Multi-view cooperative control method, system and device for power restoration and medium thereof

Abstract

This invention discloses a multi-perspective collaborative control method, system, device, and medium for power restoration, belonging to the field of power system automation technology. The method includes: acquiring initial state data from multiple monitoring terminals deployed on power lines after a power restoration command is issued; selecting collaborative nodes with communication and sensing capabilities based on the initial state data to construct a multi-perspective collaborative group; utilizing each collaborative node within the multi-perspective collaborative group to collect electrical parameters and environmental image data in a time-division multiplexing manner, and performing spatiotemporal alignment and fusion of the collected data to generate a power restoration situational awareness matrix; inputting the power restoration situational awareness matrix into a power restoration decision model to output power restoration control strategies for different line sections, including immediate power restoration, delayed power restoration, combined switching, and manual intervention maintenance. This invention, through multi-perspective collaborative sensing and fusion decision-making, solves the problem of misjudgment caused by incomplete data from a single point during power restoration, improving the accuracy and safety of power restoration operations.

Description

Technical Field

[0001] This invention relates to the field of power system automation technology, and in particular to a multi-perspective collaborative control method, system, device and medium for power restoration. Background Technology

[0002] With the deepening of smart grid construction, the automation level of distribution and transmission networks has been significantly improved. Among them, rapid power restoration after a fault (referred to as power restoration) is a key link in measuring power supply reliability. In traditional power restoration operations, when a line fault occurs and is isolated, dispatchers or automation systems usually determine whether the non-faulty section is ready for power restoration based on single or limited monitoring data reported by the fault location system, and then remotely or locally operate circuit breakers, load switches and other equipment to complete the power restoration.

[0003] Currently, to support power restoration decisions, various types of monitoring terminals have been widely deployed along power lines. These mainly include: intelligent fault indicators with electrical quantity acquisition functions, distribution automation terminals (DTU / FTU), and video surveillance devices, infrared thermal imagers, and micro-meteorological sensors for environmental perception. These terminals upload data to the main station system or edge computing nodes via fiber optics, 4G / 5G wireless networks, or low-power wide-area networks (LPWAN). Relevant technical literature (such as the article "A Review of Distribution Network Fault Handling Technology Based on Multi-Source Data Fusion" in the journal "Automation of Electric Power Systems," and the State Grid Corporation of China's enterprise standard Q / GDW 11357-2014 "Technical Guidelines for Distribution Automation") points out that utilizing multi-source data to improve the perception capability of power grid operation status has become an industry development trend.

[0004] However, in actual power restoration scenarios, existing technical solutions still have the following technical problems and shortcomings: First, power restoration decisions rely on single-point or local data, lacking a global, collaborative perspective. Existing systems typically receive data independently from each monitoring terminal, with each terminal reflecting only the local state of its installation point. After a power restoration command is issued, the decision-making system often uses data from a critical node (such as the first automated switch upstream of the fault point) as the primary basis for judgment. When the sensor at that node experiences measurement errors, communication interference, or brief transient disturbances, anomalies in single-point data can easily lead the decision-making system to misjudge the overall state of the line. For example, if a terminal falsely reports a voltage dip due to electromagnetic interference, the system may incorrectly determine that a fault still exists in that section, delaying power restoration and affecting power supply recovery efficiency. Conversely, if a terminal fails to detect a real line hazard due to a fault (such as a weak grounding caused by a tree obstruction), the system may incorrectly issue a power restoration command, triggering secondary faults or equipment damage.

[0005] Second, electrical quantity data and environmental sensing data are fragmented and fail to be effectively integrated. Although current power lines are equipped with both electrical quantity sensors and environmental monitoring equipment (such as cameras), under the existing technological framework, these two types of data are typically uploaded independently via different channels and with different time scales, and stored and displayed separately in the main station system. Electrical quantity data reflects the electromagnetic transient characteristics inside the line, while environmental data reflects physical hazards outside the line. Due to the lack of a unified data acquisition and coordination mechanism and spatiotemporal alignment and fusion methods, decision-makers or automated systems find it difficult to correlate voltage and current anomalies at a certain moment with environmental images of the same location at the same time for analysis. This fragmentation prevents power restoration decisions from comprehensively utilizing the dual verification of "internal electrical state" and "external environmental state," resulting in an incomplete assessment of line safety.

[0006] Third, the lack of time-series coordination in multi-terminal data acquisition leads to information fragmentation. Existing monitoring terminals typically report data according to their own preset cycles (e.g., uploading every 15 minutes) or event-triggered methods, resulting in random differences in the acquisition times between different terminals. In the time-sensitive scenario of power restoration, the data received by the decision-making system from multiple terminals are not actually state information at the same time segment. For example, node A reports current data at time t1, while node B reports image data at time t2, but the line status may have changed between t1 and t2. This temporal asynchrony makes it impossible for the system to construct accurate transient situational awareness and to make a consistent assessment of the true state of the line section before power restoration.

[0007] Fourth, existing power restoration control strategies are simplistic and lack differentiated decision-making capabilities based on multimodal situational awareness. Current technologies typically only distinguish between "power restoration permitted" and "power restoration prohibited," or simply employ a "one-time full-line closing" approach. This approach fails to adequately consider the differentiated states of different line sections after a fault. For example, the main line may have fully recovered, while a branch line may still have a weak grounding fault; or a section may have normal electrical parameters but be susceptible to external damage. The lack of refined power restoration strategies (such as phased and combined switching, delayed power restoration, and tiered manual intervention) can easily lead to unnecessary expansion of the outage area or damage to equipment from the power restoration inrush current.

[0008] In summary, the existing technology lacks a technical solution that can collaboratively organize multi-view monitoring terminals, synchronously collect and fuse electrical parameters and environmental image data in time and space, and then output a refined power restoration control strategy. This makes it difficult to meet the comprehensive requirements of the current smart grid for the safety, accuracy and efficiency of the power restoration process. Summary of the Invention

[0009] In order to overcome the above-mentioned defects of the prior art, the present invention provides a multi-source time-domain cooperative control method, system and device for virtual synchronous generator to solve the problems mentioned in the background art.

[0010] To achieve the above objectives, the present invention provides the following technical solution: A multi-perspective collaborative control method for power restoration, characterized by comprising: The initial status data of multiple monitoring terminals deployed on power lines after the power restoration command is issued is obtained. The initial status data includes the remaining power of each monitoring terminal, the communication signal-to-noise ratio, the status of the sensing module, and the topological correlation of the line segment in which it is located. Based on the initial state data, collaborative nodes with communication and sensing capabilities are selected from the multiple monitoring terminals to construct a multi-view collaborative group. Specifically, this includes: identifying monitoring terminals with remaining power greater than a first threshold, communication signal-to-noise ratio greater than a second threshold, normal sensing module status, and topological correlation meeting preset hierarchical conditions as collaborative nodes; and establishing a multi-view collaborative group with data relay and collaborative sensing functions based on the physical topology of the line segment where the collaborative node is located. Using the collaborative nodes within the multi-view collaborative group, electrical quantity parameters and environmental image data of their corresponding line sections are collected in a time-division multiplexing manner. The electrical quantity parameters and environmental image data are then spatiotemporally aligned and fused to generate a power restoration situation awareness matrix. Specifically, this includes: assigning time slot labels to each collaborative node within the multi-view collaborative group, enabling different collaborative nodes to sequentially collect electrical quantity parameters and environmental image data within a preset time window; aligning the timestamps of the data collected by each collaborative node using a common time reference; performing spatial registration using the geographical location identifier of the line section as a spatial anchor point; and stitching the aligned and registered electrical quantity parameters and environmental image data according to preset feature dimensions to form a power restoration situation awareness matrix. The power restoration situation awareness matrix is ​​input into a preset power restoration decision model, and power restoration control strategies for different line sections are output. The power restoration control strategies include immediate power restoration, delayed power restoration, combined switching, and manual intervention for maintenance.

[0011] Furthermore, the step of collecting electrical quantity parameters and environmental image data of the corresponding line sections in a time-division multiplexing manner, and performing spatiotemporal alignment and fusion of the electrical quantity parameters and environmental image data to generate a power restoration situation awareness matrix, further includes: The time slot length is determined based on the data type of each collaborative node. Nodes that upload only electrical quantity parameters are allocated the first time slot length, nodes that upload electrical quantity parameters plus visible light images are allocated the second time slot length, nodes that upload electrical quantity parameters plus infrared thermal images are allocated the third time slot length, and nodes that upload electrical quantity parameters plus video stream keyframes are allocated the fourth time slot length. A protection interval is set between each time slot. Based on the historical data reporting delay statistics of each collaborative node, the time slot length is dynamically adjusted. Nodes with reporting delays greater than the preset delay threshold are allocated longer time slots, while nodes with reporting delays less than the preset delay threshold are allocated shorter time slots. The dynamically adjusted time slot allocation table is broadcast to each collaborative node through the communication link within the collaborative group, and each collaborative node performs the acquisition operation within its allocated time slot.

[0012] Furthermore, the step of spatiotemporally aligning and fusing the electrical quantity parameters with environmental image data to generate a power restoration situational awareness matrix further includes: Using UTC time provided by the GPS timing module or Precision Time Protocol as the common time reference, a timestamp accurate to the microsecond level is added to each data packet. For data with time deviation exceeding the preset deviation threshold, a linear interpolation algorithm is used for time correction. Based on the tower number and latitude and longitude coordinates of the installation location of each collaborative node, the data is mapped to a pre-built feeder geographic information system model to determine the line segment number corresponding to each data point. The line segment is the line segment between two adjacent towers. For data belonging to two collaborative nodes at both ends of the same line segment, associate and bind them; For the collected environmental image data, a pre-trained convolutional neural network is called to extract image feature vectors. The convolutional neural network has been pre-trained on a typical power line scene image dataset, which includes tree obstacles, bird nests, insulator damage, and foreign objects hanging. The electrical quantity parameters of each collaborative node are formed into an electrical quantity feature vector. The electrical quantity feature vector is then vertically concatenated with the image feature vector to form the multimodal feature vector of that node. The multimodal feature vectors of the nodes at both ends of the same line segment are horizontally concatenated or weighted and fused to form the joint feature vector of the line segment. If the line segment has only one cooperating node, the feature at the other end is filled with zeros. Arrange the joint feature vectors of all line segments in the order of line topology to construct a power situation awareness matrix. Each column of the matrix corresponds to a line segment, and each row corresponds to a feature dimension. At the same time, construct an adjacency matrix. In the adjacency matrix, the value is 1 when two line segments are directly connected in topology, and 0 otherwise.

[0013] Furthermore, the power restoration decision model is a classification model based on a graph neural network. The step of inputting the power restoration situational awareness matrix into the preset power restoration decision model and outputting power restoration control strategies for different line sections specifically includes: Construct a power grid topology graph, where the node set corresponds to the line segment, the edge set corresponds to the electrical connection relationship between the line segments, and the adjacency matrix is ​​constructed based on the power grid topology graph; The complex situational awareness matrix is ​​used as the input feature matrix of the graph neural network, and the adjacency matrix is ​​used as the structural information input of the graph neural network. The input features are transformed layer by layer through at least two graph convolutional layers or graph attention layers. The calculation method of each graph convolutional layer is H^(l+1) = σ( D^(-1 / 2) AD^(-1 / 2) H^(l) W^(l) ), where H^(l) is the node feature matrix of the l-th layer, H^(0) is the input complex electrical situational awareness matrix, A is the adjacency matrix after adding self-loops, D is the degree matrix, W^(l) is the learnable weight matrix of the l-th layer, and σ is the activation function. The node features are mapped to a four-dimensional output through the last fully connected layer, and the predicted probability of the four types of power restoration control strategies for each line segment is obtained after passing through the Softmax activation function. The strategy with the highest predicted probability for each line segment is selected as the power restoration control strategy for that segment.

[0014] Furthermore, after outputting the power restoration control strategy for different line sections, it also includes: Traverse all line segments, collect the power restoration control strategy for each segment, and arrange them according to the line topology to generate a power restoration operation sequence; Line sections marked as separate closing groups are grouped into the same closing group, and the closing sequence within the group is determined according to the upstream and downstream topology. Set delay parameters for line sections marked as delayed power restoration, and dynamically adjust the delay time according to the prediction probability confidence level output by the model. Set the first delay time when the confidence level is greater than the first confidence level threshold, set the second delay time when the confidence level is less than the first confidence level threshold but greater than the second confidence level threshold, and set the third delay time when the confidence level is less than the second confidence level threshold. The line section marked for manual intervention maintenance is marked as prohibited from operation, and an alarm report is generated. The alarm report includes the abnormal section number, abnormal feature data and on-site image thumbnail. The closing command is sent to the intelligent switching equipment of the corresponding line section through the communication protocol. For the group closing strategy, the closing command is sent in sequence according to the order within the group, and the interval between adjacent commands is a preset interval time. Monitor the action feedback signals of each switching device, retry commands that fail to execute, record the execution results and report them to the dispatch center.

[0015] Furthermore, the method is applied to an edge computing gateway, which is deployed at the beginning of a feeder and connected to monitoring terminals within its jurisdiction via a local communication network. The method further includes: After receiving the power restoration command from the dispatch center, the edge computing gateway sends a status query command to the connected monitoring terminal and receives and stores the initial status data returned by each terminal. The edge computing gateway performs collaborative node screening and builds multi-view collaborative groups based on the screened collaborative nodes. When building multi-view collaborative groups, the K-means clustering algorithm is used to divide the terminal into multiple clusters according to the geographical coordinates of the terminal. In each cluster, the terminal that is closest to the cluster center and has the highest perception capability score is selected as the collaborative node of that cluster. The edge computing gateway assigns time slot labels to each collaborative node in the collaborative group and issues collection commands through the local communication network. The edge computing gateway receives electrical quantity parameters and environmental image data uploaded by each collaborative node, performs spatiotemporal alignment and fusion locally, and generates a power restoration situation awareness matrix. The edge computing gateway loads a locally deployed lightweight graph neural network model, inputs the power restoration situation awareness matrix into the model for inference, and outputs a power restoration control strategy. The edge computing gateway generates an operation sequence based on the power restoration control strategy, sends a closing command to the intelligent switching device through the local communication network, and reports the execution result to the dispatch center.

[0016] Furthermore, the multi-perspective collaborative control method for power restoration also includes the training and updating of the power restoration decision model, specifically including: Power restoration operation records are extracted from the historical power restoration case database. Each power restoration operation record includes electrical waveform data, environmental image data, line topology information, actual execution results, and the actual line status reported by maintenance personnel before power restoration. Construct the power restoration situation awareness matrix and adjacency matrix for each power restoration operation according to the multi-perspective collaborative control method for power restoration; Based on the actual execution results and the real line status reported by the operation and maintenance personnel, a power restoration control strategy label is marked for each line section. The power restoration control strategy label includes immediate power restoration, delayed power restoration, combination of switches and manual intervention maintenance. The labeled sample set is divided into a training set, a validation set, and a test set. The graph neural network model is trained using the training set, hyperparameters are tuned using the validation set, and the model performance is evaluated using the test set. When the number of new samples added to the incremental sample library exceeds the preset threshold, incremental training is triggered. The elastic weight consolidation method is used for incremental training, and a regularization term is added to the loss function to constrain the change of model parameters. The trained model is evaluated on the validation set. If the accuracy of the new model is not lower than the preset percentage threshold of the accuracy of the old model, the new model is deployed to the production environment using a blue-green deployment strategy. The new and old models run in parallel for a preset period of time before a complete switch is performed.

[0017] On the other hand, the present invention provides a multi-perspective collaborative control system for power restoration, comprising: The status acquisition module is used to acquire the initial status data of multiple monitoring terminals deployed on the power line after the power restoration command is issued. The initial status data includes the remaining power of each monitoring terminal, the communication signal-to-noise ratio, the status of the sensing module, and the topological correlation of the line segment where it is located. The collaborative construction module is used to select collaborative nodes with communication and sensing capabilities from the multiple monitoring terminals based on the initial state data, and construct a multi-view collaborative group. Specifically, it includes: determining the monitoring terminals with remaining power greater than a first threshold, communication signal-to-noise ratio greater than a second threshold, normal sensing module status, and topological correlation degree meeting the preset hierarchical conditions as collaborative nodes, and establishing a multi-view collaborative group with data relay and collaborative sensing functions based on the physical topology relationship of the line segment where the collaborative node is located. The time slot allocation unit is used to allocate time slot labels to each collaborative node in the multi-view collaborative group, so that different collaborative nodes collect data sequentially within a preset time window. The collaborative acquisition module is used to control each collaborative node to acquire electrical parameters and environmental image data of its corresponding line section within the allocated time slot; The spatiotemporal alignment and fusion module is used to align the timestamps of the data collected by each collaborative node with a common time reference, and to perform spatial registration with the geographical location identifier of the line section as the spatial anchor point. The aligned and registered electrical parameters and environmental image data are then used for feature extraction and stitching. The matrix generation module is used to arrange the feature vectors of each segment processed by the spatiotemporal alignment and fusion module according to the line topology order to construct the power restoration situation awareness matrix. The decision model module is used to load the pre-trained power restoration decision model, input the power restoration situation awareness matrix into the power restoration decision model, and output the power restoration control strategy for different line sections. The power restoration control strategy includes immediate power restoration, delayed power restoration, combined switching and manual intervention maintenance. The strategy execution module is used to generate a power restoration operation sequence based on the power restoration control strategy output by the decision model module, and send the instructions to the intelligent switchgear to perform the power restoration operation through the communication interface. The model management module is used to manage the accumulation of training data, model training, model validation, model updates and deployment of the power restoration decision model.

[0018] On the other hand, the present invention also provides a multi-view collaborative control device for power restoration, comprising: At least one processor; at least one memory storing a computer program; When the computer program is executed by the at least one processor, the device performs the following steps: The initial status data of multiple monitoring terminals deployed on power lines after the power restoration command is issued is obtained. The initial status data includes the remaining power of each monitoring terminal, the communication signal-to-noise ratio, the status of the sensing module, and the topological correlation of the line segment in which it is located. Based on the initial state data, collaborative nodes with communication and sensing capabilities are selected from the multiple monitoring terminals to construct a multi-view collaborative group. Specifically, this includes: identifying monitoring terminals with remaining power greater than a first threshold, communication signal-to-noise ratio greater than a second threshold, normal sensing module status, and topological correlation meeting preset hierarchical conditions as collaborative nodes; and establishing a multi-view collaborative group with data relay and collaborative sensing functions based on the physical topology of the line segment where the collaborative node is located. Using the collaborative nodes within the multi-view collaborative group, electrical quantity parameters and environmental image data of their corresponding line sections are collected in a time-division multiplexing manner. The electrical quantity parameters and environmental image data are then spatiotemporally aligned and fused to generate a power restoration situation awareness matrix. Specifically, this includes: assigning time slot labels to each collaborative node within the multi-view collaborative group, enabling different collaborative nodes to sequentially collect electrical quantity parameters and environmental image data within a preset time window; aligning the timestamps of the data collected by each collaborative node using a common time reference; performing spatial registration using the geographical location identifier of the line section as a spatial anchor point; and stitching the aligned and registered electrical quantity parameters and environmental image data according to preset feature dimensions to form a power restoration situation awareness matrix. The power restoration situation awareness matrix is ​​input into a preset power restoration decision model, and power restoration control strategies for different line sections are output. The power restoration control strategies include immediate power restoration, delayed power restoration, combined switching, and manual intervention for maintenance.

[0019] On the other hand, the present invention also provides a computer-readable storage medium for implementing a multi-view cooperative control method for power restoration, wherein a computer program is stored thereon, and the computer program, when executed by a processor, performs the following steps: The initial status data of multiple monitoring terminals deployed on power lines after the power restoration command is issued is obtained. The initial status data includes the remaining power of each monitoring terminal, the communication signal-to-noise ratio, the status of the sensing module, and the topological correlation of the line segment in which it is located. Based on the initial state data, collaborative nodes with communication and sensing capabilities are selected from the multiple monitoring terminals to construct a multi-view collaborative group. Specifically, this includes: identifying monitoring terminals with remaining power greater than a first threshold, communication signal-to-noise ratio greater than a second threshold, normal sensing module status, and topological correlation meeting preset hierarchical conditions as collaborative nodes; and establishing a multi-view collaborative group with data relay and collaborative sensing functions based on the physical topology of the line segment where the collaborative node is located. Using the collaborative nodes within the multi-view collaborative group, electrical quantity parameters and environmental image data of their corresponding line sections are collected in a time-division multiplexing manner. The electrical quantity parameters and environmental image data are then spatiotemporally aligned and fused to generate a power restoration situation awareness matrix. Specifically, this includes: assigning time slot labels to each collaborative node within the multi-view collaborative group, enabling different collaborative nodes to sequentially collect electrical quantity parameters and environmental image data within a preset time window; aligning the timestamps of the data collected by each collaborative node using a common time reference; performing spatial registration using the geographical location identifier of the line section as a spatial anchor point; and stitching the aligned and registered electrical quantity parameters and environmental image data according to preset feature dimensions to form a power restoration situation awareness matrix. The power restoration situation awareness matrix is ​​input into a preset power restoration decision model, and power restoration control strategies for different line sections are output. The power restoration control strategies include immediate power restoration, delayed power restoration, combined switching, and manual intervention for maintenance.

[0020] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention solves the problem of data loss or false alarms caused by poor terminal status in existing technologies by constructing a multi-view collaborative group to select collaborative nodes with communication and sensing capabilities from multiple monitoring terminals. This significantly improves the robustness and reliability of data acquisition. Specifically, after the power restoration command is issued, this invention first acquires multi-dimensional initial state data of each monitoring terminal, including remaining power, communication signal-to-noise ratio, sensing module status, and topological correlation. A screening threshold is then set, identifying only terminals with remaining power greater than 20%, communication signal-to-noise ratio greater than 15dB, normal sensing module status, and topological correlation meeting preset hierarchical conditions as collaborative nodes. This screening mechanism avoids including terminals with insufficient power, poor communication quality, or sensor malfunctions in the collaborative group, preventing interference with power restoration decisions caused by erroneous data reported by these terminals.

[0021] 2. This invention employs a time-division multiplexing method to collect data, combined with spatiotemporal alignment fusion technology, to fuse electrical parameters and environmental image data under a unified time reference and spatial coordinates. This solves the problem of information fragmentation caused by asynchronous time and spatial misalignment of multi-source data in existing technologies, forming a perception matrix reflecting the comprehensive situation of the line section. In existing technologies, each monitoring terminal independently reports data according to its own preset cycle, resulting in the decision system receiving data from different time segments, making it impossible to construct the line status at the same moment. This invention assigns time slot labels to each node in the collaborative group, enabling all nodes to complete data collection sequentially within a preset 500-millisecond time window. Using the UTC time provided by the GPS timing module as a common time reference, a timestamp accurate to 1 microsecond is added to each data packet, achieving time alignment of multi-source data.

[0022] 3. This invention constructs a power restoration situational awareness matrix and inputs it into a graph neural network model. Utilizing the topological relationships between line segments for joint inference, it solves the problem of misjudgments in power restoration decisions caused by incomplete data from single points in existing technologies, significantly improving the accuracy of power restoration decisions. This invention uses the multimodal features of each line segment as node features in the graph neural network and the electrical connections between line segments as edges. Through graph convolutional layers or graph attention layers, each node's decision not only depends on its own features but also incorporates feature information from neighboring nodes. This mechanism allows the model to correct for abnormal data at a single node by using normal data from its neighbors, avoiding misjudgments caused by abnormal data at a single point.

[0023] 4. This invention solves the problem of existing technologies having a single power restoration strategy and being unable to adapt to the differentiated states of line sections by outputting differentiated power restoration control strategies, including immediate power restoration, delayed power restoration, combined switching, and manual intervention maintenance. This achieves refined power restoration operations. Based on the output of the power restoration situational awareness matrix and graph neural network model, this invention independently outputs four types of refined control strategies for each line section. For sections with normal electrical parameters and no environmental anomalies, an "immediate power restoration" strategy is output; for sections with transient fluctuations but no persistent faults, a "delayed power restoration" strategy is output, with a delay of 60 to 200 milliseconds dynamically set according to the confidence level, waiting for the transient process to subside before automatically closing the circuit; for multi-branch lines, a "combined switching" strategy is output, closing switches within the same group sequentially at 200-millisecond intervals according to the topological order, effectively reducing the impact of inrush current on equipment; for sections with detected clear fault characteristics or severe environmental anomalies, a "manual intervention maintenance" strategy is output, locking the remote power restoration function and pushing alarm information. This differentiated and precise control strategy makes the power restoration operation more closely match the actual condition of the line, avoiding secondary faults and equipment damage caused by blind power restoration.

[0024] 5. This invention solves the problem of existing decision-making models being unable to adapt to changes in power grid operating conditions by constructing an incremental update mechanism for the model, continuously accumulating power restoration case data and performing incremental training, thus ensuring the model maintains high accuracy over the long term. This invention establishes a complete model lifecycle management mechanism, continuously recording complete data for each power restoration operation, including the situational awareness matrix before power restoration, the model's output strategy, the actual execution results, and the real line status reported by maintenance personnel. Incremental training is automatically triggered when the number of new samples added to the incremental sample library exceeds 50, or when the model's prediction accuracy falls below 85% in the past month. This invention uses an elastic weight consolidation method for incremental training, adding a regularization term to the loss function to constrain changes in important parameters, thus learning the characteristics of new samples while retaining existing knowledge and avoiding catastrophic forgetting problems.

[0025] 6. This invention, through an edge computing gateway deployment scheme, pushes collaborative data acquisition, spatiotemporal fusion, and model inference to the edge side closer to the terminal, solving the problems of high communication latency and bandwidth pressure in the centralized architecture of existing technologies, and achieving low-latency response for power restoration control. The edge computing deployment scheme provided by this invention deploys an edge computing gateway at the beginning of each feeder. The gateway and the monitoring terminal of the same feeder are connected via Wi-Fi Mesh or LoRa self-organizing network, completing the entire process of collaborative node selection, time slot allocation, data acquisition, spatiotemporal alignment fusion, and model inference locally. The edge gateway uses a lightweight graph neural network model (model file only 2.3MB), with a single inference time of less than 50 milliseconds. The end-to-end latency from power restoration command issuance to execution is controlled within 200 milliseconds, far lower than the 2-3 second latency of the centralized architecture.

[0026] 7. This invention achieves multimodal fusion of environmental image data and electrical quantity data, utilizing an image feature extraction network to identify external hazards such as tree obstructions, bird nests, foreign object entrapment, and insulator damage. This solves the problem in existing technologies where relying solely on electrical quantity data fails to detect external environmental risks, significantly improving the safety of power restoration operations. Before power restoration decisions, this invention collects environmental image data of the line section using cameras and infrared thermal imagers within a collaborative group, and uses a pre-trained convolutional neural network (ResNet-50 or MobileNet-V3) to extract features from the images and identify various external hazards. This convolutional neural network has been pre-trained on image datasets containing typical scenarios such as tree obstructions, bird nests, insulator damage, and foreign object entrapment, accurately extracting visual features related to line safety. When the image feature vector and electrical quantity feature vector are fused and input into the graph neural network model, the model can comprehensively determine whether there are external hazards on the line.

[0027] 8. This invention achieves a unified representation of electrical quantity feature dimensions and environmental image feature dimensions by constructing a power restoration situational awareness matrix as a standardized input. This solves the problem of heterogeneous multi-source data formats and difficulty in joint modeling in existing technologies, providing high-quality input data for power restoration decision-making models. This invention uses feature extraction and fusion mechanisms to convert electrical quantity parameters into fixed-dimensional feature vectors (8-32 dimensions) and image data into fixed-dimensional feature vectors (512-1024 dimensions) through a convolutional neural network. The two are then concatenated to form a unified feature representation for each line segment. Furthermore, the feature vectors of all line segments are arranged in topological order to construct a standardized power restoration situational awareness matrix. This matrix has a fixed dimension D×K, where D is the joint feature dimension and K is the number of line segments. It can serve as a unified input for various machine learning models such as graph neural networks, support vector machines, and random forests. This standardized representation method not only solves the heterogeneity problem of multi-source data but also provides a unified input format for power restoration decision-making models of different feeders and regions, facilitating model transfer learning and large-scale deployment.

[0028] 9. This invention solves the problem of communication resource waste or acquisition timeout caused by fixed time slot allocation in the prior art by dynamically adjusting the acquisition time slot length based on the historical data reporting delay statistics of each cooperating node through a dynamic time slot allocation mechanism, thereby improving communication efficiency. In wireless communication environments, the data transmission delay of different monitoring terminals varies. Fixed time slot allocation may lead to insufficient time slots for some nodes, resulting in acquisition timeouts, or excessively long time slots, causing waste of communication resources. The time slot allocation unit of this invention adopts a dynamic allocation mechanism, which calculates the historical data reporting delay of each cooperating node in real time. Nodes with larger delays are allocated longer time slots (e.g., dynamically adjusted from 30 milliseconds to 50 milliseconds), while nodes with smaller delays are allocated shorter time slots (e.g., dynamically adjusted from 30 milliseconds to 20 milliseconds).

[0029] 10. This invention introduces a "grouping and switching" strategy into the power restoration decision model. Based on the line topology and section conditions, multiple branch lines are grouped and closed sequentially. This solves the problem of inrush current superposition impacting equipment caused by simultaneous full-line closing in existing technologies, effectively protecting power equipment. When multiple branch lines close simultaneously, the inrush currents from each branch superimpose on the main line, generating an inrush current several times the rated current, potentially causing circuit breaker malfunctions, transformer winding deformation, and other equipment damage. This invention specifically designs a "grouping and switching" strategy in the power restoration decision model. For sections with multiple branch lines, the model groups branches with similar conditions and topological connections into the same closing group based on the electrical parameters and environmental conditions of each branch. The group order is set according to the upstream and downstream relationships of the topology, with each group closing sequentially at 200 millisecond intervals. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the multi-perspective collaborative control system for power restoration according to the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Example 1: This example provides a multi-perspective collaborative control method for power restoration. This method is applied to a distribution network master station system to perform safe and precise power restoration control on non-faulty sections after a fault occurs on a 10kV feeder and fault isolation is completed. Please refer to... Figure 1 The method includes the following detailed steps: Step 1: Obtain the initial status data of multiple monitoring terminals deployed on the power line after the power restoration command is issued.

[0033] When the dispatch automation system issues a power restoration command for a specific feeder, the master station system immediately broadcasts a status query message to all registered monitoring terminals on that feeder via the 104 protocol or MQTT protocol. The monitoring terminals include: intelligent fault indicators, distribution automation terminals, spherical high-definition cameras, infrared thermal imagers, micro-meteorological sensors, and tower tilt sensors.

[0034] Upon receiving a query message, each monitoring terminal returns an initial status data message within 100 milliseconds. This message uses a JSON data structure and includes the following fields: Terminal Identifier: represented by a 16-bit hexadecimal number, including terminal type and installation location codes; Remaining Battery Percentage: ranging from 0 to 100, accurate to 0.1%; Communication Signal-to-Noise Ratio: in dB, measured by the terminal's built-in 4G communication module or wireless self-organizing network module; Sensing Module Status: including subfields such as camera pan-tilt self-test status, sensor calibration status, and remaining memory card space; Topology Association Information: including the tower number to which the terminal is attached, the feeder branch level, and the identifiers of upstream and downstream adjacent terminals.

[0035] The main system stores all returned initial state data into a real-time database and marks the query timestamp, forming a state dataset S = {s1, s2, …, s}. n}, where n is the total number of monitoring terminals deployed on the feeder.

[0036] Step 2: Based on the initial state data, select collaborative nodes with communication and sensing capabilities from multiple monitoring terminals to construct a multi-view collaborative group.

[0037] The main station system executes a collaborative node selection algorithm, the processing flow of which is as follows: First, extract the remaining battery power E_i, communication signal-to-noise ratio SNR_i, sensing module status flag F_i (1 for normal, 0 for abnormal) and topology correlation level L_i (level 1 for main line, level 2 for first-level branch, level 3 for second-level branch, and so on) for each terminal from the status dataset S.

[0038] Secondly, set the filtering thresholds: remaining battery power threshold E_th = 20%, communication signal-to-noise ratio threshold SNR_th = 15dB. The filtering conditions are: E_i ≥ E_th; SNR_i ≥ SNR_th; F_i = 1; L_i ≤ 3 (that is, only terminals on the main line and first and second-level branches are filtered to avoid excessive communication overhead caused by filtering too many nodes on the terminal branches).

[0039] Monitoring terminals that meet all the above conditions are identified as collaborative nodes. Assuming there are 24 monitoring terminals on the feeder, after screening, 7 collaborative nodes are obtained and deployed at 3 key locations on the main line, 2 locations on the first-level branches, and 2 locations on the second-level branches.

[0040] Then, the master station system constructs a multi-view collaborative group based on the topology association information of the collaborative nodes. Specifically, the collaborative nodes on the backbone line are used as the core, and each backbone node and its downstream branches are divided into a subgroup. Within each subgroup, nodes establish communication links via the LoRa wireless ad hoc network protocol, with the backbone node simultaneously serving as both data aggregation and relay within the subgroup. The resulting multi-view collaborative group G = {g1, g2, g3}, where g1 covers the first segment of the feeder and branch A, g2 covers the middle segment of the feeder and branch B, and g3 covers the last segment of the feeder.

[0041] Step 3: Utilize each collaborative node in the multi-view collaborative group to collect electrical quantity parameters and environmental image data of their corresponding line sections in a time-division multiplexing manner, and perform spatiotemporal alignment and fusion of electrical quantity parameters and environmental image data to generate a power restoration situation awareness matrix.

[0042] The master station system sends a collaborative acquisition command to the multi-view collaborative group G. The command includes a preset time window length T_window = 500 milliseconds and a time slot allocation scheme.

[0043] The time slot allocation unit assigns a unique time slot label to each collaborative node. The time slot length is determined based on the node's data type: nodes uploading only electrical quantity parameters are allocated a 20-millisecond time slot, nodes uploading electrical quantity parameters plus visible light images are allocated an 80-millisecond time slot, and nodes uploading electrical quantity parameters plus infrared thermal images are allocated a 100-millisecond time slot. A 5-millisecond guard interval is set between time slots to avoid collisions. The time slots of the 7 collaborative nodes are arranged sequentially according to the topology, and the total acquisition time does not exceed 500 milliseconds.

[0044] Each coordinating node performs data acquisition within its allocated time slot. Electrical parameters are acquired as follows: the node continuously acquires voltage and current waveform data for one power frequency cycle (20 milliseconds) at a sampling rate of 12.8 kHz, and calculates the effective value, phase angle, 2nd to 19th harmonic content, and three-phase unbalance, forming an electrical characteristic vector E = [U_a, U_b, U_c, I_a, I_b, I_c, THD, unbalance], with a total of 8 dimensions. Environmental image data is acquired as follows: the node controls a camera or infrared thermal imager to capture a JPEG image with a resolution of 1920×1080, or extracts the temperature matrix data (32×32 pixels) from the thermal imager.

[0045] All collected data is uploaded to the main station system in time slot order. The main station system uses the UTC time provided by the GPS timing module as the common time reference and adds a timestamp accurate to 1 microsecond to each data packet. At the same time, the data is bound to the line section using the tower number and latitude and longitude coordinates of each cooperative node's installation location as spatial anchor points.

[0046] The spatiotemporal alignment and fusion module performs the following fusion operations: First, for the image data uploaded by each collaborative node, a pre-trained ResNet-50 convolutional neural network is used to extract image feature vectors. For visible light images, the network outputs a 1024-dimensional feature vector; for infrared thermal images, the network outputs a 512-dimensional temperature distribution feature vector. This convolutional neural network has been pre-trained on image datasets of typical power line scenarios (tree obstructions, bird nests, insulator damage, and foreign object suspension), and can effectively extract visual features related to line safety. Second, the electrical quantity feature vector of the node is concatenated with the image feature vector to form the multimodal feature vector F_i of the node, with a dimension of 8 + 1024 (or 512) = 1032 dimensions (or 520 dimensions). Then, taking a line segment as a unit, the multimodal feature vectors of two collaborative nodes (located at opposite ends of the segment) belonging to the same line segment (the segment between two adjacent towers) are horizontally concatenated to form the joint feature vector of the segment. If a segment has only one collaborative node, the feature vector at the other end is padded with zeros. Finally, the joint feature vectors of all line segments are arranged in order of line topology to construct a complex power situation awareness matrix M. Each column of this matrix corresponds to a line segment, and each row corresponds to a feature dimension. The matrix size is D × K, where D is the total dimension of the joint feature vectors and K is the number of line segments.

[0047] In this embodiment, the feeder is divided into 15 line segments, so the dimension of the power restoration situation awareness matrix M is 1032 × 15.

[0048] Step 4: Input the power restoration situation awareness matrix into the preset power restoration decision model and output the power restoration control strategy for different line sections.

[0049] The power restoration decision model employs a graph neural network architecture, specifically a graph convolutional network. The model's construction process is as follows: First, construct the power grid topology graph G_topo = (V, E, A), where V is the set of nodes, each node corresponds to a line segment; E is the set of edges, representing the electrical connection relationship between line segments; A is the adjacency matrix with dimension K×K, where A[i][j] = 1 when two line segments are directly connected, and 0 otherwise.

[0050] The graph convolutional network consists of three graph convolutional layers, with output dimensions of 64, 32, and 4 for each layer. The calculation formula for each graph convolutional layer is as follows: H^{(l+1)} = σ( D^{-1 / 2} AD^{-1 / 2} H^{(l)} W^{(l)} ) Where H^{(l)} is the node feature matrix of the l-th layer, H^{(0)} is the input complex situational awareness matrix M; A is the adjacency matrix after adding self-loops; D is the degree matrix; W^{(l)} is the learnable weight matrix of the l-th layer; σ is the ReLU activation function.

[0051] The final convolutional layer outputs the predicted probability of each node corresponding to four types of power restoration control strategies. These four strategies are: Immediate Power Restoration: This indicates that the electrical parameters of the line section are normal and the environment is normal, so the closing operation can be performed immediately; Delayed Power Restoration: This indicates that there are transient fluctuations in electrical parameters (such as a brief exceedance of harmonic content) or slight environmental anomalies (such as a slight swaying of trees due to a light breeze), and the closing operation should be performed after 3 to 5 power frequency cycles; Grouped Closing: This indicates that the line section is related to adjacent sections and the closing operation should be performed sequentially according to a preset group order to avoid the superposition of closing inrush currents. Manual intervention for maintenance: This indicates that a clear fault characteristic has been detected (such as continuous zero-sequence current, high temperature of insulators, foreign object hanging) or a serious environmental anomaly (such as fallen tree, fire). Remote power restoration is prohibited, and maintenance personnel must be on-site to handle the situation.

[0052] The master station system selects the strategy with the highest probability as the power restoration control strategy for each segment based on the predicted probability output by the model, and generates a power restoration operation sequence. This sequence is arranged according to the line topology, grouping segments with the "grouping and closing" strategy into the same closing group, setting delay time parameters for segments with the "delayed power restoration" strategy, and marking segments with the "manual intervention maintenance" strategy as prohibited from operation.

[0053] Finally, the master station system sends the power restoration operation sequence to the corresponding intelligent switch controller for execution via the 104 protocol. For the "immediate power restoration" and "separate and combined switching" strategies, the system automatically performs the closing operation after confirming that all preconditions are met; for the "delayed power restoration" strategy, the system starts a timer and automatically executes the operation after the delay; for the "manual intervention maintenance" strategy, the system pushes alarm information to the handheld terminal of the maintenance personnel and locks the remote power restoration function of that section.

[0054] Through the above steps, this embodiment realizes a complete closed loop from multi-view collaborative data acquisition to refined power restoration control, effectively improving the accuracy and safety of power restoration decisions.

[0055] Example 2: This example provides a multi-perspective collaborative control method for power restoration. The difference from Example 1 is that the method in this example is applied to an edge computing gateway device deployed on the substation side or key nodes of the line, and adopts a distributed collaborative architecture to reduce the communication pressure on the master station and improve the real-time performance of decision-making.

[0056] The method in this embodiment includes the following steps: Step 1: The edge computing gateway acquires the initial status data of multiple monitoring terminals within its jurisdiction.

[0057] At the beginning of each 10kV feeder, an edge computing gateway is deployed. This gateway is equipped with an ARM architecture quad-core processor, 4GB of RAM, 128GB of solid-state storage, and a 4G / 5G communication module. The gateway connects to 35 monitoring terminals deployed on the feeder via an RS485 bus or Ethernet interface, forming a local area data network.

[0058] After the dispatch center sends a power-back command to the feeder, the command is transmitted to the edge computing gateway via the 5G network. Within 10 milliseconds of receiving the command, the gateway sends a status query command to all connected monitoring terminals. Each terminal returns an initial status data packet within 50 milliseconds. The data packet uses a binary protocol format to reduce overhead and includes the terminal ID, battery voltage value, received signal strength indication, sensor self-test status code, and terminal location coordinates.

[0059] The edge computing gateway stores all returned data in a local SQLite database and calculates the terminal response rate. In this embodiment, all 35 terminals responded, resulting in a 100% response rate.

[0060] Step 2: The edge computing gateway filters collaborative nodes and builds multi-view collaborative groups.

[0061] The gateway executes a collaborative node selection algorithm that comprehensively considers the terminal's energy status, communication quality, sensing capabilities, and geographical distribution uniformity.

[0062] The specific screening process is as follows: The gateway extracts the remaining battery percentage E_i, communication signal strength RSSI_i, and sensor self-test status S_i for each terminal. E_th is set to 15% (allowing for lower battery levels due to the short communication distance at the edge) and RSSI_th to -85dBm.

[0063] Initial screening identified 28 terminals that simultaneously satisfy E_i ≥ E_th, RSSI_i ≥ RSSI_th, and S_i = normal state.

[0064] For the remaining 28 terminals, the gateway further optimized and selected them based on the evenness of their geographical distribution. Using the K-means clustering algorithm, the terminals were divided into 6 clusters according to their latitude and longitude coordinates, with each cluster representing a group of line segments. Within each cluster, the terminal closest to the cluster center and possessing the most comprehensive sensing capabilities was selected as the collaborative node for that cluster. The evaluation index for sensing capabilities was: 1 point for electrical quantity acquisition, 2 points for visible light image acquisition, and 3 points for infrared thermal imaging; higher scores indicated more comprehensive sensing capabilities.

[0065] After clustering optimization and screening, six collaborative nodes were finally determined, each responsible for the collaborative sensing task of a segment group. The gateway organizes these six collaborative nodes into a multi-view collaborative group and assigns each collaborative node a group number from 1 to 6. Communication within the group adopts the Wi-Fi Mesh protocol, and the nodes can communicate directly with each other, forming a peer-to-peer network structure.

[0066] Step 3: Collaborative data acquisition and spatiotemporal alignment fusion to generate a power restoration situational awareness matrix.

[0067] The edge computing gateway broadcasts a collaborative acquisition start command to the multi-view collaborative group. The command includes a acquisition window length of 400 milliseconds and specifies the acquisition time slot for each collaborative node.

[0068] This embodiment employs a dynamic time slot allocation mechanism: the gateway dynamically adjusts the time slot length based on the historical data reported by each cooperating node regarding latency statistics. Nodes with historically lower reported latency are allocated shorter time slots (30 milliseconds), while nodes with higher latency are allocated longer time slots (50 milliseconds). The time slot allocation results are broadcast, and each cooperating node starts data collection according to the specified time slot after receiving the allocation table.

[0069] During the data acquisition process, each collaborative node performs the following operations: Electrical quantity acquisition: The node continuously acquires three-phase voltage and three-phase current waveforms for two power frequency cycles (40 milliseconds) at a sampling rate of 25.6kHz, calculates parameters such as the fundamental effective value, phase difference, 2nd to 31st harmonic content, voltage sag depth, and current surge, forming a 16-dimensional electrical quantity feature vector. Environmental image acquisition: The node controls a pan-tilt-zoom camera to capture a set of images, including a wide-angle panoramic image (resolution 2592×1944) and a zoom detail image (resolution 1920×1080). For nodes equipped with infrared thermal imagers, a thermal imaging temperature matrix (64×64 pixels) is also acquired simultaneously. After acquisition, each node immediately compresses and packages the data and sends it to the edge computing gateway via the Wi-Fi Mesh network. After receiving the data from all nodes, the gateway performs spatiotemporal alignment and fusion.

[0070] The specific operations of spatiotemporal alignment and fusion include: Time alignment: The gateway uses its own system clock as a reference and calibrates the timestamp of each data packet using the Network Time Protocol (NAT). For data with a time deviation exceeding 10 milliseconds, an interpolation algorithm is used for correction. Spatial registration: The gateway maps the data to a pre-built feeder geographic information system model based on the latitude and longitude coordinates of each cooperating node, determining the line segment number corresponding to each data point. Feature extraction and fusion: For image data, the gateway calls a locally deployed lightweight convolutional neural network model (MobileNet-V3) to extract image features. This model has been accelerated by TensorRT on the edge computing platform, with a single image feature extraction time of less than 15 milliseconds. The model outputs a 512-dimensional image feature vector. Matrix construction: The 16-dimensional electrical quantity features of each line segment are concatenated with the 512-dimensional image features to form a 528-dimensional feature vector. If a line segment does not have a cooperating node deployed, its feature vector is obtained through weighted interpolation of the features of adjacent segments. Ultimately, this embodiment covers a total of 22 line sections, forming a 528×22 power restoration situational awareness matrix.

[0071] Step 4: The edge computing gateway runs the power restoration decision model, outputs and executes the power restoration control strategy.

[0072] A pruned and quantized lightweight graph neural network model is deployed locally on the edge computing gateway. This model uses the same graph convolutional network architecture as in Example 1, but to adapt to the computing power of edge devices, the number of graph convolutional layers is reduced to 2, and the hidden layer dimension is reduced to 32. The model file size is approximately 2.3MB, and the single inference time on an ARM processor is less than 50 milliseconds.

[0073] The model takes a power restoration situational awareness matrix and an adjacency matrix as input, and outputs a probability distribution of four power restoration control strategies for each of the 22 line segments. The gateway selects the strategy with the highest probability for each segment as the final strategy and generates a power restoration operation sequence.

[0074] The power restoration control strategy in this embodiment differs from that in Embodiment 1, employing a distributed execution architecture: For the "immediate power restoration" strategy: the gateway directly sends a closing command to the smart switches in the section, using Goose messages to ensure a transmission delay of less than 4 milliseconds. For the "delayed power restoration" strategy: the gateway sets a local timer, with the delay time dynamically adjusted based on the confidence level of the model output. A delay of 3 power frequency cycles (60 milliseconds) is applied for a confidence level higher than 0.9; a delay of 5 power frequency cycles (100 milliseconds) is applied for a confidence level between 0.7 and 0.9; and a delay of 10 power frequency cycles (200 milliseconds) is applied for a confidence level lower than 0.7. For the "grouped closing" strategy: the gateway groups the switches in the same section and sends closing commands sequentially at 200 millisecond intervals according to the group's order to reduce inrush current impact during closing. For the "manual intervention maintenance" strategy: the gateway generates a detailed alarm report, including the abnormal section number, abnormal feature data, and on-site image thumbnail, and pushes it to the mobile application of the maintenance personnel through the 5G network, while simultaneously blocking the remote closing circuit of the section.

[0075] During execution, the gateway monitors the action feedback signals of each switch in real time. For commands that fail to execute, the gateway automatically retry (up to 3 times) and reports the execution results to the dispatch center.

[0076] This embodiment achieves low-latency response for power restoration control by offloading decision-making calculations to the edge side, while reducing the computational burden on the main station system. It is suitable for application scenarios with limited communication conditions or high real-time requirements.

[0077] Example 3: This example provides a multi-perspective collaborative control method for power restoration. This example focuses on a detailed description of the training process, verification method, and model update mechanism of the power restoration decision model, ensuring that those skilled in the art can understand and implement the complete lifecycle management of the model.

[0078] I. The power restoration decision model adopts a graph neural network architecture, specifically a graph attention network. The structure of this model is as follows: Input layer: Receives the complex power situation awareness matrix M ∈ ℝ^{D×K}, where D is the feature dimension of each line segment (D=1024 in this embodiment), and K is the number of line segments (K=30 in this embodiment, taking a typical distribution network as an example). Simultaneously, it receives the adjacency matrix A ∈ ℝ^{K×K}, representing the electrical connection relationships between line segments.

[0079] Graph Attention Layer: The model contains two graph attention layers, each using a multi-head attention mechanism (8 heads). Each attention head is calculated as follows: for node i and its neighbor node j, the attention coefficient e_{ij} = LeakyReLU( a^T[W h_i || W h_j] ); where h_i and h_j are node features, W is the weight matrix, a is the learning vector of the attention mechanism, and || represents the concatenation operation.

[0080] The normalized attention coefficient α_{ij} = softmax_j(e_{ij}) The updated feature of node i is h'i = σ( Σ{j∈N(i)} α_{ij} W h_j ) Multi-head attention mechanisms concatenate or average the outputs of multiple heads.

[0081] Output layer: After two graph attention layers, the node feature dimension is transformed to 32-dimensional. Finally, a fully connected layer maps the node features to a 4-dimensional output, and after passing through the Softmax activation function, the probability distribution of each node corresponding to the four types of complex power control strategies is obtained.

[0082] The training dataset for the model is constructed as follows: First, records of 1,247 power restoration operations that occurred in the past three years were extracted from the historical power restoration case database. Each record includes: electrical waveform data, environmental image data, and line topology information collected by each monitoring terminal before power restoration, as well as the actual execution result of the power restoration (whether it was successful, whether a secondary fault occurred, and the actual line status confirmed on-site by maintenance personnel).

[0083] For each power restoration case, following the process described in Example 1, a power restoration situation awareness matrix M_t is constructed, and an adjacency matrix A_t is also built. Based on the actual execution results, each line segment is labeled. The labeling rules are as follows: If the section operates normally after power restoration without any abnormal alarms, it is marked as "Power Restoration Immediately"; if the section experiences transient disturbances after power restoration but subsequently returns to normal, it is marked as "Delayed Power Restoration"; if the section is reconnected in coordination with other sections during power restoration, it is marked as "Separate and Combined Reconnection"; if the section has been identified as a faulty section by maintenance personnel before power restoration, or trips immediately after power restoration, it is marked as "Manual Intervention Maintenance".

[0084] After data cleaning and labeling, a total of 892 valid samples were obtained. Each sample contains a complex situational awareness matrix, an adjacency matrix, and a label vector (length K, each element being an integer from 0 to 3).

[0085] The sample set was divided into a training set (624 samples), a validation set (179 samples), and a test set (89 samples) in a ratio of 7:2:1.

[0086] Model training was implemented using the PyTorch framework on a server equipped with an NVIDIA Tesla T4 GPU. The training parameters were set as follows: Optimizer: Adam, initial learning rate 0.001; Loss function: cross-entropy loss function, which is the average of the summation of the predicted loss of each node; Batch size: 16; Maximum number of training rounds: 200 rounds, with an early stopping mechanism, which stops training when the validation set loss no longer decreases for 10 consecutive rounds.

[0087] During training, each sample's input is M_t and A_t, and the output is the predicted probability distribution for each node. The cross-entropy loss between predictions and labels is calculated, and backpropagation updates the model parameters. After 160 rounds of training, the validation set loss converges to 0.087, at which point training stops. Evaluation results on the test set are as follows: overall classification accuracy is 91.3%, with 94.2% accuracy for "immediate power restoration," 87.6% for "delayed power restoration," 88.1% for "separate switch combinations," and 92.7% for "manual intervention maintenance."

[0088] II. To verify the effectiveness of the model, a comparative experiment was designed in this embodiment. Using the same test set, the graph attention network model of this invention was compared with the following three benchmark models: Benchmark Model 1: A fully connected neural network using only electrical quantity features (excluding environmental image features), with the input dimension being the electrical quantity feature dimension. Benchmark Model 2: A convolutional neural network using only environmental image features (excluding electrical quantity features), with the input being image features. Benchmark Model 3: A random forest classifier using multimodal features but not considering topological structure, classifying each segment independently.

[0089] The results of the comparative experiment are shown in the table below:

[0090] Experimental results show that the model of this invention, by fusing multimodal features and using graph neural networks to capture the topological relationships between line segments, significantly outperforms the benchmark model in all indicators, verifying the effectiveness of the technical solution.

[0091] III. In order to adapt to the changes in the power grid operation status and the emergence of new fault types, this embodiment establishes a model incremental update mechanism. The specific process is as follows: (1) Data accumulation: During the operation, the system continuously records the complete data of each power restoration operation, including the power restoration situation awareness matrix used in decision-making, the strategy output by the model, the actual execution results, and the real line status reported by the maintenance personnel. These data are added to the incremental sample library every week. (2) Trigger condition judgment: The system checks the number of samples in the incremental sample library every week. When the number of new samples exceeds 50, the incremental training process is triggered. At the same time, when the prediction accuracy of the model is less than 85% in the past month, retraining is also automatically triggered. (3) Incremental training: The elastic weight consolidation method is used for incremental training, which learns the features of new samples while retaining the original knowledge. Specifically, a regularization term is added to the loss function to constrain the change range of important parameters. The identification of important parameters is based on the Fisher information matrix. (4) Model verification and deployment: The incrementally trained model is evaluated on the validation set. If the accuracy is not less than 95% of the original model, the new model is deployed to the production environment. The deployment adopts a blue-green deployment strategy, which means running the old and new models simultaneously for one cycle, comparing the prediction results, and completely switching to the new model after confirming that there are no anomalies.

[0092] Through the aforementioned model construction, training, verification, and updating mechanisms, this embodiment ensures that the power restoration decision model can maintain high accuracy and adapt to changes in the power grid's operating status.

[0093] Example 4: Figure 1 A schematic diagram of the multi-perspective collaborative control system for power restoration, as presented in this invention, is provided. This system is deployed in the distribution network master station system or edge computing gateway, and is used to execute power restoration control of non-faulty sections after fault isolation. The system adopts a modular architecture design, with modules interacting with each other through standard interfaces to collaboratively complete the entire process control from obtaining terminal status to executing the power restoration strategy. The system consists of the following nine functional modules:

[0094] II. Detailed Description of Functions of Each Module Module 100: Status Acquisition Module; responsible for sending status query commands to all monitoring terminals deployed on the power line after the power restoration command is issued, and receiving, parsing, and storing the initial status data returned by each terminal. This module is the data foundation for the system to select collaborative nodes.

[0095] Specific functional implementation: Command Issuance: Upon receiving a power restoration command from the dispatch center or manually, this module broadcasts a status query message to all registered monitoring terminals on the target feeder via the 104 protocol, MQTT protocol, or Modbus protocol. The message includes a query timestamp and the expected response time.

[0096] Data Reception and Parsing: Receive initial status data packets returned by each terminal. The data packets are in JSON or binary protocol format. After parsing, extract the following fields: terminal identifier (including device type and installation location code); remaining battery percentage; communication signal-to-noise ratio (unit: dB); sensing module status (camera PTZ self-test, sensor calibration status, remaining storage space); topology association information (tower number, feeder branch level, upstream and downstream adjacent terminal identifiers); and geographic location coordinates (latitude and longitude). Data storage: The parsed data is stored in a real-time database to create a terminal status dataset for use by the collaborative construction module.

[0097] Module 200: Collaborative Construction Module; Based on the initial status data provided by the status acquisition module, collaborative nodes with communication and sensing capabilities are selected from all monitoring terminals, and multi-view collaborative groups are constructed according to the line topology relationship to lay the organizational foundation for collaborative data acquisition.

[0098] Specific functional implementation: Collaborative Node Selection: Execute a multi-condition collaborative node selection algorithm, which includes: extracting the remaining battery power E_i, communication signal-to-noise ratio SNR_i, sensing module status flag F_i, and topological correlation level L_i for each terminal from the terminal status dataset; setting the selection thresholds: remaining battery power threshold E_th (default 20%) and communication signal-to-noise ratio threshold SNR_th (default 15dB); selecting terminals that simultaneously satisfy E_i ≥ E_th, SNR_i ≥ SNR_th, F_i = 1 (sensing module is normal), and L_i ≤ 3 (main trunk and first and second level branches) as collaborative nodes.

[0099] Geographical distribution optimization screening (optional): In edge computing scenarios, the K-means clustering algorithm is used to optimize the geographical distribution uniformity of the initially screened terminals to ensure that each line segment group has at least one collaborative node.

[0100] Collaborative Group Construction: Based on the topological association information of collaborative nodes, a multi-view collaborative group is established. The construction rules are as follows: the collaborative nodes on the backbone are used as the backbone; each backbone node and its downstream branches are divided into a subgroup; a self-organizing communication network (LoRa Mesh or Wi-Fi Mesh) is established within the subgroup; the backbone node simultaneously undertakes the data aggregation and relay functions within the subgroup.

[0101] Collaborative group information storage: The completed collaborative group information (group number, member node list, communication link information, topology relationship) is stored in an in-memory database for subsequent modules to call.

[0102] Module 300: Time Slot Allocation Unit; responsible for allocating data acquisition time slots to each collaborative node in the multi-view collaborative group, ensuring that multiple nodes acquire data sequentially within a preset time window, avoiding channel conflicts, and achieving the temporal order of collaborative acquisition.

[0103] Specific functional implementation: Time slot length calculation: The time slot length is determined based on the data type of each coordinating node: 20 millisecond time slot is allocated for uploading only electrical quantity parameters; 80 millisecond time slot is allocated for uploading electrical quantity parameters plus visible light images; 100 millisecond time slot is allocated for uploading electrical quantity parameters plus infrared thermal images. Upload electrical quantity parameters and video stream keyframes: allocate a 120-millisecond time slot.

[0104] Time slot allocation algorithm: Static allocation mode: Time slots are allocated sequentially according to the order of the cooperating nodes in the topology (from upstream to downstream), with a 5-millisecond protection interval between time slots; Dynamic allocation mode (edge ​​computing scenario): The time slot length is dynamically adjusted based on the historical data reporting latency statistics of each cooperating node, and nodes with larger latency are allocated longer time slots.

[0105] Time slot label distribution: Generate a time slot allocation table (including node ID, time slot start time, and time slot length), and broadcast it to all cooperative nodes through the communication link within the cooperative group.

[0106] Data Acquisition Window Management: Set the preset time window length (default 500 milliseconds) to ensure that all collaborative nodes complete data acquisition and uploading within this time window.

[0107] Module 400: Collaborative Acquisition Module; according to the scheduling instructions of the time slot allocation unit, it controls each collaborative node to perform data acquisition tasks within the specified time slot. The acquired content includes electrical quantity parameters and environmental image data, and the acquired data is uploaded to the system.

[0108] Specific functional implementation: Acquisition command issuance: Send acquisition start command to each node in the collaborative group. The command includes parameters such as time slot allocation table, sampling rate configuration, and image resolution configuration.

[0109] Electrical quantity acquisition and control: The control node continuously acquires the three-phase voltage and three-phase current waveforms for one or two power frequency cycles (20-40 milliseconds) at a sampling rate of 12.8kHz or 25.6kHz. After acquisition, the node's built-in edge computing unit calculates the following parameters: fundamental effective value, phase angle, 2nd to 31st harmonic content, three-phase unbalance, voltage sag depth, current surge, etc., forming an electrical quantity feature vector with a dimension range of 8 to 32.

[0110] Environmental image acquisition and control: The high-definition camera of the control node captures visible light images with a resolution supporting 1920×1080 to 2592×1944; controls the infrared thermal imager to acquire temperature matrix data with a resolution supporting 32×32 to 64×64 pixels; controls the pan-tilt unit to perform preset position rotation and capture multi-angle images.

[0111] Data Upload Management: After data collection, each node compresses and packages the electrical quantity feature vectors and environmental image data, and uploads them to the system via a wireless ad hoc network or 4G / 5G network. The upload process proceeds according to the time slot allocation order to ensure no channel conflicts.

[0112] Module 500: Spatiotemporal alignment and fusion module; performs time synchronization, spatial registration, and multimodal feature fusion on multi-source data uploaded by the collaborative acquisition module to eliminate data fragmentation and provide aligned high-quality data for generating the power situation awareness matrix.

[0113] Specific functional implementation: Time alignment: UTC time provided by GPS timing module or precise time protocol is used as common time reference; a timestamp accurate to 1 microsecond is added to each data packet; for data with time deviation exceeding 10 milliseconds, a linear interpolation algorithm is used for correction to ensure that all data reflect the line status at the same time segment.

[0114] Spatial registration: Based on the tower number and latitude and longitude coordinates of the installation location of each collaborative node, the data is mapped to the pre-built feeder geographic information system model; the line segment number corresponding to each data (the line segment between two adjacent towers) is determined; and the data of two collaborative nodes belonging to the two ends of the same line segment are associated and bound.

[0115] Multimodal Feature Fusion: Image Feature Extraction: A pre-trained convolutional neural network is used to extract features from image data. The network architecture includes: Visible light images: ResNet-50 or MobileNet-V3 is used to output 1024-dimensional or 512-dimensional image feature vectors; Infrared thermal images: A lightweight thermal imaging feature extraction network is used to output a 512-dimensional temperature distribution feature vector.

[0116] Feature concatenation: The electrical quantity feature vector of the node is vertically concatenated with the image feature vector to form the multimodal feature vector of the node; Segment feature aggregation: The multimodal feature vectors of the nodes at both ends of the same line segment are horizontally concatenated or weighted and fused to form the joint feature vector of the segment; if a segment has only one cooperating node, the features at the other end are filled with zeros.

[0117] Module 600: Matrix generation module; arranges the joint feature vectors of each section output by the spatiotemporal alignment and fusion module according to the line topology order to construct a unified power restoration situation awareness matrix, which serves as the standardized input for the power restoration decision model.

[0118] Specific functional implementation: Feature Dimension Normalization: The joint feature vectors of all line segments are checked for dimension to ensure that the dimension of each segment vector is consistent (this system sets D=1024 dimensions). For vectors with insufficient dimension, zero-padding or mean-padding is used to make up the difference.

[0119] Topology order arrangement: Based on the feeder topology, the feature vectors of each segment are arranged into columns of a matrix in order from the power supply side to the load side. Each column of the matrix corresponds to a line segment, and each row corresponds to a feature dimension.

[0120] Adjacency Matrix Construction: An adjacency matrix A ∈ ℝ^{K×K} is constructed synchronously, where K is the number of line segments. A[i][j] = 1 when two line segments are directly connected topologically, and 0 otherwise. The adjacency matrix is ​​used in graph neural network models.

[0121] Matrix storage and output: The generated complex electrical situational awareness matrix M (dimension D×K) and adjacency matrix A (dimension K×K) are stored in memory for use by the decision model module. Simultaneously, the generation timestamps of the matrices and the corresponding feeder identifiers are recorded.

[0122] Module 700: Decision Model Module; Loads the pre-trained power restoration decision model, inputs the power restoration situation awareness matrix and adjacency matrix output by the matrix generation module into the model, performs inference calculations, and outputs the probability distribution of power restoration control strategies for each line segment.

[0123] Specific functional implementation: Model Loading: Loads pre-trained graph neural network models. The model architecture supports graph convolutional networks or graph attention networks. The model file format is ONNX or PyTorch. During deployment, CPU inference or GPU inference is selected based on the hardware platform.

[0124] Model Inference: Input: Complex situational awareness matrix M (D×K), adjacency matrix A (K×K); Computational flow: First layer: Graph convolution / graph attention: Input feature dimension D, output dimension 64; Second layer: Graph convolution / graph attention: Input dimension 64, output dimension 32; Fully connected layer: Input dimension 32, output dimension 4; Softmax activation: Outputs the probability distribution of each node corresponding to the 4 policies.

[0125] Output: Probability matrix P ∈ ℝ^{4×K}, where P[c][i] represents the probability of adopting the c-th strategy in the i-th segment.

[0126] Strategy Selection: For each section, the strategy with the highest probability is selected as the power restoration control strategy for that section. The four strategies are defined as follows: Strategy 0: Immediate power restoration; Strategy 1: Delayed power restoration; Strategy 2: Combined switching; Strategy 3: Manual intervention for maintenance. Confidence Calculation: Record the probability value of the selected strategy for each segment as the confidence level for reference by the strategy execution module.

[0127] Module 800: Strategy Execution Module; Based on the power restoration control strategy output by the decision model module, it generates a specific power restoration operation sequence and sends the instructions to the intelligent switchgear via the communication interface to execute the power restoration operation. Simultaneously, this module is responsible for monitoring the execution results and handling any anomalies.

[0128] Specific functional implementation: Operation sequence generation: Traverse all line sections and collect the control strategy for each section; arrange the operation sequence according to the line topology; group the sections marked as "separate grouping" into the same closing group, and the order within the group is determined according to the upstream and downstream relationship of the topology; set the delay parameter for the "delayed power restoration" strategy (dynamically adjusted according to confidence level: 60 milliseconds for confidence level > 0.9, 100 milliseconds for 0.7-0.9, and 200 milliseconds for < 0.7); mark the sections of the "manual intervention maintenance" strategy as prohibited from operation and generate alarm information.

[0129] Command issuance: Centralized architecture: The closing command is sent to the remote terminal unit of the corresponding smart switch via the 104 protocol or IEC 61850 Goose message; Edge computing architecture: The command is sent to the smart switch controller via Modbus TCP or local bus; For grouped switches, the command is sent sequentially at 200 millisecond intervals according to the group order.

[0130] Execution monitoring and feedback: Monitor the action feedback signals of each switch (success / failure of closing, position status); for commands that fail to execute, automatically retry, up to 3 times, with a retry interval of 100 milliseconds; record the execution results (success, failure, timeout) and report to the dispatch center.

[0131] Alarm push: For the "manual intervention maintenance" strategy, a detailed alarm report is generated, including the abnormal section number, abnormal feature data, and on-site image thumbnail, and sent to the handheld terminal of the operation and maintenance personnel through the message push service.

[0132] Module 900: Model Management Module; responsible for the complete lifecycle management of the power restoration decision model, including training data accumulation, model training, model validation, model updates and deployment, to ensure that the model continues to maintain high accuracy and adapts to changes in the power grid's operating status.

[0133] Specific functional implementation: Training data accumulation: Continuously record complete data for each power restoration operation, including: electrical waveform data collected by each terminal before power restoration, environmental image data, line topology information, strategies output by the decision model, actual execution results, and real line status reported by maintenance personnel; add new data to the incremental sample library every week and label the samples (label with 4 types of strategy tags based on actual execution results).

[0134] Model training: Model training is triggered when the number of new samples added to the incremental sample library exceeds 50, or when the model's prediction accuracy is below 85% in the past month. Incremental training is performed using an elastic weight consolidation method, and a regularization term is added to the loss function to constrain changes in important parameters. Training is performed using the PyTorch framework on a GPU server, with a maximum of 200 training rounds and an early stopping mechanism to prevent overfitting.

[0135] Model validation: The trained model is evaluated on the validation set, and accuracy, precision, recall, and F1 score are calculated. The performance of the new and old models is compared. If the accuracy of the new model is not less than 95% of that of the old model, the model passes the validation.

[0136] Model deployment: A blue-green deployment strategy is adopted, running the old and new models simultaneously for one cycle (7 days); the prediction results are compared, and the new model is fully switched after confirming that there are no anomalies; the model files are distributed to the main station system or edge computing gateway through a secure transmission protocol.

[0137] Model version management: Maintains model version history, supports model rollback function, and can quickly switch back to a stable version when a new model has an anomaly.

[0138] III. The data flow and collaboration relationships between the modules of this system are as follows: Status Acquisition Module (100) acquires the initial status data of the terminal after the power restoration command is triggered and passes it to the Collaborative Construction Module (200). Collaborative Construction Module (200) filters collaborative nodes and constructs collaborative groups, and passes the collaborative group information to the time slot allocation unit (300). Time Slot Allocation Unit (300) allocates acquisition time slots and passes the scheduling command to the Collaborative Acquisition Module (400). Collaborative Acquisition Module (400) performs data acquisition and passes the electrical quantity parameters and environmental image data to the Spatiotemporal Alignment and Fusion Module (500). Spatiotemporal Alignment and Fusion Module (500) performs spatiotemporal alignment and feature fusion and passes the joint feature vector of each segment to the Matrix Generation Module (600). Matrix Generation Module (600) constructs the power restoration situation awareness matrix and passes it to the Decision Model Module (700). Decision Model Module (700) performs model inference, outputs the probability distribution of the power restoration control strategy, and passes it to the strategy execution module (800). The strategy execution module (800) generates an operation sequence and issues instructions to complete the power restoration operation. At the same time, it feeds back the execution results to the model management module (900) for data accumulation. The model management module (900) periodically performs model training and updates, and deploys the updated model files to the decision model module (700).

[0139] It should be noted that this invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting various hardware environments and usage requirements.

[0140] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A multi-perspective collaborative control method for power restoration, characterized in that, include: The initial status data of multiple monitoring terminals deployed on power lines after the power restoration command is issued is obtained. The initial status data includes the remaining power of each monitoring terminal, the communication signal-to-noise ratio, the status of the sensing module, and the topological correlation of the line segment in which it is located. Based on the initial state data, collaborative nodes with communication and sensing capabilities are selected from the multiple monitoring terminals to construct a multi-view collaborative group. Specifically, this includes: identifying monitoring terminals with remaining power greater than a first threshold, communication signal-to-noise ratio greater than a second threshold, normal sensing module status, and topological correlation meeting preset hierarchical conditions as collaborative nodes; and establishing a multi-view collaborative group with data relay and collaborative sensing functions based on the physical topology of the line segment where the collaborative node is located. Using the collaborative nodes within the multi-view collaborative group, electrical quantity parameters and environmental image data of their corresponding line sections are collected in a time-division multiplexing manner. The electrical quantity parameters and environmental image data are then spatiotemporally aligned and fused to generate a power restoration situation awareness matrix. Specifically, this includes: assigning time slot labels to each collaborative node within the multi-view collaborative group, enabling different collaborative nodes to sequentially collect electrical quantity parameters and environmental image data within a preset time window; aligning the timestamps of the data collected by each collaborative node using a common time reference; performing spatial registration using the geographical location identifier of the line section as a spatial anchor point; and stitching the aligned and registered electrical quantity parameters and environmental image data according to preset feature dimensions to form a power restoration situation awareness matrix. The power restoration situation awareness matrix is ​​input into a preset power restoration decision model, and power restoration control strategies for different line sections are output. The power restoration control strategies include immediate power restoration, delayed power restoration, combined switching, and manual intervention for maintenance.

2. The multi-view collaborative control method for power restoration according to claim 1, characterized in that, The method of collecting electrical quantity parameters and environmental image data of the corresponding line section in a time-division multiplexing manner, and performing spatiotemporal alignment and fusion of the electrical quantity parameters and environmental image data to generate a power restoration situation awareness matrix, further includes: The time slot length is determined based on the data type of each collaborative node. Nodes that upload only electrical quantity parameters are allocated the first time slot length, nodes that upload electrical quantity parameters plus visible light images are allocated the second time slot length, nodes that upload electrical quantity parameters plus infrared thermal images are allocated the third time slot length, and nodes that upload electrical quantity parameters plus video stream keyframes are allocated the fourth time slot length. A protection interval is set between each time slot. Based on the historical data reporting delay statistics of each collaborative node, the time slot length is dynamically adjusted. Nodes with reporting delays greater than the preset delay threshold are allocated longer time slots, while nodes with reporting delays less than the preset delay threshold are allocated shorter time slots. The dynamically adjusted time slot allocation table is broadcast to each collaborative node through the communication link within the collaborative group, and each collaborative node performs the acquisition operation within its allocated time slot.

3. The multi-view collaborative control method for power restoration according to claim 1, characterized in that, The step of performing spatiotemporal alignment and fusion of the electrical quantity parameters and environmental image data to generate a power restoration situation awareness matrix further includes: Using UTC time provided by the GPS timing module or Precision Time Protocol as the common time reference, a timestamp accurate to the microsecond level is added to each data packet. For data with time deviation exceeding the preset deviation threshold, a linear interpolation algorithm is used for time correction. Based on the tower number and latitude and longitude coordinates of the installation location of each collaborative node, the data is mapped to a pre-built feeder geographic information system model to determine the line segment number corresponding to each data point. The line segment is the line segment between two adjacent towers. For data belonging to two collaborative nodes at both ends of the same line segment, associate and bind them; For the collected environmental image data, a pre-trained convolutional neural network is called to extract image feature vectors. The convolutional neural network has been pre-trained on a typical power line scene image dataset, which includes tree obstacles, bird nests, insulator damage, and foreign objects hanging. The electrical quantity parameters of each collaborative node are formed into an electrical quantity feature vector. The electrical quantity feature vector is then vertically concatenated with the image feature vector to form the multimodal feature vector of that node. The multimodal feature vectors of the nodes at both ends of the same line segment are horizontally concatenated or weighted and fused to form the joint feature vector of the line segment. If the line segment has only one cooperating node, the feature at the other end is filled with zeros. Arrange the joint feature vectors of all line segments in the order of line topology to construct a power situation awareness matrix. Each column of the matrix corresponds to a line segment, and each row corresponds to a feature dimension. At the same time, construct an adjacency matrix. In the adjacency matrix, the value is 1 when two line segments are directly connected in topology, and 0 otherwise.

4. The multi-view collaborative control method for power restoration according to claim 1, characterized in that, The power restoration decision model is a classification model based on a graph neural network. The step of inputting the power restoration situation awareness matrix into the preset power restoration decision model and outputting power restoration control strategies for different line sections specifically includes: Construct a power grid topology graph, where the node set corresponds to the line segment, the edge set corresponds to the electrical connection relationship between the line segments, and the adjacency matrix is ​​constructed based on the power grid topology graph; The complex situational awareness matrix is ​​used as the input feature matrix of the graph neural network, and the adjacency matrix is ​​used as the structural information input of the graph neural network. The input features are transformed layer by layer through at least two graph convolutional layers or graph attention layers. The calculation method of each graph convolutional layer is H^(l+1) = σ( D^(-1 / 2) AD^(-1 / 2) H^(l) W^(l) ), where H^(l) is the node feature matrix of the l-th layer, H^(0) is the input complex electrical situational awareness matrix, A is the adjacency matrix after adding self-loops, D is the degree matrix, W^(l) is the learnable weight matrix of the l-th layer, and σ is the activation function. The node features are mapped to a four-dimensional output through the last fully connected layer, and the predicted probability of the four types of power restoration control strategies for each line segment is obtained after passing through the Softmax activation function. The strategy with the highest predicted probability for each line segment is selected as the power restoration control strategy for that segment.

5. The multi-view collaborative control method for power restoration according to claim 1, characterized in that, After outputting the power restoration control strategy for different line sections, it also includes: Traverse all line segments, collect the power restoration control strategy for each segment, and arrange them according to the line topology to generate a power restoration operation sequence; Line sections marked as separate closing groups are grouped into the same closing group, and the closing sequence within the group is determined according to the upstream and downstream topology. Set delay parameters for line sections marked as delayed power restoration, and dynamically adjust the delay time according to the prediction probability confidence level output by the model. Set the first delay time when the confidence level is greater than the first confidence level threshold, set the second delay time when the confidence level is less than the first confidence level threshold but greater than the second confidence level threshold, and set the third delay time when the confidence level is less than the second confidence level threshold. The line section marked for manual intervention maintenance is marked as prohibited from operation, and an alarm report is generated. The alarm report includes the abnormal section number, abnormal feature data and on-site image thumbnail. The closing command is sent to the intelligent switching equipment of the corresponding line section through the communication protocol. For the group closing strategy, the closing command is sent in sequence according to the order within the group, and the interval between adjacent commands is a preset interval time. Monitor the action feedback signals of each switching device, retry commands that fail to execute, record the execution results and report them to the dispatch center.

6. The multi-view collaborative control method for power restoration according to claim 1, characterized in that, The method is applied to an edge computing gateway, which is deployed at the beginning of a feeder and connected to monitoring terminals within its jurisdiction via a local communication network. The method further includes: After receiving the power restoration command from the dispatch center, the edge computing gateway sends a status query command to the connected monitoring terminal and receives and stores the initial status data returned by each terminal. The edge computing gateway performs collaborative node screening and builds multi-view collaborative groups based on the screened collaborative nodes. When building multi-view collaborative groups, the K-means clustering algorithm is used to divide the terminal into multiple clusters according to the geographical coordinates of the terminal. In each cluster, the terminal that is closest to the cluster center and has the highest perception capability score is selected as the collaborative node of that cluster. The edge computing gateway assigns time slot labels to each collaborative node in the collaborative group and issues collection commands through the local communication network. The edge computing gateway receives electrical quantity parameters and environmental image data uploaded by each collaborative node, performs spatiotemporal alignment and fusion locally, and generates a power restoration situation awareness matrix. The edge computing gateway loads a locally deployed lightweight graph neural network model, inputs the power restoration situation awareness matrix into the model for inference, and outputs a power restoration control strategy. The edge computing gateway generates an operation sequence based on the power restoration control strategy, sends a closing command to the intelligent switching device through the local communication network, and reports the execution result to the dispatch center.

7. The multi-view collaborative control method for power restoration according to claim 1, characterized in that, The method also includes training and updating the power restoration decision model, specifically including: Power restoration operation records are extracted from the historical power restoration case database. Each power restoration operation record includes electrical waveform data, environmental image data, line topology information, actual execution results, and the actual line status reported by maintenance personnel before power restoration. Construct the power restoration situation awareness matrix and adjacency matrix for each power restoration operation according to the method described in claim 1; Based on the actual execution results and the real line status reported by the operation and maintenance personnel, a power restoration control strategy label is marked for each line section. The power restoration control strategy label includes immediate power restoration, delayed power restoration, combination of switches and manual intervention maintenance. The labeled sample set is divided into a training set, a validation set, and a test set. The graph neural network model is trained using the training set, hyperparameters are tuned using the validation set, and the model performance is evaluated using the test set. When the number of new samples added to the incremental sample library exceeds the preset threshold, incremental training is triggered. The elastic weight consolidation method is used for incremental training, and a regularization term is added to the loss function to constrain the change of model parameters. The trained model is evaluated on the validation set. If the accuracy of the new model is not lower than the preset percentage threshold of the accuracy of the old model, the new model is deployed to the production environment using a blue-green deployment strategy. The new and old models run in parallel for a preset period of time before a complete switch is performed.

8. A multi-view cooperative control system for power restoration, used to implement the multi-view cooperative control method for power restoration as described in any one of claims 1-7, characterized in that, include: The status acquisition module is used to acquire the initial status data of multiple monitoring terminals deployed on the power line after the power restoration command is issued. The initial status data includes the remaining power of each monitoring terminal, the communication signal-to-noise ratio, the status of the sensing module, and the topological correlation of the line segment where it is located. The collaborative construction module is used to select collaborative nodes with communication and sensing capabilities from the multiple monitoring terminals based on the initial state data, and construct a multi-view collaborative group. Specifically, it includes: determining the monitoring terminals with remaining power greater than a first threshold, communication signal-to-noise ratio greater than a second threshold, normal sensing module status, and topological correlation degree meeting the preset hierarchical conditions as collaborative nodes, and establishing a multi-view collaborative group with data relay and collaborative sensing functions based on the physical topology relationship of the line segment where the collaborative node is located. The time slot allocation unit is used to allocate time slot labels to each collaborative node in the multi-view collaborative group, so that different collaborative nodes collect data sequentially within a preset time window. The collaborative acquisition module is used to control each collaborative node to acquire electrical parameters and environmental image data of its corresponding line section within the allocated time slot; The spatiotemporal alignment and fusion module is used to align the timestamps of the data collected by each collaborative node with a common time reference, and to perform spatial registration with the geographical location identifier of the line section as the spatial anchor point. The aligned and registered electrical parameters and environmental image data are then used for feature extraction and stitching. The matrix generation module is used to arrange the feature vectors of each segment processed by the spatiotemporal alignment and fusion module according to the line topology order to construct the power restoration situation awareness matrix. The decision model module is used to load the pre-trained power restoration decision model, input the power restoration situation awareness matrix into the power restoration decision model, and output the power restoration control strategy for different line sections. The power restoration control strategy includes immediate power restoration, delayed power restoration, combined switching and manual intervention maintenance. The strategy execution module is used to generate a power restoration operation sequence based on the power restoration control strategy output by the decision model module, and send the instructions to the intelligent switchgear to perform the power restoration operation through the communication interface. The model management module is used to manage the accumulation of training data, model training, model validation, model updates and deployment of the power restoration decision model.

9. A multi-view cooperative control device for power restoration, used to implement the multi-view cooperative control method for power restoration as described in any one of claims 1-7, characterized in that, include: At least one processor; At least one memory, wherein a computer program is stored; When the computer program is executed by the at least one processor, the device performs the following steps: The initial status data of multiple monitoring terminals deployed on power lines after the power restoration command is issued is obtained. The initial status data includes the remaining power of each monitoring terminal, the communication signal-to-noise ratio, the status of the sensing module, and the topological correlation of the line segment in which it is located. Based on the initial state data, collaborative nodes with communication and sensing capabilities are selected from the multiple monitoring terminals to construct a multi-view collaborative group. Specifically, this includes: identifying monitoring terminals with remaining power greater than a first threshold, communication signal-to-noise ratio greater than a second threshold, normal sensing module status, and topological correlation meeting preset hierarchical conditions as collaborative nodes; and establishing a multi-view collaborative group with data relay and collaborative sensing functions based on the physical topology of the line segment where the collaborative node is located. Using the collaborative nodes within the multi-view collaborative group, electrical quantity parameters and environmental image data of their corresponding line sections are collected in a time-division multiplexing manner. The electrical quantity parameters and environmental image data are then spatiotemporally aligned and fused to generate a power restoration situation awareness matrix. Specifically, this includes: assigning time slot labels to each collaborative node within the multi-view collaborative group, enabling different collaborative nodes to sequentially collect electrical quantity parameters and environmental image data within a preset time window; aligning the timestamps of the data collected by each collaborative node using a common time reference; performing spatial registration using the geographical location identifier of the line section as a spatial anchor point; and stitching the aligned and registered electrical quantity parameters and environmental image data according to preset feature dimensions to form a power restoration situation awareness matrix. The power restoration situation awareness matrix is ​​input into a preset power restoration decision model, and power restoration control strategies for different line sections are output. The power restoration control strategies include immediate power restoration, delayed power restoration, combined switching, and manual intervention for maintenance.

10. A computer-readable storage medium having a computer program stored thereon for implementing the multi-view cooperative control method for power restoration as described in any one of claims 1-7, characterized in that, When the computer program is executed by the processor, it performs the following steps: The initial status data of multiple monitoring terminals deployed on power lines after the power restoration command is issued is obtained. The initial status data includes the remaining power of each monitoring terminal, the communication signal-to-noise ratio, the status of the sensing module, and the topological correlation of the line segment in which it is located. Based on the initial state data, collaborative nodes with communication and sensing capabilities are selected from the multiple monitoring terminals to construct a multi-view collaborative group. Specifically, this includes: identifying monitoring terminals with remaining power greater than a first threshold, communication signal-to-noise ratio greater than a second threshold, normal sensing module status, and topological correlation meeting preset hierarchical conditions as collaborative nodes; and establishing a multi-view collaborative group with data relay and collaborative sensing functions based on the physical topology of the line segment where the collaborative node is located. Using the collaborative nodes within the multi-view collaborative group, electrical quantity parameters and environmental image data of their corresponding line sections are collected in a time-division multiplexing manner. The electrical quantity parameters and environmental image data are then spatiotemporally aligned and fused to generate a power restoration situation awareness matrix. Specifically, this includes: assigning time slot labels to each collaborative node within the multi-view collaborative group, enabling different collaborative nodes to sequentially collect electrical quantity parameters and environmental image data within a preset time window; aligning the timestamps of the data collected by each collaborative node using a common time reference; performing spatial registration using the geographical location identifier of the line section as a spatial anchor point; and stitching the aligned and registered electrical quantity parameters and environmental image data according to preset feature dimensions to form a power restoration situation awareness matrix. The power restoration situation awareness matrix is ​​input into a preset power restoration decision model, and power restoration control strategies for different line sections are output. The power restoration control strategies include immediate power restoration, delayed power restoration, combined switching, and manual intervention for maintenance.

Images