Petal type net rack fault traceability and dynamic reconstruction method based on graph neural network

Abstract

The invention provides a petal-shaped net rack fault tracing and dynamic reconstruction method based on a graph neural network, and the method comprises the steps: firstly obtaining multi-dimensional electrical data of a four-stage hierarchical structure of a petal-shaped net rack, and carrying out the preprocessing of the data; inputting the preprocessed multi-dimensional electrical data into a four-level graph neural network to obtain a fault traceability result; according to a fault traceability result, generating a hierarchical dynamic reconstruction strategy and executing the strategy so as to isolate the fault and restore power supply; wherein the four-level graph neural network comprises a branch micro-graph layer, a partition sub-graph layer, a global interconnection layer and a path tracking layer. According to the method, the four-level graph neural network matched with the petal-shaped net rack physical hierarchy is constructed, local-to-global refined fault traceability is realized, and a hierarchical dynamic reconstruction strategy is driven accordingly, so that the defects of the traditional technology in the aspects of fault positioning precision, traceability depth, decision-making intelligence and cooperation efficiency are overcome, and the fault positioning accuracy, traceability depth, decision-making intelligence and cooperation efficiency are improved. And finally, optimization of fault isolation and power supply recovery is realized.

Description

Technical Field

[0001] This invention relates to the field of fault tracing technology, and in particular to a method for fault tracing and dynamic reconstruction of petal-shaped network structures based on graph neural networks. Background Technology

[0002] With the transformation of the energy structure and the advancement of "dual carbon" goals, new elements such as distributed generation (DG), energy storage systems, and electric vehicles are being integrated into the distribution network in large numbers, driving the evolution of the traditional unidirectional radial distribution network towards a proactive, multi-source, and complex distribution network form. Among them, the petal-shaped grid structure has become an important development direction for future smart distribution networks, especially in scenarios such as high-density urban parks and important load centers, due to its high power supply reliability, flexible operation mode, and good scalability.

[0003] Petal-shaped power grids typically consist of multiple structurally similar "petal" sections, each operating radially and interconnected via connecting lines between core nodes. This design combines the advantages of a simple, easily controllable radial network with the high reliability of a ring network. However, this complex hierarchical structure also presents unprecedented challenges, particularly in fault handling.

[0004] Traditional network reconfiguration schemes are mostly based on pre-set fixed logic or offline generated plans, resulting in a rigid decision-making process. Faced with the variable operating modes, uncertain distributed power output, and dynamic load distribution in a petal-shaped network structure, these strategies are often not optimal solutions. They may cause unnecessary expansion of power outage areas, excessive switching operations, or introduce new safety issues (such as line overload or voltage exceeding limits) during power restoration, failing to achieve refined and intelligent dynamic reconfiguration. Summary of the Invention

[0005] This invention provides a method for fault tracing and dynamic reconstruction of petal-shaped network structures based on graph neural networks, which is used to solve the problem of not being able to achieve refined and intelligent dynamic reconstruction.

[0006] A first aspect of this invention provides a method for fault tracing and dynamic reconstruction of petal-shaped network structures based on graph neural networks, comprising: Obtain multi-dimensional electrical data of the four-level structure of the petal-shaped space frame, and preprocess the data; The preprocessed multi-dimensional electrical data is input into a four-level graph neural network to obtain fault tracing results; Based on the fault tracing results, a hierarchical dynamic reconfiguration strategy is generated and executed to isolate the fault and restore power supply; The four-level graph neural network includes branch micro-layers, partition sub-layers, global interconnection layers, and path tracing layers.

[0007] In one possible implementation, the preprocessed multi-dimensional electrical data is input into a four-level graph neural network to obtain fault tracing results, including: Multi-dimensional electrical data is input into the branch micro-layer, and a micro-map is constructed with a single radial branch as the unit. A time-series graph convolutional network is used for analysis to output the branch fault probability, preliminary fault location, and preliminary fault type judgment. By inputting branch fault probability, preliminary fault location, preliminary fault type judgment and corresponding multi-dimensional electrical data of the sub-layer of the sub-zone, the fault confidence heat map, fault influence range and fault spread rate within the sub-zone are obtained. Input the fault confidence heatmap, fault impact range and fault propagation speed of each partition sub-layer into the global interconnection layer to obtain the cross-region fault risk level and the list of adjacent petal partitions that may be affected. The outputs of the branch micro-layer, the partition sub-layer, and the global interconnection layer are fused according to time series and input into the path tracing layer to obtain a list of fault propagation paths sorted by probability, fault source nodes, and fault occurrence timestamps of each node, which serve as the fault tracing results.

[0008] In one possible implementation, the method further includes: Using a single radial branch as the physical prototype, nodes are defined as end load nodes, distributed power access points, and branch nodes within the branch. The weight of the edge connecting the node is set to the reciprocal of the line impedance to quantify the electrical distance between nodes and complete the construction of the branch micro-layer. Using a single petal partition as the physical prototype, a tree-like topology is constructed with the core node as the root, the branch nodes as the middle layer, and the terminal nodes as the leaf nodes. Partition ownership labels and hierarchy labels are embedded in the node feature vectors to complete the construction of partition sub-layers. Using the connection lines between core nodes as the physical prototype, each core node is abstracted into a macro node, and the weight of the edge is zero in a non-faulty state, thus completing the construction of the global interconnection layer. Extract the overall structural features and local structural features of the branch micro-layer, partition sub-layer, and global interconnection layer. Based on the overall and local structural features, determine the parameters of the dynamic temporal attention mechanism of the temporal prediction model to complete the construction of the path tracing layer.

[0009] In one possible implementation, a single radial branch is used as the physical prototype. Nodes are defined as end-load nodes, distributed power supply access points, and branch nodes within that branch. The weights of the edges connecting nodes are set to the reciprocal of the line impedance to quantify the electrical distance between nodes, thus completing the construction of the branch micro-layer, including: The terminal load nodes, distributed power supply access points, and branch nodes within the branch are used as modeling nodes, and corresponding monitoring dimensions are configured for different types of nodes. The monitoring dimensions for terminal load nodes and distributed power supply access points include current distortion, temperature change, and voltage fluctuation, while the monitoring dimensions for branch nodes include current RMS value, switch on / off status, and load fluctuation coefficient. Based on the line impedance values ​​between modeling nodes, the weight of the edge connecting each node is set to the reciprocal of the line impedance to achieve a negative correlation quantitative matching between the edge weight and the electrical distance between nodes. A temporal graph convolutional network is used as the core computational unit, and a temporal-spatial two-dimensional convolutional kernel is configured to complete the construction of branch micro-layers.

[0010] In one possible implementation, a tree-like topology is constructed using a single petal partition as the physical prototype. This topology has core nodes as the root, branch nodes as the middle layer, and terminal nodes as leaf nodes. Partition affiliation and hierarchy labels are embedded in the node feature vectors to complete the construction of the partition sub-layers, including: The core node, branch node, and terminal node of a single petal partition are used as modeling nodes to construct a tree-like topology with the core node as the root, the branch node as the middle layer, and the terminal node as the leaf node. The connection relationship between the nodes strictly corresponds to the electrical connection relationship within the partition. Embed partition affiliation label and level label in the feature vector of each node. The partition affiliation label is a unique code that identifies the petal partition to which the node belongs, and the level label is an identifier that represents the level of the node in the tree topology. The weights of the edges connecting nodes in the tree topology are set as the fault propagation coefficients. The fault propagation coefficients are calculated by normalization based on the electrical distance between the nodes and the core nodes, so as to quantify the longitudinal propagation trend of the fault under fault conditions. A gated graph attention network is used as the core computational unit to complete the construction of partitioned sub-layers.

[0011] In one possible implementation, the connection lines between core nodes are used as the physical prototype. Each core node is abstracted into a macro node, and the weight of the edges is zero in a non-faulty state, thus completing the construction of the global interconnection layer, including: Construct a macro-node graph consisting of core nodes of each petal partition, where the edges between macro nodes correspond to the connecting lines between core nodes to represent the global architecture of the petal-shaped network; in a non-faulty state, the weight of the edges is zero; A cross-regional propagation attenuation factor is introduced to correct the edge weights after activation; A graph convolutional network is used as the core computational unit to complete the construction of the global interconnection layer.

[0012] In one possible implementation, the parameters of the dynamic temporal attention mechanism of the temporal prediction model are determined based on the overall structural features and local structural features to complete the construction of the path tracing layer, including: Based on the topological hierarchical relationship in the overall structural features, the hierarchical weight allocation rules of the dynamic temporal attention mechanism are determined; Based on the edge weight rules in the local structural features, the edge weights are dynamically correlated with the temporal window size of the attention mechanism; Using the physical topology in the overall structural features as the skeleton, the nodes of each layer are mapped to the time axis according to the electrical connection relationship to construct a time-space fusion graph. In this graph, the node attributes inherit the level and partition labels in the structural features, and the edge attributes inherit the edge weight rules of the corresponding layer and are labeled with the temporal information of the topological association. A combination of dynamic temporal attention mechanism and temporal long short-term memory network model is used as the core computing unit to complete the construction of the path tracing layer; the parameters of the dynamic temporal attention mechanism are dynamically adjusted according to the node hierarchy relationship in the overall structural features.

[0013] In one possible implementation, the method further includes: Differentiated initial attention weights are assigned to nodes at different levels, with core nodes having higher initial weights than branch nodes, and branch nodes having higher initial weights than terminal nodes. Based on the node's hierarchical label, configure different numbers of attention heads for nodes at different levels; The node's hierarchical label is input as part of the positional encoding into the dynamic temporal attention mechanism.

[0014] In one possible implementation, based on the fault tracing results, a hierarchical dynamic reconfiguration strategy is generated and executed to isolate the fault and restore power supply, including: Based on the fault tracing results, the corresponding decision levels of the hierarchical decision model are dynamically activated; the hierarchical structure of the hierarchical decision model matches the hierarchical structure of the petal-shaped grid. Within the activated decision level, a dynamic reconstruction strategy is generated based on a preset multi-objective optimization function; The space frame is reconfigured according to the dynamic reconfiguration strategy.

[0015] In one possible implementation, the method further includes: A hierarchical decision model is constructed using a hierarchical reinforcement learning framework. The model contains multiple decision levels, each corresponding to a physical level of a petal-shaped grid. Configure an independent intelligent agent for each decision level, and define the action space and state space that match the physical scope and control authority of that level; A hierarchical reward function is designed for each level of intelligent agent. The hierarchical reward function performs differentiated weighting of load recovery amount, response time, operating cost and system risk based on the fault type and reconstruction goal of the level. Establish a collaborative communication mechanism between levels.

[0016] Compared to traditional technologies, this invention provides a method for fault tracing and dynamic reconstruction of petal-shaped grid structures based on graph neural networks. First, multi-dimensional electrical data of the four-level hierarchical structure of the petal-shaped grid structure is acquired and preprocessed. The preprocessed multi-dimensional electrical data is then input into a four-level graph neural network to obtain fault tracing results. Based on the fault tracing results, a hierarchical dynamic reconstruction strategy is generated and executed to isolate the fault and restore power supply. The four-level graph neural network includes branch micro-layers, partition sub-layers, a global interconnection layer, and a path tracing layer. This invention, by constructing a four-level graph neural network that matches the physical hierarchy of the petal-shaped grid structure, achieves refined fault tracing from local to global levels and drives hierarchical dynamic reconstruction strategies accordingly. This addresses the shortcomings of traditional technologies in terms of fault location accuracy, tracing depth, decision-making intelligence, and collaborative efficiency, ultimately achieving optimal fault isolation and power restoration. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the implementation of the petal-shaped grid fault tracing and dynamic reconstruction method based on graph neural networks provided in this embodiment of the invention. Detailed Implementation

[0018] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0019] Figure 1 This is a flowchart illustrating the implementation of the petal-shaped network structure fault tracing and dynamic reconstruction method based on graph neural networks provided in this embodiment of the invention. Figure 1 As shown, the method includes: S110: Obtain multi-dimensional electrical data of the four-level structure of the petal-shaped space frame and preprocess the data; S120, the preprocessed multi-dimensional electrical data is input into a four-level graph neural network to obtain the fault tracing results; S130, based on the fault tracing results, generates and executes a hierarchical dynamic reconfiguration strategy to isolate the fault and restore power supply; The four-level graph neural network includes branch micro-layers, partition sub-layers, global interconnection layers, and path tracing layers.

[0020] In this embodiment of the invention, to accurately acquire electrical data at the fourth level of the petal-shaped grid structure, the solution adopts a differentiated sensor deployment strategy of "high precision at the core, strong coverage in the zones, practicality in the branches, and low cost at the end": At the core node (substation hub), two sets of high-precision power quality monitoring devices are deployed at both the busbar outgoing end and the tie-line interface end, totaling 8-12 sets per core node, monitoring transient data such as voltage sags and current distortions; at the zone aggregation nodes (petal radiation starting point), one intelligent aggregation terminal (sampling rate 4.8kHz, supporting IEC61850) is deployed. The system is divided into petal-shaped sections to monitor switch status, branch load power, etc.; branch nodes rely on the existing FTU / DTU integrated monitoring function (sampling rate 2kHz, temperature accuracy ±0.5℃), with 1 monitoring point set for every 2-3 end nodes, for a total of 4-6 sets / petal-shaped sections; end nodes (load / DG access points) are equipped with low-cost sensors (sampling rate 1kHz, LoRa wireless communication), covering all end nodes, for a total of 15-20 sets / petal-shaped sections, to monitor current fluctuations, DG output, and other data, forming a full-level, differentiated data acquisition network.

[0021] Data preprocessing relies on a "local aggregation within a partition + cross-regional collaboration" architecture, with fine-grained processing performed by the edge computing nodes of the core nodes in the petal partition: The first step is data cleaning and completion, using the 3σ criterion to remove outliers caused by sensor malfunctions, and filling in data with a missing rate ≤5% through linear interpolation to ensure data integrity ≥98%; the second step is feature extraction and enhancement, using the db4 wavelet basis (5-level decomposition) to denoise the current and voltage data (improving the signal-to-noise ratio to over 40dB), and extracting the 3rd / 5th / 7th harmonic content, current distortion rate (accurate to 0.1%), voltage sag amplitude (accurate to 0.1kV), etc. through FFT transformation. The first step is to analyze fault characteristics and calculate the power fluctuation rate (accurate to 0.1% / s) and grid connection point voltage fluctuation frequency for the DG access point. The second step is data compression and filtering, which compresses the pre-processed feature data by 10 times, retaining only key features with a fault confidence of ≥50% and removing redundant data from normal operation. The third step is local backup and triggering mechanism, which retains the original data for 72 hours and the pre-processed data for 30 days. When the monitoring data exceeds the threshold (such as current distortion rate ≥5% or temperature ≥80℃), the "fault data priority upload" is automatically triggered, transmitting only fault event flags, confidence levels and core features within 10KB, reducing cross-regional communication by 70%.

[0022] The preprocessed feature data is input into a four-level GNN model hierarchically: the branch micro-layer uses a single radiating branch as a micro-graph (nodes include branches / ends / DG access points, edge weight = 1 / line impedance), and uses a temporal GCN (3×3 convolution kernel, LeakyReLU activation) to analyze data from 3 sampling periods, outputting the branch fault probability (threshold 70% for fault identification), preliminary location, and type (short circuit / overload / disturbance); the partition sub-layer constructs a petal-shaped partitioned tree subgraph (edge ​​weight = fault propagation coefficient, normal ≈ 0, fault increases with distance), and uses a gated graph attention network (4 attention heads, 64-dimensional hidden layers) to filter non-faulty branch interference, outputting a fault confidence heatmap and influence range ( The system analyzes the local / multi-branch / core faults at the end of the fault and their propagation speed; the global interconnection layer abstracts core nodes into macro nodes (non-fault edge weight = 0, fault time = real-time power of tie line / rated capacity), uses GCN (Sigmoid activation) to assess cross-regional risks, and outputs risk levels (low < 30% / medium 30%-70% / high > 70%) and a list of affected partitions; the path tracing layer splices the features of 30 time steps (100ms / step) from the first three layers to construct a "spatiotemporal fusion graph", combines temporal attention (2 attention heads) and Bi-LSTM (64-dimensional hidden layer), and outputs fault propagation paths, source nodes (error ≤ 100ms) and timestamps sorted by probability to complete accurate source tracing.

[0023] Based on the source tracing results, a hierarchical reconstruction strategy is generated through a three-tiered intelligent agent system: The branch intelligent agent (deployed at the edge) identifies local faults such as short circuits in the terminal lines within 100ms, generates a solution within 300ms, disconnects the faulty branch switch within 500ms, and transfers ≤50% of the rated load through healthy branches in the same zone, prioritizing DG output adjustment rather than disconnection. The zone intelligent agent, for multi-branch faults, disconnects the faulty node connection within 3 seconds, initiates full-power energy storage discharge, transfers ≤80% of the rated load according to the principle of "core load (hospitals, weight 5) > ordinary load (weight 1)," and requests cross-regional support within 2 seconds when resources are insufficient. The global coordination intelligent agent, for cross-regional faults, matches 10+ pre-generated emergency solutions within 5 seconds, instructs the disconnection of high-risk interconnection lines, allocates cross-regional load according to interconnection line capacity, schedules DG / energy storage to support the global core load (transportation hubs, weight 10), and completes global reconstruction within 10 seconds. Simultaneously, a hierarchical composite reward function optimizes decision-making, balancing recovery accuracy and cost control.

[0024] In some embodiments, the preprocessed multi-dimensional electrical data is input into a four-level graph neural network to obtain fault tracing results, including: inputting the multi-dimensional electrical data into a branch micro-layer, constructing a micro-map with a single radial branch as the unit, and using a time-series graph convolutional network for analysis to output branch fault probability, preliminary fault location, and preliminary fault type judgment; inputting the branch fault probability, preliminary fault location, preliminary fault type judgment, and corresponding multi-dimensional electrical data of the partition into the partition sub-layer to obtain a fault confidence heatmap, fault impact range, and fault propagation speed within the partition; inputting the fault confidence heatmap, fault impact range, and fault propagation speed of each partition sub-layer into the global interconnection layer to obtain the cross-region fault risk level and a list of potentially affected adjacent petal partitions; fusing the outputs of the branch micro-layer, partition sub-layer, and global interconnection layer according to the time series and inputting them into the path tracing layer to obtain a fault propagation path list sorted by probability, fault source nodes, and fault occurrence timestamps of each node, as the fault tracing results.

[0025] In this embodiment of the invention, the preprocessed multi-dimensional electrical data is first input into a branch micro-layer of a four-level graph neural network. This layer constructs a micro-graph using a single radial branch of a petal-shaped grid as an independent modeling unit. The nodes in the micro-graph precisely correspond to the branch nodes, all end load nodes, and distributed power supply access points within the branch. The edge weights between nodes are set to the reciprocal of the line impedance to quantify the electrical distance between nodes; the closer the electrical distance, the higher the edge weight, thus accurately characterizing the electrical correlation characteristics of the radial branch. The layer uses a time-series graph convolutional network as the core analysis unit, inputting preprocessed feature data such as current, temperature, and voltage for three consecutive sampling periods. The time-series convolutional kernel captures the dynamic trend of fault changes over time, and the graph convolutional kernel extracts the spatial correlation features between nodes. Finally, it outputs the branch fault probability, preliminary fault location results, and preliminary fault type judgment. The fault probability is determined by a 70% threshold to identify the faulty branch. The fault type initially distinguishes between short circuit, overload, distributed power supply disturbance, and normal state, laying a microscopic foundation for subsequent fault analysis.

[0026] The branch fault probability, preliminary fault location, and preliminary fault type judgment results output by the branch micro-layer will be fused with the multi-dimensional electrical data of the corresponding petal partition and jointly input into the partition sub-layer. This layer uses a single petal partition as the modeling scope, constructing a tree-like subgraph with core nodes as the root, branch nodes as the middle layer, and terminal nodes as leaf nodes. The edge weights between nodes are set as the fault propagation coefficient. Under normal operation, the propagation coefficient approaches 0, and under fault conditions, it increases with the propagation distance between the node and the core node, accurately reproducing the vertical propagation law of the fault within the partition. The layer uses a gated graph attention network for analysis, filtering out interference data from non-faulty branches through a gating mechanism, focusing on the radiating branch where the fault is located. At the same time, the partition topology morphology factor is introduced as an additional node feature to improve the model's adaptability. Finally, a fault confidence heatmap, fault influence range, and fault propagation speed are generated within the partition. The fault influence range clearly defines the specific type of fault, such as terminal local fault, single-branch fault, multi-branch fault, or core node fault. The fault propagation speed is divided into three levels: slow, medium, and fast, according to the node propagation rate, to achieve accurate location of regional faults.

[0027] The fault confidence heatmaps, fault impact ranges, and fault propagation speeds output by each sub-layer are uniformly aggregated and input into the global interconnection layer. This layer focuses on assessing the risk of fault propagation between sub-layers. During modeling, only the core nodes of each petal sub-layer are abstracted as macro nodes. The edges between macro nodes correspond to the tie lines between core nodes. In non-fault states, the edge weights are set to 0 to adapt to the weak isolation characteristics between sub-layers. When the fault confidence of a core node exceeds the 80% threshold, the edge weights are dynamically activated and calculated based on the ratio of the real-time transmission power of the tie line to its rated capacity. A weight ≥ 0.5 indicates the existence of cross-regional propagation risk. The layer uses a graph convolutional network for analysis. The input features cover the fault confidence, voltage frequency deviation, and tie line power data of each core node. At the same time, a cross-regional propagation attenuation factor is introduced to correct the edge weights. The final output includes the cross-regional fault risk level, a list of potentially affected adjacent petal sub-layers, the number of high-risk tie lines requiring emergency disconnection, and the upper limit of cross-regional load transfer. The cross-regional fault risk level is divided into three levels: low, medium, and high, based on probability, providing a decision-making basis for the collaborative handling of cross-regional faults.

[0028] To achieve accurate tracing of fault propagation paths, the path tracing layer first performs time-series fusion preprocessing on the output features of the branch micro-layer, partition sub-layer, and global interconnection layer. The fusion is performed within a time range of 0-3 seconds after the fault occurs, divided into 30 time nodes at 100ms intervals. The fault probability, confidence level, and risk level features output from the first three layers are concatenated according to these time nodes to construct a "time-space" fusion graph. Nodes in the fusion graph are selected as candidate fault nodes with a fault probability ≥50% at each level. Edges between nodes represent the temporal relationship of fault propagation; that is, a fault node in a previous time step points to a node affected by it in a subsequent time step. The edge weights are calculated using the fault propagation speed and electrical distance to determine the propagation probability, thus completely preserving the temporal evolution trajectory of the fault from the microscopic end to the macroscopic global level, providing comprehensive spatiotemporal feature support for subsequent path tracing.

[0029] After completing the temporal feature fusion, the path tracing layer introduces a combination model of temporal attention mechanism and bidirectional long short-term memory network for in-depth analysis. The temporal attention mechanism automatically identifies key nodes and time nodes in fault propagation through path attention weights, accurately capturing the temporal order and causal relationship of fault features at each node, such as the complete process of a distributed power supply access point fault gradually spreading to the terminal node, branch node, and then to the core node. The bidirectional long short-term memory network assists in extracting the bidirectional dependency relationship of temporal features, further improving the accuracy of path tracing. The model quantifies the probability of different fault propagation paths by analyzing the temporal correlation between each node and edge in the fused graph, and finally outputs a list of fault propagation paths sorted by probability, clearly presenting the propagation order of the fault from the initial node to each affected node.

[0030] After the path tracing layer analysis is completed, a complete fault tracing result will be output. The core results include a list of fault propagation paths sorted by probability, the fault source node, and the fault occurrence timestamps for each node. The fault propagation path list is arranged from highest to lowest probability, clearly defining the propagation order and corresponding probability of nodes on each path, providing a basis for reconstructing the fault evolution process. The fault source node is the initial fault node with the highest probability in the propagation path, achieving precise location of the fault root cause. The error of each node's fault occurrence timestamp is controlled within 100ms, accurately recording the time of fault occurrence at different nodes, and also including explanations of fault propagation termination conditions, such as key information like fault propagation stopping after cutting off the fault branch. These results together constitute a comprehensive and accurate fault tracing conclusion, providing core decision support for the subsequent generation of hierarchical dynamic reconfiguration strategies.

[0031] In some embodiments, the method further includes: using a single radial branch as a physical prototype, defining nodes as end load nodes, distributed power access points, and branch nodes within the branch, and setting the weight of the edges connecting the nodes to the reciprocal of the line impedance to quantify the electrical distance between nodes, thus completing the construction of the branch micro-layer; using a single petal partition as a physical prototype, constructing a tree topology with core nodes as roots, branch nodes as middle layers, and end nodes as leaf nodes, and embedding partition affiliation labels and hierarchy labels into the node feature vectors, thus completing the construction of the partition sub-layer; using the connecting lines between core nodes as a physical prototype, abstracting each core node into a macro node, and setting the edge weight to zero in a non-fault state, thus completing the construction of the global interconnection layer; extracting the overall structural features and local structural features of the branch micro-layer, partition sub-layer, and global interconnection layer, and determining the parameters of the dynamic temporal attention mechanism of the temporal prediction model based on the overall structural features and local structural features, thus completing the construction of the path tracing layer.

[0032] In this embodiment of the invention, the branch micro-layer is constructed using a single radial branch of a petal-shaped grid as the physical prototype, precisely matching the microscopic electrical connection characteristics. First, the node definitions are clearly defined, including all end load nodes, distributed power supply access points, and branch nodes within the branch in the modeling scope, ensuring coverage of all critical electrical nodes within the branch. Next, edge weights are configured, setting the weight of the edges connecting each node to the reciprocal of the line impedance. This design achieves a quantitative representation of the electrical distance between nodes. The smaller the line impedance, the tighter the electrical connection between nodes, and the larger the edge weight, thus accurately depicting the electrical coupling relationship between nodes within the radial branch. After completing the node definitions and edge weight settings, a branch micro-layer adapted to a single radial branch is constructed, laying the topological foundation for subsequent capture of microscopic fault characteristics.

[0033] The construction of the partitioned sublayers uses a single petal partition as the physical prototype, focusing on adapting to the vertical propagation pattern of faults within the partition. The core of the modeling is to construct a tree-like topology, clearly defining the core node as the root node, branch nodes as mid-level nodes, and terminal nodes (including terminal load nodes and distributed power supply access points) as leaf nodes, strictly corresponding to the physical structure of the petal partition's "core aggregation, partition radiation" principle. Simultaneously, two types of key labels are embedded in the feature vectors of each node: a partition affiliation label to uniquely identify the petal partition to which the node belongs, avoiding cross-partition feature confusion; and a hierarchy label to clarify the node's hierarchical position in the tree topology, strengthening the model's recognition of hierarchical propagation characteristics. Through tree topology construction and label embedding, the partitioned sublayers are constructed, achieving accurate adaptation to the fault propagation pattern within the partition.

[0034] The global interconnection layer is constructed using the interconnecting lines between the core nodes of the petal-shaped network as its physical prototype, focusing on the assessment and adaptation of fault propagation risks between different zones. A macro-node abstraction strategy is employed during modeling, abstracting the core nodes of each petal zone as macro nodes of the global interconnection layer, eliminating all non-core nodes, and retaining only the interconnecting lines between macro nodes as modeling edges, accurately reproducing the "weak isolation between zones" global architecture of the petal-shaped network. Edge weights are designed with state-differentiated configuration: in non-fault states, edge weights are set to zero, weakening unnecessary electrical connections between zones and adapting to the lateral isolation characteristics of different zones; in fault states, weights are dynamically activated based on the fault confidence of the core nodes, ensuring the accuracy of cross-zone risk assessment. Through macro-node abstraction and state-based edge weight configuration, the global interconnection layer is constructed, providing macro-topological support for cross-zone fault risk analysis.

[0035] The construction of the path tracing layer is based on branch micro-layers, partition sub-layers, and a global interconnection layer. The core first step is feature extraction. First, the overall structural features of each layer are extracted, including the single-branch radial topology of the branch micro-layers, the tree-like hierarchical topology of the partition sub-layers, and the macro-node connection topology of the global interconnection layer. Then, the local structural features of each layer are extracted, covering the "edge weight = reciprocal of line impedance" rule for the branch micro-layers, the label embedding rules for the partition sub-layers, and the stateful configuration rules for the edge weights of the global interconnection layer. This comprehensive extraction of overall and local structural features provides the core basis for the parameter configuration of the subsequent time-series prediction model, achieving accurate matching between structural features and model parameters.

[0036] After structural feature extraction, the core of path tracing layer construction is determining the parameters of the dynamic temporal attention mechanism for the time-series prediction model. Parameter determination requires close integration with the extracted overall and local structural features: based on the topological hierarchy relationships in the overall structural features, attention weight allocation rules for nodes at different levels are set; combined with the edge weight rules in the local structural features, the temporal window size of the attention mechanism is correlated to ensure the model can accurately capture the fault propagation characteristics under different topological structures. Once the parameters of the dynamic temporal attention mechanism are configured, the path tracing layer is constructed.

[0037] In some embodiments, a single radial branch is used as the physical prototype. Nodes are defined as end load nodes, distributed power supply access points, and branch nodes within the branch. The weights of the edges connecting the nodes are set to the reciprocal of the line impedance to quantify the electrical distance between nodes, thus completing the construction of the branch micro-layer. This includes: using the end load nodes, distributed power supply access points, and branch nodes within the branch as modeling nodes, and configuring corresponding monitoring dimensions for different types of nodes. The monitoring dimensions for end load nodes and distributed power supply access points include current distortion, temperature change, and voltage fluctuation. The monitoring dimensions for branch nodes include the effective value of current, switch on / off status, and load fluctuation coefficient. Based on the line impedance values ​​between the modeling nodes, the weights of the edges connecting each node are set to the reciprocal of the line impedance to achieve a negative correlation between edge weights and the electrical distance between nodes. A time-series graph convolutional network is used as the core computational unit, and a time-space two-dimensional convolutional kernel is configured to complete the construction of the branch micro-layer.

[0038] In this embodiment of the invention, the construction of the branch micro-layer is based on a single radial branch in a petal-shaped network structure as the physical prototype. The core objective is to accurately characterize the microscopic electrical correlation characteristics within the branch, providing topological support for the early identification of minor faults at the end. The construction process revolves around three core aspects: "precise node definition, scientific configuration of edge weights, and adaptation of computing units." It strictly follows the actual electrical connection relationships of the branch to ensure that the layer modeling not only fits the physical scenario but also meets the technical requirements for fault feature extraction, ultimately achieving refined perception and analysis of the operating status of a single radial branch.

[0039] The first step in the construction is to define the scope of the modeling nodes and configure dedicated monitoring dimensions. All end-load nodes, distributed power supply access points, and branch nodes within a single radial branch are included in the modeling nodes, achieving full coverage of the branch's key electrical nodes. Simultaneously, differentiated monitoring dimensions are designed based on the functional characteristics of different nodes: end-load nodes and distributed power supply access points focus on the electrical anomalies in the early stages of faults, with monitoring dimensions set as current distortion, temperature changes, and voltage fluctuations, accurately capturing early signals of end-of-line faults and power disturbances; branch nodes, as branch connection hubs, have monitoring dimensions set as current RMS value, switch on / off status, and load fluctuation coefficient, focusing on understanding the overall operating status of the branch and the working conditions of key control nodes.

[0040] After the nodes are defined, the electrical distance between nodes is accurately quantified by scientifically configuring edge weights. Based on the actual line impedance values ​​between the modeled nodes, the weights of the edges connecting each node are strictly set to the reciprocal of the line impedance. This design achieves a negative correlation between edge weights and the electrical distance between nodes. The smaller the line impedance, the tighter the electrical connection and the closer the electrical distance between nodes, and the larger the corresponding edge weight; conversely, the larger the line impedance, the smaller the edge weight. This allows for the accurate numerical reconstruction of the electrical coupling relationships between nodes within a branch, providing a foundation for subsequent spatial correlation analysis of fault characteristics.

[0041] To adapt to the temporal evolution and spatial correlation characteristics of branch faults, a temporal graph convolutional network is selected as the core computational unit for the branch micro-layer. This network combines the capabilities of temporal feature capture and spatial correlation analysis, and can simultaneously process the temporal dynamic changes of branch electrical data and the spatial coupling relationships between nodes. A temporal-spatial two-dimensional convolutional kernel is configured as the core computational component. The temporal dimension convolutional kernel is used to capture the dynamic evolution of faults over time (such as stepwise temperature increases and sudden current surges), while the spatial dimension convolutional kernel is used to extract the fault feature propagation correlations between different nodes. Comprehensive extraction of fault features is achieved through two-dimensional convolution operations.

[0042] After completing node definition, monitoring dimension configuration, edge weight setting, and computation unit configuration, the complete construction of the branch micro-layer is achieved. This layer can input preprocessed multi-dimensional electrical data of the branches into a time-series graph convolutional network. Through the collaborative operation of two-dimensional convolutional kernels, it accurately outputs the branch fault probability, preliminary fault location, and preliminary fault type judgment results. The fault probability is determined with a 70% threshold to identify faulty branches, providing accurate initial information on micro-faults for further analysis of subsequent sub-layers, enabling early identification and preliminary diagnosis of faults in single radial branches.

[0043] In some embodiments, using a single petal partition as a physical prototype, a tree-like topology is constructed with core nodes as roots, branch nodes as middle layers, and terminal nodes as leaf nodes. Partition affiliation labels and hierarchy labels are embedded in the node feature vectors to complete the construction of the partition sub-layer. This includes: using the core nodes, branch nodes, and terminal nodes of a single petal partition as modeling nodes to construct a tree-like topology with core nodes as roots, branch nodes as middle layers, and terminal nodes as leaf nodes, where the connection relationships between nodes strictly correspond to the electrical connection relationships within the partition; embedding partition affiliation labels and hierarchy labels in the feature vectors of each node, where the partition affiliation label is a unique identifier of the petal partition to which the node belongs, and the hierarchy label is an identifier representing the level of the node in the tree-like topology; setting the weights of the edges connecting nodes in the tree-like topology as fault propagation coefficients, which are calculated by normalization based on the electrical distance between the node and the core node to quantify the vertical propagation trend of the fault under fault conditions; and using a gated graph attention network as the core computational unit to complete the construction of the partition sub-layer.

[0044] In this embodiment of the invention, the construction of the partitioned sublayer is based on a single petal partition as a physical prototype. The core objective is to accurately adapt to the structural characteristics of the petal partition, which is characterized by "core aggregation and partition radiation," and to capture the longitudinal propagation pattern of faults from the end to the core within the partition. This provides topological support for the range locking and in-depth analysis of regional faults. The construction process revolves around four key steps: "topology restoration, feature label embedding, edge weight quantization, and computational unit adaptation." It strictly follows the actual electrical connection relationships within the partition to ensure that the layer not only fits the physical scene but also meets the technical requirements for accurate fault feature extraction.

[0045] The first step in the construction is to define the scope of the modeling nodes and build a tree-like topology. All core nodes, branch nodes, and terminal nodes (including terminal load nodes and distributed power supply access points) within a single petal partition are included in the modeling nodes, achieving full coverage of key electrical nodes within the partition. Based on this, a tree-like topology is constructed with core nodes as root nodes, branch nodes as middle-level nodes, and terminal nodes as leaf nodes. The connections between nodes strictly correspond to the actual electrical connection paths within the partition, accurately recreating the radial structure of "core → branch → terminal," laying the topological foundation for subsequent analysis of the vertical propagation patterns of faults.

[0046] To enhance the model's ability to identify partition attributes and hierarchical characteristics, two types of key labels are embedded in the feature vectors of each node. The partition affiliation label uniquely identifies the petal partition to which a node belongs, effectively preventing confusion between node features from different petal partitions and ensuring the model focuses solely on fault analysis within the current partition. The hierarchical label identifies the node's level within the tree topology, clarifying the hierarchical differences between core nodes, branch nodes, and terminal nodes, helping the model accurately capture the propagation characteristics of faults across different levels. The embedding of these two types of labels makes node features more distinctive, significantly improving the model's accuracy in locating faults within partitions.

[0047] After the tree-like topology is constructed, the vertical propagation trend of faults is accurately quantified by configuring edge weights. The weights of the edges connecting each node are set as the fault propagation coefficient, which is calculated based on the electrical distance between the node and the core node. The closer a node is to the core node, the smaller the electrical distance, and the larger the fault propagation coefficient, meaning that the probability of a fault propagating from that node to the core is higher; conversely, the farther away a node is, the smaller the propagation coefficient. This accurately matches the propagation pattern of faults within a partition, which is "gradually spreading from the end to the core, with the degree of impact increasing with distance," providing a numerical basis for the model to quantify the risk of fault propagation.

[0048] To meet the needs of fault analysis within partitions, a gated graph attention network was selected as the core computational unit for the partition sublayer. This network effectively filters out interference data from non-faulty branches through a gating mechanism, focusing on the characteristic propagation of the radiating branch where the fault is located. Simultaneously, the attention mechanism assigns higher weights to nodes along the fault propagation path, further improving the accuracy of fault range localization. After completing the tree-like topology construction, label embedding, edge weight setting, and computational unit configuration, the partition sublayer is officially formed. It can receive the output results of branch micro-layers and multi-dimensional electrical data from the partition, generating key analytical results such as fault confidence heatmaps, fault impact ranges, and fault propagation speeds within the partition.

[0049] In some embodiments, using the connection lines between core nodes as physical prototypes, each core node is abstracted into a macro node, and the weight of the edges is zero in a non-faulty state, to complete the construction of the global interconnection layer. This includes: constructing a macro node graph composed of core nodes of each petal partition, wherein the edges between macro nodes correspond to the connection lines between core nodes to represent the global architecture of the petal-shaped network; in a non-faulty state, the weight of the edges is zero; introducing a cross-regional propagation attenuation factor to correct the weight of the activated edges; and using a graph convolutional network as the core computing unit to complete the construction of the global interconnection layer.

[0050] In this embodiment of the invention, the global interconnection layer is constructed using the connecting lines between the core nodes of the petal-shaped network as a physical prototype. The core objective is to accurately adapt to the "weakly interconnected and isolated" global architecture across different zones, focusing on assessing the risk of cross-zone fault propagation. The first step in the construction is to perform macro-node abstract modeling. The core nodes of each petal zone are extracted individually as macro nodes for the global interconnection layer, all non-core nodes are removed, and only the connecting lines between macro nodes are retained as modeling edges. Through this design, a macro node graph composed of the core nodes of each petal zone is constructed, with each edge corresponding one-to-one with the connecting lines between the core nodes. This accurately represents the "zone-independent, core-interconnected" global topology of the petal-shaped network, providing a simple and practical topology foundation for cross-zone fault analysis.

[0051] To adapt to the characteristics of fault propagation between sections, the global interconnection layer adopts a differentiated edge weight configuration strategy. In non-fault states, the weights of the connecting edges between macro nodes are set to zero. This design precisely matches the "weak isolation" operation characteristics of the petal-shaped network structure sections, weakening unnecessary electrical connections between sections and avoiding risk misjudgments caused by cross-section data interference during normal operation, ensuring the model focuses on monitoring the operational status within its own section. When the fault confidence of a core node exceeds a preset threshold, the edge weights of the corresponding tie lines are dynamically activated. The weight value is calculated based on the ratio of the tie line's real-time transmission power to its rated capacity, achieving accurate quantification of the cross-section correlation strength under fault conditions.

[0052] To more accurately assess the actual risk of fault propagation across regions, the global interconnect layer introduces a cross-region propagation attenuation factor to correct the activated edge weights. Since faults exhibit energy attenuation when propagating through tie lines between regions, the impact of a single region fault decreases with increasing propagation distance. The attenuation factor precisely matches this physical law. This factor has a default value of 0.3 and decreases linearly with increasing tie line impedance. By combining its calculation with the activated edge weights, it effectively corrects the cross-region fault risk assessment results, avoiding overestimation of cross-region risk due to neglecting energy attenuation and improving the accuracy of cross-region fault analysis.

[0053] To efficiently process the topological features and fault data of macro-node graphs, a graph convolutional network is selected as the core computational unit for the global interconnection layer. This network can accurately extract the spatial correlation features between macro nodes and, combined with corrected edge weight data, quickly analyze the potential impact of the fault status of each core node on adjacent partitions. After completing macro node modeling, edge weight configuration, introduction of attenuation factors, and deployment of computational units, the global interconnection layer is formally established. This layer can receive fault information output from each partition sublayer and ultimately output the cross-regional fault risk level, a list of potentially affected adjacent petal partitions, and high-risk connection line numbers, providing core decision-making basis for collaborative prevention and control of cross-regional faults.

[0054] In some embodiments, the parameters of the dynamic temporal attention mechanism of the temporal prediction model are determined based on the overall structural features and local structural features to complete the construction of the path tracing layer. This includes: determining the hierarchical weight allocation rules of the dynamic temporal attention mechanism based on the topological hierarchy in the overall structural features; dynamically associating the edge weights with the temporal window size of the attention mechanism according to the edge weight rules in the local structural features; using the physical topology in the overall structural features as the skeleton, mapping the nodes of each layer to the time axis according to their electrical connections to construct a time-space fusion graph, wherein the node attributes inherit the hierarchy and partition labels in the structural features, and the edge attributes inherit the edge weight rules of the corresponding layer and label the temporal information of the topological association; using a combination of the dynamic temporal attention mechanism and the temporal long short-term memory network model as the core computing unit to complete the construction of the path tracing layer; wherein the parameters of the dynamic temporal attention mechanism are dynamically adjusted according to the node hierarchy in the overall structural features.

[0055] In this embodiment of the invention, the core of constructing the path tracing layer is to combine the overall and local structural features of the branch micro-layer, the partition sub-layer, and the global interconnection layer to accurately configure the dynamic temporal attention mechanism parameters of the temporal prediction model, ultimately achieving full temporal tracing of the fault propagation path. The overall structural features cover the topological morphology of the first three layers, including the single-branch radial topology of the branch micro-layer, the tree-like hierarchical topology of the partition sub-layer, and the macro-node connection topology of the global interconnection layer; the local structural features focus on the edge weight configuration rules and node attribute design of each layer. These features together provide the core basis for determining the attention mechanism parameters in accordance with the physical characteristics of the network structure and the fault propagation law.

[0056] The hierarchical weight allocation rules of the dynamic temporal attention mechanism are strictly formulated based on the topological hierarchy of the overall structural features. For the four-level hierarchical structure of the petal-shaped network ("core-branch-terminal"), differentiated attention weights are assigned to nodes at different levels: core nodes, as the network hub, undertake global coordination functions and are assigned the highest attention weight to enhance the capture of global temporal features; branch nodes, as intermediate transmission links, are assigned medium weights to balance local and global features; and terminal nodes focus on the initial signals of micro-faults and are assigned appropriate basic weights to ensure that small fault features are not ignored. This hierarchical weight design ensures that the model can accurately match the pattern of faults "propagating step by step from the terminal to the core," improving the targeting of temporal feature extraction.

[0057] To accommodate the varying electrical correlation strengths between different nodes, a dynamic correlation is established between edge weights and the timing window size of the attention mechanism, based on the edge weight rules in the local structural features. In the branch micro-layer, the edge weight is the reciprocal of the line impedance; a larger weight indicates a closer electrical distance between nodes and faster fault propagation, corresponding to a smaller timing window to accurately capture rapidly changing fault characteristics. In the partition sub-layer, the edge weight is the fault propagation coefficient; a larger coefficient means a higher probability of longitudinal fault propagation, balancing propagation speed and feature integrity with a medium timing window. In the global interconnection layer, the edge weight is the dynamically activated cross-regional correlation value. After correction with the cross-regional propagation attenuation factor, a larger timing window is configured for high-weighted interconnects to avoid missing cross-regional fault features.

[0058] Using the physical topology of the overall structural features as the skeleton, the nodes of the branch micro-layers, partition sub-layers, and global interconnection layers are mapped to the time axis according to their actual electrical connections, constructing a complete time-space fusion diagram. Node attributes fully inherit the structural features of the first three layers, retaining hierarchical labels (core / branch / terminal) and partition affiliation labels, ensuring that the model can accurately identify the functional positioning and partition attributes of nodes; edge attributes not only inherit the edge weight rules of the corresponding layer, but also additionally annotate the temporal information of topological associations, including the temporal order of fault propagation between nodes and the dynamic change trend of edge weights over time. Through spatiotemporal feature fusion, the evolution trajectory of faults from the micro-terminal to the macro-global is fully restored.

[0059] The path tracing layer employs a combination of a dynamic temporal attention mechanism and a temporal long short-term memory (LSTM) network as its core computational units. The dynamic temporal attention mechanism, through the hierarchical weight rules and temporal window association strategy defined earlier, accurately focuses on key nodes and time segments in fault propagation. The temporal LSTM, with its powerful temporal feature capture capabilities, deeply mines the dependencies between faults at nodes in the fusion graph, improving the accuracy of path tracing. The parameters of the dynamic temporal attention mechanism are dynamically adjusted in real time based on the node hierarchy relationships in the overall structural features. When a fault propagates to nodes at different levels, the corresponding attention weights and temporal window sizes are automatically adapted to ensure the model always conforms to the dynamic characteristics of fault propagation. After completing the above parameter configuration and computational unit deployment, the path tracing layer is officially constructed and can output a list of fault propagation paths sorted by probability, the fault source node, and the timestamps of fault occurrence at each node.

[0060] In some embodiments, the method further includes: assigning differentiated initial attention weights to nodes at different levels, wherein the initial weight of core nodes is higher than that of branch nodes, and the initial weight of branch nodes is higher than that of terminal nodes; configuring different numbers of attention heads for nodes at different levels according to the node's hierarchical label; and inputting the node's hierarchical label as part of the positional encoding into the dynamic temporal attention mechanism.

[0061] In this embodiment of the invention, the method, taking into account the hierarchical characteristics of the petal-shaped network structure, assigns differentiated initial attention weights to nodes at different levels, constructing a weight gradient that conforms to the fault propagation law. Core nodes, as the control and communication hub of the network structure, have a wide fault impact range and strong correlation, and are assigned the highest initial attention weights to enhance the model's capture of global core temporal features. Branch nodes, as intermediate transmission links in fault propagation, play a connecting role between local and global features, and their initial weights are set to a medium level to balance the attention given to local fault details and global propagation trends. Terminal nodes focus on the initial signals of micro-faults; although their impact range is small, they are crucial for early fault identification, and are assigned basic initial weights to ensure that small fault features are not ignored. Through weight differentiation, accurate adaptation of temporal features of nodes at different levels is achieved.

[0062] By combining the hierarchical labels of nodes, different numbers of attention heads are configured for nodes at different levels to optimize the efficiency and accuracy of temporal feature extraction. Core nodes need to process multi-dimensional temporal information such as cross-partition communication and global resource scheduling simultaneously, and are configured with the most attention heads to achieve comprehensive capture of complex global temporal correlations. Branch nodes are mainly responsible for the temporal feature analysis of fault propagation within their own partitions, and are configured with an appropriate number of attention heads to control computational overhead while ensuring the accuracy of local feature extraction. Terminal nodes have relatively simple temporal features, focusing on local signals such as current distortion and temperature abrupt changes, and are configured with the fewest attention heads to achieve lightweight and efficient processing. Through the hierarchical configuration of the number of attention heads, the model can achieve a balance between performance and efficiency in the temporal analysis of nodes at different levels.

[0063] By incorporating node hierarchical labels as part of the location encoding and synchronously inputting them into the dynamic temporal attention mechanism, the model's ability to recognize the temporal features of node hierarchical attributes is enhanced. By converting hierarchical labels into standardized numerical codes (e.g., dedicated codes for core nodes, branch nodes, and terminal nodes), and fusing them with the temporal information from traditional location encoding, the model can simultaneously perceive the hierarchical location of nodes when processing temporal data, accurately identifying the temporal differences in fault characteristics across different node levels. Examples include the rapid initial fluctuations of terminal node faults and the delayed global impact of core node faults. This design allows the dynamic temporal attention mechanism to consider both temporal order and hierarchical attribute differences when allocating attention resources, further improving the accuracy of fault propagation path tracing.

[0064] In some embodiments, a hierarchical dynamic reconfiguration strategy is generated and executed based on the fault tracing results to isolate the fault and restore power supply. This includes: dynamically activating the corresponding decision level of the hierarchical decision model based on the fault tracing results; wherein the hierarchical structure of the hierarchical decision model matches the hierarchical structure of the petal-shaped grid; generating a dynamic reconfiguration strategy based on a preset multi-objective optimization function within the activated decision level; and performing a reconfiguration operation on the grid according to the dynamic reconfiguration strategy.

[0065] In this embodiment of the invention, based on the fault tracing results output by the four-level graph neural network, the first step is to dynamically activate the corresponding decision level of the hierarchical decision model to ensure that the reconstruction strategy accurately matches the actual impact range of the fault. Core information in the fault tracing results, such as the fault propagation path, source node, and impact range, is directly used as the basis for determining the hierarchical activation level: if it is a branch-level fault affecting only a single radiating branch, only the lowest-level branch decision level is activated; if the fault spreads to a single petal partition, the partition decision level is activated; if there is a cross-partition cascading risk, the highest-level global coordination decision level is activated. This dynamic activation mechanism avoids a "one-size-fits-all" global operation, achieving a precise decision-making logic of "local fault handling and global fault collaborative response."

[0066] The hierarchical structure of the layered decision-making model perfectly matches the four-level physical structure of the petal-shaped grid structure: "core-zone-branch-terminal," ensuring that the decision-making logic aligns with the electrical connection characteristics and operational rules of the grid structure. The branch decision-making level corresponds to the radial branches of the grid structure, focusing on the rapid handling of minor faults at the terminal; the zone decision-making level corresponds to a single petal zone, responsible for the overall handling of multiple branch faults within that area; the global coordination decision-making level corresponds to core nodes and cross-zone connecting lines, undertaking the coordinated scheduling of cross-zone faults. Information exchange between levels is achieved through standardized interfaces, but decision-making authority is strictly defined. Lower levels prioritize autonomous processing, only requesting support from higher levels when resources are insufficient or faults propagate, ensuring the efficiency and independence of decision-making.

[0067] Within the activated decision-making level, an optimal dynamic reconfiguration strategy is generated based on a preset multi-objective optimization function, achieving a comprehensive balance between load restoration, response speed, risk control, and cost optimization. The optimization function assigns weights differently for different decision-making levels: the branch level emphasizes response speed and operating costs, prioritizing rapid fault isolation and minimizing switching actions; the zone level emphasizes core load restoration and network loss control, while also considering the efficient utilization of local resources such as energy storage and distributed power sources; the global level focuses on cross-regional resource balancing and fault propagation prevention, ensuring global power supply reliability. Through the quantitative calculation of this function, the execution scheme with the best overall benefits is selected from multiple candidate reconfiguration schemes, avoiding decision-making biases caused by a single objective.

[0068] After determining the dynamic reconfiguration strategy, the network reconfiguration operation is carried out based on a three-level execution architecture of "edge-region-global": branch-level strategies are executed rapidly locally by edge computing nodes, ensuring that fault branch isolation and load transfer are completed within 0.5 seconds; partition-level strategies are executed by partition control units, and the operation results are simultaneously uploaded to the regional control center for record-keeping; global-level strategies are distributed from the regional control center to each partition for collaborative execution, and the cross-regional resource scheduling status is monitored in real time. During the operation, real-time data such as network voltage, current, and power are continuously collected to verify the fault isolation effect and load recovery. If abnormalities such as voltage exceeding limits or line overload occur, the reconfiguration strategy is dynamically adjusted based on a multi-objective optimization function until the network returns to stable operation, realizing closed-loop management of "decision-execution-verification-optimization".

[0069] In some embodiments, the method further includes: constructing a hierarchical decision model using a hierarchical reinforcement learning framework, wherein the model contains multiple decision levels, each corresponding to a physical level of a petal-shaped network; configuring an independent agent for each decision level, and defining an action space and state space that match the physical scope and control authority of that level; designing a hierarchical reward function for the agent at each level, wherein the hierarchical reward function performs differentiated weighting on load recovery amount, response time, operating cost and system risk according to the fault type and reconstruction goal handled by that level; and establishing a collaborative communication mechanism between levels.

[0070] In this embodiment of the invention, the method constructs a hierarchical decision-making model using a hierarchical reinforcement learning framework. The core principle is to ensure a precise correspondence between the model's decision-making levels and the physical levels of the petal-shaped network structure. The framework divides multiple decision-making levels according to a "branch-partition-global" model, corresponding to the radial branches of the network structure, individual petal partitions, and the core nodes and interconnections of the entire network structure, respectively. The decision-making scope of each level is strictly limited to the electrical boundaries of the corresponding physical level. This hierarchical adaptation design ensures that the model conforms to the network structure's characteristics of "local independent operation and global coordinated response," providing targeted decision support for faults of different ranges and avoiding efficiency losses and accuracy deviations caused by cross-level decision-making.

[0071] Each decision-making level is configured with an independent intelligent agent, whose action space and state space are strictly matched to the physical scope and control authority of its level. The state space of a branch intelligent agent covers local data such as current, temperature, and switch status of a single branch, while its action space is limited to small-scale operations such as opening and closing faulty branch switches and local load transfer. The state space of a zone intelligent agent extends to the operating data of all nodes within the zone, as well as the status of energy storage and distributed power sources, while its action space includes cross-branch load scheduling and energy storage charging and discharging control within the zone. The state space of the global intelligent agent covers the operating parameters of each core node and tie-line power, while its action space focuses on global operations such as cross-regional resource allocation and high-risk tie-line management. The independent intelligent agent design ensures the professionalism and efficiency of decision-making at each level.

[0072] Differentiated hierarchical reward functions are configured for each level of intelligent agents. Based on the fault type and reconstruction goal responsible for each level, different weights are assigned to load recovery, response time, operating cost, and system risk: branch-level reward functions emphasize response speed and operating cost, prioritizing rapid fault isolation; zone-level functions emphasize core load recovery and network loss control; and global-level functions focus on cross-zone risk prevention and resource balancing. Simultaneously, a collaborative communication mechanism between levels is established. When a lower-level intelligent agent lacks sufficient processing capacity, it sends a support request to the upper level. The upper-level intelligent agent coordinates resources through a command interface, achieving an efficient operation mode of "local autonomous decision-making + global on-demand collaboration."

[0073] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0074] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A petal type net rack fault tracing and dynamic reconstruction method based on a graph neural network, characterized in that, The method comprises the following steps: acquiring multi-dimensional electrical data of a four-level hierarchical structure of a petal-type grid, and preprocessing the data; inputting the preprocessed multi-dimensional electrical data into a four-level graph neural network to obtain a fault tracing result; generating a hierarchical dynamic reconstruction strategy according to the fault tracing result and executing the strategy to isolate the fault and restore power supply; wherein the four-level graph neural network comprises a branch micrograph layer, a partition subgraph layer, a global interconnection layer and a path tracking layer.

2. The petal type net rack fault tracing and dynamic reconstruction method based on a graph neural network according to claim 1, characterized in that, inputting the preprocessed multi-dimensional electrical data into a four-level graph neural network to obtain a fault tracing result, comprising: inputting the multi-dimensional electrical data into the branch micrograph layer to construct a micrograph in units of a single radial branch, and using a time series graph convolution network for analysis to output a branch fault probability, a fault preliminary positioning and a fault type preliminary judgment; inputting the branch fault probability, the fault preliminary positioning and the fault type preliminary judgment and the multi-dimensional electrical data of the corresponding partition into the partition subgraph layer to obtain a partition internal fault confidence heat map, a fault influence range and a fault diffusion speed; inputting the fault confidence heat map, the fault influence range and the fault diffusion speed of each partition subgraph layer into the global interconnection layer to obtain a cross-zone fault risk level and a list of adjacent petal partitions that may be affected; fusing the outputs of the branch micrograph layer, the partition subgraph layer and the global interconnection layer in time sequence and inputting them into the path tracking layer to obtain a list of fault propagation paths sorted by probability, a fault source node and a fault occurrence time stamp of each node as a fault tracing result.

3. The petal type net rack fault tracing and dynamic reconstruction method based on a graph neural network according to claim 2, characterized in that, The method further comprises: taking a single radial branch as a physical prototype, defining nodes as end load nodes, distributed power access points and branch nodes within the branch, and setting the weight of the edge connecting the nodes as the inverse of the line impedance to quantify the electrical distance between the nodes, thereby completing the construction of the branch micrograph layer; taking a single petal partition as a physical prototype, constructing a tree topology with core nodes as roots, branch nodes as middle layers and end nodes as leaf nodes, and embedding partition ownership labels and hierarchical labels in node feature vectors, thereby completing the construction of the partition subgraph layer; taking the tie line between core nodes as a physical prototype, abstracting each core node into a macro node, and setting the weight of the edge to zero in a non-fault state, thereby completing the construction of the global interconnection layer; extracting the overall structural features and local structural features of the branch micrograph layer, the partition subgraph layer and the global interconnection layer, and determining the parameters of the dynamic time sequence attention mechanism of the time sequence prediction model according to the overall structural features and the local structural features, thereby completing the construction of the path tracking layer.

4. The petal type net rack fault tracing and dynamic reconstruction method based on a graph neural network according to claim 3, characterized in that, taking a single radial branch as a physical prototype, defining nodes as end load nodes, distributed power access points and branch nodes within the branch, and setting the weight of the edge connecting the nodes as the inverse of the line impedance to quantify the electrical distance between the nodes, thereby completing the construction of the branch micrograph layer, comprising: The end load node, the distributed power access point and the branch node in the branch are taken as modeling nodes, and corresponding monitoring dimensions are configured for different types of nodes, wherein the monitoring dimensions of the end load node and the distributed power access point include current distortion, temperature change and voltage fluctuation, and the monitoring dimension of the branch node includes current effective value, switch on-off state and load fluctuation coefficient; According to the line impedance value between the modeling nodes, the weight of the edge connecting the nodes is set as the inverse of the line impedance, so as to realize the negative correlation quantitative matching of the edge weight and the electrical distance between the nodes; A time sequence diagram convolution network is adopted as a core calculation unit, and a time sequence-space two-dimensional convolution kernel is configured to complete the construction of the branch micrograph layer.

5. The petal type net rack fault tracing and dynamic reconstruction method based on a graph neural network according to claim 3, characterized in that, Taking a single petal partition as a physical prototype, a tree topology with core nodes as roots, branch nodes as middle layers and end nodes as leaf nodes is constructed, and a partition attribution label and a hierarchical label are embedded in a node feature vector to complete the construction of a partition subgraph layer, including: The core node, the branch node and the end node of the single petal partition are taken as modeling nodes, and a tree topology structure with the core node as the root, the branch node as the middle layer and the end node as the leaf node is constructed, wherein the connection relationship between the nodes strictly corresponds to the electrical connection relationship in the partition; A partition attribution label and a hierarchical label are embedded in the feature vector of each node, the partition attribution label is a code for uniquely identifying the petal partition to which the node belongs, and the hierarchical label is an identification representing the level of the node in the tree topology; The weight of the edge connecting the nodes in the tree topology is set as a fault diffusion coefficient, the fault diffusion coefficient is normalized calculated according to the electrical distance between the nodes and the core node, so as to quantize the longitudinal propagation trend of the fault in the fault state; A gated graph attention network is adopted as a core calculation unit to complete the construction of the partition subgraph layer.

6. The petal type net rack fault tracing and dynamic reconstruction method based on a graph neural network according to claim 3, characterized in that, Taking the tie line between the core nodes as a physical prototype, each core node is abstracted as a macro node, and the weight of the edge is zero in the non-fault state, to complete the construction of the global interconnection layer, including: A macro node graph composed of the core nodes of each petal partition is constructed, wherein the edges between the macro nodes correspond to the tie lines between the core nodes to represent the global architecture of the petal-type net rack; in the non-fault state, the weight of the edge is zero; A cross-zone propagation attenuation factor is introduced to correct the activated edge weight; A graph convolution network is adopted as a core calculation unit to complete the construction of the global interconnection layer.

7. The petal type net rack fault tracing and dynamic reconstruction method based on a graph neural network according to claim 3, characterized in that, According to the overall structure feature and the local structure feature, parameters of a dynamic time sequence attention mechanism of a time sequence prediction model are determined to complete the construction of a path tracking layer, including: Based on the topological hierarchical relationship in the overall structure feature, a hierarchical weight distribution rule of the dynamic time sequence attention mechanism is determined; According to the edge weight rule in the local structure feature, the edge weight is dynamically associated with the time sequence window size of the attention mechanism; Taking the physical topology in the overall structure feature as a skeleton, the nodes of each graph layer are mapped to a time axis according to the electrical connection relationship, to construct a time-space fusion graph, wherein the node attributes inherit the hierarchical and partition labels in the structure feature, and the edge attributes inherit the edge weight rule of the corresponding graph layer and mark the time sequence information of the topological association; The combination of a dynamic time sequence attention mechanism and a time sequence long short-term memory network model is adopted as a core calculation unit to complete the construction of the path tracking layer.

8. The petal type net rack fault tracing and dynamic reconstruction method based on a graph neural network according to claim 7, characterized in that, The method further includes: Differentiated initial attention weights are assigned to nodes at different levels, wherein the initial weight of a core node is higher than that of a branch node, and the initial weight of a branch node is higher than that of a terminal node; Different numbers of attention heads are configured for nodes at different levels according to the level labels of the nodes. The level label of a node is input into the dynamic time sequence attention mechanism as part of the position encoding.

9. The petal type net rack fault tracing and dynamic reconstruction method based on a graph neural network according to claim 1, characterized in that, According to the fault tracing result, a hierarchical dynamic reconstruction strategy is generated and executed to isolate the fault and restore power supply, including: According to the fault tracing result, the corresponding decision level of the hierarchical decision model is dynamically activated; wherein the hierarchical structure of the hierarchical decision model matches the hierarchical structure of the petal-shaped grid; Within the activated decision level, a dynamic reconstruction strategy is generated based on a pre-set multi-objective optimization function; According to the dynamic reconstruction strategy, the grid is reconstructed.

10. The petal-type grid fault tracing and dynamic reconstruction method based on a graph neural network according to claim 9, characterized in that, The method further includes: A hierarchical reinforcement learning framework is used to build the hierarchical decision model, wherein the model includes multiple decision levels, each level corresponding to a physical level of the petal-shaped grid; An independent agent is configured for each decision level, and an action space and a state space matching the physical range and control authority of the level are defined; A hierarchical reward function is designed for the agent of each level, which differentially weights the load recovery amount, response time, operation cost and system risk according to the fault type and reconstruction target responsible by the level; A collaborative communication mechanism between levels is established.

Images