Panoramic intelligent monitoring system and monitoring method for transformer substation

Abstract

The invention discloses a substation panoramic intelligent monitoring system and a substation panoramic intelligent monitoring method, and relates to the technical field of power system monitoring, a sensing layer collects an equipment state, a topology state and digital twin simulation data, and an edge calculation layer performs secondary weight reduction on a lightweight anomaly detection model based on the data, namely, adapts to substation topology / equipment difference, and also performs secondary weight reduction on the lightweight anomaly detection model based on the data. The model complexity is reduced, real-time anomaly monitoring of an edge end is realized, and a short-term result is output; the cloud computing layer depends on AI and a digital twinborn model and fuses multi-source data to perform state prediction to generate a long-term result, and the execution control layer formulates a decision scheme according to the result. According to the scheme, the limitation of an existing fixed framework is broken through, the problems of adaptability and edge real-time response are solved, and meanwhile operation and maintenance mode transformation of the transformer substation from passive fault response to active risk prevention is supported.

Description

Technical Field

[0001] This invention relates to the field of power system monitoring technology, and in particular to a panoramic intelligent monitoring system and method for substations. Background Technology

[0002] As power systems accelerate their upgrade towards intelligence and large-scale operation, substation equipment configurations are becoming increasingly diverse and complex. Their topology is frequently and dynamically adjusted in response to daily maintenance operations such as switching operations and line commissioning / decommissioning. Meanwhile, extreme natural environments such as strong light, rain, and fog can easily interfere with the accuracy of on-site monitoring data. Traditional substation monitoring methods are gradually becoming insufficient to meet actual safety and efficiency requirements when dealing with such highly dynamic and complex maintenance scenarios.

[0003] To alleviate this problem, early substation automation monitoring solutions were gradually applied. These solutions achieved a certain degree of automated operation and maintenance assistance by building monitoring models. However, due to limitations in technical design, their core relied on a fixed model architecture for data processing and analysis, without fully considering the dynamic characteristics of substations in actual operation.

[0004] This fixed model architecture has several drawbacks. First, it struggles to adapt to the topological heterogeneity and individual equipment differences between different substations, resulting in insufficient model universality. Second, due to the high complexity of the model itself, data processing latency is difficult to control, making it challenging to achieve low-latency real-time responses at the edge. Consequently, the existing system has significant shortcomings in key areas such as accurate prediction of long-term substation operating trends, effective simulation of fault propagation paths, and quantitative evaluation of operation and maintenance plans. It cannot support the transformation of substations from a "passive fault response" to a "proactive risk prevention" operation and maintenance model. Therefore, it is urgent to achieve deep integration of topology adaptation, model lightweighting, and virtual-physical collaboration through technological improvements. Summary of the Invention

[0005] This invention provides a panoramic intelligent monitoring system and method for substations, which solves the technical problems of insufficient adaptability and real-time response at the edge in existing automated monitoring schemes with fixed model architectures under operation and maintenance scenarios with complex equipment and dynamic topology changes in substations.

[0006] The first aspect of this invention provides a substation panoramic intelligent monitoring system, comprising:

[0007] The perception layer is used to acquire equipment status data and topology status data of the substation, as well as digital twin simulation data of the digital twin model associated with the substation.

[0008] An edge computing layer is used to perform secondary weight reduction on the lightweight anomaly detection model based on the topology state data and the digital twin simulation data, and to perform anomaly monitoring in conjunction with the device state data to determine short-term monitoring results.

[0009] The cloud computing layer is used to predict the status based on the pre-built AI intelligent model and the digital twin model, using the device status data, the topology status data and the short-term monitoring results, and to determine the long-term monitoring results.

[0010] The execution control layer is used to determine the execution decision scheme based on the long-term monitoring results.

[0011] Optionally, the digital twin simulation data includes topology dynamic evolution data and device state simulation deviation values, and the edge computing layer is used for:

[0012] The initial equipment connection diagram of the substation is constructed based on the topology status data and the topology dynamic evolution data;

[0013] The initial device connection diagram is weighted according to the device state simulation deviation value to obtain the target device connection diagram.

[0014] Configure the lightweight anomaly detection model as a dynamic graph neural network structure;

[0015] Using the target device connection graph as the topology input and the device status data as the node feature input, the dynamic graph neural network structure is mapped to topology and state to obtain a pre-configured dynamic graph neural network structure.

[0016] The topology dynamic evolution data and the equipment state simulation deviation value are used as inputs to the pre-configured dynamic graph neural network structure for dynamic adaptation and optimization, resulting in an optimized dynamic graph neural network structure.

[0017] The optimized dynamic graph neural network structure is subjected to topology and state association graph convolution processing to obtain a secondary lightweight anomaly detection model;

[0018] The device status data is input into the secondary lightweight anomaly detection model for feature extraction to obtain device status features;

[0019] The device status characteristics are input into a preset anomaly detection classifier or threshold judgment module, and a short-term warning signal for the device is output.

[0020] The device status features associated with the short-term warning signal of the device are matched as key feature data;

[0021] The short-term monitoring results include the short-term warning signals of the equipment and the key feature data used to interpret the short-term warning signals.

[0022] Optionally, the dynamic adaptation optimization includes:

[0023] The topological dynamic evolution data is mapped to the topological change features of the pre-configured dynamic graph neural network structure;

[0024] The device state simulation deviation value is mapped to the node feature deviation weight of the pre-configured dynamic graph neural network structure;

[0025] The graph convolution computation links of the pre-configured dynamic graph neural network structure are adjusted according to the topology change characteristics, and the node feature input dimensions of the pre-configured dynamic graph neural network structure are adjusted according to the node feature deviation weights to obtain an optimized dynamic graph neural network structure.

[0026] Optionally, the cloud computing layer is used for:

[0027] Based on a pre-built AI intelligent model, the device status data, the topology status data, and the short-term monitoring results are used to perform trend prediction and determine long-term trend prediction information.

[0028] The long-term trend prediction information is input into the digital twin model to perform fault simulation, and the fault simulation results are obtained.

[0029] The fault simulation results include fault propagation path evolution data and operation and maintenance scheme simulation data.

[0030] By integrating the long-term trend prediction information, the fault propagation path evolution data, and the operation and maintenance scheme simulation results, long-term monitoring results are obtained.

[0031] Optionally, the pre-built AI intelligent model includes a dynamic feature fusion module, a causal reasoning module, and a multi-scale prediction module, wherein the trend prediction includes:

[0032] Obtain the external operating variables of the substation;

[0033] The device status data, the topology status data, and the short-term monitoring results are input into the dynamic feature fusion module for feature fusion to obtain a fused feature vector.

[0034] The fused feature vector and the preset fault causal knowledge graph are input into the causal reasoning module to perform fault source tracing and deduction, and the fault causal analysis results are obtained.

[0035] The fused feature vector, the fault causal analysis results, and the external operating variables are input into the multi-scale prediction module for time-series dimension extrapolation to obtain the multi-scale operating state prediction uncertainty results.

[0036] By integrating the results of the fault causal analysis and the results of the multi-scale operational state prediction uncertainty, long-term trend prediction information is obtained.

[0037] Optionally, the fault simulation includes:

[0038] The long-term trend prediction information is used to initialize the digital twin model at multiple scales, and a virtual working condition corresponding to the long-term trend prediction information is constructed based on the initialized digital twin model.

[0039] Based on the virtual working conditions, multiphysics coupling simulation is performed using the initialized digital twin model to obtain multiphysics coupling simulation data.

[0040] Based on the multiphysics coupling simulation data, the fault propagation path is deduced using the initialized digital twin model to obtain the fault propagation path evolution data.

[0041] Based on the fault propagation path evolution data, the operation and maintenance scheme is simulated using the initialized digital twin model to obtain the operation and maintenance scheme simulation data;

[0042] By integrating the fault propagation path evolution data and the operation and maintenance scheme simulation data, fault simulation results are obtained.

[0043] Optionally, the multiphysics coupling simulation includes:

[0044] Match the corresponding multiphysics configuration parameters according to the virtual working conditions;

[0045] The multiphysics configuration parameters are input into the initialized digital twin model to perform multiphysics iterative coupling and obtain initial simulation results;

[0046] Calculate the dynamic error between the initial simulation results and the device status data;

[0047] The multiphysics configuration parameters are dynamically adjusted based on the dynamic error until the dynamic error converges, and then the current initial simulation results are used as multiphysics coupling simulation data.

[0048] Optionally, the edge computing layer is further used for:

[0049] The topology dynamic evolution data is parsed to identify nodes, and a model topology pruning strategy is determined based on the node parsing results.

[0050] The deviation values ​​of the equipment state simulation are analyzed for deviation. The dominant deviation features of high-sensitivity equipment with deviation rates exceeding a preset deviation rate threshold are retained to obtain a set of dominant deviation features of high-sensitivity equipment.

[0051] Redundant input features that do not belong to the set of dominant features of high-sensitivity equipment deviation in the secondary lightweight anomaly detection model are removed, and the model parameter weights associated with the set of dominant features of high-sensitivity equipment deviation are adjusted to obtain an anomaly detection model with optimized input features and parameters.

[0052] The aforementioned model topology pruning strategy is used to perform topology pruning on the anomaly detection model after optimizing the input features and parameters;

[0053] The anomaly detection model after topology trimming is used as a new secondary lightweight anomaly detection model, and the process jumps to execute the step of inputting the device status data into the secondary lightweight anomaly detection model for feature extraction to obtain device status features.

[0054] Optionally, the node resolution includes:

[0055] Based on the dynamic evolution data of the topology, the device connection nodes corresponding to the model topology are quantitatively evaluated to identify core device nodes and redundant device nodes.

[0056] Based on the quantitative evaluation results of the core device nodes, the core device nodes are hierarchically divided.

[0057] Based on the hierarchical division results and the redundant device node identification results, a model topology pruning strategy is formulated. Specifically, the model topology pruning strategy involves pruning the entire model topology link corresponding to the redundant device node and optimizing and retaining the model topology link corresponding to the core device node.

[0058] A second aspect of the present invention provides a monitoring method applied to the aforementioned substation panoramic intelligent monitoring system, comprising:

[0059] Acquire equipment status data and topology status data of the substation, as well as digital twin simulation data of the substation's associated digital twin model;

[0060] Based on topology status data and digital twin simulation data, the lightweight anomaly detection model is further lightweighted, and anomaly monitoring is performed in conjunction with equipment status data to determine short-term monitoring results.

[0061] Based on pre-built AI intelligent models and digital twin models, status prediction is performed using equipment status data, topology status data, and short-term monitoring results to determine long-term monitoring results;

[0062] Based on long-term monitoring results, an implementation decision plan is determined.

[0063] As can be seen from the above technical solutions, the present invention has the following advantages:

[0064] This invention provides a panoramic intelligent monitoring system and method for substations. It acquires equipment status data, topology status data, and digital twin simulation data of the associated digital twin model of the substation through a perception layer. An edge computing layer further lightweights the anomaly detection model based on the topology status data and digital twin simulation data, and combines this with equipment status data to conduct anomaly monitoring to determine short-term monitoring results. Simultaneously, a cloud computing layer, based on a pre-built AI intelligent model and digital twin model, integrates equipment status data, topology status data, and short-term monitoring results to perform status prediction and obtain long-term monitoring results. Finally, an execution control layer determines the execution decision scheme based on the long-term monitoring results. In this invention, the multi-source data acquisition of the perception layer provides real-time data support for adapting to the dynamically changing topology and complex equipment types of substations. The edge computing layer performs a secondary lightweighting operation on the model, dynamically adjusting the model architecture based on real-time topology status and digital twin simulation data. This breaks the limitations of existing fixed-architecture monitoring solutions, effectively adapting to the topological heterogeneity and individual equipment differences of different substations. Furthermore, the secondary lightweighting further reduces model complexity, ensuring efficient operation at the edge and enabling real-time response to anomaly monitoring. Meanwhile, the cloud computing layer, through collaborative computation of multi-source data, digital twins, and AI models, can accurately generate monitoring results reflecting the long-term operational status of substations. Combined with targeted decision-making solutions output by the execution control layer, this not only addresses the technical shortcomings of existing solutions in terms of adaptability and real-time response at the edge, but also provides comprehensive and reliable technical support for the transformation of substations from passive fault response to proactive risk prevention in operation and maintenance. Attached Figure Description

[0065] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0066] Figure 1 This is a structural block diagram of the substation panoramic intelligent monitoring system according to an embodiment of the present invention;

[0067] Figure 2 This is a flowchart illustrating the steps of a monitoring method applied to a substation panoramic intelligent monitoring system, according to an embodiment of the present invention. Detailed Implementation

[0068] This invention provides a panoramic intelligent monitoring system and method for substations, which addresses the technical problems of insufficient adaptability and real-time response at the edge in existing automated monitoring schemes with fixed model architectures under operation and maintenance scenarios involving complex substation equipment and dynamic topology changes.

[0069] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0070] Please see Figure 1 The present invention provides a substation panoramic intelligent monitoring system, comprising:

[0071] The perception layer is used to acquire equipment status data and topology status data of the substation, as well as digital twin simulation data of the substation's associated digital twin model.

[0072] In this embodiment of the invention, the perception layer is equipped with a variety of sensors and data interfaces to acquire equipment status data and topology status data of the substation, as well as digital twin simulation data of the substation-related digital twin model. As the data acquisition foundation of the entire monitoring system, the perception layer deploys various types of sensors, such as electrical parameter sensors (e.g., current transformers, voltage transformers), physical status sensors (e.g., infrared temperature sensors, partial discharge sensors), and environmental sensors (e.g., humidity sensors, wind speed sensors), to collect data such as oil temperature, current, and insulation status of core equipment such as transformers and circuit breakers in real time, and integrates all the collected data to obtain equipment status data. Meanwhile, the perception layer connects to the substation's SCADA system and relay protection devices through standardized data interfaces to synchronously acquire data such as equipment connection relationships and switch opening and closing status. It integrates all collected data to obtain topology status data. In addition, the digital twin model associated with the substation continuously simulates the substation's topology changes and equipment operation, outputting topology dynamic evolution data such as changes in equipment connection relationships and power flow distribution adjustments after switching operations, as well as equipment status simulation deviation values ​​such as the difference between the model's simulated oil temperature and the actual collected oil temperature. These two types of data together constitute digital twin simulation data, which is synchronously acquired by the perception layer, providing complete original data support for subsequent analysis at each level.

[0073] The edge computing layer is used to further lightweight the lightweight anomaly detection model based on topology status data and digital twin simulation data, and to perform anomaly monitoring in conjunction with device status data to determine short-term monitoring results.

[0074] In this embodiment of the invention, the edge computing layer undertakes the core tasks of real-time data processing and short-term monitoring, focusing on the dynamic optimization and application of a lightweight anomaly detection model: it receives topology state data and digital twin simulation data transmitted from the perception layer, and performs secondary lightweighting processing on the lightweight anomaly detection model.

[0075] This process is not a fixed optimization, but rather combines the dynamic evolution data of the topology in the digital twin simulation data. Based on the current substation's equipment connection relationships, operating load distribution, and other topological characteristics, it identifies core equipment nodes (such as buses with concentrated power flow) and redundant equipment nodes (such as equipment associated with out-of-operation lines). It dynamically adjusts the model's computational links, trims computational links corresponding to redundant nodes, and retains the key feature calculations of core nodes. At the same time, it combines the equipment state simulation deviation values ​​in the digital twin simulation data to select high-sensitivity equipment deviation dominant features that contribute significantly to anomaly detection (such as temperature features with large simulation deviations). It dynamically adapts the model's feature input dimensions, optimizes the corresponding parameter weights, and eliminates redundant feature inputs. After completing the second lightweighting, a second lightweight anomaly detection model is obtained.

[0076] Subsequently, the edge computing layer inputs the equipment status data collected by the perception layer into the secondary lightweight anomaly detection model. Through feature extraction, anomaly identification and other operations, it quickly outputs short-term early warning signals for whether the equipment is abnormal, as well as key feature data supporting the early warning (such as the circuit breaker's opening and closing time exceeding the tolerance). These signals and data together constitute the short-term monitoring results, meeting the substation's need for rapid response to sudden anomalies.

[0077] The cloud computing layer is used to predict the status and determine the long-term monitoring results based on pre-built AI intelligent models and digital twin models, using equipment status data, topology status data and short-term monitoring results.

[0078] In this embodiment of the invention, the cloud computing layer focuses on long-term trend analysis and in-depth simulation. It integrates equipment status data and topology status data from the perception layer, as well as short-term monitoring results output by the edge computing layer, and invokes pre-built AI intelligent models and digital twin models. Through the deep feature fusion and causal reasoning capabilities of the pre-built AI intelligent models, it mines equipment aging trends and fault evolution patterns to generate long-term trend prediction information. Utilizing the multi-scale, multi-physics simulation capabilities of the digital twin models, it simulates the propagation path of faults between equipment, substations, and the regional power grid, evaluates the implementation effects of different operation and maintenance schemes, and obtains fault propagation path evolution data and operation and maintenance scheme simulation data. Finally, the cloud computing layer integrates the aforementioned long-term trend prediction information, fault propagation path evolution data, and operation and maintenance scheme simulation data to form a result that includes the remaining lifespan of the equipment, the fault risk level, and the optimal operation and maintenance recommendations. This result, namely the long-term monitoring result, provides a basis for the medium- and long-term operation and maintenance planning of substations.

[0079] The execution control layer is used to determine execution decision-making schemes based on long-term monitoring results.

[0080] In this embodiment of the invention, the execution control layer, as the hub connecting the system and actual operation and maintenance actions, receives long-term monitoring results output by the cloud computing layer, interprets and prioritizes these results (e.g., classifying handling levels according to the scope of fault impact and urgency), and combines the substation's operation and maintenance resources (e.g., spare parts inventory, personnel scheduling) and power grid operation constraints (e.g., no power outages during peak load periods) to transform the long-term monitoring results into actionable decision-making solutions, including but not limited to emergency fault handling solutions (e.g., immediately arranging bus joint repairs) and preventive operation and maintenance solutions (e.g., conducting insulation tests on the main transformer within 3 months). At the same time, the execution time, steps, and responsible parties of the solutions are clearly defined to ensure that the monitoring results are transformed into actual operation and maintenance actions.

[0081] Perception Layer: The foundational data acquisition layer of the system, this module enables multi-dimensional data acquisition by deploying various sensors and connecting with existing equipment systems. Equipment Status Data: This refers to the real-time operational parameters of core substation equipment (such as transformers and circuit breakers) collected by sensors, including electrical parameters (current, voltage) and physical status parameters (oil temperature, insulation status), reflecting the real-time operating status of the equipment. Topology Status Data: This refers to data reflecting the connection relationships and operating topology of substation equipment, including physical wiring methods between equipment, switch opening and closing status, and basic information on power flow distribution. Substation Associated Digital Twin Model: This is a virtual simulation model that accurately maps to the actual physical entity of the substation, capable of simulating the operating status of substation equipment, topology changes, and fault evolution processes. Digital Twin Simulation Data: This refers to the simulation data output during the operation of the digital twin model, used to support model optimization and analysis decisions at various levels of the system. Edge Computing Layer: The computing layer closest to the data acquisition end in the system, responsible for real-time data processing, lightweight model optimization, and short-term anomaly monitoring, meeting low-latency response requirements. Lightweight anomaly detection model: Refers to an anomaly detection algorithm model that has undergone architectural simplification and parameter optimization to adapt to the limited computing resources of the edge. Secondary lightweighting: Based on the lightweight anomaly detection model, this involves further optimizing the model architecture and eliminating redundant computational links by combining real-time substation topology status data and digital twin simulation data, thereby improving model adaptability and operational efficiency. Short-term monitoring results: Refers to the results data output by the edge computing layer through anomaly monitoring, reflecting the abnormal status of equipment in the substation within a short period (e.g., from real-time to several hours). Cloud computing layer: A remote computing layer with powerful computing capabilities in the system, relying on AI intelligent models and digital twin models to perform complex calculations such as long-term trend prediction and fault simulation. Pre-built AI intelligent model: Refers to an artificial intelligence algorithm model pre-deployed in the cloud computing layer, possessing functions such as feature fusion, causal reasoning, and multi-scale prediction. Long-term monitoring results: Refers to the comprehensive results output by the cloud computing layer through trend prediction and fault simulation, reflecting the medium- to long-term (e.g., several months to several years) equipment health trends, fault risks, and operation and maintenance recommendations of the substation. Execution Control Layer: This layer connects monitoring results with actual operational actions within the system, responsible for transforming long-term monitoring results into actionable operational decisions. Execution Decisions: These refer to the operational action plans formulated by the Execution Control Layer based on long-term monitoring results, including emergency fault handling measures and preventative operational plans, clearly defining the execution time, steps, and responsible parties.

[0082] It should be noted that the secondary lightweighting strategy relies on the dynamic updating of digital twin simulation data output by the digital twin model, which not only meets the computing power requirements of the edge but also ensures the accuracy of anomaly detection, effectively avoiding the problems of poor adaptability and wasted computing power of traditional fixed models.

[0083] This invention provides a substation panoramic intelligent monitoring system. The digital twin simulation data includes topology dynamic evolution data and equipment state simulation deviation values. An edge computing layer is used for: constructing an initial equipment connection diagram of the substation based on the topology state data and topology dynamic evolution data; weighting the initial equipment connection diagram according to the equipment state simulation deviation values ​​to obtain a target equipment connection diagram; configuring a lightweight anomaly detection model as a dynamic graph neural network structure; using the target equipment connection diagram as the topology structure input and the equipment state data as the node feature input, performing topology-state mapping on the dynamic graph neural network structure to obtain a pre-configured dynamic graph neural network structure; and using the topology dynamic evolution data... The simulated deviation values ​​of the equipment status are input into a pre-configured dynamic graph neural network structure for dynamic adaptation and optimization, resulting in an optimized dynamic graph neural network structure. The optimized dynamic graph neural network structure undergoes topology and state association graph convolution processing to obtain a secondary lightweight anomaly detection model. Equipment status data is input into the secondary lightweight anomaly detection model for feature extraction, yielding equipment status features. These features are then input into a pre-defined anomaly detection classifier or threshold judgment module, outputting a short-term warning signal for the equipment. Equipment status features associated with the short-term warning signal are matched as key feature data. The short-term monitoring results include the short-term warning signal and the key feature data used to interpret it.

[0084] Furthermore, the dynamic adaptation optimization includes: mapping the topology dynamic evolution data to the topology change features of the pre-configured dynamic graph neural network structure; mapping the device state simulation deviation value to the node feature deviation weights of the pre-configured dynamic graph neural network structure; adjusting the graph convolution calculation links of the pre-configured dynamic graph neural network structure according to the topology change features; and adjusting the node feature input dimension of the pre-configured dynamic graph neural network structure according to the node feature deviation weights to obtain the optimized dynamic graph neural network structure.

[0085] Topology Dynamic Evolution Data: A core component of digital twin simulation data, referring to simulation data depicting the temporal changes in the substation topology, including information such as changes in equipment connection relationships after switching operations, adjustments in power flow distribution, and potential fault propagation links. Equipment State Simulation Deviation Value: A core component of digital twin simulation data, referring to the difference between the equipment state parameters generated by the digital twin model simulation and the actual equipment state data collected by the sensing layer, used to reflect the degree of deviation between simulation and actual operation. Initial Equipment Connection Graph: A graphical data structure constructed by the edge computing layer based on topology state data and topology dynamic evolution data, reflecting the current basic connection relationships of substation equipment. Weight Configuration: The process by which the edge computing layer assigns dynamic weights to each connection edge in the initial equipment connection graph based on the equipment state simulation deviation value. The weights comprehensively reflect the power transmission efficiency, functional dependence, and fault propagation probability between equipment. Target Equipment Connection Graph: A graphical data structure that, after weight configuration, simultaneously reflects the equipment connection relationships and the importance of the connections, providing a foundation for model topology input. Dynamic Graph Neural Network (Graph Neural Network) Architecture: Refers to a graph neural network model whose architecture can dynamically adjust to changes in the input topology, possessing the ability to adapt to dynamic topological changes. Topology-State Mapping: Refers to the process of mapping the topology of the target device connection graph to the node features of the device state data, corresponding to the input of the dynamic graph neural network structure, enabling the network model to establish topology-state relationships. Pre-configured Dynamic Graph Neural Network Architecture: Refers to a dynamic graph neural network model architecture that is initially adapted to the current substation topology and device states after completing the topology-state mapping. Dynamic Adaptation Optimization: Refers to the process of adjusting and optimizing the computational links and feature input dimensions of the pre-configured dynamic graph neural network structure based on the dynamic evolution data of the topology and the simulation deviation values ​​of the device states. Optimized Dynamic Graph Neural Network Architecture: Refers to a dynamic graph neural network model architecture that has improved both topology adaptability and feature specificity after dynamic adaptation optimization. Topology-State Relationship Graph Convolution Processing: Refers to the graph convolution operation performed on the optimized dynamic graph neural network structure, the core of which is to capture the relationship between topology changes and device states, while simplifying redundant calculations to reduce the model's computational power requirements. Equipment Status Features: These refer to the feature vector obtained after the secondary lightweight anomaly detection model extracts features from equipment status data, integrating topological relationships, individual equipment status, and temporal change patterns. Pre-defined Anomaly Detection Classifier: This refers to a classification model pre-deployed in the edge computing layer, built based on training data, used to determine and classify anomalies in equipment status features. Threshold Judgment Module: This refers to a functional module pre-deployed in the edge computing layer, setting thresholds based on industry standards or equipment rated parameters to determine anomalies in equipment status features. Short-Term Equipment Warning Signal: This refers to the signal output by the anomaly detection classifier or threshold judgment module, reflecting whether the equipment is abnormal and the degree of abnormality (e.g., normal, slightly abnormal, severely abnormal).Key Feature Data: This refers to equipment status feature data directly related to short-term warning signals, used to explain the triggering reasons of the warning signals. Topology Change Features: This refers to feature data obtained by mapping the dynamic evolution data of the topology to reflect topology changes such as the addition or removal of equipment nodes, changes in connection edges, and the association of concentrated power flow regions. Node Feature Deviation Weights: This refers to the weight values ​​corresponding to each equipment status feature after weighting the equipment status simulation deviation values. The higher the deviation rate, the greater the corresponding node feature weight. Graph Convolution Computation Link: This refers to the computational path in the dynamic graph neural network used to process the inter-node association features, directly affecting the model's ability to capture topology association information. Node Feature Input Dimension: This refers to the number of dimensions in the dynamic graph neural network that receive equipment status feature data, corresponding to different types of equipment status parameters.

[0086] It should be noted that the secondary lightweighting strategy of the lightweight anomaly detection model is not statically set, but dynamically updated based on the output of the substation-related digital twin model. This digital twin model continuously simulates the substation's topology changes and equipment operation, outputting topology dynamic evolution data (such as changes in equipment connection relationships after switching operations and adjustments in power flow distribution) and equipment state simulation deviation values ​​(such as the difference between the oil temperature simulated by the model and the actual collected oil temperature), thereby ensuring that the model always adapts to the real-time operating status of the substation.

[0087] In this embodiment of the invention, the edge computing layer first performs deep analysis on the topology state data transmitted by the perception layer, extracting the real-time operating status of each device (such as circuit breaker opening and closing, transformer commissioning and decommissioning), the physical connection relationship between devices (such as the wiring method between the bus and the circuit breaker, the association between the line and the disconnecting switch), and the functional linkage logic in the operation of the devices (such as the switching relationship between the backup line and the main power supply line after a certain line is out of service). Combined with the above-mentioned topology dynamic evolution data, an initial device connection relationship diagram is constructed. Then, the device state simulation deviation value is called to assign a dynamic weight to each connection edge in the initial device connection relationship diagram. The determination of the weight value comprehensively considers the power transmission efficiency between devices (such as the higher the proportion of the power transmitted by a connection edge to the total power of the whole station, the greater the weight), the degree of functional dependence (such as the connection edge between the main transformer and the bus, which directly affects the power supply of the whole station, has a higher weight than the connection edge of ordinary feeder), and the possibility of fault propagation (such as the connection edge located in a high-fault area, the weight will be appropriately increased). This allows the connection relationship diagram to not only reflect whether the devices are connected, but also to reflect the importance of the connection. After completing the weight configuration, the target device connection relationship diagram is obtained.

[0088] Next, to enable the lightweight anomaly detection model to better adapt to dynamic topology changes and accurately capture the correlation between topology and device status, the edge computing layer configures the lightweight anomaly detection model as a dynamic graph neural network structure. The core advantage of this structure lies in its dynamic adaptability:

[0089] Unlike traditional fixed-architecture models (such as static CNN convolutional neural network models), its computational logic is adjusted as the device connection graph is updated. When nodes or edges in the connection graph are added or removed, the network automatically adjusts the computational dimension of the graph convolution. When the weights of the edges are updated, the network simultaneously optimizes the attention allocation mechanism to ensure that it always focuses on key associations. For example, if the circuit breaker corresponding to a high-weight edge is disconnected, the dynamic graph neural network will immediately remove the computation link of that edge and increase the computational priority of the backup edge. This avoids the problem that traditional models cannot adapt to topology changes in a timely manner due to their fixed architecture, which leads to anomaly detection bias.

[0090] Subsequently, the edge computing layer uses the target device connection graph as the topology input and transforms the device status data (such as transformer oil temperature and current, circuit breaker opening and closing time) collected by the perception layer into node features of the dynamic graph neural network (each device corresponds to a node in the network, and the various state parameters of the device are integrated into the multi-dimensional feature vector of the node). At the same time, the weighted connection edges in the target device connection graph are used as dynamic weights and input into the dynamic graph neural network to complete the topology and state mapping, thus obtaining the pre-configured dynamic graph neural network structure.

[0091] Subsequently, the edge computing layer performs dynamic adaptation and optimization operations: First, the topology dynamic evolution data is mapped to the topology change features of the pre-configured dynamic graph neural network structure (including the addition and removal of device nodes, changes in connection edges, and structural association features corresponding to power flow concentration areas, etc.). The device state simulation deviation values ​​are mapped to the node feature deviation weights of the structure (i.e., the weight values ​​corresponding to the degree of deviation of each device state feature; the higher the deviation rate, the greater the corresponding node feature weight). Then, the graph convolution calculation links of the pre-configured structure are adjusted according to the topology change features. Based on the structural association features corresponding to the power flow concentration areas, the core device nodes in the current topology (such as the bus in the power flow concentration area) are identified, and their corresponding graph convolution calculation links are retained. At the same time, based on the redundancy addition and removal features of device nodes, redundant device nodes (such as equipment associated with out-of-operation lines) are identified, and their corresponding graph convolution calculation links are pruned. Simultaneously, the node feature input dimensions of the pre-configured structure are adjusted according to the node feature deviation weights. Sensitive features with weights higher than a preset threshold (such as temperature features with large simulation deviations) are selected, and their corresponding input dimensions are retained. The input dimensions corresponding to non-sensitive features with weights lower than the threshold (such as environmental temperature and humidity data with low correlation to anomaly detection) are removed. After the adjustment is completed, the optimized dynamic graph neural network structure is obtained.

[0092] Next, the edge computing layer performs topology and state association graph convolution processing on the optimized dynamic graph neural network structure: based on node features, it calculates the feature interactions between adjacent nodes by combining dynamic weights (such as the bus voltage fluctuations corresponding to high-weight connection edges, which will quickly affect the current state of downstream circuit breakers). At the same time, it simplifies or even omits the feature interaction calculations of redundant connection edges with extremely low weights (such as idle connections of out-of-service equipment), and reduces the weight of node parameters with low feature contribution in the calculation. By focusing on key associations and simplifying redundant calculations, it not only ensures the model's accuracy in capturing topology and equipment state associations, but also further compresses the computational load of the model. Finally, it completes the secondary lightweighting of the lightweight anomaly detection model, resulting in the secondary lightweight anomaly detection model in this system.

[0093] Finally, the edge computing layer inputs the equipment status data into the secondary lightweight anomaly detection model for feature extraction to obtain equipment status features. These features are then input into the edge computing layer's preset anomaly detection classifier or threshold judgment module to output short-term equipment warning signals. Simultaneously, the equipment status features associated with the warning are matched as key feature data (such as the circuit breaker's opening and closing time exceeding the tolerance). Together, they constitute the short-term monitoring results in this system, meeting the substation's need for rapid response to sudden anomalies.

[0094] Through this dynamic update mechanism, the lightweight anomaly detection model can meet the computing power requirements of the edge while ensuring the accuracy of anomaly detection, effectively avoiding the problems of poor adaptability and wasted computing power of traditional fixed models.

[0095] It should be noted that the edge computing layer calls a second-lightweight anomaly detection model (i.e., a dynamic graph neural network after second-lightweight processing) to extract features from the real-time device status data input from the perception layer:

[0096] This process does not simply extract isolated state parameters of a single device, but rather combines the topological relationships and temporal change patterns recorded in the model. For example, when extracting the current characteristics of a circuit breaker, the network will simultaneously associate the voltage fluctuation data of its upstream bus and the load change data of its downstream line, while also incorporating the temporal change trend of the circuit breaker's current over the past 5 minutes. The device's own state, the state of the topologically associated devices, and the historical temporal state are fused together to finally generate a device state feature vector that combines individual device characteristics, topological association characteristics, and temporal evolution characteristics. The device state feature vector is a device state feature that integrates topological temporal information, ensuring that the extracted features can fully reflect the actual operating state of the device in the current topological environment.

[0097] After feature extraction, the edge computing layer inputs the generated equipment status feature vector into a preset anomaly detection classifier or threshold judgment module. If an anomaly detection classifier is used, it compares the current equipment status features with normal feature templates and typical fault feature templates based on a model trained using historical substation fault data and normal operation data. This determines whether the equipment exhibits an anomaly and its type (e.g., overheating or insulation anomaly). If a threshold judgment module is used, it compares key parameters in the feature vector (e.g., transformer oil temperature, circuit breaker opening and closing time) with preset industry standard thresholds and equipment rated thresholds. An anomaly is determined when a parameter exceeds the threshold range. Regardless of the method used, a clear short-term equipment warning signal is ultimately output. The signal format includes graded indicators such as normal, minor anomaly, and severe anomaly, facilitating maintenance personnel to quickly grasp the degree of equipment anomaly.

[0098] In addition to the aforementioned short-term warning signals from the equipment, the short-term monitoring results output by the edge computing layer also include key characteristic data to explain these warning signals. This data clearly indicates the core reason for triggering the warning. For example, if the warning signal is "circuit breaker severe abnormality," the key characteristic data will detail that "the opening and closing time exceeds the rated threshold by 15ms three times consecutively" and "the voltage fluctuation of the associated bus reaches 8%." It will also indicate the circuit breaker's position in the topology connection (e.g., "connecting bus #2 and line #3") and the timing curve characteristics of the opening and closing time over the past 10 minutes. By providing this key characteristic data, maintenance personnel can not only know whether the equipment is abnormal, but also understand why it is abnormal. This provides a direct basis for quickly locating the anomaly and making a preliminary judgment on the cause of the fault, avoiding the problem of being unable to trace the root cause of the anomaly based solely on the warning signal.

[0099] This invention provides a substation panoramic intelligent monitoring system. The cloud computing layer is used for: based on a pre-set AI intelligent model, using equipment status data, topology status data, and short-term monitoring results to perform trend prediction and determine long-term trend prediction information; using the long-term trend prediction information as input to a digital twin model to perform fault simulation and obtain fault simulation results; wherein, the fault simulation results include fault propagation path evolution data and operation and maintenance scheme simulation data; integrating the long-term trend prediction information, fault propagation path evolution data, and operation and maintenance scheme simulation results to obtain long-term monitoring results.

[0100] Trend Prediction: This refers to the process by which the cloud computing layer, based on a pre-built AI intelligent model, integrates multi-source data to predict the long-term operating status and fault risk development trends of substation equipment. Long-Term Trend Prediction Information: This refers to the information output by the trend prediction, covering different time dimensions (e.g., short-term 1-30 days, medium-term 1-6 months, long-term 6-12 months), including equipment health status degradation trends, fault risk escalation probabilities, and equipment remaining life assessments. Fault Simulation: This refers to the simulation process of inputting long-term trend prediction information into a digital twin model to simulate the occurrence, development, and propagation of faults, and to evaluate the effectiveness of different operation and maintenance solutions. Fault Simulation Results: This refers to the comprehensive result data output by the fault simulation, including fault propagation path evolution data and operation and maintenance solution simulation data, providing core support for long-term monitoring results. Fault Propagation Path Evolution Data: This refers to the dynamic data generated during the fault simulation process, recording the time nodes, affected equipment, and triggering conditions for the fault's spread from its initial occurrence point to other equipment and levels. Operation and maintenance scheme simulation data: refers to the data used in fault simulation to quantitatively evaluate the implementation effects of different operation and maintenance schemes (such as emergency repair, backup replacement, and flow-limiting operation), including indicators such as fault elimination efficiency, economic cost, and operational risk. Operation and maintenance scheme simulation results: refers to the evaluation conclusions on the advantages, disadvantages, and applicable scenarios of each scheme obtained after organizing and analyzing the operation and maintenance scheme simulation data.

[0101] In this embodiment of the invention, the cloud computing layer first uses a pre-built AI intelligent model, along with device status data and topology status data collected by the perception layer, and short-term monitoring results output by the edge computing layer, to perform trend prediction and determine long-term trend prediction information.

[0102] This analysis process is not a single-dimensional data processing, but is achieved through the collaboration of multiple modules within the pre-set AI intelligent model. The pre-set AI intelligent model first integrates equipment status data (such as changes in the insulation resistance of transformers and the decay of the mechanical characteristics of circuit breakers), topology status data (such as long-term power flow distribution trends and normal adjustment patterns of topology structures) with short-term monitoring results (such as periodic mild anomaly warnings) across dimensions to extract core features that reflect the essence of equipment operation.

[0103] Then, the causal reasoning module built into the pre-set AI intelligent model traces the potential correlations behind the features, such as identifying the causal relationship between long-term bus power overload and continuous high transformer oil temperature, eliminating the interference of unrelated factors; finally, with the help of the multi-scale prediction module, combined with the rated parameters of the equipment, years of operation and historical fault data, it outputs long-term trend prediction information covering different time dimensions, including the short-term (1-3 months) fault risk escalation probability, the medium-term (6-12 months) equipment health status decay curve, and the long-term (1-3 years) equipment remaining life assessment, providing clear trend basis for subsequent simulation and deduction.

[0104] Subsequently, the cloud computing layer uses this long-term trend prediction information to input into the digital twin model for fault simulation, obtaining fault simulation results (including fault propagation path evolution data and operation and maintenance scheme simulation data). The fault simulation process embodies the characteristics of "multi-scale" and "multi-physics field". Among them, multi-scale refers to the simulation scope covering multiple levels from the component level to the regional power grid level. For example, when the long-term trend prediction information indicates that "there is a risk of insulation aging in the transformer winding", the simulation will first start from the component level, simulating the development process of partial discharge after the winding insulation is damaged; then it will be extended to the equipment level to deduce the impact of the winding temperature rise caused by partial discharge on the entire transformer. The impact of transformer failures on the overall operation is analyzed, extending to the station level to examine the cascading effects of transformer failures on bus voltage and adjacent circuit breakers. If necessary, the analysis will be extended to the regional power grid level to simulate the impact of load transfer on surrounding substations. Multi-physics simulation refers to the simultaneous simulation of multiple physical fields associated with the fault. For example, in the case of a winding fault, the electromagnetic field (electric field distortion caused by partial discharge), temperature field (spatial distribution of winding heating), flow field (convective heat transfer changes in transformer oil), and mechanical field (winding deformation caused by temperature rise) will be calculated simultaneously to reconstruct the coupling effect of each physical field when the fault occurs. The results generated by the above process are the fault propagation path evolution data.

[0105] Meanwhile, the cloud computing layer loads different operation and maintenance schemes (such as emergency winding repair, replacement of backup transformer, and current-limited operation during power outage windows) into the digital twin model. By simulating the changes in the physical field, fault termination, and power grid operation parameters after the execution of each scheme, the fault elimination efficiency (such as how long it takes to restore normal operation after maintenance), economic costs (such as spare parts consumption and labor costs), and operational risks (such as the load carrying capacity of the power grid during maintenance) of each scheme are quantitatively evaluated. These evaluation data are the operation and maintenance scheme simulation data.

[0106] Finally, the cloud computing layer integrates long-term trend prediction information, fault propagation path evolution data, and operation and maintenance scheme simulation data to obtain long-term monitoring results. This integration process first prioritizes various information items, such as listing key information like equipment remaining life of less than 6 months and fault propagation affecting the main power supply line as high priority, serving as the core basis for decision-making. Then, it performs weighted analysis on the evaluation indicators of the operation and maintenance scheme, combined with the actual operation and maintenance needs of the substation (such as whether long-term power outages are permissible and whether spare parts inventory is sufficient), to select the optimal operation and maintenance scheme. The final long-term monitoring results not only include the long-term health trend of the equipment, the fault risk level, and the key nodes of fault propagation, but also clearly define the recommended operation and maintenance scheme, the best time window for implementation, and precautions during scheme execution (such as grid dispatch requirements that need to be coordinated and spare parts models that need to be prepared). It may even estimate the long-term benefits after the scheme is implemented (such as how much the remaining life of the equipment can be extended after implementation and how much the probability of failure can be reduced), becoming a comprehensive conclusion that can directly guide the medium- and long-term operation and maintenance planning of the substation, providing operation and maintenance personnel with complete decision support from trend judgment to scheme implementation.

[0107] This invention provides a substation panoramic intelligent monitoring system. The pre-built AI intelligent model includes a dynamic feature fusion module, a causal reasoning module, and a multi-scale prediction module. Trend prediction includes: acquiring external operating variables of the substation; inputting equipment status data, topology status data, and short-term monitoring results into the dynamic feature fusion module for feature fusion to obtain a fused feature vector; inputting the fused feature vector and a pre-set fault causal knowledge graph into the causal reasoning module for fault source tracing and deduction to obtain fault causal analysis results; inputting the fused feature vector, fault causal analysis results, and external operating variables into the multi-scale prediction module for time-series dimension deduction to obtain multi-scale operating state prediction uncertainty results; and integrating the fault causal analysis results and multi-scale operating state prediction uncertainty results to obtain long-term trend prediction information.

[0108] External operating variables: These refer to the external environment and scheduling factors affecting the operation of substation equipment, including data such as future ambient temperature forecasts, power grid load scheduling plans, and equipment maintenance arrangements. Dynamic feature fusion module: This refers to a pre-built AI intelligent model that integrates features from multiple data sources, enabling cross-dimensional integration of features from equipment status data, topology status data, and short-term monitoring results. Fusion feature vector: This refers to a unified dimensional feature vector output by the dynamic feature fusion module, integrating individual equipment features, topological correlation features, and features from previous anomaly evolution. Pre-built fault causal knowledge graph: This refers to a structured knowledge system built based on power industry fault mechanisms and historical substation fault cases, storing hierarchical relationships of "fault phenomenon - direct cause - root cause - scope of impact". Causal reasoning module: This refers to a pre-built AI intelligent model that uses fault tracing and deduction, enabling the identification of the root cause and scope of impact of a fault based on the fused feature vector and the fault causal knowledge graph. Fault tracing and deduction: This refers to the process by which the causal reasoning module traces the root cause of a fault and deduces its scope of impact by matching fault phenomena, eliminating false associations. Fault Causal Analysis Results: These refer to the analytical conclusions output by fault tracing and deduction, including the root cause of the fault and the potential scope of equipment impact. Multi-Scale Prediction Module: This refers to the functional module in the pre-built AI intelligent model used for time-series dimension deduction, capable of using differentiated algorithms for trend prediction at different time scales. Time-Series Dimension Deduction: This refers to the multi-scale prediction module combining fused feature vectors, fault causal analysis results, and external operating variables to perform time-dimensional deduction calculations of the long-term operating status of the equipment. Multi-Scale Operating Status Prediction Uncertainty Results: This refers to the results output by the multi-scale prediction module, including prediction conclusions for different time dimensions and corresponding confidence intervals. The confidence intervals quantify the uncertainty of the prediction results (such as deviations caused by load fluctuations and data errors).

[0109] In this embodiment of the invention, the cloud computing layer first obtains the external operating variables of the substation, specifically including the environmental temperature forecast for the next 7 days, the power grid load dispatch plan, and the equipment maintenance schedule.

[0110] Subsequently, the cloud computing layer synchronously inputs the device status data (covering real-time operating parameters of each core device, such as transformer oil temperature and circuit breaker opening and closing time), topology status data (providing device connection relationships and power flow distribution information) collected by the perception layer, and short-term monitoring results output by the edge computing layer (including identified minor abnormal signals and key features) into the dynamic feature fusion module of the pre-built AI intelligent model for feature fusion. This module does not simply splice the three types of data, but first associates the topology relationship with the device status (such as binding and analyzing the voltage data of a bus with the current data of all circuit breakers connected to that bus), and then combines the abnormal features in the short-term monitoring results to extract the topology correlation parameters and time series change patterns related to the abnormality. Finally, it compresses these multi-dimensional information into a unified-dimensional fusion feature vector, which includes both individual device operating characteristics and topology correlation characteristics and early abnormal evolution characteristics.

[0111] After generating the fused feature vector, the cloud computing layer inputs this vector along with a pre-defined fault causal knowledge graph (built based on power industry fault mechanisms and historical substation fault cases, storing hierarchical relationships of "fault phenomenon - direct cause - root cause - scope of impact," such as "circuit breaker opening and closing time deviation" corresponding to the direct cause "insufficient operating voltage," the root cause "power module aging," and the scope of impact "downstream line power supply reliability") into the causal reasoning module for fault tracing and deduction.

[0112] The causal reasoning module first matches the abnormal features in the fused feature vector (such as "excessive opening and closing time + low operating voltage") with the fault phenomena in the knowledge graph to initially identify possible causal chains. Then, it uses causal analysis algorithms to eliminate false associations (such as verifying the correlation between "low ambient temperature" and "excessive opening and closing time" to eliminate interference from non-related factors) and finally determines the root cause of the fault (such as "power module aging"). At the same time, it combines the device connection relationships in the topology status data to infer the range of devices that may be affected by the root cause (such as the other 3 circuit breakers powered by the same power module and their corresponding lines). The root cause and the range of influence together constitute the causal analysis result of the fault.

[0113] Subsequently, the cloud computing layer will integrate feature vectors, fault causal analysis results, and the aforementioned acquired external operating variables, and input them into the multi-scale prediction module for time-series inference. This module adopts differentiated prediction methods for different time dimensions. For the short term (1-30 days), it predicts the escalation trend of abnormal features based on the evolution rate of the root cause of the fault (such as the operating voltage decreasing by 2% per month due to power module aging) and combined with short-term load changes in external operating variables. For the medium term (1-6 months), it assesses the rate of decline in the health status of the equipment by combining the equipment's operating years and historical aging curves. For the long term (6-12 months), it predicts whether the fault may develop into a serious fault before maintenance by referring to the annual maintenance plan in the external operating variables.

[0114] Meanwhile, the multi-scale prediction module quantifies the uncertainty of the prediction results: by analyzing the fluctuation range of external operating variables (such as load prediction error ±5%) and the randomness of fault evolution, it marks the confidence interval for the prediction results of each time dimension (e.g., "the probability of circuit breaker opening and closing time exceeding the tolerance after 30 days is 85%, confidence interval [80%, 90%]"). The core sources of uncertainty include the suddenness of resource supply, the volatility of the characteristics of alternative resources, and the cumulative bias caused by data acquisition errors and model simplification assumptions. The final output of the prediction results and their uncertainties in different time dimensions is the multi-scale operating state prediction uncertainty result.

[0115] Finally, the cloud computing layer integrates and sorts out the results of fault causal analysis and the uncertainty results of multi-scale operational status prediction to obtain long-term trend prediction information:

[0116] During the integration process, the core impact of the root cause on the long-term operation of the equipment is first identified. Then, short-, medium-, and long-term predictions are arranged in chronological order, with the uncertainty range of each conclusion marked. A "risk escalation path when the root cause is not resolved" is also added (e.g., if the power module is not replaced in time, it may cause two circuit breakers to fail to operate after 3 months, and the impact may expand to the entire busbar after 6 months). The final long-term trend prediction information is presented in a structured form, containing both clear prediction conclusions and explanations of the reasoning behind them, providing a clear trend guide for subsequent simulation and deduction of the digital twin model.

[0117] This invention provides a substation panoramic intelligent monitoring system. Fault simulation includes: initializing a digital twin model using long-term trend prediction information at multiple scales, and constructing virtual operating conditions corresponding to the long-term trend prediction information based on the initialized digital twin model; performing multi-physics coupling simulation using the initialized digital twin model based on the virtual operating conditions to obtain multi-physics coupling simulation data; performing fault path propagation deduction using the initialized digital twin model based on the multi-physics coupling simulation data to obtain fault propagation path evolution data; performing operation and maintenance scheme simulation using the initialized digital twin model based on the fault propagation path evolution data to obtain operation and maintenance scheme simulation data; and integrating the fault propagation path evolution data and the operation and maintenance scheme simulation data to obtain fault simulation results.

[0118] Multi-scale initialization: This refers to the process of initializing and configuring parameters at multiple levels of the digital twin model, including component level, equipment level, station level, and regional power grid level, using long-term trend prediction information to ensure that the initial state of the model closely matches the predicted fault evolution starting point. Virtual operating conditions: This refers to a virtual operating scenario built based on the initialized digital twin model, corresponding to the long-term trend prediction information, capable of reproducing the predicted initial fault state and substation-related operating characteristics. Multi-physics coupling simulation: This refers to the simulation process of simultaneously simulating multiple physical fields such as electromagnetic field, temperature field, flow field, and mechanical field in the digital twin model, considering the interactions between these physical fields (e.g., electromagnetic field losses converting into heat sources in the temperature field, and temperature gradients inducing flow field convection). Multi-physics coupling simulation data: This refers to the accurate data output from the multi-physics coupling simulation, reflecting the spatial distribution and temporal evolution of each physical field, providing a physical field foundation for fault propagation and deduction. Fault path propagation simulation: refers to the simulation process based on multi-physics field coupled simulation data to simulate the propagation of a fault from its initial occurrence level to other devices and levels. The core is to restore the temporal logic and impact relationship of fault propagation.

[0119] In this embodiment of the invention, firstly, the cloud computing layer uses long-term trend prediction information to perform multi-scale initialization of the digital twin model, and constructs corresponding virtual operating conditions based on the initialized digital twin model. Here, the multi-scale coverage includes multiple levels such as component level, equipment level, station level, and even regional power grid level. For example, if the long-term trend prediction information indicates that "the transformer winding has an insulation aging risk, and the initial partial discharge reaches 200pC", the cloud computing layer will initialize the insulation parameters of the winding to a value that matches the aging state in the component-level sub-model of the digital twin model, and set the initial intensity of the partial discharge. In the equipment-level sub-model, the overall heat dissipation parameters and electrical parameters of the transformer are adjusted synchronously to conform to the equipment operating state corresponding to the aging winding. In the station-level sub-model, the power flow distribution of the associated bus and the protection settings of adjacent circuit breakers are initialized according to the transformer topology connection relationship, ensuring that the constructed virtual operating conditions can not only restore the predicted initial fault state, but also reflect the associated operating characteristics of each level of the substation under this state, laying a realistic foundation for subsequent simulations.

[0120] Subsequently, the cloud computing layer performs multi-physics coupling simulation based on the virtual operating conditions using the initialized digital twin model, obtaining multi-physics coupling simulation data: the multi-physics field covers electromagnetic field, temperature field, flow field and mechanical field—the electromagnetic field simulation calculates the electromagnetic induction intensity and eddy current loss at the fault point (such as aging winding), and inputs the loss data as a heat source into the temperature field; the temperature field simulation calculates the temperature distribution changes of the winding, iron core and transformer oil according to the heat conduction law, and then inputs the temperature gradient data into the flow field to simulate the convective heat transfer process of the transformer oil; the mechanical field combines the electromagnetic force generated by the electromagnetic field and the thermal stress generated by the temperature field to calculate the deformation and stress distribution of the winding, forming a coupling closed loop of interaction between the various physical fields. During the simulation, the cloud computing layer receives real-time data on the device status (such as transformer oil top temperature and winding DC resistance) collected by the perception layer through the data interface, calculates the error between the simulation results and the real-time data, and immediately locates the physical field dominated by the error (such as the temperature field error originating from inaccurate heat dissipation parameters) by adjusting the key parameters of the physical field (such as correcting the thermal conductivity of the transformer oil) and resubmitting it into the simulation process until the error between the simulation results and the real-time data converges to a reasonable range. The final output multi-physics dynamic evolution data is the multi-physics coupled simulation data.

[0121] Next, the cloud computing layer uses the initialized digital twin model to perform fault path propagation simulation based on multiphysics coupling simulation data, obtaining fault propagation path evolution data. The simulation process starts from the level where the fault initially occurs (such as winding insulation aging at the component level) and tracks the diffusion of fault characteristics to other levels. For example, an increase in partial discharge in the winding leads to a rise in temperature, which first causes the overall oil temperature of the transformer to exceed the standard at the equipment level, and then causes fluctuations in the bus voltage connected to the transformer at the station level. If the fluctuation exceeds the threshold, it will also affect the load distribution at the regional power grid level. The simulation records key information for each propagation step, including the time node of propagation (such as oil temperature exceeding the standard 30 minutes after partial discharge), the name and status change of the affected equipment (such as bus voltage dropping from 110kV to 105kV), and key triggering conditions during the propagation process (such as oil temperature exceeding 85℃ triggering cooler overload protection), forming a complete dynamic evolution dataset of the propagation path, i.e., fault propagation path evolution data.

[0122] Subsequently, the cloud computing layer uses the initialized digital twin model to simulate the operation and maintenance scheme based on the fault propagation path evolution data, and obtains the operation and maintenance scheme simulation data. The cloud computing layer will simulate and execute various preset operation and maintenance schemes in the digital twin model (such as "emergency shutdown for winding maintenance", "live replacement of cooler", "current limiting operation to the planned power outage window"). During the simulation, the cloud computing layer will adjust the status parameters of the corresponding equipment in the model according to the operation process of each scheme (such as simulating shutdown to disconnect the transformer power supply connection, simulating maintenance to update the winding insulation parameters), and then rerun the multiphysics coupling simulation to observe whether the fault propagation terminates and whether the equipment status returns to normal. Meanwhile, the implementation effects of each scheme are quantitatively evaluated: the fault elimination efficiency is measured by the "time for the fault to be completely terminated after the scheme is implemented" (e.g., emergency repair requires 4 hours, current-limited operation requires 24 hours); the economic cost covers spare parts consumption costs, labor costs and power outage losses (e.g., spare parts cost of 50,000 yuan for emergency repair, power outage loss of 200,000 yuan); the operational risk is evaluated by "whether new anomalies occur in the power grid during the scheme implementation period" (e.g., whether the bus voltage is continuously low during current-limited operation) and "fault recurrence probability" (e.g., 5% probability of fault recurrence within 1 month after cooler replacement). The final quantitative evaluation results of each scheme are the operation and maintenance scheme simulation data.

[0123] Finally, the cloud computing layer integrates fault propagation path evolution data and operation and maintenance scheme simulation data to obtain fault simulation results. During the integration process, the propagation path data is first correlated with the evaluation results (e.g., although a certain scheme has high fault elimination efficiency, it has high economic cost and high operational risk; its applicable scenarios are determined by combining the fault impact range in the propagation path). Then, the evaluation results of all schemes are prioritized according to the three dimensions of "efficiency-cost-risk", and the advantages and disadvantages of each scheme are marked (e.g., scheme A is efficient but costly, scheme B is costly but time-consuming). The final integrated information is the fault simulation result.

[0124] The present invention provides a substation panoramic intelligent monitoring system, wherein multi-physics coupling simulation includes: matching corresponding multi-physics configuration parameters according to virtual operating conditions; performing multi-physics iterative coupling using the multi-physics configuration parameters as input to the initialized digital twin model to obtain initial simulation results; calculating the dynamic error between the initial simulation results and equipment status data; dynamically adjusting the multi-physics configuration parameters based on the dynamic error until the dynamic error converges, and then using the current initial simulation results as multi-physics coupling simulation data.

[0125] Multiphysics configuration parameters: These refer to the boundary conditions and initial values ​​of each coupled physical field (electromagnetic field, temperature field, flow field, mechanical field) that match the virtual operating conditions. The boundary conditions reflect the constraints of the equipment's operating environment, and the initial values ​​match the initial state of the fault. Multiphysics iterative coupling: This refers to the simulation process in the digital twin model where, according to preset inter-field coupling rules, the state changes of individual physical fields are iteratively calculated and transmitted to associated physical fields, achieving dynamic interaction among the physical fields. Initial simulation results: These refer to the dynamic evolution data of each physical field output by the multiphysics iterative coupling calculation, without error calibration. Dynamic error: This refers to the difference between the initial simulation results and the real-time equipment state data collected by the sensing layer. It is real-time and updates synchronously as the simulation progresses. Parameter adjustment: This refers to the process of optimizing and correcting the key configuration parameters of the dominant error physical field based on the sensitivity analysis of dynamic error. The parameter adjustment range does not violate the physical properties of the equipment. Error convergence: This refers to the state where the dynamic error is stabilized within a preset threshold range through repeated execution of the "parameter adjustment-iterative coupling-error calculation" process.

[0126] In this embodiment of the invention, firstly, the cloud computing layer matches the corresponding multi-physics configuration parameters according to the virtual operating conditions: the virtual operating conditions correspond to the initial state of the fault in the long-term trend prediction information (such as "partial discharge of transformer windings", "overheating of circuit breaker contacts", etc.), and the matched multi-physics configuration parameters are specifically the boundary conditions and initial values ​​of each coupled physical field (electromagnetic field, temperature field, flow field, mechanical field). Taking the electromagnetic field as an example, the boundary condition is set as the electromagnetic shielding property of the metal shell of the equipment, and the initial value is set according to the initial state of the fault to the electromagnetic induction intensity and current density of the fault point; the boundary condition of the temperature field refers to the heat dissipation environment of the equipment (such as air convection velocity, cooler operating status), and the initial value is substituted into the initial temperature deviation of the fault point (such as 10°C higher than the normal operating temperature); the boundary condition of the flow field (such as the flow of transformer oil) is set as the geometric constraint of the oil tank, and the initial value is determined according to the oil temperature gradient around the fault point to determine the initial distribution of the flow velocity; the boundary conditions and initial values ​​of the mechanical field are also set in accordance with the actual operating characteristics of the equipment. Through this matching process, it is ensured that the initial state of each physical field can accurately reflect the fault characteristics of the virtual operating conditions.

[0127] Subsequently, the cloud computing layer uses multiphysics configuration parameters as input to the initialized digital twin model to perform multiphysics iterative coupling to obtain initial simulation results. Specifically, it calls the multiphysics coupling simulation engine and performs iterative coupling calculations according to preset inter-field coupling rules (based on the physical operation mechanism of power equipment). For example, eddy current losses generated by the electromagnetic field in the equipment conductors will be directly transmitted to the temperature field as a heat source, driving up the temperature around the fault point; the temperature gradient formed by the temperature field will cause density differences in the medium (such as transformer oil) in the flow field, thereby generating natural convection, which will carry away heat and feed back into the heat dissipation calculation of the temperature field; at the same time, the thermal expansion of equipment components caused by the temperature field and the electromagnetic force generated by the electromagnetic field will be superimposed on the mechanical field, affecting the deformation state of the equipment structure. The iterative coupling calculation is a dynamic loop process: the simulation engine first calculates the state change of a single physical field, and then, according to the coupling rules, transmits the calculation result of that field to the associated physical field, updates the parameters of the associated field, and performs the next round of calculation. This process is repeated iteratively to gradually simulate the dynamic changes of each physical field over time, and finally generates initial simulation results containing the spatial distribution and temporal evolution of each physical field.

[0128] Next, the cloud computing layer calculates the dynamic error between the initial simulation results and the equipment status data collected by the perception layer. During the simulation, the cloud computing layer continuously acquires the equipment status data fed back by the perception layer through the real-time data interface (such as the real-time temperature of the transformer winding collected by the infrared sensor, the real-time stress of the equipment shell collected by the strain gauge, and the partial discharge signal intensity collected by the ultrasonic sensor). It compares these real-time data with the initial simulation results of the same period point by point, and calculates the dynamic error through the preset error calculation algorithm (such as root mean square error and mean absolute error). This error is real-time and will be updated synchronously as the simulation progresses. For example, if the simulated oil temperature is 75°C at a certain moment, and the actual collected oil temperature is 72°C, the dynamic error is 3°C. The error value will be adjusted synchronously as the simulation and actual states change.

[0129] Finally, the cloud computing layer dynamically adjusts the multiphysics configuration parameters based on the dynamic error until the dynamic error converges, and uses the current initial simulation results as multiphysics coupled simulation data: when the dynamic error exceeds the preset threshold (e.g., the temperature field error threshold is set to 2℃ and the mechanical field stress error threshold is set to 5MPa), the correction mechanism is immediately triggered—first, a sensitivity analysis of the dynamic error is performed to calculate the error contribution of each physics field (i.e., the proportion of the error of that physics field to the total dynamic error), and the physics field that dominates the error is located (e.g., if 60% of the total error comes from the simulation deviation of the temperature field, then the temperature field is the dominant error physics field); then, optimization algorithms (e.g., Bayesian optimization, gradient descent algorithm) are used to adjust the key configuration parameters of the dominant error physics field (taking the temperature field as an example, if the error originates from inaccurate simulation of the heat dissipation effect, then parameters such as the heat dissipation coefficient and the heat exchange efficiency of the cooler are adjusted, and the adjustment process will constrain the parameter range to ensure that it does not violate the physical properties of the equipment).

[0130] After adjustment, the multiphysics iterative coupling calculation is restarted, and the dynamic error is calculated again. If the error still does not converge to the preset threshold, the process of "sensitivity analysis - parameter adjustment - coupling calculation" is repeated until the dynamic error stabilizes within the threshold range. Finally, the dynamic evolution process of the physical fields output by the cloud computing layer after real-time data calibration (including accurate data of the changes of each physical field over time and the coupling relationship between physical fields) is the multiphysics coupling simulation data of this system, providing reliable physical field data support for subsequent fault propagation simulation and operation and maintenance scheme evaluation.

[0131] This invention provides a substation panoramic intelligent monitoring system. The edge computing layer is further used for: node parsing of dynamic topology evolution data and determining a model topology pruning strategy based on the node parsing results; deviation parsing of equipment state simulation deviation values, retaining the dominant features of high-sensitivity equipment deviations associated with deviation rates exceeding a preset deviation rate threshold, and obtaining a set of dominant features of high-sensitivity equipment deviations; removing redundant input features from the secondary lightweight anomaly detection model that do not belong to the set of dominant features of high-sensitivity equipment deviations, and adjusting the model parameter weights associated with the set of dominant features of high-sensitivity equipment deviations, to obtain an anomaly detection model with optimized input features and parameters; performing topology pruning on the anomaly detection model with optimized input features and parameters using a model topology pruning strategy; using the topology-pruned anomaly detection model as a new secondary lightweight anomaly detection model, and then jumping to execute the step of inputting equipment state data into the secondary lightweight anomaly detection model for feature extraction to obtain equipment state features.

[0132] Node parsing: This refers to the process by which the edge computing layer analyzes the dynamic evolution data of the topology, identifies core and redundant device nodes, and formulates a model topology pruning strategy. Model topology pruning strategy: This refers to the rules for pruning and retaining model topology links based on the node parsing results. The core is to eliminate redundant computations and retain critical links. Deviation parsing: This refers to the process by which the edge computing layer decomposes and analyzes the deviation values ​​of device state simulations to identify highly sensitive devices and their corresponding dominant deviation features. Preset deviation rate threshold: This refers to a pre-set critical deviation rate value used to determine whether a device is highly sensitive and whether a feature is a dominant deviation feature. Dominant deviation feature of highly sensitive devices: This refers to the device state feature among highly sensitive devices (devices with deviation values ​​exceeding the preset threshold) that accounts for the highest proportion of the overall deviation and contributes the most to anomaly detection. Set of dominant deviation features of highly sensitive devices: This refers to the feature set composed of the dominant deviation features of all highly sensitive devices, providing a basis for optimizing model input features. Redundant input features: In the secondary lightweight anomaly detection model, these refer to input features that do not belong to the dominant feature set of high-sensitivity equipment deviations and have low contribution to anomaly detection (such as secondary state features of non-high-sensitivity equipment). Anomaly detection model with optimized input features and parameters: This refers to the model with improved feature specificity and computational efficiency obtained after removing redundant input features and adjusting the parameter weights corresponding to the dominant features of high-sensitivity equipment deviations. Topology pruning: This refers to the process of deleting computational links corresponding to redundant nodes in the anomaly detection model with optimized input features and parameters, based on a model topology pruning strategy.

[0133] In this embodiment of the invention, firstly, the edge computing layer performs node parsing on the topology dynamic evolution data output by the digital twin model to determine the model topology pruning strategy: the topology dynamic evolution data includes the time-series change records of device connection relationships (such as the adjustment of circuit breaker opening and closing status caused by switching operations, and the change of connection links caused by line commissioning and decommissioning), real-time power flow distribution information (input and output power values ​​of each device node, power transmission efficiency of the connection edges between nodes), and potential fault propagation links based on long-term trend prediction annotations. During the analysis process, the edge computing layer first organizes the data to form a complete topology information framework, and then identifies the core nodes and redundant nodes in the current topology through multi-dimensional indicators. When identifying core nodes, the following criteria are used: power flow contribution (the proportion of node transmission power to the total power of the entire station), fault propagation criticality (whether the node is located on the potential fault propagation path and the scope of its influence), and equipment functional importance (whether it is a core device of the substation). Nodes with high contribution and wide influence are identified as core nodes (which can also be divided into primary cores such as main transformers and buses, and secondary cores such as feeder circuit breakers). When identifying redundant nodes, the following criteria are used: equipment operating status (whether it is out of service or on standby), power flow ratio (long-term lower than the threshold and without fluctuation), and fault path correlation (not on the potential fault propagation path and not a core device). Redundant nodes that do not play a critical role in monitoring are selected. Based on this identification result, the edge computing layer determines the model topology pruning strategy: the core nodes retain the complete computing links, the first-level core nodes maintain full feature calculation and connection edge weight update, and the second-level core nodes can appropriately simplify the calculation of non-critical features; for redundant nodes, their corresponding model computing links are completely pruned, including the node feature input channel and the connection edge weight calculation module with other nodes.

[0134] Next, the edge computing layer analyzes the equipment state simulation deviation values ​​output by the digital twin model to obtain the set of dominant deviation features for highly sensitive equipment. The equipment state simulation deviation value is the difference between the equipment state parameters simulated by the digital twin model (such as simulated transformer oil temperature and circuit breaker opening / closing time) and the actual equipment state data collected by the sensing layer. The edge computing layer first calculates the overall deviation value for each device, marking devices with deviation values ​​exceeding a preset threshold (e.g., a deviation threshold of 8% for a certain type of equipment) as highly sensitive equipment. Then, it breaks down the deviations of various state characteristics of the highly sensitive equipment, statistically analyzes the proportion of each type of deviation to the overall deviation, and identifies the feature with the highest proportion (e.g., the oil temperature deviation of a highly sensitive transformer accounts for 65% of the overall deviation), determining it as the dominant deviation state feature. Finally, the edge computing layer retains these dominant deviation features of highly sensitive equipment, forming the set of dominant deviation features for highly sensitive equipment.

[0135] Subsequently, the edge computing layer removes redundant input features from the secondary lightweight anomaly detection model that do not belong to the dominant feature set of the high-sensitivity equipment deviation: features with low deviation values ​​and small contributions to anomaly detection are deleted (such as an equipment with an environmental wind speed deviation of only 3% and low correlation with anomalies, i.e., no longer used as model input), retaining only the dominant deviation features of the high-sensitivity equipment and the key state features of the core equipment; at the same time, the model parameter weights associated with the dominant deviation feature set of the high-sensitivity equipment are adjusted, such as increasing the parameter weight of the oil temperature feature by 20% to enhance the model's sensitivity to changes in such features, while freezing the parameters corresponding to small deviation and non-dominant features to avoid unnecessary parameter update calculations, thereby obtaining the anomaly detection model with optimized input features and parameters.

[0136] Subsequently, the edge computing layer adopts the aforementioned model topology pruning strategy to prune the anomaly detection model after optimizing the input features and parameters: according to the pruning strategy, the computation branches corresponding to redundant nodes are deleted in the model structure, and the process of collecting state data and extracting features from redundant nodes is terminated; for core nodes, the computation links are retained hierarchically, with first-level core nodes maintaining the input and computation of all features, and second-level core nodes retaining only the computation of key features such as power and temperature, simplifying the processing of non-key features such as environmental humidity.

[0137] Finally, the edge computing layer uses the topology-pruned anomaly detection model as a new, secondary lightweight anomaly detection model and verifies its performance: by using real-time device status data collected at the edge, it checks whether the model's inference time meets the edge's computing power requirements (e.g., controlled within 100 milliseconds), and confirms whether the anomaly detection accuracy is maintained at a preset standard (e.g., not less than 90%). If the standard is not met, the pruning ratio or parameter weights are fine-tuned until the model simultaneously meets the requirements of lightweightness and accuracy. Then, it jumps to the step of "inputting the device status data into the secondary lightweight anomaly detection model for feature extraction to obtain device status features".

[0138] This invention provides a substation panoramic intelligent monitoring system. Node parsing includes: quantitatively evaluating the equipment connection nodes corresponding to the model topology based on topology dynamic evolution data, and identifying core equipment nodes and redundant equipment nodes; classifying the core equipment nodes into hierarchical levels based on the quantitative evaluation results of the core equipment nodes; and formulating a model topology pruning strategy based on the hierarchical classification results and the redundant equipment node identification results. Specifically, the model topology pruning strategy involves full-link pruning of the model topology links corresponding to redundant equipment nodes, while optimizing and retaining the model topology links corresponding to core equipment nodes.

[0139] Quantitative Assessment: This refers to the evaluation process where the edge computing layer scores equipment connection nodes based on topology dynamic evolution data using multi-dimensional indicators (such as power flow contribution, fault propagation criticality, and equipment functional importance). Core Equipment Nodes: These are equipment nodes (such as main transformers and buses) that achieve a preset threshold in quantitative assessment and play a crucial role in substation power transmission, fault propagation, and functional implementation. Redundant Equipment Nodes: These are equipment nodes that simultaneously meet the following criteria: out of service / standby status, power flow proportion below a low threshold, and not on potential fault propagation paths; they have no critical role in monitoring. Hierarchical Classification: This refers to the process of classifying core equipment nodes into different priority levels (such as primary core and secondary core) based on their quantitative assessment scores and performance of various indicators. Nodes with the highest comprehensive score and directly related to the entire station's power supply (such as main transformers and buses) are classified as primary cores, while nodes undertaking regional power supply or equipment connection functions (such as feeder circuit breakers) are classified as secondary cores. Different levels correspond to differentiated retention strategies for subsequent model topology links. Full-link pruning: This refers to the complete removal of redundant device nodes from the model topology links. This includes deleting node feature input channels and connection edge weight calculation modules, while simultaneously stopping data acquisition and feature extraction at that node, thus completely eliminating redundant computing power consumption. Optimized retention: This refers to the differentiated retention of core device nodes from the model topology links—first-level core nodes retain the complete computing link and full feature processing, while second-level core nodes retain core feature calculations and simplify non-critical feature processing, reducing computing power requirements while ensuring monitoring accuracy.

[0140] In this embodiment of the invention, firstly, the edge computing layer establishes a complete set of indexes for determining core device nodes and redundant device nodes, providing a standardized basis for subsequent quantitative evaluation:

[0141] The criteria for determining core device nodes include three categories:

[0142] Power flow contribution: measures the proportion of power transmitted by a node in the current topology to the total power of the entire station. The higher the proportion, the more critical the role of the node in energy transmission.

[0143] Fault propagation criticality: This is assessed by analyzing whether a node is on a potential fault propagation path and the scope of its impact on downstream equipment after a node failure (such as the number of affected devices and the power supply area involved). Nodes on the critical path with a wide impact range score higher.

[0144] Equipment functional importance: The score is based on the core functional positioning of the equipment in the substation (such as the main transformer undertaking the core function of voltage transformation, and the busbar being the power collection hub). The more core the function of the equipment, the higher the score of the corresponding node.

[0145] The criteria for determining redundant device nodes include three categories:

[0146] Operational status indicators: Monitor whether the equipment corresponding to the node is in a shutdown, standby, or idle state;

[0147] Power flow ratio indicator: The long-term power transmission ratio of a statistical node. If it is consistently below a low threshold and does not fluctuate significantly, it tends to be redundant.

[0148] Fault path correlation index: Determines whether a node is not on any potential fault propagation path and is not a core functional device.

[0149] Subsequently, the edge computing layer quantitatively evaluates the device connection nodes corresponding to the model topology based on the topology dynamic evolution data (including real-time power distribution, device connection relationship change records, and potential fault propagation links). This involves extracting specific parameters for each device connection node from the topology dynamic evolution data—such as obtaining the real-time transmission power and percentage of a node from the power flow distribution, confirming the current connection status of a node from the connection relationship change records, and determining whether a node is a critical link from the fault propagation link. Then, each node is scored according to the aforementioned indicator system (e.g., 30 points for power flow contribution exceeding 20%, 40 points for being on the main fault propagation path, and 30 points for being a core functional device). These scores are then accumulated to form a comprehensive score for the node. Nodes with scores reaching or exceeding a set threshold (e.g., 60 points) are identified as core device nodes; nodes that simultaneously meet the following criteria—being out of service, having a power flow percentage below 3%, and not being on a fault propagation path—are identified as redundant device nodes.

[0150] Next, the edge computing layer classifies the core equipment nodes according to the quantitative evaluation results: the classification is based on the comprehensive score of the core equipment nodes and the specific performance of each indicator. The nodes with the highest comprehensive score (e.g., 80 points and above) and a power flow contribution of more than 20%, which are directly related to the power supply of the entire station (e.g., main transformers, busbars) are classified as first-level core equipment nodes. These nodes are the core hubs of substation operation, and any abnormality may cause a wide range of impacts. The nodes with a comprehensive score between 60 and 80 points, a power flow contribution between 10% and 20%, and which mainly undertake regional power supply or equipment connection functions (e.g., feeder circuit breakers, sectionalizing switches) are classified as second-level core equipment nodes. Their abnormal impact range is relatively limited, but they still need to be closely monitored.

[0151] Finally, based on the hierarchical partitioning results of core device nodes and the identification results of redundant device nodes, the edge computing layer formulates a model topology pruning strategy:

[0152] A full-link pruning strategy is implemented for the model topology links corresponding to redundant device nodes: the computational branches corresponding to redundant device nodes are completely removed from the network model, including the feature input channels of the node and the weight calculation modules of the connection edges with other nodes. At the same time, the real-time collection and feature extraction of the state data of redundant device nodes are stopped, and the ineffective computing power consumption caused by redundant nodes is completely eliminated.

[0153] An optimization and retention strategy is implemented for the model topology links corresponding to core device nodes: Level 1 core device nodes retain the complete computing links, maintain the input and processing of all state features (such as electrical parameters, temperature, and mechanical characteristics), and ensure comprehensive monitoring of their operating status; Level 2 core device nodes retain the calculation of core features (such as parameters directly related to power supply safety such as current and voltage), while simplifying the processing flow of non-critical features (such as parameters with less impact such as ambient temperature and humidity), reducing some secondary calculation links, and appropriately reducing computing power requirements while ensuring monitoring accuracy.

[0154] Please see Figure 2 The present invention provides a monitoring method for a substation panoramic intelligent monitoring system, comprising: step 101, acquiring equipment status data and topology status data of the substation, as well as digital twin simulation data of the substation associated digital twin model; step 102, based on the topology status data and digital twin simulation data, performing secondary lightweighting of the lightweight anomaly detection model, and combining it with equipment status data to perform anomaly monitoring and determine short-term monitoring results; step 103, based on a pre-set AI intelligent model and digital twin model, using equipment status data, topology status data, and short-term monitoring results to perform status prediction and determine long-term monitoring results; step 104, based on the long-term monitoring results, determining an execution decision scheme.

[0155] In this embodiment of the invention, step 101 involves data acquisition by the perception layer of the monitoring system: by deploying electrical parameter sensors, physical state sensors, and environmental sensors, real-time data on equipment status such as oil temperature, current, and insulation status of core equipment such as transformers and circuit breakers are collected. Simultaneously, through standardized interfaces, the system connects to the substation SCADA system and relay protection devices to synchronously acquire topology status data such as equipment connection relationships and switch opening and closing status. Furthermore, the digital twin model associated with the substation continuously simulates topology changes and equipment operation, outputting topology dynamic evolution data such as connection relationship changes and power flow adjustments after switching operations, as well as equipment status simulation deviation values ​​such as the difference between simulated oil temperature and actual collected values. These two together constitute digital twin simulation data, providing complete multi-source data support for subsequent steps. Step 102, executed by the edge computing layer, involves first constructing an initial equipment connection diagram of the substation in real time by combining topology state data and topology dynamic evolution data. Then, dynamic weights are assigned to the connection edges based on the equipment state simulation deviation values ​​to obtain the target equipment connection diagram. Subsequently, the lightweight anomaly detection model is configured as a dynamic graph neural network structure. The topology and state mapping is completed using the target equipment connection diagram as the topology structure input and the equipment state data as the node feature input, resulting in a pre-configured dynamic graph neural network. Next, the topology dynamic evolution data is mapped to topology change features, and the equipment state simulation deviation values ​​are mapped to node feature deviation weights, adjusting the pre-configured structure. The graph convolution computation link (pruning redundant node links and retaining core node links) and node feature input dimension (retaining high-sensitivity features and eliminating redundant features) are dynamically adapted and optimized. Finally, a secondary lightweight anomaly detection model is obtained through topology and state association graph convolution processing. Then, the device status data is input into the model to extract device status features that integrate topology, device status, and time-series features. These features are input into a preset anomaly detection classifier or threshold judgment module to output short-term device warning signals such as "normal / mild anomaly / severe anomaly". At the same time, the key feature data that triggers the warning are matched. The two together constitute the short-term monitoring results to meet the needs of rapid response to sudden anomalies.Step 103 is executed by the cloud computing layer: First, device status data, topology status data, and short-term monitoring results are input into the dynamic feature fusion module to obtain a fused feature vector covering individual devices, topology relationships, and anomaly evolution characteristics. Then, combined with a preset fault causal knowledge graph, the causal reasoning module identifies the root cause and scope of the fault. After supplementing external operating variables, the data is input into the multi-scale prediction module, which outputs multi-scale prediction results and confidence intervals covering 1-30 days (fault risk escalation), 1-6 months (health degradation), and 6-12 months (remaining lifespan). These are integrated to obtain long-term trend prediction information. Subsequently, this long-term trend prediction information is used to perform multi-scale initialization of the digital twin model (covering the entire...). A virtual operating condition is constructed at the component-equipment-station-regional power grid level. The fault evolution process is reconstructed through multi-physics coupling simulation. Simulation errors are corrected by combining real-time data from the sensing layer to obtain multi-physics coupling simulation data. Based on this data, the fault propagation path is deduced, and operation and maintenance schemes such as "emergency maintenance / backup replacement" are simulated to quantitatively evaluate the fault elimination efficiency, economic cost, and operational risk, resulting in fault propagation path evolution data and operation and maintenance scheme simulation data. Finally, by integrating long-term trend prediction information, fault propagation data, and operation and maintenance scheme evaluation results, key information is prioritized and the optimal operation and maintenance scheme is selected to form a long-term monitoring result that includes equipment health trends, fault risk levels, recommended operation and maintenance schemes, and execution windows. Step 104 is executed by the execution control layer: The long-term monitoring results are prioritized according to the scope of the fault impact and the degree of urgency. Combined with the substation operation and maintenance resources (spare parts inventory, personnel scheduling) and grid constraints (no power outages during peak load), the long-term monitoring results are transformed into implementable decision-making plans, including emergency fault handling plans (such as "immediately arrange bus joint repairs") and preventive operation and maintenance plans (such as "complete the main transformer insulation test within 3 months"). At the same time, the execution time, steps and responsible parties of the plans are clearly defined to ensure that the monitoring results are transformed into actual operation and maintenance actions.

[0156] In step 102, the "node analysis based on topology dynamic evolution data + feature selection of equipment state simulation deviation values" achieves secondary model lightweighting, dynamically adapting to the scenario of complex substation equipment and frequent topology changes, solving the problem of insufficient adaptability of traditional fixed model architecture automated monitoring solutions. At the same time, step 102 compresses the model's computing power requirements by pruning redundant node links and simplifying non-critical feature calculations. Combined with the adaptability design of dynamic graph neural networks, it achieves low-latency real-time monitoring at the edge while ensuring the accuracy of anomaly detection, solving the technical problem of insufficient real-time response at the edge in existing solutions. In addition, step 103 achieves fault root cause tracing, long-term trend prediction, and quantitative evaluation of operation and maintenance solutions through the collaborative operation of pre-built AI intelligent models and digital twin models. Combined with the implementation decision-making solution in step 104, it fills the gap in long-term operation and maintenance planning in existing solutions, helping substations shift from "passive fault response" to "proactive risk prevention" operation and maintenance mode. It can also accurately match operation and maintenance resources with fault requirements through operation and maintenance solution simulation and priority decision-making, avoiding cost waste and grid risks caused by blind maintenance, and further improving the utilization efficiency of operation and maintenance resources.

[0157] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0158] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between apparatuses or units through some interfaces, and may be electrical, mechanical, or other forms.

[0159] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A panoramic intelligent monitoring system for substations, characterized in that, include: The perception layer is used to acquire equipment status data and topology status data of the substation, as well as digital twin simulation data of the digital twin model associated with the substation. An edge computing layer is used to perform secondary weight reduction on the lightweight anomaly detection model based on the topology state data and the digital twin simulation data, and to perform anomaly monitoring in conjunction with the device state data to determine short-term monitoring results. The cloud computing layer is used to predict the status based on the pre-built AI intelligent model and the digital twin model, using the device status data, the topology status data and the short-term monitoring results, and to determine the long-term monitoring results. The execution control layer is used to determine the execution decision scheme based on the long-term monitoring results.

2. The substation panoramic intelligent monitoring system according to claim 1, characterized in that, The digital twin simulation data includes topology dynamic evolution data and device state simulation deviation values. The edge computing layer is used for: The initial equipment connection diagram of the substation is constructed based on the topology status data and the topology dynamic evolution data; The initial device connection diagram is weighted according to the device state simulation deviation value to obtain the target device connection diagram. Configure the lightweight anomaly detection model as a dynamic graph neural network structure; Using the target device connection graph as the topology input and the device status data as the node feature input, the dynamic graph neural network structure is mapped to topology and state to obtain a pre-configured dynamic graph neural network structure. The topology dynamic evolution data and the equipment state simulation deviation value are used as inputs to the pre-configured dynamic graph neural network structure for dynamic adaptation and optimization, resulting in an optimized dynamic graph neural network structure. The optimized dynamic graph neural network structure is subjected to topology and state association graph convolution processing to obtain a secondary lightweight anomaly detection model; The device status data is input into the secondary lightweight anomaly detection model for feature extraction to obtain device status features; The device status characteristics are input into a preset anomaly detection classifier or threshold judgment module, and a short-term warning signal for the device is output. The device status features associated with the short-term warning signal of the device are matched as key feature data; The short-term monitoring results include the short-term warning signals of the equipment and the key feature data used to interpret the short-term warning signals.

3. The substation panoramic intelligent monitoring system according to claim 2, characterized in that, The dynamic adaptation optimization includes: The topological dynamic evolution data is mapped to the topological change features of the pre-configured dynamic graph neural network structure; The device state simulation deviation value is mapped to the node feature deviation weight of the pre-configured dynamic graph neural network structure; The graph convolution computation links of the pre-configured dynamic graph neural network structure are adjusted according to the topology change characteristics, and the node feature input dimensions of the pre-configured dynamic graph neural network structure are adjusted according to the node feature deviation weights to obtain an optimized dynamic graph neural network structure.

4. The substation panoramic intelligent monitoring system according to claim 1, characterized in that, The cloud computing layer is used for: Based on a pre-built AI intelligent model, the device status data, the topology status data, and the short-term monitoring results are used to perform trend prediction and determine long-term trend prediction information. The long-term trend prediction information is input into the digital twin model to perform fault simulation, and the fault simulation results are obtained. The fault simulation results include fault propagation path evolution data and operation and maintenance scheme simulation data. By integrating the long-term trend prediction information, the fault propagation path evolution data, and the operation and maintenance scheme simulation results, long-term monitoring results are obtained.

5. The substation panoramic intelligent monitoring system according to claim 4, characterized in that, The pre-built AI intelligent model includes a dynamic feature fusion module, a causal reasoning module, and a multi-scale prediction module. The trend prediction includes: Obtain the external operating variables of the substation; The device status data, the topology status data, and the short-term monitoring results are input into the dynamic feature fusion module for feature fusion to obtain a fused feature vector. The fused feature vector and the preset fault causal knowledge graph are input into the causal reasoning module to perform fault source tracing and deduction, and the fault causal analysis results are obtained. The fused feature vector, the fault causal analysis results, and the external operating variables are input into the multi-scale prediction module for time-series dimension extrapolation to obtain the multi-scale operating state prediction uncertainty results. By integrating the results of the fault causal analysis and the results of the multi-scale operational state prediction uncertainty, long-term trend prediction information is obtained.

6. The substation panoramic intelligent monitoring system according to claim 4, characterized in that, The fault simulation includes: The long-term trend prediction information is used to initialize the digital twin model at multiple scales, and a virtual working condition corresponding to the long-term trend prediction information is constructed based on the initialized digital twin model. Based on the virtual working conditions, multiphysics coupling simulation is performed using the initialized digital twin model to obtain multiphysics coupling simulation data. Based on the multiphysics coupling simulation data, the fault propagation path is deduced using the initialized digital twin model to obtain the fault propagation path evolution data. Based on the fault propagation path evolution data, the operation and maintenance scheme is simulated using the initialized digital twin model to obtain the operation and maintenance scheme simulation data; By integrating the fault propagation path evolution data and the operation and maintenance scheme simulation data, fault simulation results are obtained.

7. The substation panoramic intelligent monitoring system according to claim 6, characterized in that, The multiphysics coupling simulation includes: Match the corresponding multiphysics configuration parameters according to the virtual working conditions; The multiphysics configuration parameters are input into the initialized digital twin model to perform multiphysics iterative coupling and obtain initial simulation results; Calculate the dynamic error between the initial simulation results and the device status data; The multiphysics configuration parameters are dynamically adjusted based on the dynamic error until the dynamic error converges, and then the current initial simulation results are used as multiphysics coupling simulation data.

8. The substation panoramic intelligent monitoring system according to claim 2, characterized in that, The edge computing layer is also used for: The topology dynamic evolution data is parsed to identify nodes, and a model topology pruning strategy is determined based on the node parsing results. The deviation values ​​of the equipment state simulation are analyzed for deviation. The dominant deviation features of high-sensitivity equipment with deviation rates exceeding a preset deviation rate threshold are retained to obtain a set of dominant deviation features of high-sensitivity equipment. Redundant input features that do not belong to the set of dominant features of high-sensitivity equipment deviation in the secondary lightweight anomaly detection model are removed, and the model parameter weights associated with the set of dominant features of high-sensitivity equipment deviation are adjusted to obtain an anomaly detection model with optimized input features and parameters. The aforementioned model topology pruning strategy is used to perform topology pruning on the anomaly detection model after optimizing the input features and parameters; The anomaly detection model after topology trimming is used as a new secondary lightweight anomaly detection model, and the process jumps to execute the step of inputting the device status data into the secondary lightweight anomaly detection model for feature extraction to obtain device status features.

9. The substation panoramic intelligent monitoring system according to claim 8, characterized in that, The node parsing includes: Based on the dynamic evolution data of the topology, the device connection nodes corresponding to the model topology are quantitatively evaluated to identify core device nodes and redundant device nodes. Based on the quantitative evaluation results of the core device nodes, the core device nodes are hierarchically divided. Based on the hierarchical division results and the redundant device node identification results, a model topology pruning strategy is formulated. Specifically, the model topology pruning strategy involves pruning the entire model topology link corresponding to the redundant device node and optimizing and retaining the model topology link corresponding to the core device node.

10. A monitoring method applied to the substation panoramic intelligent monitoring system according to any one of claims 1-9, characterized in that, include: Acquire equipment status data and topology status data of the substation, as well as digital twin simulation data of the substation's associated digital twin model; Based on topology status data and digital twin simulation data, the lightweight anomaly detection model is further lightweighted, and anomaly monitoring is performed in conjunction with equipment status data to determine short-term monitoring results. Based on pre-built AI intelligent models and digital twin models, status prediction is performed using equipment status data, topology status data, and short-term monitoring results to determine long-term monitoring results; Based on long-term monitoring results, an implementation decision plan is determined.

Images