Method for monitoring operation state of transformer substation of electric power system

Abstract

The invention provides a transformer substation operation state monitoring method of an electric power system, and relates to the technical field of transformer substation monitoring, and the method comprises the steps: collecting the real-time state data and historical state data of a transformer substation and corresponding fault labels, taking each device of the transformer substation as a node, taking the causal relationship between device pairs as an edge, constructing a fault causal model, and carrying out the fault causal model; and training the fault causal model by using the historical state data and the corresponding fault tag to obtain a trained fault causal model, and inputting the real-time state data into the trained fault causal model to obtain a transformer substation real-time fault tag. According to the method, the causal relationship model is constructed, the causal verification graph is constructed through the nonlinear Granger causal test method to verify the existence of the novel fault mode, and the fault causal model is adjusted, so that the flexibility of the model facing the novel fault mode is improved, and the purpose of improving the fault diagnosis accuracy of the transformer substation is achieved.

Description

Technical Field

[0001] This application relates to the technical field of substation monitoring, and particularly to a method for monitoring the operating status of a substation in a power system. Background Art

[0002] In modern power systems, substations, as the core nodes for power conversion, dispatching, and distribution, the security and stability of their operation are of crucial importance. With the complexity of the power grid structure and the growth of load demand, traditional substation control systems face many challenges, such as slow response speed, insufficient data processing capabilities, limited information sharing, and other issues. Therefore, how to improve the intelligent management level of substations and achieve efficient and accurate status monitoring and fault diagnosis has become an important research direction.

[0003] The Chinese invention patent with the application publication number CN119651923A provides a substation control system based on cloud control and local area network. This patent uses a dynamic multi-variate autoregressive tree algorithm to construct a multi-dimensional causal relationship graph to obtain the non-linear dependence relationship between multi-variables, so as to accurately identify potential abnormal and fault patterns, and uses a sparse distributed optimization algorithm for global load scheduling, and realizes fast local response control based on the Markov random field model.

[0004] For the above technical solution, the dynamic multi-variate autoregressive tree algorithm is a data-driven statistical learning method. Therefore, the multi-dimensional causal relationship graph obtained based on the dynamic multi-variate autoregressive tree algorithm is a causal hypothesis extracted from data patterns, which does not represent the actual causal results between the data. When encountering new types of fault patterns, it is impossible to correctly infer the causal associations generated by the new types of fault patterns, affecting the accuracy of fault diagnosis. Summary of the Invention

[0005] In order to improve the accuracy of substation fault diagnosis, this application provides a method for monitoring the operating status of a substation in a power system.

[0006] In the first aspect, this application provides a method for monitoring the operating status of a substation in a power system, adopting the following technical solution: A method for monitoring the operating status of a substation in a power system includes the following steps: Data collection: Collect the real-time status data of the substation, where the real-time status data includes real-time device status data, real-time device location, and real-time health data of the device; collect the historical status data of the substation and the corresponding fault labels, where the historical status data includes historical device status data, historical device location, and historical health data of the device; Model construction: Construct a fault causal model with each device in the substation as a node and the causal relationship between devices as an edge; Model training: including the first judgment, the first update, the first processing, the second processing, the third processing, and the first training; The first judgment: Based on historical state data, construct a causal verification graph among devices using the non-linear Granger causality test method, and judge whether the causal relationship in the causal verification graph is consistent with the edges between device pairs in the fault causal model: If so, execute the steps of the first processing; If not, execute the steps of the first update; The first update: Based on the causal verification graph, update the fault causal model, and use the updated fault causal model as the new fault causal model; The first processing: Obtain the time stamps of the historical state data. Based on the fault labels, merge the historical state data with the same fault label and continuous time stamps in the order of time stamps, and record the merged historical state data as the first data; The second processing: Calculate the Euclidean distance between device pairs based on the historical device locations, and collect the Euclidean distances between each device pair to obtain the first feature set; The third processing: Obtain the time series of the first data, and judge whether the historical health data of the device changes within the time series: If not, execute the steps of the first training; If so, obtain the time points when the historical health data of the device changes, divide the first data based on the obtained time points to get the divided first data, and update the divided first data as the first data; The first training: Concatenate the first feature set with the first data, update the concatenated data as the first data, and use the first data and the corresponding fault labels to train the new fault causal model to obtain the trained fault causal model; Fault discrimination: Input the real-time state data into the trained fault causal model to obtain the real-time fault labels of the substation.

[0007] By adopting the above technical solutions, constructing a causal verification graph among devices using the non-linear Granger causality test method can not only capture the linear causal relationships among devices, but also identify more complex non-linear causal relationships. When the causal relationship in the model is inconsistent with that in the causal verification graph, the fault causal model is dynamically adjusted, improving the flexibility of the model in the face of new fault patterns. In addition, merging historical state data based on time stamps and forming the corresponding time series helps the model capture the occurrence pattern of faults and their time dynamics in the face of new faults, thereby improving the accuracy of substation fault diagnosis. Moreover, dividing historical state data based on the life cycle data of devices helps the model adapt to new fault patterns that appear in different life cycle stages of devices, contributing to further improving the accuracy of model fault diagnosis.

[0008] Optionally, after performing the step of building the model and before performing the step of model training, the following steps are further included: First construction: Denote the nodes in the fault causal model as first nodes, denote the fault causal model as the first causal layer, obtain the functional attributes of each device in the substation, classify the devices based on the functional attributes to obtain several subsystems, use the subsystems as nodes, and use the causal relationships between the subsystems as edges to construct a second causal layer; Second construction: Denote the nodes of the second causal layer as second nodes, use the substation as a node, and use the causal relationships between the substations as edges to construct a third causal layer; Third construction: Denote the nodes of the third causal layer as third nodes, and update the first causal layer, the second causal layer, and the third causal layer as the fault causal model.

[0009] By adopting the above technical solution, hierarchical modeling is carried out on the causal relationships from the device, subsystem to the entire substation in sequence. The fault diagnosis problem of the entire substation is extended from the single device level to a higher-level systematic level, enabling the model to more precisely capture the fault patterns at each level. This helps the model to infer and analyze through hierarchical causal relationships in complex systematic faults or new fault patterns, discover potential new fault patterns across subsystems and devices, and reduce the situation where the diagnostic accuracy is reduced due to new fault patterns not being captured. In addition, setting up a hierarchical fault diagnosis model not only enhances the model's ability to identify new local faults but also helps the model predict the chain reaction caused by device faults, achieving the technical effect of early warning before the occurrence of new fault patterns and improving the response speed of substation fault diagnosis.

[0010] Optionally, after performing the step of the third construction and before performing the step of model training, the following steps are further included: First transformation: Based on the first causal layer, use each device as a first graph node, transform the causal relationship between each pair of devices into an embedding vector, and use the obtained embedding vector as the node feature of the corresponding first graph node to construct a first causal graph; Second transformation: Based on the second causal layer, use the subsystem as a second graph node, adopt the embedding vector of the first causal graph, calculate the embedding vector between each pair of subsystems, and use the obtained embedding vector as the node feature of the second graph node to construct a second causal graph; Third transformation: Based on the third causal layer, use the substation as a third graph node, adopt the embedding vector of the second causal graph, calculate the embedding vector between each pair of substations, and use the obtained embedding vector as the node feature of the third graph node to construct a third causal graph; Fourth construction: Based on the first causal graph, the second causal graph, and the third causal graph, construct a multi-causal relationship graph, and use the constructed multi-causal relationship graph as the new fault causal model; In the first update step, update the new fault causal model based on the causal verification graph.

[0011] By adopting the above technical solution, converting the causal relationship into an embedded vector and using the embedded vector as the feature of the graph node helps capture the complex non-linear relationship between graph nodes. Moreover, by propagating the embedded vector layer by layer, the technical effect of gradually establishing a complex causal chain between devices, subsystems, and substations is achieved, enhancing the model's recognition ability for new fault modes. When the model faces new and complex fault modes, it can infer potential causal associations from existing data through the multi-layer embedded vector propagation mechanism, which helps improve the diagnostic accuracy of the model for new fault modes. In addition, by integrating the causal graphs of each layer into a multi-causal relationship graph, it helps to grasp the complex causal relationship between each level from a global perspective, avoid information fragmentation between levels, further enhance the model's recognition ability for complex fault modes, and improve the diagnostic accuracy of the model for new fault modes.

[0012] Optionally, after the step of performing the first judgment and before the step of performing the first update, it further includes: Fourth processing: Add lag features to the historical state data to obtain the first data set; Fifth construction: Construct a causal forest, train the causal forest using the historical state data to obtain the trained causal forest, and input the first data set and the causal verification graph into the trained causal forest to obtain the causal relationship prediction results between each pair of devices; Second judgment: Judge whether the causal relationship prediction results between each pair of devices are exactly the same as the edges between the first pair of graph nodes in the fault causal model: If so, perform the steps of the first processing; If not, based on the causal relationship prediction results between each pair of devices, construct a new causal verification graph, update the new causal verification graph as the causal verification graph, and perform the first update step.

[0013] By adopting the above technical solution, adding lag features to the historical state data enables the causal forest to not only predict causal relationships based on the device state at a certain moment but also helps the causal forest infer potential causal connections between devices based on the state changes before a certain moment, providing time-series context information for the causal forest, improving the accuracy of the causal forest in inferring causal relationships between devices, and facilitating further fault mode judgment.

[0014] Optionally, after the step of performing the first judgment and before the step of performing the first processing, it further includes: Third judgment: Based on historical state data, use the non-linear Granger causality test method to construct a causality verification graph between each pair of subsystems, denoted as the first verification graph, and judge whether the causal relationship between each pair of subsystems in the first verification graph is consistent with the edge between the second node pairs in the fault causality model: If so, execute the steps of the fourth judgment; If not, use the causality verification graph and the first verification graph as the new causality verification graph, and execute the steps of the first update; Fourth judgment: Based on historical state data, use the non-linear Granger causality test method to construct a causality verification graph between each pair of substations, denoted as the second verification graph, and judge whether the causal relationship between each pair of substations in the second verification graph is consistent with the edge between the third node pairs in the fault causality model: If so, execute the steps of the first processing; If not, update the causality verification graph, the first verification graph and the second verification graph to the causality verification graph, and execute the steps of the first update.

[0015] By adopting the above technical solutions, the causal relationships between each layer are gradually verified from the equipment to the subsystems and then to the substations. The refined causality verification graph helps to more accurately identify new fault modes, reduce misdiagnosis and missed diagnosis. In addition, the multi-level causal relationship test enhances the adaptability of the model to new fault modes, reduces the limitations caused by fixed causality graphs, improves the flexibility and scalability of model fault diagnosis, and at the same time improves the speed of model fault detection and response.

[0016] Optionally, after executing the steps of the first update and before executing the steps of the first processing, it further includes: First collection: Input the first data set and the causality verification graph into the trained causal forest to obtain the confidence level of the causal relationship between the first graph node pairs; First annotation: Based on the fault causality model before update, annotate the added causal relationships and deleted causal relationships in the updated fault causality model, and collect the annotated causal relationships as the updated causal relationships; First screening: Obtain the confidence level of the updated causal relationships to obtain a confidence level data set; First calculation: Obtain the first graph node pairs corresponding to the updated causal relationships, calculate the embedding vectors of the first graph node pairs corresponding to the updated causal relationships before and after update, and based on the calculated embedding vectors, calculate the difference between the embedding vectors before and after update, and use the calculated difference as the vector change amount of the updated causal relationships; Second calculation: Perform a weighted sum of the vector change amount of the updated causal relationships and the corresponding confidence levels, and use the calculation result as the association degree of the updated causal relationships; First discrimination: Determine whether the correlation degree of the updated causal relationship is higher than the preset correlation degree threshold: If so, execute the steps of the second discrimination; If not, execute the steps of the third discrimination; Second discrimination: Determine whether the current updated causal relationship is an increasing causal relationship: If so, do nothing; If not, increase the causal relationship of the first graph node pair corresponding to the current updated causal relationship; Third discrimination: Determine whether the current updated causal relationship is a deleted causal relationship: If so, do nothing; If not, delete the causal relationship of the first graph node pair corresponding to the current updated causal relationship.

[0017] By adopting the above technical solution, based on the difference between the embedding vectors before and after the update, the change of the causal relationship before and after the model update is quantified. Combining with the confidence of the causal relationship, the importance of the model causal relationship is calculated, which helps to reduce the noise and redundant information introduced by the model update, optimize the modeling of the causal relationship, improve the interpretability of the model and the effectiveness of dynamic update, and enhance the accuracy and robustness of the model. In addition, the above dynamic adjustment and verification mechanism helps to enhance the adaptability of the model, helps the model to timely discover new fault modes caused by faults or changes in the operating environment in the substation, and improves the accuracy and effectiveness of fault diagnosis of the model in complex and dynamic environments.

[0018] Optionally, each device in the substation includes a device ID. After executing the steps of the first processing and before executing the steps of the second processing, it further includes: Second annotation: Based on each device in the substation, label the collected historical status data with the corresponding device ID, and record the labeled data as the new first data; First association: Obtain the edges between each node pair, convert the edges between each node pair into causal relationship strength values, and record them as the first feature; Position encoding: Based on the timestamp of the historical status data, perform position encoding on the timestamp of each node, and record the encoded timestamp as the time feature; Data update: Concatenate the time feature, the first feature and the new first data, and update the concatenated data as the first data.

[0019] By adopting the above technical solutions, the causal relationship is quantified. Based on the quantified features, it helps to enhance the model's ability to capture the causal chain within the substation. Moreover, adding positional encoding to each node based on the timestamps of historical state data helps the model better understand the temporal sequence and time dependence of time series data, improving the model's ability to capture trends and periodic changes over time and enhancing the model's processing ability for time series data. In addition, based on the concatenation of the quantified causal relationship and time features, it helps to enhance the expressive ability of the feature space of historical state data, improving the model's ability to capture new fault patterns and potential complex relationships when facing various fault patterns of equipment, and enhancing the accuracy of substation fault diagnosis.

[0020] Optionally, after the steps of the second processing and before the steps of the third processing, it further includes: First acquisition: Acquire the environmental parameters of the location where each device is located; Difference calculation: Calculate the difference in environmental parameters between each pair of devices; Third calculation: Based on the difference in environmental parameters between each pair of devices and the Euclidean distance, use the environmental space difference formula to calculate the environmental space difference between each pair of devices, and take the calculated environmental space difference between each pair of devices as the new first feature set; The calculation model of the environmental space difference formula is: ; Where, is the environmental space difference between device i and device j, is the adjustment factor, is the difference in environmental parameters between device i and device j, is the Euclidean distance between device i and device j.

[0021] By adopting the above technical solutions, by combining the difference in environmental parameters and the Euclidean distance between each device, the environmental space difference features are constructed, which helps the model consider the influence of the environment and spatial location where the device is located on its behavior and state, enhances the model's learning ability for spatial correlation and environmental factors, improves the model's ability to capture the potential influence of environmental factors on equipment faults, helps the model automatically adapt to different operating environments and working conditions during the training process, improves the accuracy and stability of the model's prediction ability for various devices in different environments and spaces, and further improves the comprehensiveness and accuracy of the model in diagnosing various fault patterns.

[0022] Optionally, the historical state data further includes the historical usage duration of the device, historical load data, and historical resource data. After the steps of the second processing and before the steps of the third processing, it further includes: First Statistics: Obtain the historical load data of each device and the maximum load value of each device, and calculate the total duration for which the historical load data of each device exceeds the maximum load value, the difference between the historical load data that exceeds the maximum load value and the maximum load value, and the corresponding duration; Second Statistics: Obtain the historical resource data of each device and the minimum resource requirement value of each device, and calculate the total duration for which the historical resource data of each device is lower than the minimum resource requirement value, the difference between the historical resource data that is lower than the minimum resource requirement value and the minimum resource requirement value, and the corresponding duration; Load Calculation: Use the load impact formula to calculate the impact factor of the load fluctuation of each device on the device state, denoted as the first impact; Resource Calculation: Use the resource impact formula to calculate the impact factor of the resource fluctuation of each device on the device state, denoted as the second impact; Fourth Calculation: Based on the historical health data of the device, the first impact, and the second impact, use the device health formula to calculate the actual health data of the device, and use the calculated actual health data as the new historical health data of the device; The calculation model of the load impact formula is: ; Wherein, is the number of historical load data that exceeds the maximum load value in the historical load data, a is the a-th historical load data that exceeds the maximum load value, is the difference between the a-th historical load data that exceeds the maximum load value and the maximum load value, is the maximum load value of the device, is the duration of the a-th historical load data that exceeds the maximum load value, is the historical usage duration of the device; The calculation model of the resource impact formula is: ; Wherein, is the number of historical resource data that is lower than the minimum resource requirement value in the historical resource data, b is the b-th historical resource data that is lower than the minimum resource requirement value, is the difference between the b-th historical resource data that is lower than the minimum resource requirement value and the minimum resource requirement value, is the minimum resource requirement value of the device, is the duration of the b-th historical resource data that is lower than the minimum resource requirement value; The calculation model of the device health formula is: ; Wherein, is the impact weight of the load fluctuation on the device state, is the influence weight of resource fluctuation on the device state, is the historical health data of the device.

[0023] By adopting the above technical solution, the dynamic characteristics of load fluctuation and resource fluctuation are introduced, the influence of load fluctuation and resource fluctuation on device health is quantified, and the adaptability of the model to devices under different conditions is improved, which helps to enhance the fault prediction ability of the model under unstable operating conditions. In addition, based on the introduced dynamic characteristics of load fluctuation and resource fluctuation, considering the situation of the device state when the device is under abnormal load and resource conditions for a long time, the recognition ability of the model for potential fault modes of devices under abnormal operating conditions for a long time is enhanced, the sensitivity of the model to changes in device state is improved, and the fault prediction accuracy of the substation is enhanced.

[0024] Optionally, after the step of performing the third process and before the step of performing the first training, it further includes: Second acquisition: Obtain the time point of the change in the historical health data of the device, denoted as the first time point; Third acquisition: Obtain the historical health data of the device at a time point before the first time point, denoted as the first device data; obtain the historical health data of the device at a time point after the first time point, denoted as the second device data; Rate calculation: Calculate the time difference between the time point before the first time point and the time point after the first time point, and based on the calculated time difference, the first device data and the second device data, calculate the change rate of the historical health data of the device, denoted as the first change rate; Fifth judgment: Judge whether the first change rate is less than a preset rate threshold: If so, merge the first data segmented based on the first time point, and update the merged data as the first data; If not, do not process.

[0025] By adopting the above technical solution, introducing the dynamic characteristics reflecting the change speed of device health helps to improve the model's ability to capture the long-term trend changes of the device, reduce the training error caused by noise data or irrelevant changes, thereby enhancing the model's ability to predict the device state and further improving the accuracy of the model for substation fault identification.

[0026] In summary, the present application includes at least one of the following beneficial technical effects: 1. The non - linear Granger causality test method is used to construct the causality verification graph between devices. It can not only capture the linear causality between devices, but also identify more complex non - linear causality. When the causality in the model is inconsistent with that in the causality verification graph, the fault causality model is dynamically adjusted, improving the flexibility of the model in the face of new fault patterns.

[0027] 2. The causality is transformed into an embedding vector, and the embedding vector is used as the feature of the graph node, which helps to capture the complex non - linear relationship between graph nodes. Moreover, by propagating the embedding vector layer by layer, the technical effect of gradually establishing a complex causal chain between devices, subsystems and substations is achieved, enhancing the model's ability to identify new fault patterns. When facing new and complex fault patterns, the model can infer potential causal associations from existing data through the multi - layer embedding vector propagation mechanism, which helps to improve the diagnostic accuracy of the model for new fault patterns.

[0028] 3. Based on the difference between the embedding vectors before and after update, the change of causality before and after model update is quantified. Combining with the confidence of causality, the importance of the model's causality is calculated, which helps to reduce the noise and redundant information introduced by model update, optimize the causality modeling, improve the interpretability of the model and the effectiveness of dynamic update, and enhance the precision and robustness of the model.

[0029] 4. The dynamic characteristics of load fluctuation and resource fluctuation are introduced to quantify the impact of load fluctuation and resource fluctuation on device health, improving the model's adaptability to devices under different conditions, helping to enhance the model's fault prediction ability under unstable operating conditions, increasing the sensitivity of the model to changes in device status, and enhancing the fault prediction accuracy of the substation. Description of the Drawings

[0030] Figure 1 is the flowchart of Embodiment 1 of this application; Figure 2 is the flowchart of S3 model training in Embodiment 1 of this application; Figure 3 is the flowchart of the first reconstruction of S21 in Embodiment 2 of this application; Figure 4 is the flowchart of the second reconstruction of S22 in Embodiment 4 of this application; Figure 5 is the flowchart of S321 update verification in Embodiment 4 of this application. Detailed Embodiment

[0031] The following is a further detailed description of this application in conjunction with Figures 1 to 5 to further illustrate this application in detail.

[0032] Embodiment 1: This embodiment discloses a method for monitoring the operation status of a substation in a power system. As Figure 1 shown, the method includes: collecting real-time status data, historical status data, and corresponding fault tags of the substation, constructing a fault causality model with each device in the substation as a node and the causal relationship between devices as edges, training the fault causality model using the historical status data and corresponding fault tags to obtain a trained fault causality model, and inputting the real-time status data into the trained fault causality model to obtain real-time fault tags of the substation. This embodiment includes the following steps: S1 Data collection: Collect real-time status data of the substation; collect historical status data of the substation and corresponding fault tags, where the historical status data includes historical device status data, historical device location, and historical health data of the device.

[0033] The real-time status data includes real-time device status data, real-time device location, real-time health data of the device, real-time device usage duration, real-time load data, and real-time resource data. The real-time device status data includes real-time switch status data of the device, real-time working mode of the device, real-time electrical parameter data of the device, and real-time communication status of the device. The real-time electrical parameter data includes real-time current of the device, real-time voltage of the device, real-time frequency of the device, and real-time power of the device. The real-time device location includes real-time geographical location of the device and real-time configuration information of the device, where the real-time configuration information of the device is the location of the device in the substation system and its connection relationship with other devices. Each device is provided with a device ID. The real-time load data includes, but is not limited to, real-time computing load, real-time working load, and real-time energy load, etc. The real-time resource data includes, but is not limited to, real-time temperature, real-time power capacity, real-time power network traffic, and real-time electricity quantity, etc. The real-time health data of the device is the current remaining life of the device.

[0034] The historical status data includes historical device status data, historical device location, historical health data of the device, historical device usage duration, historical load data, and historical resource data. The historical device status data includes historical switch status data of the device, historical working mode of the device, historical electrical parameter data of the device, and historical communication status of the device. The historical electrical parameter data includes historical current of the device, historical voltage of the device, historical frequency of the device, and historical power of the device. The historical device location includes historical geographical location of the device and historical configuration information of the device. The historical load data includes, but is not limited to, historical computing load, historical working load, and historical energy load, etc. The historical resource data includes, but is not limited to, historical temperature, historical power capacity, historical power network traffic, and historical electricity quantity, etc. The historical health data of the device is the historical remaining life of the device.

[0035] The fault tags include, but are not limited to, short - circuit fault, overload fault, over - voltage fault, low - voltage fault, grounding fault, current imbalance, power failure fault, equipment mechanical fault, equipment over - temperature fault, low oil level fault, vibration fault, communication loss, circuit breaker fault, distribution line fault, frequency fluctuation fault, low power factor fault, and harmonic fault, etc.

[0036] In this embodiment, it is necessary to pre - process the collected data. The steps of the data pre - processing include data cleaning and data normalization.

[0037] Data cleaning: Perform operations such as denoising the collected data, handling missing values, and removing outliers.

[0038] Data normalization: Normalize the collected data.

[0039] S2 Construct a model: Taking each device in the substation as a node and the causal relationship between device pairs as an edge, construct a fault causal model.

[0040] In this embodiment, the basic model of the fault causal model can be a Bayesian network model. Taking each device in the substation as a node and the status data of the substation as node features, set the causal relationship between devices through a physical model and current domain expert experience, and take the causal relationship between device pairs as an edge to construct a fault causal model. The fault causal model has the function of outputting real - time fault tags according to the input real - time status data of the substation.

[0041] S3 Model training: Includes S31 First judgment, S32 First update, S33 First processing, S34 Feature splicing, S35 Second processing, S36 First compensation, S37 Second compensation, S38 Third processing, S39 First verification, and S310 First training, as Figure 2 shown.

[0042] S31 First judgment: Based on historical status data, use the non - linear Granger causality test method to construct a causal verification graph between each device. Corresponding the causal relationship of each device pair in the causal verification graph with the edges between each device pair in the fault causal model, and judge whether the causal relationship in the causal verification graph is consistent with the edges between each device pair in the fault causal model.

[0043] If so, execute S33 First processing.

[0044] If not, execute S32 First update.

[0045] The non - linear Granger causality test methods include the Hiemstra - Jones test method, the Diks - Panchenko test method, and the test method based on neural networks, etc. In this embodiment, the test method based on neural networks is used to construct a causality verification graph between devices. The specific steps are as follows: Select a multi - layer perceptron as the basic model to construct an experimental model and a control model respectively. Obtain all device - pair combinations, select one device - pair combination as device A and device B, and obtain the historical state data of device A and device B respectively. Label the historical state data of device B as the experimental label of the historical state data of device A, input the labeled historical state data of device A into the experimental model, output the experimental fault prediction result of device B, calculate the error of the experimental model based on the experimental fault prediction result of device B and the true label of device B, and record the calculated error as the experimental error. Input the historical state data of device B into the control model, output the control fault prediction result of device B, calculate the error of the control model based on the control fault prediction result of device B and the true label of device B, and record the calculated error as the control error. Determine whether the experimental error of device B is less than the control error of device B: if so, determine that there is a causal relationship between device A and device B; otherwise, there is no causal relationship. Repeat the above operations to obtain the causality test results of all device - pair combinations, and construct a causality verification graph based on the causality test results of all device - pair combinations.

[0046] In the above steps, in the experimental model, calculate the gradient of the data input by device A and the data output by device B through the back - propagation algorithm, and use the calculated gradient as the causality strength value between device A and device B. Calculate the causality strength values of all device - pair combinations according to the above steps. When constructing the causality verification graph, label the corresponding causality strength values for the causal relationships of all device - pair combinations.

[0047] S32 First update: Based on the causal relationships in the causality verification graph, perform addition and deletion operations on the edges in the fault causality model, so that the edges between each device - pair in the processed fault causality model are consistent with the causal relationships in the causality verification graph, and update the processed fault causality model as the fault causality model.

[0048] S33 First processing: Obtain the timestamps of the historical state data. Based on the fault labels, in the order of timestamps, merge the historical state data with the same fault label and continuous timestamps, and record the merged historical state data as the first data. Example, a set of historical state data tables, see Table 1.

[0049] Serial number Time Historical status data Fault label 1 2-11 10:00 Historical status data A Short - circuit fault 2 2-11 11:00 Historical status data B Normal 3 2-11 12:00 Historical status data C Normal 4 2-11 13:00 Historical status data D Low - voltage fault 5 2-11 14:00 Historical status data E Low - voltage fault Among them, the fault labels of the historical status data with serial numbers 2 and 3 are the same, and the timestamps are consecutive. The fault labels of the historical status data with serial numbers 4 and 5 are the same, and the timestamps are consecutive. Therefore, the historical status data with serial numbers 2 and 3 are merged, and the historical status data with serial numbers 4 and 5 are merged. All the merged historical status data are denoted as the first data. The data in the above historical status data table are merged to obtain the first data table, as shown in Table 2.

[0050] Serial number Time Historical status data Fault label 1 2-11 10:00 Historical status data A Short - circuit fault 2 2-11 11:00-12:00 Historical status data B Historical status data C Normal 3 2-11 13:00-14:00 Historical status data D Historical status data E Low - voltage fault S34 Feature Concatentation: including S341 Second Annotation, S342 First Association, S343 Position Encoding, and S344 Data Update.

[0051] S341 Second Annotation: Based on each device in the substation, obtain the device ID of each device, associate the historical status data of different sources of the same device collected through the device ID, and label the corresponding device ID on the historical status data of the device. The labeled data is denoted as the new first data.

[0052] S342 First Association: Obtain the edges between each pair of nodes. Based on the causal verification graph, obtain the causal relationship intensity values corresponding to the edges between each pair of nodes. For each device, construct a causal relationship intensity value matrix between the device and all other devices. The constructed matrix is denoted as the first matrix of the device. Aggregate the first matrices of all devices to obtain the first feature.

[0053] S343 Position Encoding: Based on the timestamps of the historical status data, extract the device ID of each device and the timestamps of the corresponding historical status data, perform position encoding on the timestamps of the historical status data of each device ID, store the encoded data in the time feature matrix corresponding to each device, and aggregate the time feature matrices of all devices to obtain the time feature.

[0054] S344 Data Update: Add and concatenate the obtained time feature, first feature, and new first data, and update the concatenated data as the first data.

[0055] S35 Second Processing: Based on the historical device positions, calculate the Euclidean distances between each pair of devices using the Euclidean distance formula, and aggregate the Euclidean distances between each pair of devices to obtain the first feature set.

[0056] In this embodiment, the position of each device in the substation is defaulted to be fixed, so the historical device positions in each historical status data are the same.

[0057] S36 First Compensation: including S361 First Acquisition, S362 Difference Calculation, and S363 Third Calculation.

[0058] S361 First Acquisition: Acquire the environmental parameters at the location of each device, and the environmental parameter data includes temperature and humidity.

[0059] S362 Difference Calculation: Based on the temperature and humidity at the location of each acquired device, calculate the temperature difference and humidity difference between each pair of devices respectively.

[0060] S363 Third Calculation: Based on the temperature difference, humidity difference and Euclidean distance between each pair of devices, use the environmental space difference formula to calculate the environmental space difference between each pair of devices respectively, and take the calculated environmental space difference between each pair of devices as the new first feature set. In this embodiment, the environmental space difference between each pair of devices includes temperature space difference and humidity space difference.

[0061] The calculation model of the environmental space difference formula is: ; Where, is the environmental space difference between device and device , is the adjustment factor, is the environmental parameter difference between device and device , is the Euclidean distance between device and device . The value of the adjustment factor is obtained through cross-validation calculation. In this embodiment, = 0.5.

[0062] Example, set the temperature difference between device A and device B to -2 °C, the humidity difference between device A and device B to 5%, and the Euclidean distance between device A and device B to 14.14. Then the temperature space difference between device A and device B is , and the humidity space difference between device A and device B is . Then the environmental space difference between device A and device B is , .

[0063] In this embodiment, the unit of humidity is %, and when calculating, the humidity represented by percentage is converted into decimal form to participate in the calculation of the environmental space difference formula.

[0064] S37 Second Compensation: Includes S371 First Statistics, S372 Second Statistics, S373 Load Calculation, S374 Resource Calculation and S375 Fourth Calculation.

[0065] S371 First Statistics: Obtain the historical load data of each device and the maximum load value of each device, and count the total duration for which the historical load data of each device exceeds the maximum load value, the difference between the historical load data that exceeds the maximum load value and the maximum load value, and the corresponding duration.

[0066] S372 Second Statistics: Obtain the historical resource data of each device and the minimum resource requirement value of each device, and count the total duration for which the historical resource data of each device is lower than the minimum resource requirement value, the difference between the historical resource data that is lower than the minimum resource requirement value and the minimum resource requirement value, and the corresponding duration.

[0067] S373 Load Calculation: Use the load impact formula to calculate the impact factor of the load fluctuation of each device on the device state, denoted as the first impact.

[0068] S374 Resource Calculation: Use the resource impact formula to calculate the impact factor of the resource fluctuation of each device on the device state, denoted as the second impact.

[0069] S375 Fourth Calculation: Based on the historical health data of the device, the first impact, and the second impact, use the device health formula to calculate the actual health data of the device, and use the calculated actual health data as the new historical health data of the device.

[0070] S375 Fourth Calculation: Based on the historical health data of the device, the first impact, and the second impact, use the device health formula to calculate the actual health data of the device, and use the calculated actual health data as the new historical health data of the device.

[0071] The calculation model of the load impact formula is: .

[0072] Among them, is the number of historical load data that exceeds the maximum load value in the historical load data, a is the a-th historical load data that exceeds the maximum load value, is the difference between the a-th historical load data that exceeds the maximum load value and the maximum load value, is the maximum load value of the device, is the duration of the a-th historical load data that exceeds the maximum load value, is the historical usage duration of the device.

[0073] The calculation model of the resource impact formula is: .

[0074] Among them, Let \(a\) be the number of historical resource data in the historical resource data that is lower than the minimum resource requirement value, and \(b\) be the \(b\)th historical resource data that is lower than the minimum resource requirement value. Let \(\Delta b\) be the difference between the \(b\)th historical resource data that is lower than the minimum resource requirement value and the minimum resource requirement value. Let \(R\) be the minimum resource requirement value of the device. Let \(t_b\) be the duration of the \(b\)th historical resource data that is lower than the minimum resource requirement value.

[0075] The calculation model of the device health formula is: .

[0076] Among them, Let \(w_1\) be the influence weight of load fluctuation on the device state. Let \(w_2\) be the influence weight of resource fluctuation on the device state. Let \(H\) be the historical health data of the device. Among them, the values of and are obtained by performing regression analysis on historical state data. In this embodiment, \(w_1 = 0.6\), .

[0077] Example: Set the historical health data of device A to 2000h, the historical usage duration of device A to 5h (this is only an example, only a small part of the historical state data is listed. By default, the first sampling time is the initial usage time of the device, so the historical usage duration of device A is the interval between the first sampling time and the last sampling time), the maximum load value of device A to 100L, the minimum resource requirement value of device A to 40R, and set the sampling time and sampling interval of the historical load data to be the same as those of the historical state data. The historical load data and historical resource data of device A are shown in Table 3.

[0078] Time Historical load data (L) Historical resource data (R) 2-11 10:00 80 50 2-11 11:00 120 40 2-11 12:00 110 35 2-11 13:00 130 45 2-11 14:00 90 60 2-11 15:00 140 30 According to Table 3 above, the number of historical load data in the historical load data of device A that exceeds the maximum load value is 4, so the first influence of device A is .

[0079] The number of historical resource data in the historical resource data of device A that is lower than the minimum resource requirement value is 2, so the first influence of device A is .

[0080] Therefore, the actual health data of device A is \(H = 1999.85\). So the new historical health data of device A is 1999.85h.

[0081] S38 Third Processing: Obtain each time series in the first data, and determine whether the historical health data of the device has changed within the time series.

[0082] If not, execute S310 First Training.

[0083] If so, obtain the time points at which the historical health data of the device has changed, split the first data based on the obtained time points to get the split first data, update the split first data as the first data, and then execute S39 First Verification.

[0084] S39 First Verification: Includes S391 Second Obtaining, S392 Third Obtaining, S393 Rate Calculation, and S394 Fifth Judgment.

[0085] S391 Second Obtaining: Obtain the time points at which the historical health data of the device has changed, denoted as the first time points.

[0086] S392 Third Obtaining: Obtain the historical health data of the device at the time point before the first time point, denoted as the first device data. Obtain the historical health data of the device at the time point after the first time point, denoted as the second device data.

[0087] S393 Rate Calculation: Calculate the time difference between the time point before the first time point and the time point after the first time point, and calculate the change rate of the historical health data of the device based on the calculated time difference, the first device data, and the second device data, denoted as the first change rate.

[0088] S394 Fifth Judgment: Judge whether the first change rate is less than the preset rate threshold.

[0089] If so, merge the first data split based on the first time point, update the merged data as the first data, and execute S310 First Training.

[0090] If not, execute S310 First Training.

[0091] S310 First Training: Add and splice the first feature set and the first data, update the spliced data as the first data, and train the new fault causality model with the first data and the corresponding fault labels to obtain the trained fault causality model.

[0092] S4 Fault Discrimination: Input the real-time status data into the trained fault causality model to obtain the real-time fault label of the substation.

[0093] In this embodiment, a non-linear Granger causality test method is adopted to construct a causality verification graph between devices. It can not only capture the linear causality between devices, but also identify more complex non-linear causality. When the causality in the model is inconsistent with that in the causality verification graph, the fault causality model is dynamically adjusted, improving the flexibility of the model in the face of new fault modes. Moreover, by introducing the dynamic characteristics of load fluctuations and resource fluctuations, the impact of load fluctuations and resource fluctuations on device health is quantified, enhancing the adaptability of the model to devices under different conditions, contributing to strengthening the fault prediction ability of the model under unstable operating conditions, improving the sensitivity of the model to changes in device states, and enhancing the fault prediction accuracy of the substation.

[0094] Embodiment 2: The difference from Embodiment 1 is that: As Figure 3 shown, after performing S2 to build the model and before performing S3 model training, it further includes S21 the first reconstruction, and the S21 the first reconstruction includes S211 the first construction, S212 the second construction, and S213 the third construction.

[0095] S211 the first construction: Denote the nodes in the fault causality model as the first nodes, and denote the fault causality model as the first causality layer. Obtain the functional attributes of each device in the substation, classify the devices based on the functional attributes to obtain several subsystems. In this embodiment, based on the functional attributes, the devices are classified into a power transmission subsystem (such as busbars, cables, overhead lines, etc.), a switching subsystem (such as circuit breakers, disconnectors, etc.), a protection subsystem (such as relay protection devices, lightning arresters, etc.), a voltage transformation subsystem (such as transformers, reactors, etc.), and a measurement and monitoring subsystem (such as current transformers, voltage transformers, etc.). Based on the causal relationships between the devices in the first causality layer, obtain the causal strength between the devices. In this embodiment, since the first causality layer sets the causal relationships between the devices according to the physical model and current domain expert experience, the causal strength between the devices is also known. Construct a device causal matrix based on the causal strength between the devices. Based on the constructed device causal matrix and the devices included in each subsystem, aggregate the causal strength of the devices included in each subsystem as the causal strength between subsystem pairs, obtain the causal relationships and causal strengths between subsystem pairs, use the subsystems as nodes, and use the causal relationships between subsystem pairs as edges to construct the second causality layer.

[0096] S212 the second construction: Denote the nodes of the second causality layer as the second nodes, use the substation as a node, and based on the causal relationships between substations set according to current domain expert experience, use the causal relationships between substations as edges to construct the third causality layer.

[0097] S213 Third construction: Denote the nodes in the third causal layer as third nodes, and update the first causal layer, the second causal layer, and the third causal layer into a fault causal model.

[0098] In this embodiment, hierarchical modeling is performed on the causal relationships from devices, subsystems to the entire substation in sequence, expanding the fault diagnosis problem of the entire substation from the single device level to a higher-level systematic level, enabling the model to capture the fault modes at each level more meticulously, helping the model to infer and analyze through hierarchical causal relationships in complex systematic faults or new fault modes, discover potential new fault modes across subsystems and devices, and reduce the situation where the diagnostic accuracy is reduced due to new fault modes not being captured.

[0099] Embodiment 3: The difference from Embodiment 2 is that: After performing S31 First Judgment and before performing S33 First Processing, it further includes S311 Model Judgment, and S311 Model Judgment includes S3111 Third Judgment and S3112 Fourth Judgment.

[0100] S3111 Third Judgment: Based on historical state data, use the non-linear Granger causality test method to construct a causal verification graph between each pair of subsystems, denoted as the first verification graph, and judge whether the causal relationship between each pair of subsystems in the first verification graph is consistent with the edge between the second node pairs in the fault causal model.

[0101] If so, then perform S3112 Fourth Judgment.

[0102] If not, then use the causal verification graph and the first verification graph as the new causal verification graph, and perform S32 First Update.

[0103] S3112 Fourth Judgment: Based on historical state data, use the non-linear Granger causality test method to construct a causal verification graph between each pair of substations, denoted as the second verification graph, and judge whether the causal relationship between each pair of substations in the second verification graph is consistent with the edge between the third node pairs in the fault causal model: If so, then perform S33 First Processing.

[0104] If not, then update the causal verification graph, the first verification graph, and the second verification graph into the causal verification graph, and perform S32 First Update.

[0105] In this embodiment, the causal relationships between layers are gradually verified from the device to the subsystem and then to the substation. The refined causal verification diagram helps to more accurately identify new existing fault modes, reducing the occurrence of misdiagnosis and missed diagnosis. In addition, the multi-level causal relationship test enhances the adaptability of the model to new fault modes, reduces the limitations caused by fixed causal diagrams, improves the flexibility and scalability of model fault diagnosis, and at the same time increases the speed of model fault detection and response.

[0106] Embodiment 4: The difference from Embodiment 2 is as follows: As Figure 4 shown, after the third construction of S213 is executed and before the model training of S3 is executed, it further includes the second reconstruction of S22. The second reconstruction of S22 includes the first conversion of S221, the second conversion of S222, the third conversion of S223, and the fourth construction of S224.

[0107] The first conversion of S221: Based on the first causal layer, each device is used as the first graph node, the causal relationship between each pair of devices is converted into an embedding vector, and the obtained embedding vector is used as the node feature of the corresponding first graph node to construct the first causal graph.

[0108] In this embodiment, the Node2Vec algorithm is used to calculate the embedding vector. The specific steps are as follows: Based on the first causal layer, the parameters of the Node2Vec algorithm are defined. The parameters include the vector dimension, the random walk length, the number of random walks, the return parameter, and the in-out parameter. According to the defined parameters, starting from the first graph node of the first causal layer, the next graph node is selected for random walk according to the neighbor graph node of the first graph node and the corresponding causal relationship strength value until the preset random walk length is reached, obtaining a random walk sequence. The above steps are executed for each graph node in the new medical knowledge graph diagram to generate multiple random walk sequences. The generated random walk sequences are used as sample data, and the Skip-Gram model is trained using the above sample data to obtain the trained Skip-Gram model. A random first graph node is input into the trained Skip-Gram model to obtain the embedding vector of the first graph node.

[0109] The second conversion of S222: Based on the third causal layer, the substation is used as the third graph node, and the embedding vector of the second causal graph and the Node2Vec algorithm are used to calculate the embedding vector between substations. The obtained embedding vector is used as the node feature of the third graph node to construct the third causal graph.

[0110] S223 Third Transformation: Based on the third causal layer, taking the substation as the third graph node, using the embedding vector of the second causal graph and the Node2Vec algorithm, calculate the embedding vectors between substations, and use the obtained embedding vectors as the node features of the third graph node to construct the third causal graph.

[0111] S224 Fourth Construction: Based on the first causal graph, the second causal graph, and the third causal graph, construct a multi-causal relationship graph, and use the constructed multi-causal relationship graph as the new fault causal model.

[0112] In S32 First Update, based on the causal verification graph, update the new fault causal model.

[0113] After performing S31 First Judgment and before performing S32 First Update, it also includes S312 Second Verification, and S312 Second Verification includes S3121 Fourth Processing, S3122 Fifth Construction, and S3123 Second Judgment.

[0114] S3121 Fourth Processing: Add a lag value to the historical state data as a lag feature to obtain the first data set.

[0115] S3122 Fifth Construction: Construct a causal forest, train the causal forest using the historical state data to obtain the trained causal forest, input the first data set and the causal verification graph into the trained causal forest to obtain the causal relationship prediction results between each pair of devices.

[0116] S3123 Second Judgment: Judge whether the causal relationship prediction results between each pair of devices are exactly the same as the edges between the first graph nodes in the fault causal model.

[0117] If so, perform S33 First Processing.

[0118] If not, construct a new causal verification graph based on the causal relationship prediction results between each pair of devices, update the new causal verification graph as the causal verification graph, and perform S32 First Update.

[0119] As Figure 5 shown, after performing S32 First Update and before performing S33 First Processing, it also includes S321 Update Verification, and S321 Update Verification includes S3211 First Collection, S3212 First Annotation, S3213 First Screening, S3214 First Calculation, S3215 Second Calculation, S3216 First Discrimination, S3217 Second Discrimination, and S3218 Third Discrimination.

[0120] S3211 First Collection: Input the first data set and the causal verification graph into the trained causal forest to obtain the confidence of the causal relationship between the first graph nodes.

[0121] In this embodiment, since the first data set is obtained from historical state data, the confidence level of the causal relationship between the first graph node pairs obtained here actually refers to the confidence level of the causal relationship between the first graph node pairs before the model update.

[0122] S3212 First annotation: Based on the fault causal model before the update, annotate the added causal relationships and deleted causal relationships in the updated fault causal model, and collect the annotated causal relationships as the updated causal relationships.

[0123] S3213 First screening: Obtain the confidence levels of the updated causal relationships to obtain a confidence level data set.

[0124] S3214 First calculation: Obtain the first graph node pairs corresponding to the updated causal relationships, calculate the embedding vectors of the first graph node pairs corresponding to the updated causal relationships before and after the update, and based on the calculated embedding vectors, calculate the difference between the embedding vectors before and after the update, and use the calculated difference as the vector change amount of the updated causal relationship.

[0125] In this embodiment, the confidence levels of the updated causal relationships obtained actually refer to the confidence levels of the added causal relationships and the deleted causal relationships. Since there may be a situation where causal relationships are deleted after the model update, the confidence level of the causal relationship deleted after the update is the confidence level of this causal relationship before deletion, and the embedding vector of the causal relationship deleted after the update is 0. For the added causal relationships, the corresponding confidence levels can be set for the added causal relationships by methods of domain knowledge or prior assumptions, and the embedding vector of the added causal relationships before the model update is 0, and the vector change amount of the updated causal relationship takes a positive value.

[0126] S3215 Second calculation: Perform a weighted sum of the vector change amount of the updated causal relationship and the corresponding confidence level, and use the calculation result as the association degree of the updated causal relationship.

[0127] In this embodiment, the weights of the vector change amount of the updated causal relationship and the weights of the confidence levels of the updated causal relationships are both set according to expert knowledge or domain experience. Among them, the weights of the vector change amount of the updated causal relationship and the weights of the confidence levels of the updated causal relationships are both 0.5.

[0128] S3216 First discrimination: Judge whether the association degree of the updated causal relationship is higher than the preset association degree threshold.

[0129] If so, execute S3217 Second discrimination.

[0130] If not, execute S3218 Third discrimination.

[0131] S3217 Second discrimination: Judge whether the current updated causal relationship is an added causal relationship.

[0132] If so, perform the first process of S33.

[0133] If not, increase the causal relationship of the first graph node pair corresponding to the current updated causal relationship.

[0134] S3218 Third determination: Determine whether the current updated causal relationship is a deleted causal relationship: If so, perform the first process of S33.

[0135] If not, delete the causal relationship of the first graph node pair corresponding to the current updated causal relationship.

[0136] In this embodiment, converting the causal relationship into an embedding vector and using the embedding vector as the feature of the graph node helps to capture the complex non - linear relationship between graph nodes. Moreover, by propagating the embedding vector layer by layer, the technical effect of gradually establishing a complex causal chain between devices, subsystems, and substations is achieved, enhancing the model's ability to identify new fault modes. When the model faces new and complex fault modes, it can infer potential causal associations from existing data through the multi - layer embedding vector propagation mechanism, which helps to improve the diagnostic accuracy of the model for new fault modes. Based on the difference between the embedding vectors before and after the update, quantifying the change in the causal relationship before and after the model update, and combining with the confidence of the causal relationship, calculating the importance of the model's causal relationship helps to reduce the noise and redundant information introduced by the model update, optimize the modeling of the causal relationship, improve the interpretability of the model and the effectiveness of dynamic update, and enhance the precision and robustness of the model.

[0137] The above are all preferred embodiments of this application. The protection scope of this application is not limited accordingly. Therefore, all equivalent changes made according to the structure, shape, and principle of this application should be covered within the protection scope of this application.

Claims

1. A method for monitoring the operating state of a substation in a power system, characterized in that, Including: Data collection: Collect real-time status data of the substation. The real-time status data includes real-time device status data, real-time device location, and real-time health data of the device; collect historical status data of the substation and corresponding fault labels. The historical status data includes historical device status data, historical device location, and historical health data of the device; Model construction: Taking each device in the substation as a node and the causal relationship between devices as edges, construct a fault causal model; Model training: Including the first judgment, the first update, the first processing, the second processing, the third processing, and the first training; The first judgment: Based on the historical status data, use the non-linear Granger causality test method to construct a causal verification graph between devices, and judge whether the causal relationship in the causal verification graph is consistent with the edge between the device pairs in the fault causal model: If so, execute the steps of the first processing; If not, execute the steps of the first update; The first update: Based on the causal verification graph, update the fault causal model, and use the updated fault causal model as the new fault causal model; The first processing: Obtain the time stamps of the historical status data. Based on the fault labels, in the order of time stamps, merge the historical status data with the same fault label and continuous time stamps, and record the merged historical status data as the first data; The second processing: Calculate the Euclidean distance between device pairs based on the historical device location, and collect the Euclidean distances between each device pair to obtain the first feature set; The third processing: Obtain the time series of the first data, and judge whether the historical health data of the device changes within the time series: If not, execute the steps of the first training; If so, obtain the time points when the historical health data of the device changes, split the first data based on the obtained time points to get the split first data, and update the split first data to the first data; The first training: Concatenate the first feature set with the first data, update the concatenated data to the first data, and use the first data and the corresponding fault labels to train the new fault causal model to obtain the trained fault causal model; Fault discrimination: Input the real-time status data into the trained fault causal model to obtain the real-time fault label of the substation.

2. The substation operation state monitoring method of the power system according to claim 1, characterized in that After executing the steps of model construction and before executing the steps of model training, it also includes: The first construction: Denote the nodes in the fault causal model as the first nodes, denote the fault causal model as the first causal layer, obtain the functional attributes of each device in the substation, classify the devices based on the functional attributes to get several subsystems, take the subsystems as nodes, and take the causal relationship between subsystem pairs as edges to construct the second causal layer; The second construction: Denote the nodes of the second causal layer as the second nodes, take the substation as a node, and take the causal relationship between substation pairs as edges to construct the third causal layer; The third construction: Denote the nodes of the third causal layer as the third nodes, and update the first causal layer, the second causal layer, and the third causal layer to the fault causal model.

3. The substation operation status monitoring method of the power system according to claim 2, characterized in that After executing the steps of the third construction and before executing the steps of model training, it also includes: First transformation: Based on the first causal layer, each device is regarded as a first graph node, the causal relationship between each pair of devices is transformed into an embedding vector, and the obtained embedding vector is used as the node feature of the corresponding first graph node to construct a first causal graph; Second transformation: Based on the second causal layer, the subsystem is regarded as a second graph node, the embedding vectors of the first causal graph are used to calculate the embedding vectors between pairs of subsystems, and the obtained embedding vectors are used as the node features of the second graph node to construct a second causal graph; Third transformation: Based on the third causal layer, the substation is regarded as a third graph node, the embedding vectors of the second causal graph are used to calculate the embedding vectors between pairs of substations, and the obtained embedding vectors are used as the node features of the third graph node to construct a third causal graph; Fourth construction: Based on the first causal graph, the second causal graph and the third causal graph, a multi-causal relationship graph is constructed, and the constructed multi-causal relationship graph is used as a new fault causal model; In the first update step, the new fault causal model is updated based on the causal verification graph.

4. The substation operation status monitoring method of the power system according to claim 3, characterized in that, After performing the first judgment step and before performing the first update step, it also includes: Fourth processing: Add lag features to the historical state data to obtain a first data set; Fifth construction: Construct a causal forest, train the causal forest using the historical state data to obtain a trained causal forest, input the first data set and the causal verification graph into the trained causal forest to obtain the causal relationship prediction results between each pair of devices; Second judgment: Judge whether the causal relationship prediction results between each pair of devices are exactly the same as the edges between the first graph node pairs in the fault causal model: If so, perform the first processing step; If not, based on the causal relationship prediction results between each pair of devices, construct a new causal verification graph, update the new causal verification graph as the causal verification graph, and perform the first update step.

5. The method for monitoring the operation status of a substation in a power system according to claim 2, characterized in that, After performing the first judgment step and before performing the first processing step, it also includes: Third judgment: Based on the historical state data, use the non-linear Granger causality test method to construct a causal verification graph between each pair of subsystems, denoted as the first verification graph, and judge whether the causal relationship between each pair of subsystems in the first verification graph is the same as the edges between the second node pairs in the fault causal model: If so, perform the fourth judgment step; If not, use the causal verification graph and the first verification graph as the new causal verification graph, and perform the first update step; Fourth judgment: Based on the historical state data, use the non-linear Granger causality test method to construct a causal verification graph between each pair of substations, denoted as the second verification graph, and judge whether the causal relationship between each pair of substations in the second verification graph is the same as the edges between the third node pairs in the fault causal model: If so, perform the first processing step; If not, update the causal verification graph, the first verification graph and the second verification graph as the causal verification graph, and perform the first update step.

6. The substation operation state monitoring method of the power system according to claim 4, characterized in that, After performing the first update step and before performing the first processing step, it also includes: First collection: Input the first data set and the causal verification graph into the trained causal forest to obtain the confidence level of the causal relationship between the first graph node pairs; First annotation: Based on the fault causal model before update, annotate the added causal relationships and deleted causal relationships in the updated fault causal model, and collect the annotated causal relationships as updated causal relationships; First screening: Obtain the confidence levels of the updated causal relationships to obtain a confidence level data set; First calculation: Obtain the first graph node pairs corresponding to the updated causal relationships, calculate the embedding vectors of the first graph node pairs corresponding to the updated causal relationships before and after the update, and based on the calculated embedding vectors, calculate the difference between the embedding vectors before and after the update, and use the calculated difference as the vector change amount of the updated causal relationships; Second calculation: Perform weighted summation on the vector change amount of the updated causal relationships and the corresponding confidence levels, and use the calculation result as the correlation degree of the updated causal relationships; First discrimination: Determine whether the correlation degree of the updated causal relationships is higher than a preset correlation degree threshold: If so, execute the steps of the second discrimination; If not, execute the steps of the third discrimination; Second discrimination: Determine whether the current updated causal relationship is an added causal relationship: If so, do not process; If not, add the causal relationship of the first graph node pair corresponding to the current updated causal relationship; Third discrimination: Determine whether the current updated causal relationship is a deleted causal relationship: If so, do not process; If not, delete the causal relationship of the first graph node pair corresponding to the current updated causal relationship.

7. The method for monitoring the operation status of a substation in a power system according to claim 1, characterized in that, Each device in the substation contains a device ID. After executing the steps of the first processing and before executing the steps of the second processing, it further includes: Second annotation: Based on each device in the substation, annotate the corresponding device ID for the collected historical status data, and record the annotated data as new first data; First association: Obtain the edges between each node pair, convert the edges between each node pair into causal relationship intensity values, and record them as the first feature; Position encoding: Based on the time stamps of the historical status data, perform position encoding on the time stamps of each node, and record the encoded time stamps as time features; Data update: Concatenate the time features, the first feature, and the new first data, and update the concatenated data as the first data.

8. The method for monitoring the operation status of a substation in a power system according to claim 7, wherein After executing the steps of the second processing and before executing the steps of the third processing, it further includes: First acquisition: Obtain the environmental parameters of the location where each device is located; Difference calculation: Calculate the environmental parameter differences between each device pair; Third calculation: Based on the environmental parameter differences between each device pair and the Euclidean distance, use the environmental space difference formula to calculate the environmental space differences between each device pair, and use the calculated environmental space differences between each device pair as a new first feature set; The calculation model of the environmental space difference formula is: ； Among them, is the environmental space difference between device i and device j, is the adjustment factor, is the difference in environmental parameters between device i and device j, is the Euclidean distance between device i and device j.

9. The substation operation status monitoring method of the power system according to claim 1, characterized in that, The historical status data also includes the historical usage duration of the device, the historical load data, and the historical resource data. After executing the steps of the second processing and before executing the steps of the third processing, it further includes: First statistics: Obtain the historical load data of each device and the maximum load value of each device, and count the total duration for which the historical load data of each device exceeds the maximum load value, the difference between the historical load data exceeding the maximum load value and the maximum load value, and the corresponding duration; Second Statistics: Obtain the historical resource data of each device and the minimum resource requirement value of each device, and count the total duration during which the historical resource data of each device is lower than the minimum resource requirement value, the difference between the historical resource data lower than the minimum resource requirement value and the minimum resource requirement value, and the corresponding duration; Load Calculation: Use the load impact formula to calculate the impact factor of the load fluctuation of each device on the device state, denoted as the first impact; Resource Calculation: Use the resource impact formula to calculate the impact factor of the resource fluctuation of each device on the device state, denoted as the second impact; Fourth Calculation: Based on the historical health data of the device, the first impact and the second impact, use the device health formula to calculate the actual health data of the device, and use the calculated actual health data as the new historical health data of the device; The calculation model of the load impact formula is: ； Among them, is the number of historical load data exceeding the maximum load value in the historical load data, a is the a-th historical load data exceeding the maximum load value, is the difference between the a-th historical load data exceeding the maximum load value and the maximum load value, is the maximum load value of the device, is the duration of the a-th historical load data exceeding the maximum load value, is the historical usage duration of the device; The calculation model of the resource impact formula is: ； Among them, is the quantity of historical resource data in the historical resource data that is lower than the minimum resource requirement value, b is the b-th historical resource data lower than the minimum resource requirement value, is the difference between the b-th historical resource data lower than the minimum resource requirement value and the minimum resource requirement value, is the minimum resource requirement value of the device, is the duration of the b-th historical resource data lower than the minimum resource requirement value; The calculation model of the device health formula is: ； Among them, is the influence weight of load fluctuation on the device state, is the influence weight of resource fluctuation on the device state, is the historical health data of the device.

10. The method for monitoring the operation state of a substation in a power system according to claim 1, wherein, After performing the steps of the third processing and before performing the steps of the first training, it further includes: Second Obtaining: Obtain the time point of the change in the historical health data of the device, denoted as the first time point; Third Obtaining: Obtain the historical health data of the device at the time point one before the first time point, denoted as the first device data; obtain the historical health data of the device at the time point one after the first time point, denoted as the second device data; Rate Calculation: Calculate the time difference between the time point one before the first time point and the time point one after the first time point, and based on the calculated time difference, the first device data and the second device data, calculate the change rate of the historical health data of the device, denoted as the first change rate; Fifth Judgment: Judge whether the first change rate is less than the preset rate threshold: If so, merge the first data segmented based on the first time point, and update the merged data as the first data; If not, do not perform any processing.

Images