphs within the time point set (the nodes are the same, but the edges may be different) to reduce structural differences caused by differences in feature selection.

[0158] The frequency of occurrence of each edge in the key causal graph is obtained. Key edges are obtained based on the frequency of occurrence and a preset frequency. The core causal graph is obtained based on the key edges. Causal edges that appear stably in the pattern are selected to form a pattern-level core graph.

[0159] Obtain the causal intermediateness (e.g., the weighted sum of in-degree and out-degree) of each node in the core causal graph, and obtain the initial node (a node with an in-degree of 0 or very low, usually an external driving factor or initial cause of the pattern), key node (a point corresponding to high causal intermediateness) and result node (a node with an out-degree of 0 or very low, usually the final manifestation or result of the pattern) based on the causal intermediateness.

[0160] Several key paths are obtained based on the aforementioned key nodes;

[0161] The causal description of the critical path is generated based on the feature-physical quantity mapping dictionary (used to map higher-order feature dimensions back to the interpretation of specific physical quantities); for example, the critical path: f_123-f_456-f_789, causal description: total regional solar irradiance - total active power output of photovoltaic power station - 110kV bus voltage.

[0162] Based on the aforementioned operating mode, the core causal graph, and the causal narrative, several pattern-causal knowledge pairs are obtained.

[0163] If any two operating modes share the same initial node, then the two operating modes have a sharing relationship; if the two modes share certain key causal nodes, then an association is established. For example, if mode 1 and mode 2 both share photovoltaic output as a cause, but lead to different results, then they can be associated as a pair of modes with the same cause but different effects.

[0164] If the initial node of any of the operating modes is the result node of another operating mode, then the two operating modes have a temporal relationship; if the result variable in the knowledge pair of one mode happens to be the cause variable of another mode, then a temporal evolutionary association is established.

[0165] The pattern evolution map is obtained based on the pattern-causal knowledge pairs, the sharing relationship, and the temporal relationship.

[0166] In this embodiment, the system may further include:

[0167] Visualization module: Used to display the dynamic causal graph, the feature-node topology graph, the data-node topology graph, the candidate causal graph, the operating mode, the core causal graph, and the mode evolution graph based on visualization technology, and to highlight key nodes, key nodes, result nodes, and key paths in red.

[0168] Example 2

[0169] Based on Example 1, this example also provides a method for processing simulation results of a digital distribution network system, the method including:

[0170] A fused data package is obtained based on simulation results data from the simulation model and distribution network data;

[0171] The fused data package is input into the spatiotemporal graph convolutional feature extractor to obtain a high-order feature matrix;

[0172] The high-order feature matrix is ​​input into the causal discovery engine to obtain a dynamic causal graph;

[0173] The operating mode is obtained based on the higher-order feature matrix and the dynamic causal graph;

[0174] Based on the operating mode and the dynamic causal graph, a pattern evolution map is obtained.

[0175] 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.

[0176] 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 simulation result processing system for a digital distribution network system, characterized in that, The system includes: Data module: used to obtain fused data packets based on simulation results data from the simulation model and distribution network data; Feature module: used to input the fused data package into the spatiotemporal graph convolutional feature extractor to obtain a high-order feature matrix; Causal module: used to input the high-order feature matrix into the causal discovery engine to obtain a dynamic causal graph; Mode module: used to obtain the operating mode based on the high-order feature matrix and the dynamic causal graph; Knowledge module: used to obtain a pattern evolution graph based on the operating mode and the dynamic causal graph, and to obtain a knowledge base based on the pattern evolution graph; The knowledge module is specifically used for: Based on the operating mode, a set of time points is obtained, the dynamic causal graph corresponding to the set of time points is obtained, several key causal graphs are obtained, and the key causal graphs are aligned. Obtain the occurrence frequency of each edge in the key causal graph, obtain key edges based on the occurrence frequency and a preset frequency, obtain the core causal graph based on the key edges, obtain the causal intermediateness of each node in the core causal graph, and obtain the initial node, key node, and result node based on the causal intermediateness. Several key paths are obtained based on the aforementioned key nodes; The causal description of the critical path is generated based on the feature-physical quantity mapping dictionary; Based on the aforementioned operating mode, the core causal graph, and the causal narrative, several pattern-causal knowledge pairs are obtained. If any two running modes share the same initial node, then the two running modes have a sharing relationship; if the initial node of any one running mode is the result node of another running mode, then the two running modes have a temporal relationship. The pattern evolution map is obtained based on the pattern-causal knowledge pairs, the sharing relationship, and the temporal relationship.

2. The simulation result processing system for the digital distribution network system according to claim 1, characterized in that, The distribution network data includes distribution network topology data, equipment parameters, system logs, and external data from third parties. The fused data package includes timestamps, equipment entities, physical quantity values ​​and semantic metadata, associated events, and topology versions. The data module is specifically used for: A dynamic topology graph is obtained based on the topology data and the device parameters, and the first node of the dynamic topology graph is obtained. The steady-state features of the second node in the simulation model are extracted to obtain the feature vector; The first sub-topology graph of the simulation model and the second sub-topology graph of the dynamic topology graph are matched to obtain a matching result. Based on the matching result, the mapping relationship and the feature vector, the second node is mapped to the first node to obtain a feature-node topology graph. Metadata and semantic event tags are obtained based on the distribution network data. A time-series data stream is obtained based on the semantic event tags and a first preset time window. The metadata includes physical quantity type, associated equipment, equipment type, functional role, and security boundary threshold. A data-node topology graph is obtained based on the metadata, the time-series data stream, and the feature-node topology graph, and the fused data packet is obtained based on the data-node topology graph.

3. The simulation result processing system for the digital distribution network system according to claim 2, characterized in that, The feature module is specifically used for: A first topology graph is constructed based on the topology version, the data-node topology graph, and the second preset time window. The third node and edge of the first topology graph are the connection relationships between the device entities and the device entities, respectively. The first topology graph is input into the spatiotemporal graph convolutional feature extractor to obtain the high-order feature matrix at each time point.

4. The simulation result processing system for the digital distribution network system according to claim 3, characterized in that, The spatiotemporal graph convolutional feature extractor includes several spatiotemporal convolutional blocks. Each spatiotemporal convolutional block includes a temporal convolutional layer and a spatial graph convolutional layer. The temporal convolutional layer includes several convolutional layers, and the spatial graph convolutional layer includes several graph convolutional layers.

5. The simulation result processing system for the digital distribution network system according to claim 4, characterized in that, The causal module is specifically used for: A third preset time window is set for different causal hypothesis types. The third preset time window and the higher-order feature matrix are input into the causal discovery engine to obtain the candidate causal graph at each time point. Obtain the undirected edges of the candidate causal graph, and obtain undirected nodes based on the undirected edges; Based on the system log, operation events are obtained; based on the operation events, electrical quantity change data is obtained; based on the electrical quantity change data, a first causal direction is obtained. Obtain the information transfer metric between any two adjacent undirected nodes, and obtain the second causal direction based on the information transfer metric; The prediction residual of the undirected node is obtained based on the physical equation, and the third causal direction is obtained based on the prediction residual; The dynamic causal graph at each time point is obtained based on the candidate causal graph, the first causal direction, the second causal direction, and the third causal direction.

6. The simulation result processing system for the digital distribution network system according to claim 5, characterized in that, The third preset time window, the higher-order feature matrix, and the constraints are input into the causal discovery engine to obtain the candidate causal graph at each time point; The constraints include: Topological connectivity constraint: If any two device entities have no direct or indirect connection path in the electrical topology, then it is prohibited to establish a causal edge between the graph nodes corresponding to the two device entities. Causal timing constraint: The causal event must occur no later than the result event. Power flow direction constraint: In a radial network, first establish causal edges from upstream devices to downstream devices; Control logic constraints: Preset automatic control logic relationships are added to the graph structure as known causal edges.

7. The simulation result processing system for the digital distribution network system according to claim 6, characterized in that, The mode module is specifically used for: Align the higher-order feature matrix and the dynamic causal graph at each time point; The structural information of the dynamic causal graph is encoded to obtain a graph embedding vector, and the higher-order feature matrix and the graph embedding vector are fused to obtain several joint representation vectors; Obtain the local density of the joint representation vector, obtain the joint representation vector corresponding to the maximum value of the local density to obtain the first prototype, and obtain several candidate prototypes based on the first prototype; Based on the candidate prototype clustering of the joint representation vector, several initial clusters are obtained; Obtain the center point of each initial cluster, and obtain several global clusters based on the center points; The global clustering is clustered based on the preset distribution network operation hierarchy to obtain several regional clusters; Several operating modes are obtained based on the region clustering.

8. The simulation result processing system for the digital distribution network system according to claim 7, characterized in that, The specific steps for obtaining several candidate prototypes based on the first prototype include: A1. Obtain the first distance between the first prototype and any of the joint representation vectors. Based on the first distance, obtain the joint representation vector that satisfies the iteration condition and whose local density is at its maximum value, and obtain the second prototype. A2. Update the first prototype to the second prototype and return to A1; A3. Iterate through A1 to A2 until the iteration condition is no longer met, then the iteration ends, and the candidate prototype is obtained based on the first prototype and the second prototype. The iteration condition is: The first distance is greater than a first threshold, and the local density is greater than a second threshold; The specific steps for obtaining several global clusters based on the central points include: B1. Based on the central points, cluster the joint representation vector to obtain several first clusters; B2. Obtain the target center point of the first cluster, update the center point to the target center point, and return to B1; B3. Iterate through B1 to B2 until the termination condition is met, then the iteration ends, and the global cluster is obtained based on the first cluster.

9. A method for processing simulation results of a digital distribution network system, characterized in that, The method includes: A fused data package is obtained based on simulation results data from the simulation model and distribution network data; The fused data package is input into the spatiotemporal graph convolutional feature extractor to obtain a high-order feature matrix; The high-order feature matrix is ​​input into the causal discovery engine to obtain a dynamic causal graph; The operating mode is obtained based on the high-order feature matrix and the dynamic causal graph; Based on the operating mode and the dynamic causal graph, a pattern evolution graph is obtained, and based on the pattern evolution graph, a knowledge base is obtained. The specific steps for obtaining the pattern evolution map include: Based on the operating mode, a set of time points is obtained, the dynamic causal graph corresponding to the set of time points is obtained, several key causal graphs are obtained, and the key causal graphs are aligned. Obtain the occurrence frequency of each edge in the key causal graph, obtain key edges based on the occurrence frequency and a preset frequency, obtain the core causal graph based on the key edges, obtain the causal intermediateness of each node in the core causal graph, and obtain the initial node, key node, and result node based on the causal intermediateness. Several key paths are obtained based on the aforementioned key nodes; The causal description of the critical path is generated based on the feature-physical quantity mapping dictionary; Based on the aforementioned operating mode, the core causal graph, and the causal narrative, several pattern-causal knowledge pairs are obtained. If any two running modes share the same initial node, then the two running modes have a sharing relationship; if the initial node of any one running mode is the result node of another running mode, then the two running modes have a temporal relationship. The pattern evolution map is obtained based on the pattern-causal knowledge pairs, the sharing relationship, and the temporal relationship.

Images