Power grid master station system health state diagnosis method, system, device and storage medium

Abstract

The application relates to a power grid master station system health state diagnosis method and device, computer equipment and a storage medium. The method comprises the following steps: acquiring multi-source heterogeneous data and a mapping relationship between sensors and equipment; inputting the multi-source heterogeneous data into a physical information graph neural network according to the mapping relationship, outputting a multi-modal feature matrix after time and space alignment, and determining the edge weight between nodes of the physical information graph neural network according to the distance and the electrical connection type of the sensor-bound equipment in a physical topology graph; inputting the multi-modal feature matrix into a causal graph model, taking a device state variable as a node and a causal relationship as a directed edge, and carrying a causal strength value and a time delay parameter on each directed edge; when a target state variable deviates from a normal operation interval, performing counterfactual reasoning to determine a root cause node; and outputting a treatment instruction sequence according to the root cause node type and a causal path. The application can accurately locate a fault root cause node, and improve diagnosis accuracy and treatment pertinence.

Description

Technical Field

[0001] This application relates to the field of power system automation technology, and in particular to a method, system, computer equipment and storage medium for diagnosing the health status of a power grid master station system. Background Technology

[0002] With the continuous expansion of the power grid and the large-scale integration of new energy sources, the number of substation equipment and the amount of data are increasing dramatically. The power grid master station system needs to process multi-source heterogeneous data, including oil temperature, load, partial discharge, switch status, and inspection images. Traditional master station system health status diagnosis methods typically employ threshold-based rule-based judgment or correlation analysis models based on historical data to identify and alert on equipment anomalies. These methods can detect whether equipment status parameters deviate from normal operating ranges and output fault types or risk levels.

[0003] However, traditional methods rely on correlation analysis between variables, making it difficult to distinguish the correspondence between the causes and consequences of faults. For example, when oil temperature rises, it's impossible to determine whether the root cause is increased load or a cooling system malfunction, resulting in diagnostic results that only indicate the type of anomaly and failing to accurately pinpoint the root cause. Maintenance personnel need to manually trace the source using multi-source data, leading to low diagnostic efficiency and a high risk of ineffective measures due to misjudgment of the root cause, thus impacting the safety of power grid operation. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, device, computer equipment, and storage medium for diagnosing the health status of a power grid master station system that can accurately locate the root cause node of a fault, in order to address the above-mentioned technical problems.

[0005] Firstly, this application provides a method for diagnosing the health status of a power grid master station system. The method includes: Acquire multi-source heterogeneous data from substation equipment; Obtain the mapping relationship between sensors and devices; Based on the mapping relationship between the sensor and the device, the multi-source heterogeneous data is input into the physical information graph neural network, and the spatiotemporally aligned multimodal feature matrix is ​​output. The edge weights between nodes of the physical information graph neural network are determined by the distance between the device to which the sensor is bound in the physical topology graph and the electrical connection type between the bound devices. The multimodal feature matrix is ​​input into the causal graph model, which uses device state variables as nodes and causal relationships between variables as directed edges. Each directed edge carries a causal strength value and a time delay parameter. When the target state variable in the causal graph model deviates from the normal operating range, counterfactual reasoning is performed on the causal graph model to determine the root cause node. Based on the type of the root cause node and the causal path of the root cause node in the causal graph model, a sequence of disposal instructions is output.

[0006] In one embodiment, the causal graph model is constructed through the following steps: Based on knowledge of the physical mechanism of the equipment, the prior causal relationship between the state variables of the equipment is determined, and a prior causal framework is constructed. Historical operational data is acquired, and a causal discovery algorithm is used to correct the prior causal framework. The correction includes: estimating the causal strength of existing directed edges in the prior causal framework and deleting edges whose causal strength is lower than a preset strength threshold; searching for new edge candidate sets for directed edges that do not exist in the prior causal framework, and performing physical rationality verification on each edge in the new edge candidate set, retaining only the edges that pass the physical rationality verification. For each directed edge in the modified prior causal skeleton, the time delay parameter of each directed edge is estimated to obtain the causal graph model.

[0007] In one embodiment, performing counterfactual reasoning on the causal graph model includes: In the causal graph model, the target state variable is... The current outlier is denoted as ; make: in Candidate root dependent variable The normal range of values, To make the candidate root dependent variable The value is set to The target state variables were then calculated using the structural equations of each node. The counterfactual value; Will The corresponding candidate root dependent variable is determined as the root cause node.

[0008] In one embodiment, the step of inputting the multi-source heterogeneous data into a physical information graph neural network based on the mapping relationship between the sensor and the device, and outputting a spatiotemporally aligned multimodal feature matrix, includes: A graph structure is constructed based on the mapping relationship between the sensors and devices and the physical topology, where nodes represent sensor-device pairs. With nodes The initial edge weights between them are ,in For nodes With nodes The shortest path distance in the physical topology map to the devices corresponding to the sensors bound to the space. For nodes and nodes The electrical connection type code between the bound sensors and the corresponding devices. , and These are preset hyperparameters; Through nodes The corresponding sensor data is linearly transformed to obtain the initial hidden state vector. ; Execute on the graph structure Round-by-round message passing, in each round of message passing, nodes The hidden state is updated to ,in For nodes The set of neighboring nodes, For learnable weight matrix, This is a vector concatenation operation. For activation function, For the first Nodes in round-robin message passing The hidden state vector, For the first Nodes in round-robin message passing The hidden state vector; The hidden state vectors of each node after the last round of message passing are used as the final hidden state vectors of each node, and are reorganized into the multimodal feature matrix according to the mapping relationship between the sensor and the device.

[0009] In one embodiment, it further includes: When the causal graph model is updated, the causal strength value of each directed edge in the updated causal graph model is obtained; Update the weights of the edges between nodes in the physical information graph neural network based on the causal strength value; After the weights of the physical information graph neural network are updated, the steps of inputting the multi-source heterogeneous data into the physical information graph neural network to obtain the multimodal feature matrix and inputting the multimodal feature matrix into the causal graph model to perform counterfactual reasoning to determine the root cause node are re-executed.

[0010] In one embodiment, the step of outputting a sequence of disposal instructions based on the type of the root cause node and the causal path of the root cause node in the causal graph model includes: Establish a mapping relationship library between the types of the root cause nodes and the corresponding actions; Based on the outgoing edge direction of the root node in the causal graph model, the set of downstream variables affected by the root node is determined, and the degree coefficient of the influence of the root node on each downstream variable is calculated. The degree coefficient is determined by the product of the causal strengths on the causal path from the root node to the downstream variable. Generate the treatment instruction sequence, which includes root cause treatment instructions for the root cause node and preventive treatment instructions for downstream variables in the downstream variable set whose influence coefficient exceeds a preset influence threshold.

[0011] In one embodiment, the target state variable is calculated using the structural equations of each node. The counterfactual values ​​include: For any variable in the causal graph model Obtain the variable Parent node variable set ; According to the structural equation Calculate the variable The possible values ​​of , where It is a function fitted based on historical data. Noise term; From the candidate root dependent variables Begin by sequentially calculating the values ​​of each downstream variable along the direction of the directed edges in the causal graph model, until the calculated values ​​are obtained. .

[0012] Secondly, this application also provides a power grid master station system health status diagnosis system. The system includes: The data acquisition module is used to acquire multi-source heterogeneous data from substation equipment; The mapping acquisition module is used to acquire the mapping relationship between sensors and devices; The feature extraction module is used to input the multi-source heterogeneous data into the physical information graph neural network according to the mapping relationship between the sensor and the device, and output a spatiotemporally aligned multimodal feature matrix. The edge weights between nodes of the physical information graph neural network are determined by the distance between the sensor and the device in the physical topology graph and the electrical connection type between the bound devices. The causal reasoning module is used to input the multimodal feature matrix into the causal graph model. The causal graph model uses device state variables as nodes and causal relationships between variables as directed edges. Each directed edge carries a causal strength value and a time delay parameter. The counterfactual reasoning module is used to perform counterfactual reasoning on the causal graph model to determine the root cause node when the target state variable in the causal graph model deviates from the normal operating range. The instruction output module is used to output a sequence of disposal instructions based on the type of the root cause node and the causal path of the root cause node in the causal graph model.

[0013] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the aforementioned power grid master station system health status diagnosis method.

[0014] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the above-described method for diagnosing the health status of a power grid master station system.

[0015] The aforementioned power grid master station system health status diagnosis method, system, computer equipment, and storage medium, by constructing a physical information graph neural network, encodes the physical topological relationship between sensors and equipment as edge weights in the graph structure. This achieves spatiotemporal alignment and feature-level fusion of multi-source heterogeneous data in the physical space, providing physically interpretable fusion features for causal inference. Based on this, a causal graph model is established with equipment state variables as nodes and causal relationships as directed edges. Counterfactual reasoning is used to determine the candidate root cause variable that maximizes the intervention effect of restoring the target state variable to normal as the root cause node. This elevates fault diagnosis from correlation analysis to causal analysis, achieving precise location of the fault root cause. Based on the root cause node type and its propagation path in the causal graph, a dynamic sequence of handling instructions is output, avoiding ineffective treatments that only address the symptoms in traditional methods. This improves the accuracy of power grid master station system health status diagnosis and the targeted nature of fault handling, ensuring the safe operation of the power grid. Attached Figure Description

[0016] Figure 1 This is an application environment diagram of a power grid master station system health status diagnosis method in one embodiment; Figure 2 This is a flowchart illustrating a method for diagnosing the health status of a power grid master station system in one embodiment; Figure 3 This is a flowchart illustrating the steps involved in constructing a cause-effect graph model in one embodiment. Figure 4 This is a structural block diagram of a power grid master station system health status diagnosis system in one embodiment; Figure 5 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. First, to facilitate understanding of the technical solutions provided by the embodiments of this application, the background technology involved in the embodiments of this application will be described below.

[0018] With the continuous expansion of the power grid and the large-scale grid connection of new energy sources such as wind power and photovoltaics, the number of substation equipment has surged. The power grid master station system needs to process massive amounts of data from multiple sources, including online monitoring terminals, intelligent inspection devices, and remote monitoring and control terminals. These data types are diverse, including continuous time-series monitoring data such as oil temperature, load, partial discharge, and oil pressure; discrete status data such as circuit breaker opening and closing status and disconnector switch position; and unstructured data such as inspection images and inspection text records. Traditional master station system health status diagnosis methods typically use threshold-based rule judgment or correlation analysis models based on historical data to identify and alarm for equipment anomalies. Specifically, the system presets normal ranges for various equipment status parameters. When the real-time collected data exceeds the range, an alarm is triggered, and monitoring personnel manually determine the fault type based on the alarm information and their own experience. Some upgraded versions introduce deep learning-based classification models, which are trained using historical fault data to output the fault type or risk level.

[0019] However, the aforementioned traditional methods have significant technical limitations in practical applications. First, traditional methods essentially rely on correlation analysis between variables, that is, fitting the mapping relationship between input data and output labels through statistical learning or deep learning models. While this black-box model can determine "what is abnormal," it cannot reveal the causal chain of the abnormality. For example, when a transformer oil temperature rise is detected, it could be due to increased load, a cooling system malfunction, or increased losses caused by insulation aging. The corresponding handling strategies for these three scenarios are drastically different: the former requires adjusting load distribution, the middle requires activating the backup cooler, and the latter requires scheduling a power outage for maintenance. Traditional correlation models can only output "abnormal oil temperature" or "high risk level of excessively high oil temperature," failing to distinguish the above causal paths, resulting in diagnostic results remaining at the symptom level and making it difficult to accurately pinpoint the root cause of the fault.

[0020] Secondly, due to a lack of causal reasoning ability, maintenance personnel still need to manually trace the source of alarms by combining data from multiple sources, resulting in low diagnostic efficiency. Communication and repeated verification are required between different professional positions (maintenance, repair, and centralized control), making the emergency response process lengthy. More importantly, misjudging the root cause can lead to ineffective or even erroneous actions. For example, misjudging a cooling system malfunction causing oil temperature rise as an increase in load and implementing current limiting measures instead of activating the backup cooler not only fails to eliminate the root cause but may further restrict power supply capacity and affect the safe operation of the power grid.

[0021] Therefore, in order to break through the technical paradigm of traditional correlation analysis models, realize the autonomous mining of causal relationships from multi-source heterogeneous data, and accurately locate the root cause node of the fault, this embodiment provides a method, system, computer equipment and storage medium for diagnosing the health status of a power grid master station system.

[0022] The power grid master station system health status diagnosis method provided in this application embodiment can be applied to, for example... Figure 1 In the application environment shown, the monitoring terminal 102 communicates with the power grid master station server 104 via a network. The data storage system can store the multi-source heterogeneous data, causal graph model parameters, and diagnostic logs that the power grid master station server 104 needs to process. The data storage system can be integrated on the power grid master station server 104 or placed on the cloud or other network servers. The power grid master station server 104 acquires multi-source heterogeneous data from substation equipment and the mapping relationship between sensors and equipment. It inputs the multi-source heterogeneous data into a physical information graph neural network to obtain a multimodal feature matrix, and then inputs the multimodal feature matrix into the causal graph model. When the target state variable deviates from the normal operating range, it performs counterfactual reasoning to determine the root cause node and outputs a sequence of handling instructions based on the root cause node type and causal path. The monitoring terminal 102 is used to display diagnostic results and receive query instructions and confirmation instructions from maintenance personnel. The monitoring terminal 102 can be a workstation, personal computer, industrial tablet computer, or mobile terminal with communication capabilities in the substation control center; the power grid master station server 104 can be an independent control system server, a cloud server, or a server cluster composed of multiple servers. Furthermore, the data collection involved in this application embodiment has been authorized by the data owner and complies with relevant laws and regulations. Users can turn off data collection through the settings interface to protect their personal privacy.

[0023] Firstly, in the actual operation of the power grid master station system, traditional fault diagnosis methods based on threshold or correlation analysis can only output the fault type or risk level, failing to reveal the causal chain of the fault, making it difficult for maintenance personnel to locate the true root cause. Therefore, in one embodiment, such as... Figure 2 As shown, a method for diagnosing the health status of a power grid master station system is provided, and this method is applied to... Figure 1Taking the power grid master station server 104 as an example, the following steps are included: Step S1: Obtain multi-source heterogeneous data from substation equipment.

[0024] In this embodiment, the power grid master station server acquires multi-source heterogeneous data from substation equipment in real time through a data acquisition interface. This data includes, but is not limited to: time-series data of transformer top-level oil temperature collected by oil temperature sensors, load current data collected by current transformers, partial discharge data output by partial discharge monitoring devices, and oil pressure data collected by oil pressure sensors (these are continuous time-series monitoring data); circuit breaker opening and closing status, disconnector switch position status, and protection action signals (these are discrete status data); and infrared images captured by inspection robots or drones, and text records entered by inspection personnel (these are unstructured inspection data). The above data is transmitted to the master station server via different communication protocols (such as IEC 61850, Modbus, MQTT, etc.). The server performs timestamp alignment and format standardization on the raw data to form a dataset for subsequent processing. In practical applications, the data acquisition frequency can be dynamically adjusted according to the importance of the equipment. For example, it can be set to acquire data once per second for main transformers and once per minute for general line equipment; this embodiment does not limit this.

[0025] Step S2: Obtain the mapping relationship between the sensor and the device.

[0026] In this embodiment, each sensor is bound to a specific substation device upon deployment. The server maintains a mapping table that records the unique identifier of each sensor, its bound device name, device type, and the device's location coordinates in the physical topology diagram. The specific method for obtaining this mapping relationship can employ any existing technology, such as parsing it from a substation configuration file (e.g., an SCL file), or manual input by maintenance personnel through a human-machine interface; this embodiment does not limit this approach.

[0027] Step S3: Based on the mapping relationship between the sensor and the device, input the multi-source heterogeneous data into the physical information graph neural network, and output the spatiotemporally aligned multimodal feature matrix. The edge weights between nodes of the physical information graph neural network are determined by the distance between the sensor and the device in the physical topology graph and the electrical connection type between the devices.

[0028] In this embodiment, the physical information graph neural network is a type of graph neural network where nodes represent "sensor-device pairs." The edge weights between nodes are determined by the distance between the devices bound to the two sensors in the substation physical topology map and the electrical connection type between the devices. The server first constructs a physical topology map of the devices based on the substation's primary wiring diagram, and then constructs the specific graph structure of the graph neural network based on the sensor-device mapping relationship. The calculation method for the edge weights between nodes can employ any existing technology, such as a weighted combination of an exponential decay function based on physical distance and an electrical connection type encoding. The server inputs multi-source heterogeneous data as node features into the graph neural network. After multiple rounds of message passing, the hidden state vector of each node incorporates information from its physical neighbors, achieving spatiotemporal alignment. Finally, the server reassembles the final hidden state vectors of all nodes into a multimodal feature matrix, where rows correspond to sensor nodes and columns correspond to feature dimensions.

[0029] Step S4: Input the multimodal feature matrix into the causal graph model. The causal graph model uses device state variables as nodes and causal relationships between variables as directed edges. Each directed edge carries a causal strength value and a time delay parameter.

[0030] In this embodiment, the causal graph model is a directed graph where nodes represent device state variables (such as oil temperature, load, partial discharge, switch status, etc.), and directed edges represent causal relationships between variables. Each directed edge carries two parameters: causal strength (representing the expected change in the result variable when the causal variable changes by one unit) and a time delay parameter (representing how much time will pass after the causal variable changes before it affects the result variable). The causal graph model can be constructed using any existing technology, such as manual construction based on expert knowledge, data-driven causal discovery algorithms, or a combination of both. The server uses the multimodal feature matrix obtained in the previous step as the current observation value of each state variable in the causal graph model.

[0031] Step S5: When the target state variable in the causal graph model deviates from the normal operating range, perform counterfactual reasoning on the causal graph model to determine the root cause node.

[0032] In this embodiment, when the current value of a target state variable deviates from its normal operating range (this range can be obtained statistically from historical data, such as the mean ± 3 standard deviations, or set according to the parameters on the device nameplate), the server triggers counterfactual inference. The core idea of ​​counterfactual inference is: assuming that a candidate root cause variable is forcibly set to a normal value, and then observing whether the target variable will return to its normal value, the true root cause node is determined by maximizing the intervention effect. The specific calculation method can employ any existing technology, such as counterfactual calculation based on a structural causal model. The server ultimately outputs the root cause node.

[0033] Step S6: Output the sequence of disposal instructions based on the type of root cause node and the causal path of the root cause node in the causal graph model.

[0034] In this embodiment, after identifying the root cause node, the server matches the corresponding action from a pre-set action strategy library based on the node's type. Simultaneously, it considers the downstream impact path of the root cause node in the causal graph, generating a sequence of instructions that includes root cause actions and preventative actions. The specific implementation can employ any of the existing technologies, such as a method combining rule-based instruction mapping with causal path propagation.

[0035] Through the above steps, the method provided in this embodiment can automatically mine causal relationships from multi-source heterogeneous data, accurately locate the root cause of the fault, and output targeted handling instructions, thereby solving the technical problem that traditional methods can only identify abnormal phenomena but cannot locate the root cause, and improving the accuracy of the health status diagnosis of the power grid master station system and the effectiveness of fault handling.

[0036] The quality of causal graph models directly affects the accuracy of counterfactual reasoning. Traditional causal graph construction relies on human expert knowledge, which is inefficient and struggles to cover complex failure modes; while purely data-driven causal discovery algorithms are prone to generating false causal relationships. Therefore, in one embodiment, such as... Figure 3 As shown, the causal graph model is constructed through the following steps: Step S41: Based on the knowledge of the physical mechanism of the equipment, determine the prior causal relationship between the equipment state variables and construct the prior causal framework; Step S42: Obtain historical running data and use a causal discovery algorithm to correct the prior causal skeleton. The correction includes: estimating the causal strength of existing directed edges in the prior causal skeleton and deleting edges with causal strength lower than a preset strength threshold; searching for new edge candidate sets for directed edges that do not exist in the prior causal skeleton, and verifying the physical rationality of each edge in the new edge candidate set, retaining only the edges that pass the physical rationality verification. Step S43: For each directed edge in the corrected prior causal skeleton, estimate the time delay parameter of each directed edge to obtain the causal graph model.

[0037] In this embodiment, the server pre-stores an expert knowledge base based on the physical mechanisms of the power system. For example, it is known that "load current" affects "transformer oil temperature," "transformer oil temperature" affects "insulation aging rate," and "cooler fan speed" affects "transformer oil temperature," etc. The server represents these causal relationships as directed edges, constructing an initial causal skeleton graph. This skeleton graph only contains definite causal relationships determined according to the physical mechanisms and does not contain statistically spurious correlations. This skeleton graph can be stored in a graph database, with nodes representing variable names and edges representing directions.

[0038] The server acquires historical operating data of the substation (such as time-series data on oil temperature, load, and partial discharge over the past year) and uses a causal discovery algorithm to correct the prior framework. The causal discovery algorithm can be any existing technology, such as constraint-based PC algorithm, fraction-based NOTEARS algorithm, or function-based causal discovery algorithm; this embodiment does not limit the specific algorithm used.

[0039] The correction process includes two aspects: For directed edges already existing in the prior skeleton, the algorithm estimates the causal strength of the edge. The estimation method for causal strength can be any existing technique, such as standardized regression coefficients of linear regression, path coefficients in structural causal models, or causal measures based on conditional mutual information; this embodiment does not limit this. A preset strength threshold (e.g., 0.3) is set. If the estimated causal strength is lower than this threshold, the causal relationship is considered statistically insignificant and is removed from the skeleton.

[0040] For directed edges not present in the prior skeleton, the algorithm uses a causal discovery algorithm to search for possible candidate new edges. The search typically uses a conditional independence test: if two variables still have a dependency given other variables, then a direct causal relationship is likely. For each candidate new edge, the server calls a physical plausibility verification module to check whether the edge conforms to basic physical laws (such as energy conservation, matter conservation, etc.). Physical plausibility verification can employ any existing technology, such as rule-based verification, where rules are derived from physical formulas and industry standards; this embodiment does not limit this approach.

[0041] Furthermore, for each directed edge in the corrected skeleton, the server estimates its latency parameter using historical data. The latency parameter represents how long it takes for a change in the causal variable to affect the outcome variable. The estimation method can be any existing technique, such as choosing the lag order in Granger causality tests or calculating the maximum lag time of the cross-correlation function; this embodiment does not limit this approach. The latency parameter can be dynamically adjusted for different devices and different seasons.

[0042] Through the above steps, the final causal graph model is obtained. This model maintains the interpretability of the physical mechanism and corrects for omissions or errors through data-driven correction. The causal strength value and time delay parameter are both quantitative indicators, providing a reliable foundation for subsequent counterfactual reasoning.

[0043] After establishing the causal graph model, when an anomaly occurs in the target state variable, it is necessary to identify the true root cause from numerous candidate variables. Traditional correlation analysis cannot answer the causal question of "whether changing a certain variable will eliminate the anomaly." Therefore, in one embodiment, counterfactual reasoning is performed on the causal graph model, including: In the cause-effect graph model, the target state variable is... The current outlier is denoted as ; make: in Candidate root dependent variable The normal range of values, To select candidate root dependent variables The value is set to The target state variables were then calculated using the structural equations of each node. The counterfactual value; Will The corresponding candidate root dependent variable is determined as the root cause node.

[0044] Specifically, let the target state variable be... For example, "top oil temperature of the main transformer". The server detected this. The current real-time value is Furthermore, this value exceeds the preset normal operating range. (For example, the normal range is [50℃, 85℃], and the current temperature is 95℃). The candidate root dependent variable set contains all possible influences in the causal graphical model. Ancestor variables, such as "load current", "cooler fan speed", "ambient temperature", "insulation aging coefficient", etc. For each candidate root variable... The server obtains its normal value range. (For example, the normal range for load current is [0, 800A], and the normal range for cooler fan speed is [0, 1000rpm]). The normal range can be obtained from historical data statistics, such as the mean ± 3 times the standard deviation, or set according to the parameters on the equipment nameplate.

[0045] For candidate variables Server traversal or search optimization Each possible value in Perform the following counterfactual calculation: Assume Forced to be set as Other variables (except for) as well as The descendant variables are kept unchanged from the current observations. Then, using the structural equations in the causal graph model, the results are calculated layer by layer along the direction of the directed edges. Values ​​in the new counterfactual scenario .

[0046] Next, calculate the difference. This difference indicates that... Set as Then, the target variable The degree of movement in the normal direction; the larger the difference, the more significant the intervention effect. The server identifies the value that maximizes this difference. , recorded as Then, Corresponding candidate variables The root cause node is identified. For example, if the oil temperature returns to normal after the cooler fan speed is set to the normal value, but the oil temperature is still too high after other candidate variables (such as load current) are set to the normal value, then the difference corresponding to the cooler fan speed is the largest. Therefore, the root cause node is "abnormal cooler fan speed", which means the cooling system is faulty.

[0047] This method achieves a leap from "correlation" to "causation" through counterfactual reasoning, which can accurately locate the root cause of the anomaly and avoid misjudgment caused by confusion between correlation and causation in traditional methods.

[0048] Counterfactual reasoning requires using the structural equations of each node to calculate the counterfactual values ​​of the target variable. A structural equation is a functional relationship between each variable and its parent node variables in a causal graph model. In one embodiment, the target state variable is calculated using the structural equations of each node. The counterfactual values ​​include: For any variable in the causal graph model Get variables Parent node variable set ; According to the structural equation Calculate variables The possible values ​​of , where It is a function fitted based on historical data. Noise term; From candidate root dependent variables Initially, the values ​​of each downstream variable are calculated sequentially along the directed edges in the cause-effect graph model until the result is obtained. .

[0049] Specifically, in a causal graph model, each variable They all have a structural equation . It can be a linear function, a nonlinear function, or a neural network; this embodiment does not limit this. For example, for the variable "oil temperature"... Its parent node may include "load current". "Ambient temperature" "Cooler speed" Then the structural equation can be , where the coefficient Obtained through linear regression fitting of historical data. Noise term. It is usually assumed that the distribution follows a normal distribution with a mean of 0, reflecting unobserved factors.

[0050] In counterfactual reasoning, the server fixes the candidate root variable. Counterfactual value Meanwhile, the noise terms of other variables remain at their true observation values ​​(obtained by substituting the true observation values ​​into the inverse structural equations). Then, from... Initially, following the direction of the directed edges in the causal graph model (i.e., the direction from cause to effect), the counterfactual value of each downstream variable is calculated sequentially. Specifically: ① The value is set to .

[0051] ②For direct downstream variables its parent node set Includes And other possible parent nodes. Except for those parent nodes which are replaced with counterfactual values, all other parent nodes use true observations (or already calculated counterfactual values) and are substituted into the structural equations. Calculate The counterfactual value.

[0052] ③ Repeat the above steps until the target variable is calculated. counterfactual value .

[0053] In this way, the server can accurately simulate the propagation effect of intervention in counterfactual scenarios, providing quantitative evidence for root cause determination.

[0054] Physical information graph neural networks (PIPs) are crucial for achieving spatiotemporal alignment and fusion of multi-source heterogeneous data. Traditional feature fusion methods (such as simple concatenation and weighted averaging) neglect the physical spatial relationships between sensors, resulting in fused features that fail to reflect the causal interactions between devices. Therefore, in one embodiment, multi-source heterogeneous data is input into a PIP based on the mapping relationship between sensors and devices, and the output is a spatiotemporally aligned multimodal feature matrix, including: A graph structure is constructed based on the mapping relationship between sensors and devices and the physical topology, where nodes represent sensor-device pairs. With nodes The initial edge weights between them are ,in For nodes With nodes The shortest path distance in the physical topology map to the devices corresponding to the sensors bound to the space. For nodes and nodes The electrical connection type code between the bound sensors and the corresponding devices. , and These are preset hyperparameters; Through nodes The corresponding sensor data is linearly transformed to obtain the initial hidden state vector. ; Execute on graph structure Round-robin message passing, in each round of message passing, nodes The hidden state is updated to ,in For nodes The set of neighboring nodes, For learnable weight matrix, This is a vector concatenation operation. For activation function, For the first Nodes in round-robin message passing The hidden state vector, For the first Nodes in round-robin message passing The hidden state vector; The hidden state vectors of each node after the last round of message passing are used as the final hidden state vectors of each node, and then reorganized into a multimodal feature matrix according to the mapping relationship between sensors and devices.

[0055] Specifically, the server constructs an undirected graph based on the physical topology diagram (i.e., the substation primary wiring diagram) and the mapping relationship between sensors and devices. The vertex set Each node in the set corresponds to a "sensor-device pair," for example, the "#1 main transformer top oil temperature sensor" is a node. (Edge set) Each edge in the graph connects two nodes if the devices bound to those two nodes have a path (path length is unlimited) in the physical topology graph. For example, the distance between the transformer and the directly connected circuit breaker is 1, and the distance between the transformer and the disconnecting switch is 2. The value decreases exponentially with increasing distance. Electrical connection type coding typically assigns 1 for direct connections, 0.8 for connections via switchgear, and 0.5 for connections via transformers. (In the preset hyperparameters...) and These are weighting coefficients, typically taken as... , , This is the distance attenuation factor, typically set to 2. The above hyperparameters can be adjusted according to the actual substation topology.

[0056] For the initial hidden state vector If node This corresponds to continuous sensors (such as oil temperature sensors), whose raw data is a scalar value. ,but ,in It is the input weight matrix (mapping scalars to...) (dimensional vector) It's a bias. If the node For discrete sensors (such as circuit breaker status, with values ​​of 0 or 1), the embedding layer maps them to dense vectors. If the node... For unstructured data (such as images), feature vectors are extracted using a pre-trained convolutional neural network as the initial hidden state. The specific parameters and structure of the linear transformation, embedding layer, and pre-trained network can be any of the existing technologies, and this embodiment does not limit them.

[0057] The server executes on the constructed graph structure. Round-robin message passing ( The hyperparameters are preset and are typically set to 2-5 rounds. In each round, each node collects the hidden state vectors of all its neighboring nodes, performs weighted aggregation, and then concatenates them with its own state from the previous round. After passing through a fully connected layer and an activation function, a new hidden state vector is obtained. For neighboring nodes, it is possible to exclude itself or to choose to include itself; this embodiment does not impose any limitations on this. It is a learnable weight matrix with dimension . Because the length of the concatenated vector is . This represents concatenating two vectors along their feature dimensions. The activation function is used to compress the output value to a range of 0 to 1. After multiple rounds of message passing, the hidden state vector of each node incorporates information from its physical neighbors, achieving spatiotemporal alignment.

[0058] In the After the round of message passing is completed, the server will output the final hidden state vector of each node. Collect them and rearrange them according to the device order in the sensor-device mapping relationship to form a... matrix ,in It is the total number of sensor nodes. This is the dimension of the feature vector. This matrix is ​​the spatiotemporally aligned multimodal feature matrix, which will be input into the causal graph model later.

[0059] Through the aforementioned physical information graph neural network, the server can effectively integrate physical spatial information from multi-source heterogeneous data, providing high-quality feature input for causal inference.

[0060] A causal graph model and a physical information graph neural network can form a mutually reinforcing synergistic relationship. The causal strength values ​​in the causal graph model reflect the causal relationships between variables, and this information can be used to guide the updating of edge weights in the physical information graph neural network, thereby obtaining more accurate feature fusion; and more accurate feature fusion, in turn, can improve the quality of the causal graph model. In one embodiment, the method further includes: When the causal graph model is updated, obtain the causal strength value of each directed edge in the updated causal graph model; Update the weights of edges between nodes in the physical information graph neural network based on the causal strength value; After updating the weights of the physical information graph neural network, the steps of inputting multi-source heterogeneous data into the physical information graph neural network to obtain a multimodal feature matrix, and inputting the multimodal feature matrix into the causal graph model to perform counterfactual reasoning to determine the root cause node are re-executed.

[0061] Specifically, updates to the causal graph model can occur in various situations, such as when the server detects that the existing causal graph model cannot explain the current anomaly, triggering an incremental update, or periodically (e.g., weekly) re-running the build process based on newly accumulated historical data. This embodiment does not limit this. After the update is completed, the causal strength values ​​of each directed edge in the causal graph model may change.

[0062] For each pair of nodes in the physical information graph neural network and If the corresponding device state variables have directed edges (regardless of direction) in the causal graph model, then the causal strength value of that directed edge is used as an adjustment factor to adjust the original weights. The update process can be performed using any existing technology, such as weighted multiplication, addition, or attention-based fusion; this embodiment is not limited to any particular method. After the update, nodes with strong causal relationships receive higher weights during message passing, thereby generating feature representations that are more conducive to causal reasoning.

[0063] After updating the edge weights of the graph neural network, the server re-executes the following: the same multi-source heterogeneous data is re-inputted into the updated graph neural network to obtain a new multimodal feature matrix; the new feature matrix is ​​then input into the causal graph model to perform counterfactual inference again and determine the root cause node. If the newly obtained root cause node is consistent with the previous one, the optimization converges; if it is inconsistent, it can continue iterating or stop according to the preset number of iterations (e.g., a maximum of 3 times).

[0064] Through this iterative optimization of fusion and causal synergy, the embodiments of this application can continuously improve the performance of the diagnostic system, enabling the causal graph model and the feature fusion network to promote each other and form a virtuous cycle.

[0065] The update triggering condition of the causal graph model is a prerequisite for collaborative iterative optimization. In one embodiment, the update of the causal graph model includes: responding to the root cause node corresponding to... Make When the value exceeds a preset threshold, an incremental update of the causal graph model is triggered.

[0066] In counterfactual reasoning, the server has already calculated the optimal counterfactual value. and the corresponding difference This difference indicates that even if the best candidate root variable is forcibly set to a normal value, the target variable... It still cannot fully recover to the normal range. If A large value (e.g., oil temperature still 10°C above the normal upper limit) indicates that the current causal graph model may have missed some important causal variables, or that the existing causal strength estimates are inaccurate.

[0067] The server presets an update threshold. (For example, 30% of the normal interval width). If If the existing cause-effect graph model is insufficient to explain the current anomaly, it needs to be updated.

[0068] When the conditions are met, the server initiates the incremental update process for the causal graph model. Incremental updates do not require rebuilding the entire graph from scratch; instead, they only correct local regions relevant to the current anomaly: using the current anomalous data as new samples, the causal discovery algorithm is rerun, adding possible new edges, deleting or adjusting causal strength. The specific algorithm for incremental updates can be any existing technology, such as online causal discovery algorithms or Bayesian update-based methods; this embodiment is not limited in this regard. After the incremental update is completed, the new causal graph model will be used for subsequent diagnostics.

[0069] This mechanism ensures that the causal graph model can continuously adapt to new situations such as equipment aging and changes in operating conditions, maintaining the accuracy of diagnosis.

[0070] After identifying the root cause node, generating effective handling instructions is crucial for ensuring power grid security. Traditional handling methods typically only address the fault phenomenon itself, neglecting the root cause and its propagation path, leading to incomplete handling or causing secondary problems. Therefore, in one embodiment, based on the type of the root cause node and its causal path in the causal graph model, a sequence of handling instructions is output, including: Establish a mapping library between root cause node types and corresponding actions; Based on the outgoing edges of the root node in the causal graph model, determine the set of downstream variables affected by the root node, and calculate the degree coefficient of each downstream variable affected by the root node. The degree coefficient is determined by the product of the causal strengths on the causal path from the root node to the downstream variable. Generate a sequence of handling instructions, which includes root cause handling instructions for the root cause node and preventive handling instructions for downstream variables whose influence coefficients exceed a preset influence threshold in the downstream variable set.

[0071] Specifically, the server pre-establishes a mapping database that records the standard handling actions corresponding to each root cause node type. For example: Root cause type "Cooler fan speed too low" → Actions to take: "Start standby cooler", "Check fan power supply", "Clean filter". Root cause node type "Overload current" → Actions to handle: "Request load transfer", "Activate overload protection", "Adjust operating mode"; Root cause node type "insulation aging coefficient exceeds standard" → Remedial actions "arrange power outage for maintenance", "replace insulating oil", "strengthen partial discharge monitoring".

[0072] This mapping database can be entered by power industry experts according to operation and maintenance procedures, or it can be automatically learned from historical successful handling cases. This embodiment does not limit this.

[0073] In a causal graph model, a root cause node may have multiple directed edges pointing to its direct downstream variables, which in turn may point to even more downstream variables. The server uses a breadth-first search to find all variables reachable from the root cause node, forming a set of downstream variables. For each downstream variable, the server calculates a coefficient representing the degree of influence it receives from the root cause node. This coefficient is defined as the sum of the products of causal strengths along all causal paths from the root cause node to that downstream variable. If multiple paths exist, the maximum value or the sum is taken; this embodiment does not impose such limitations. For example, there is one path from root cause node R to downstream variable Q1. The causal strengths are respectively and Then the degree coefficient The degree coefficient reflects the intensity of the propagation of the root cause failure to the downstream variable; the larger the coefficient, the more severe the influence of the root cause on the downstream variable.

[0074] When generating the sequence of handling instructions, root cause handling instructions for the root cause node can be obtained from the mapping database. Preventive handling instructions can include phrases such as "strengthen monitoring," "inspect the device," and "prepare backup plans," aiming to prevent root cause failures from triggering downstream chain reactions. Finally, the server arranges all instructions in causal order (handling the root cause first, then the downstream), forming an instruction sequence that is output to the monitoring terminal or automated execution system.

[0075] Through the above steps, the embodiments of this application can output a targeted and forward-looking solution, effectively avoiding the drawbacks of traditional methods that focus on the symptoms and only address the symptoms without addressing the root cause.

[0076] To improve human-computer interaction efficiency, especially in the high-intensity working environment of a control center, in one embodiment, this method further includes: The system receives voice control commands, recognizes them using a voice recognition model for power control scenarios, and maps the recognition results to query or command confirmation operations. The query operation is used to query the location results of root cause nodes, while the command confirmation operation is used to confirm the output sequence of handling commands.

[0077] Specifically, the server receives voice commands from maintenance personnel via the microphone array in the control center. After noise reduction and endpoint detection, the voice signal is fed into a speech recognition model specifically designed for power control scenarios. This speech recognition model can employ any existing technology, such as a deep learning-based model (e.g., a Transformer or Conformer architecture), and is trained using a corpus containing substation terminology; this embodiment does not impose any limitations on this. The recognition result is a text string.

[0078] The server matches the identified text against a predefined set of instructions. These instructions are divided into two categories: ① Query operations: For example, "Query root cause node", "Display causal path", "View intervention effect", the server executes the corresponding database query, returns information such as the name, type, and confidence level of the root cause node, and provides feedback in voice or text form on the monitoring terminal interface.

[0079] ② Command confirmation operation: For example, "confirm execution of the sequence of handling instructions", "execute root cause handling", "skip preventive handling", the server forwards the user-confirmed instructions to the execution module, triggering the corresponding automatic control or issuing a maintenance work order.

[0080] Voice data is used only for command recognition and is not stored for extended periods. Users can disable the voice reception function at any time via a physical switch or the voice command "Exit Voice Mode". This application does not involve the collection of sensitive personal information; the function is only enabled when the user explicitly authorizes it and voice interaction is required.

[0081] As described above, voice control allows maintenance personnel to complete critical steps without manually operating the mouse and keyboard, improving emergency response efficiency.

[0082] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps.

[0083] Secondly, based on the same inventive concept, this application also provides a system for implementing the above-mentioned method for diagnosing the health status of a power grid master station system. The solution provided by this system is similar to the solution described in the above method; therefore, the specific limitations in one or more system embodiments provided below can be found in the limitations of the power grid master station system health status diagnosis method described above, and will not be repeated here.

[0084] In one embodiment, such as Figure 4 As shown, the system includes: a data acquisition module, a mapping acquisition module, a feature extraction module, a causal reasoning module, a counterfactual reasoning module, and an instruction output module. Among them: The data acquisition module is used to acquire multi-source heterogeneous data from substation equipment; The mapping acquisition module is used to acquire the mapping relationship between sensors and devices; The feature extraction module is used to input multi-source heterogeneous data into the physical information graph neural network and output a spatiotemporally aligned multimodal feature matrix. The edge weights between nodes of the physical information graph neural network are determined by the distance between the sensor-attached devices in the physical topology graph and the electrical connection type between the attached devices. The causal reasoning module is used to input the multimodal feature matrix into the causal graph model. The causal graph model uses device state variables as nodes and causal relationships between variables as directed edges. Each directed edge carries a causal strength value and a time delay parameter. The counterfactual reasoning module is used to perform counterfactual reasoning on the causal graph model to determine the root cause node when the target state variable in the causal graph model deviates from the normal operating range. The instruction output module is used to output a sequence of disposal instructions based on the type of the root cause node and the causal path of the root cause node in the causal graph model.

[0085] In one embodiment, the causal reasoning module includes a causal graph model building unit.

[0086] The causal graph model construction unit is used to determine the prior causal relationships between equipment state variables based on the knowledge of equipment physical mechanisms, and to construct a prior causal skeleton. The causal graph model construction unit is used to acquire historical operating data and use a causal discovery algorithm to correct the prior causal skeleton. The correction includes: estimating the causal strength of existing directed edges in the prior causal skeleton and deleting edges with causal strength below a preset strength threshold; searching for new edge candidate sets for non-existent directed edges in the prior causal skeleton and verifying the physical rationality of each edge in the new edge candidate set, retaining only edges that pass the physical rationality verification; the causal graph model construction unit is used to estimate the time delay parameter of each directed edge in the corrected prior causal skeleton to obtain the causal graph model.

[0087] In one embodiment, the counterfactual reasoning module includes a counterfactual reasoning unit.

[0088] Among them, the counterfactual reasoning unit is used in the causal graph model to transform the target state variable The current outlier is denoted as The counterfactual reasoning unit is used to let: in Candidate root dependent variable The normal range of values, To select candidate root dependent variables The value is set to The target state variables were then calculated using the structural equations of each node. The counterfactual value; the counterfactual reasoning unit is used to... The corresponding candidate root dependent variable is determined as the root cause node.

[0089] In one embodiment, the feature extraction module includes a feature matrix extraction unit.

[0090] The feature matrix extraction unit is used to construct a graph structure based on the mapping relationship between sensors and devices and the physical topology graph, where nodes represent sensor-device pairs. With nodes The initial edge weights between them are ,in For nodes With nodes The shortest path distance in the physical topology map to the devices corresponding to the sensors bound to the space. For nodes and nodes The electrical connection type code between the bound sensors and the corresponding devices. , and Preset hyperparameters; the feature matrix extraction unit uses nodes The corresponding sensor data is linearly transformed to obtain the initial hidden state vector. The feature matrix extraction unit performs on the graph structure. Round-robin message passing, in each round of message passing, nodes The hidden state is updated to ,in For nodes The set of neighboring nodes, For learnable weight matrix, This is a vector concatenation operation. For activation function, For the first Nodes in round-robin message passing The hidden state vector, For the first Nodes in round-robin message passing The hidden state vectors of each node after the last round of message passing are used as the final hidden state vectors of each node, and are reorganized into a multimodal feature matrix according to the mapping relationship between the sensor and the device.

[0091] In one embodiment, the power grid master station system health status diagnosis system further includes a model update module.

[0092] The model update module is used to obtain the causal strength value of each directed edge in the updated causal graph model when the causal graph model is updated; the model update module is used to update the weights of the edges between nodes in the physical information graph neural network according to the causal strength values; the model update module is used to re-execute the steps of inputting multi-source heterogeneous data into the physical information graph neural network to obtain the multimodal feature matrix after the weights of the physical information graph neural network are updated, and the steps of inputting the multimodal feature matrix into the causal graph model to perform counterfactual reasoning to determine the root cause node.

[0093] In one embodiment, the instruction output module includes a disposal instruction output unit.

[0094] The system includes: a treatment instruction output unit for establishing a mapping database between root cause node types and treatment actions; a treatment instruction output unit for determining the set of downstream variables affected by the root cause node based on the outgoing edges of the root cause node in the causal graph model, and calculating the degree coefficient of each downstream variable affected by the root cause node, which is determined by the product of causal strengths along the causal path from the root cause node to the downstream variable; and a treatment instruction output unit for generating a treatment instruction sequence, which includes root cause treatment instructions for the root cause node and preventive treatment instructions for downstream variables whose degree coefficients exceed a preset impact threshold.

[0095] In one embodiment, the counterfactual reasoning unit includes a counterfactual value calculation subunit.

[0096] The counterfactual value calculation subunit is used for any variable in the causal graph model. Get variables Parent node variable set The counterfactual value calculation subunit is used to calculate the value based on the structural equation. Calculate variables The possible values ​​of , where It is a function fitted based on historical data. For noise terms; the counterfactual value calculation subunit is used to calculate the candidate root dependent variable. Initially, the values ​​of each downstream variable are calculated sequentially along the directed edges in the cause-effect graph model until the result is obtained. .

[0097] Each module in the aforementioned power grid master station system health status diagnosis system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0098] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores all data related to system operation. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer program implements a method for diagnosing the health status of a power grid master station system.

[0099] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0100] Thirdly, in one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods described in the above method embodiments.

[0101] Fourthly, in one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described in the above method embodiments.

[0102] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0103] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0104] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for diagnosing the health status of a power grid master station system, characterized in that, The method includes: Acquire multi-source heterogeneous data from substation equipment; Obtain the mapping relationship between sensors and devices; Based on the mapping relationship between the sensor and the device, the multi-source heterogeneous data is input into the physical information graph neural network, and the spatiotemporally aligned multimodal feature matrix is ​​output. The edge weights between nodes of the physical information graph neural network are determined by the distance between the device to which the sensor is bound in the physical topology graph and the electrical connection type between the bound devices. The multimodal feature matrix is ​​input into the causal graph model, which uses device state variables as nodes and causal relationships between variables as directed edges. Each directed edge carries a causal strength value and a time delay parameter. When the target state variable in the causal graph model deviates from the normal operating range, counterfactual reasoning is performed on the causal graph model to determine the root cause node. Based on the type of the root cause node and the causal path of the root cause node in the causal graph model, a sequence of disposal instructions is output.

2. The method according to claim 1, characterized in that, The causal graph model is constructed through the following steps: Based on knowledge of the physical mechanism of the equipment, the prior causal relationship between the state variables of the equipment is determined, and a prior causal framework is constructed. Historical operational data is acquired, and a causal discovery algorithm is used to correct the prior causal framework. The correction includes: estimating the causal strength of existing directed edges in the prior causal framework and deleting edges whose causal strength is lower than a preset strength threshold; searching for new edge candidate sets for directed edges that do not exist in the prior causal framework, and performing physical rationality verification on each edge in the new edge candidate set, retaining only the edges that pass the physical rationality verification. For each directed edge in the modified prior causal skeleton, the time delay parameter of each directed edge is estimated to obtain the causal graph model.

3. The method according to claim 1, characterized in that, The process of performing counterfactual reasoning on the causal graph model includes: In the causal graph model, the target state variable is... The current outlier is denoted as ; make: in Candidate root dependent variable The normal range of values, To make the candidate root dependent variable The value is set to The target state variables were then calculated using the structural equations of each node. The counterfactual value; Will The corresponding candidate root dependent variable is determined as the root cause node.

4. The method according to claim 1, characterized in that, The step of inputting the multi-source heterogeneous data into a physical information graph neural network based on the mapping relationship between the sensor and the device, and outputting a spatiotemporally aligned multimodal feature matrix, includes: A graph structure is constructed based on the mapping relationship between the sensors and devices and the physical topology, where nodes represent sensor-device pairs. With nodes The initial edge weights between them are ,in For nodes With nodes The shortest path distance in the physical topology map to the devices corresponding to the sensors bound to the space. For nodes and nodes The electrical connection type code between the bound sensors and the corresponding devices. , and These are preset hyperparameters; Through nodes The corresponding sensor data is linearly transformed to obtain the initial hidden state vector. ; Execute on the graph structure Round-by-round message passing, in each round of message passing, nodes The hidden state is updated to ,in For nodes The set of neighboring nodes, For learnable weight matrix, This is a vector concatenation operation. For activation function, For the first Nodes in round-robin message passing The hidden state vector, For the first Nodes in round-robin message passing The hidden state vector; The hidden state vectors of each node after the last round of message passing are used as the final hidden state vectors of each node, and are reorganized into the multimodal feature matrix according to the mapping relationship between the sensor and the device.

5. The method according to any one of claims 1 to 4, characterized in that, Also includes: When the causal graph model is updated, the causal strength value of each directed edge in the updated causal graph model is obtained; Update the weights of the edges between nodes in the physical information graph neural network based on the causal strength value; After the weights of the physical information graph neural network are updated, the steps of inputting the multi-source heterogeneous data into the physical information graph neural network to obtain the multimodal feature matrix and inputting the multimodal feature matrix into the causal graph model to perform counterfactual reasoning to determine the root cause node are re-executed.

6. The method according to claim 1, characterized in that, The step of outputting a sequence of disposal instructions based on the type of the root cause node and the causal path of the root cause node in the causal graph model includes: Establish a mapping relationship library between the types of the root cause nodes and the corresponding actions; Based on the outgoing edge direction of the root node in the causal graph model, the set of downstream variables affected by the root node is determined, and the degree coefficient of the influence of the root node on each downstream variable is calculated. The degree coefficient is determined by the product of the causal strengths on the causal path from the root node to the downstream variable. Generate the treatment instruction sequence, which includes root cause treatment instructions for the root cause node and preventive treatment instructions for downstream variables in the downstream variable set whose influence coefficient exceeds a preset influence threshold.

7. The method according to claim 3, characterized in that, The target state variable is calculated using the structural equations of each node. The counterfactual values ​​include: For any variable in the causal graph model Obtain the variable Parent node variable set ; According to the structural equation Calculate the variable The possible values ​​of , where It is a function fitted based on historical data. Noise term; From the candidate root dependent variables Begin by sequentially calculating the values ​​of each downstream variable along the direction of the directed edges in the causal graph model, until the calculated values ​​are obtained. .

8. A health status diagnosis system for a power grid master station system, characterized in that, The system includes: The data acquisition module is used to acquire multi-source heterogeneous data from substation equipment; The mapping acquisition module is used to acquire the mapping relationship between sensors and devices; The feature extraction module is used to input the multi-source heterogeneous data into the physical information graph neural network according to the mapping relationship between the sensor and the device, and output a spatiotemporally aligned multimodal feature matrix. The edge weights between nodes of the physical information graph neural network are determined by the distance between the sensor and the device in the physical topology graph and the electrical connection type between the bound devices. The causal reasoning module is used to input the multimodal feature matrix into the causal graph model. The causal graph model uses device state variables as nodes and causal relationships between variables as directed edges. Each directed edge carries a causal strength value and a time delay parameter. The counterfactual reasoning module is used to perform counterfactual reasoning on the causal graph model to determine the root cause node when the target state variable in the causal graph model deviates from the normal operating range. The instruction output module is used to output a sequence of disposal instructions based on the type of the root cause node and the causal path of the root cause node in the causal graph model.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the power grid master station system health status diagnosis method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the power grid master station system health status diagnosis method according to any one of claims 1 to 7.

Images