Method for intelligently and dynamically evaluating substation equipment

Abstract

The invention discloses an intelligent dynamic evaluation method for substation equipment, which relates to the technical field of power equipment systems, and comprises the following steps: collecting multi-source heterogeneous data, obtaining a spatio-temporal data set of the substation equipment, performing feature extraction and tagging on the collected spatio-temporal data set of the substation equipment, and obtaining a tagged data set of the substation equipment; obtaining a complete substation equipment operation state label and a label set, constructing an evaluation model for the substation equipment, inputting the substation equipment operation state label into the substation equipment evaluation model, and carrying out state division on different types of substation equipment; the method comprises the following steps: receiving output evaluation states and causal interpretation of different types of substation equipment, screening the evaluation states of all the substation equipment to obtain conformed substation equipment, constructing a maintenance decision model, inputting historical data into the constructed substation equipment evaluation model for verification, setting a dynamic calculation and evaluation period, and performing dynamic calculation and evaluation on the substation equipment. And carrying out dynamic trimming on the model according to a verification result and an actual condition.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment systems, and more specifically, to a method for intelligent dynamic evaluation of substation equipment. Background Art

[0002] The process of evaluating the status of substation equipment is crucial throughout the entire life cycle of the equipment and is one of the core functional modules of the entire power equipment system. In order to maintain the reliability of substation equipment and avoid power grid failures and load losses, the equipment operation and maintenance team mainly carries out regular maintenance plans and fault repairs. According to the equipment operation regulations and evaluation guidelines, regular maintenance stipulates the maintenance work of the equipment according to different equipment types, voltage levels, and operation years. Regular maintenance is carried out regularly, which may cause a significant waste of manpower and resources; for fault repair, when the equipment fails, the operation and maintenance team participates in the repair passively. This method was mostly used in the early days. In today's continuous development of the power system, fault power outages will affect the economic development of the region; In recent years, with the widespread use of microcomputer protection in the power system and the improvement of manufacturing processes, equipment such as transformers, SF6 switches, and lightning arresters already support online monitoring. Condition-based maintenance has been gradually carried out in many places. By analyzing the records of real-time acquisition, online monitoring, and daily inspections, the inflection point where the equipment has problems is found, and corresponding maintenance and tests are carried out in advance to solve the problems. First, the operation data of the equipment parts are collected and analyzed, and then deductions and evaluations are carried out according to the deduction basis of the guidelines. Then, the situations of each part are summarized, and finally the comprehensive score of the equipment is obtained. Because there are many types of equipment, it is difficult to collect data, the equipment scale is large, and the evaluation mainly relies on manual work, which also brings three problems: a large amount of evaluation work, a long evaluation cycle, and different evaluation results for different people. The value of measurement, online monitoring, and test historical data has not been truly utilized. Either the maintenance is not carried out in time or the equipment is over-maintained. As a result, the traditional maintenance method is still adopted in many places, which is neither economical nor good for equipment maintenance; Although a certain level of interconnection can be achieved through the data middle platform, it is not thorough enough. Using the existing system to support the equipment status evaluation work still has deficiencies, which are mainly manifested in the following aspects: Difficult data fusion. Referring to the equipment evaluation basis, equipment parameters and equipment operation data are scattered and stored in various systems. The same equipment object number is not fully associated, and it is difficult to expand and analyze the data dimension; these data need to be processed and calculated before they can be used and cannot be used directly; The data quality is uneven. Due to the inconsistency of data collection standards, update frequencies, storage formats, etc. in each system, there are significant differences in data quality. The data update of some systems lags behind, and there are even cases of data missing or incorrect, which affects the accuracy and reliability of equipment status evaluation. In addition, the workload of data cleaning and preprocessing is large, increasing the complexity and difficulty of data use; The evaluation flexibility is insufficient. Different regions, climates and other environments where the equipment operates lead to different local equipment maintenance practices. When formulating corresponding rules and strategies, the existing system evaluation rules lack user participation. The evaluation items for different equipment types involve different calculation and deduction strategies and cannot be updated in a timely manner, so the functions cannot meet the actual use. Summary of the Invention

[0003] To solve the above problems, the present invention provides a method for intelligent dynamic evaluation of substation equipment.

[0004] The present invention provides a method for intelligent dynamic evaluation of substation equipment, including the following steps: S1. Collect multi-source heterogeneous data, construct a substation equipment integration model with substation equipment as the core, associate cross-system data through unified coding, achieve cross-system equipment matching, and add spatio-temporal dimension tags to each substation equipment data to obtain a spatio-temporal data set of substation equipment; S2. Extract features and label the spatio-temporal data set of the collected substation equipment to obtain substation equipment operation status labels and environmental factor labels, and integrate them to obtain a complete substation equipment operation status label and label set; S3. Construct an evaluation model for substation equipment, input the substation equipment operation status label into the substation equipment evaluation model to divide the status of different types of substation equipment; Establish a fault association network between substation equipment and the environment and input it into the substation equipment evaluation model; Establish a dynamic scoring for the substation equipment evaluation model, specifically including a basic rule layer and a causal intervention layer. The substation equipment evaluation model outputs the evaluation status and causal explanation of different types of substation equipment according to the multi-source heterogeneous data and environmental association relationship in S1; S4. Receive the evaluation status and causal explanation of different types of substation equipment output in S3, screen the evaluation status of all substation equipment to obtain compliant substation equipment, construct a maintenance decision model, dynamically generate a maintenance strategy for substation equipment, and based on the causal explanation, give priority to handling high-weight defects; S5. Input the substation equipment evaluation model constructed in S3 into historical data for verification, set the period of dynamic calculation and evaluation, support the secondary confirmation of the substation equipment owner for the output result, and dynamically revise the model according to the verification result and actual situation.

[0005] Preferably, the specific working steps of S1 are as follows: S11. Collect data from multi-source heterogeneous data sources such as equipment ledgers, defect records, maintenance records, hidden danger records, test records, inspection records, centralized control measurement data, auxiliary oil chromatographic data, SF6 on-line monitoring data, and fault recorder data; S12. Taking substation equipment as the core, associate cross-system data through unified coding, perform matching of cross-system substation equipment, and output the globally unique coding of each substation equipment; S13. Add spatio-temporal dimension tags to each substation equipment data, and output a fusion data set of substation equipment with spatio-temporal dimension tags.

[0006] Preferably, the specific steps of S13 are as follows: Connect to multi-source heterogeneous data sources within the substation; Perform format conversion and standardization processing on the accessed data, associate the same equipment data in different data sources through the globally unique coding of each substation equipment, establish a data quality evaluation index system, classify the multi-source heterogeneous data, and evaluate the accuracy, integrity, consistency, and timeliness of the data according to different classifications, identify and mark to remove low-quality data; Obtain the geographical location information of substation equipment, including longitude and latitude coordinates, and associate it with each equipment data record; Combined with terrain data and the geographical location of the substation, determine the altitude where the equipment is located, and add it as a tag to the data to reflect the impact of different altitudes on equipment performance and aging; According to the corrosiveness data of the substation environment, specifically including salt fog concentration, industrial pollution degree, and humidity; Take the geographical location information, altitude, and corrosiveness data of substation equipment as spatio-temporal dimension tags of substation equipment, add spatio-temporal dimension tags to each substation equipment data, and obtain the spatio-temporal data set of substation equipment.

[0007] Preferably, the specific steps of S2 include: S21. Connect to the spatio-temporal data set output by S1, classify the spatio-temporal data set of substation equipment, and extract the characteristics of the operation data of different types of spatio-temporal data sets according to the classification results to obtain the operation data characteristics of substation equipment; S22. According to the operation data characteristics of substation equipment, combined with the rated parameters and operation standards of substation equipment, generate operation status tags of substation equipment; Collect the spatio-temporal dimension tags in the spatio-temporal data set output by S1, extract the geographical location information, altitude, and corrosiveness data of each substation equipment, and assign corresponding environmental factor tags to each substation equipment; S23. Integrate the extracted operation status labels and environmental factor labels of substation equipment to construct a complete equipment status feature vector and label set.

[0008] Preferably, the specific steps of S21 are as follows: S211. Divide the spatio-temporal data set into three categories: numerical type, text type, and time series type, and adopt standardization, word embedding, and time-frequency transformation methods for feature extraction respectively; S212. Calculate the sliding window statistics for the numerical spatio-temporal data set, specifically including near mean, variance, extreme value, and change trend, and use the above sliding window statistics as the extracted features; S213. Extract semantic feature vectors from the text spatio-temporal data set through the Power BERT model, specifically including semantic encoding of the fault location, severity, and maintenance suggestions of the defect description text; S214. Adopt wavelet transform for the time series spatio-temporal data set to extract the frequency domain energy distribution features, and combine the dynamic time warping algorithm to match the historical fault waveform patterns.

[0009] Preferably, the specific steps of S3 are as follows: S31. The basic rule layer adopts the fuzzy comprehensive evaluation method to calculate the initial health score based on the equipment status feature vector, and the scoring formula is: ; where is the empirical weight, is the feature normalization function, is the initial score, where n is the total number of devices, and i represents the i-th device; S32. The causal intervention layer adopts the PC causal discovery algorithm to construct a fault association network, identify the key causal path and calculate the causal effect value, and adjust the initial score through the structural equation model: ; where is the causal path influence factor, is the causal effect coefficient, is the dynamic adjustment parameter, is the adjusted initial score, is the initial score; S33. Input the environmental factor labels into the graph neural network to generate the graph embedding features of the interaction between the device and the environment, and fuse them with the causally adjusted score to generate the final evaluation status.

[0010] Preferably, the specific steps of constructing the maintenance decision-making model in S4 include: S41. Establish a multi-agent game framework and define the utility functions of the three parties of maintenance resources, fault risk, and operation and maintenance costs: ; wherein 、 and are the utility weights for the three aspects of maintenance resources, failure risks, and operation and maintenance costs; is the failure risk utility function, is the operation and maintenance cost function, is the resource constraint function; S42. Solve the Pareto optimal solution set based on the NSGA-II multi-objective optimization algorithm to generate a maintenance strategy that takes into account risk reduction and resource constraints; S43. Automatically trigger a red warning for high-weight defects in the causal explanation, preferentially allocate maintenance resources, and generate a disposal work order within 48 hours.

[0011] Preferably, the specific steps of S5 are as follows: S51. Adopt a federated learning framework, and each substation node uploads the local causal structure to the cloud platform, and aggregates and generates a global causal knowledge base through knowledge distillation technology; S52. Establish a dual verification mechanism: for historical data verification, the leave-one-out cross-validation method is used to calculate the F1 value of the model, and for on-site verification, the equipment owner is required to perform a secondary confirmation on the evaluation result, and results with a confidence level lower than 85% trigger model reconstruction; S53. Set a dynamic update period, update the model once every quarter under normal working conditions, and start an emergency update when an environmental mutation is detected.

[0012] Preferably, S51 specifically includes the following steps: S511. Define a causal structure similarity measurement function; S512. Use the hierarchical clustering algorithm to divide sites with similarity greater than the threshold into homogeneous groups, and use the weighted average method to fuse causal graphs within the group; S513. Start an expert arbitration mechanism for conflicting causal paths, and determine the retained path by combining physical mechanism analysis.

[0013] Preferably, S5 further includes S54, specifically to establish a feedback closed-loop mechanism, and correct the causal knowledge graph through real-time feedback of the maintenance effect at the edge; The specific implementation of the feedback closed-loop mechanism is as follows: S101. Deploy a lightweight causal inference engine at the edge to receive the equipment status data after maintenance in real time; S102. Calculate the maintenance intervention effect index: ; wherein is the maintenance intervention effect index, is the real-time health score before repair, reflecting the current degradation state of the equipment. is the health score after maintenance, calculated from the status monitoring data. is the real-time health score before maintenance, reflecting the current degradation state of the equipment. S103 When it is greater than the threshold, it is determined as inefficient maintenance, and the causal knowledge graph backtracking and correction process is automatically triggered to update the weight parameters of the environment-fault association edge.

[0014] Beneficial effects: Optimize the operation and working efficiency of substation equipment: timely discover potential problems of the equipment, eliminate potential defects through preventive maintenance, maintain equipment performance and efficiency, reduce equipment failure rate, avoid power outages caused by sudden failures, and ensure the operation of the power grid; at the same time, realize digital state evaluation and simplify personnel carrying capacity. Control costs and improve benefits: fully explore the data value of inspections, measurements, online monitoring, etc., reasonably evaluate, optimize the allocation of maintenance resources, reduce unnecessary downtime and maintenance costs, avoid wastes such as over-maintenance or insufficient maintenance, and achieve differentiated operation and maintenance; effectively reduce the equipment wear and aging speed, extend the service life of the equipment, reduce the power outage duration, and simply improve the production efficiency of the power grid. Timely and differentially eliminate potential safety hazards, reduce safety risks caused by equipment failures (such as explosions, leaks, etc.), implement state evaluation work online, intelligently, and standardly, reduce problems of missed evaluation or incorrect evaluation caused by individuals, establish an equipment health chronicle, and provide data support for long-term maintenance and operation decisions. Description of the Drawings

[0015] Figure 1 is the flow chart of the system of the present invention. Detailed Embodiments

[0016] As Figure 1 shown: A method for intelligent dynamic evaluation of substation equipment includes the following steps: S1. Collect multi-source heterogeneous data, construct a substation equipment integration model with substation equipment as the core, associate cross-system data through unified coding, achieve cross-system equipment matching, and add spatio-temporal dimension tags to each substation equipment data to obtain the spatio-temporal data set of substation equipment; it should be noted that, such as geographical location longitude and latitude, altitude, environmental corrosion grade, the subsequent model can identify the differential aging laws of, for example, "transformers in high-salt fog areas along the coast" and "transformers in clean inland areas", special treatment and the associated design of the subsequent model, spatio-temporal tag injection. It should be noted that data is specifically collected from multi-source heterogeneous data sources such as equipment ledgers, defects, maintenance, potential hazards, test records, inspections, centralized control measurements, auxiliary oil chromatography, SF6 online monitoring, and fault recording. S2. Extract features from the spatio-temporal dataset of substation equipment and label them to obtain the operation status labels and environmental factor labels of substation equipment, and integrate them to obtain the complete operation status labels and label sets of substation equipment; S3. Build an evaluation model for substation equipment, and input the operation status labels of substation equipment into the substation equipment evaluation model to classify the status of different types of substation equipment; Establish a fault correlation network between substation equipment and the environment, and input it into the evaluation model of substation equipment; It should be noted that to obtain the multi-source heterogeneous data and environmental correlation relationships in S1, traditional methods mostly rely on manually setting scoring rules or models based on a single data source, lacking the mining of causal relationships between features; Establish a dynamic scoring for the substation equipment evaluation model, specifically including a basic rule layer and a causal intervention layer. The substation equipment evaluation model outputs the evaluation status and causal explanations of different types of substation equipment according to the multi-source heterogeneous data and environmental correlation relationships in S1; S4. Receive the evaluation status and causal explanations of different types of substation equipment output by S3, screen the evaluation status of all substation equipment to obtain compliant substation equipment, build a maintenance decision-making model, dynamically generate the maintenance strategy of substation equipment, and preferentially handle high-weight defects based on causal explanations; S5. Input the substation equipment evaluation model constructed in S3 into historical data for verification, set the cycle of dynamic calculation and evaluation, support the secondary confirmation by the owner of substation equipment for the output results, and dynamically revise the model according to the verification results and actual situations. Continuously improve the accuracy of the model to adapt to the changes in the operation status of substation equipment and the improvement of management requirements; It should be noted that this technical solution collects multi-source heterogeneous data more comprehensively, including not only conventional data such as equipment ledgers and defects, but also spatio-temporal dimension labels such as geographical location longitude and latitude, altitude, and environmental corrosion level, realizing spatio-temporal correlation modeling, enabling the model to accurately identify the differential aging laws of equipment under different environments, while traditional methods rarely pay attention to the impact of spatio-temporal factors on equipment status; Establish a fault correlation network between the substation and the environment, introduce a causal intervention layer, and combine causal discovery algorithms to mine the causal relationships between features, making up for the deficiencies of traditional methods that rely on manually setting scoring rules or single-data-source models, and being able to output more interpretable evaluation results. Traditional methods have limitations in causal relationship mining; Preferentially handle high-weight defects based on causal explanations, build a multi-agent game decision-making model, and compared with the traditional relatively fixed maintenance strategy formulation method, can more accurately dynamically generate maintenance strategies and achieve the optimal allocation of maintenance resources; By adopting the federated causal evolution and feedback closed-loop mechanism, each site uploads local causal structures to aggregate and generate a global knowledge base, and the edge side real-time feedbacks the maintenance effect to correct the knowledge graph, realizing dynamic model reduction and continuous optimization, adapting to the changes in the equipment operation status and the requirements of management improvement. Traditional methods are relatively weak in dynamic model optimization.

[0017] As an optional embodiment: The specific working steps of S1 are as follows: S11. Collect data from multi-source heterogeneous data sources such as equipment ledgers, defect records, maintenance records, hidden danger records, test records, patrol records, centralized control measurement data, auxiliary oil chromatogram data, SF6 on-line monitoring data, and fault recorder data; ensure the comprehensiveness and accuracy of the data; S12. Taking substation equipment as the core, associate cross-system data through unified coding, perform matching of cross-system substation equipment, and output the globally unique coding of each substation equipment; lay a foundation for subsequent data fusion and analysis; It should be noted that in combination with the logical node naming rules in the IEC61850 standard, a globally unique coding is generated for each device. For example, substation code + voltage level + device type + serial number. The coding "B500kV_TF_001" represents the No. 1 transformer of a certain 500kV substation; Adopt a pre-trained BERT model in the power field to extract equipment name entities; for example, input: "The cooler of the No. 3 main transformer leaks oil" and output: "No. 3, main transformer, cooler, leaks oil"; Realize the accurate matching and data association of cross-system substation equipment, lay a foundation for subsequent data fusion and analysis, and ensure the consistency and accuracy of equipment data through unified coding and the pre-trained BERT model in the power field, so as to better perform equipment status evaluation, fault prediction, and maintenance strategy formulation; S13. Add spatio-temporal dimension labels to the data of each substation equipment, and output a fusion data set of substation equipment with spatio-temporal dimension labels. So that the subsequent model can identify the differential aging laws of equipment in different environments and realize spatio-temporal correlation modeling; It should be noted that the spatio-temporal dimension labels are such as geographical location longitude and latitude, altitude, and environmental corrosion grade; Perform fusion processing on the collected multi-source heterogeneous data, and add spatio-temporal dimension labels to the data of each substation equipment, providing a high-quality fusion data basis for subsequent equipment status evaluation, fault prediction, and intelligent decision-making; It should also be noted that by fusing multi-source heterogeneous data and adding spatio-temporal dimension labels, a complete and high-quality spatio-temporal data set is formed, providing a solid data basis for subsequent analysis and modeling, enabling the solution to better adapt to the equipment operation characteristics in different environments, and improving the generality and practicality of the solution.

[0018] As an optional embodiment, the specific steps of S13 are as follows: Connect to multi-source heterogeneous data sources within the substation; including but not limited to equipment inventory systems, defect management systems, on-line monitoring devices, environmental monitoring sensors, etc., to ensure the comprehensiveness and real-time nature of the data; Perform format conversion and standardization processing on the accessed data, associate the same equipment data in different data sources through the globally unique encoding of each substation equipment, establish a data quality evaluation index system, classify the multi-source heterogeneous data, and evaluate the accuracy, integrity, consistency, and timeliness of the data according to different classifications, identify and mark and remove low-quality data; It should be noted that by using a unified encoding (such as substation code + voltage level + equipment type + serial number) to associate the same equipment data in different data sources, the accuracy of data consistency evaluation is ensured; The specific methods for evaluating the accuracy, integrity, consistency, and timeliness of the data are as follows: Classify the multi-source heterogeneous data, specifically including numerical data: Centralized control measurement, auxiliary oil chromatogram, SF6 on-line monitoring, and test records; Text data: Equipment inventory, defect management, maintenance records, hidden danger records, and patrol records; Time series data: On-line monitoring devices, fault recorders, environmental monitoring sensors, and some test records; For numerical data, compare the data recorded by the on-line monitoring device with the calibrated standard value of the equipment. If the error is within the allowable range of the equipment accuracy, which is ±5% in this embodiment, then the record is considered accurate; Suppose an on-line monitoring device records 100 pieces of data, and 95 of them have errors within a reasonable range. Then the accuracy of the data of this device is 95%. Mark and remove the data with accuracy lower than the industry standard as low-quality data; For text data, verify on-site or compare with authoritative documents to confirm whether the text content is consistent with the actual situation; Suppose an equipment inventory system has 100 records, and 98 of them are consistent with the actual situation. Then the accuracy of the data of this system is 98%. Mark and remove the data with accuracy lower than the industry standard as low-quality data; For structured data, check whether each field has a value. If all fields are filled in completely, it is regarded as a complete record; Suppose a device ledger system should have 100 records, and all fields of 95 records have values. Then the data integrity of this system is 95%. Data with accuracy lower than the industry standard is marked as low-quality data and removed; Obtain the geographical location information of substation equipment, including longitude and latitude coordinates, and associate it with each device data record; Combine the terrain data and the geographical location of the substation to determine the altitude where the equipment is located, and add it as a label to the data to reflect the impact of different altitudes on equipment performance and aging; According to the corrosivity data of the substation environment, specifically including salt spray concentration, industrial pollution degree, and humidity; Take the geographical location information, altitude, and corrosivity data of the substation equipment as the spatio-temporal dimension labels of the substation equipment, add spatio-temporal dimension labels to each substation equipment data, and obtain the spatio-temporal dataset of the substation equipment.

[0019] As an optional embodiment: The specific steps of S2 include: S21. Connect to the spatio-temporal dataset output by S1, classify the spatio-temporal dataset of substation equipment, and extract features from different types of spatio-temporal datasets according to the classification results to obtain the operation data features of substation equipment. It should be noted that due to the large number of types of multi-source heterogeneous data, conventional feature extraction methods need to be carried out one by one. The steps are cumbersome and the extracted features need to be re-integrated. In this technical solution, by classifying the spatio-temporal dataset of substation equipment and extracting features from different types of spatio-temporal datasets according to the classification results, the multi-source heterogeneous data is divided into three categories and three different extraction methods are used, which simplifies the extraction steps and improves the extraction efficiency; S22. According to the operation data features of substation equipment, combine the rated parameters and operation standards of substation equipment to generate the operation status labels of substation equipment. It should be noted that in this embodiment, the specific steps are as follows: Input the operation data of the equipment, including but not limited to numerical data such as current, voltage, power, temperature, etc.; Continuously collect the operation data of the equipment at different time periods to ensure the integrity and accuracy of the data; Set a reasonable threshold range according to the rated parameters and historical operation data of the equipment. In this embodiment, the threshold range is obtained according to the factory rated parameters of the substation equipment; For the operation data of each time period, judge whether it is within the set threshold range. If it is within the range, mark it as "normal"; if it exceeds the upper limit, mark it as "overload"; if it is lower than the lower limit, mark it as "underload"; Output the operation status labels of the equipment, such as "normal", "overload", "underload"; Collect the spatiotemporal dimension labels in the spatiotemporal data set output by S1, extract the geographic location information, altitude and corrosiveness data of each substation equipment, and assign corresponding environmental factor labels to each substation equipment; It should be noted that the geographical location information, altitude and corrosiveness data of the substation equipment are input, the geographical location information, altitude and corrosiveness data of the substation equipment are collected, and the data are standardized; Extract corrosion-related features, such as the mean and variance of salt spray concentration, the level of industrial pollution, the range of humidity, etc. Using the collected data and features, an environmental corrosion level assessment model was constructed and trained using the XGBoost machine learning algorithm. The model input was the corrosive data features and the output was the corrosion level label. Input the corrosive data features of the equipment to be evaluated into the trained model to generate corresponding environmental corrosion level labels, such as "low corrosion", "medium corrosion" and "high corrosion"; Output the corrosion level label of the environment in which the substation equipment is located, such as "low corrosion", "medium corrosion" and "high corrosion". Additional environmental factor labels are extracted to facilitate subsequent evaluation calculations; S23, integrate the extracted operation status labels and environmental factor labels of the substation equipment to construct a complete equipment status feature vector and label set. It should be noted that through feature engineering and labeling processing, the operation status and environmental information of the equipment are fully and accurately reflected, and based on the fused features and labels, the equipment status information containing the equipment status feature vector and label set is generated; It should also be noted that the specific steps for constructing a complete device state feature vector and label set are as follows: Collect features: Summarize the numerical data features, text data features, and time series data features extracted from S21. This includes the mean, variance, trend slope of the equipment operation data, the semantic vector of the text information, and the waveform distortion rate and spectrum energy distribution of the time series.

[0020] Feature alignment: Ensure that the dimensions of all feature vectors are consistent, and perform zero padding or truncation on feature vectors of different lengths to facilitate subsequent fusion operations.

[0021] Feature concatenation: connect the feature vectors of numerical, textual, and time series data end to end in a predefined order to form a complete feature vector. For example, concatenate numerical features first, then textual features, and finally time series features.

[0022] Collect the operation status labels, defect status labels and environmental factor labels generated in S22; Combine the operating status label, defect status label, and environmental factor label into a label set, adopt a data frame structure, and store various labels with the device identifier and time as the index; Through the globally unique encoding of each substation device, associate the integrated feature vector with the label set to ensure that each feature vector corresponds to the correct label set and the corresponding substation device.

[0023] As an optional embodiment: The specific steps of S21 are as follows: S211. Divide the spatio-temporal data set into three categories: numerical type, text type, and time series type, and adopt standardization, word embedding, and time-frequency transformation methods for feature extraction respectively; S212. Calculate the sliding window statistics for the numerical spatio-temporal data set, specifically including near mean, variance, extreme value, and change trend, and use the above sliding window statistics as the extracted features; S213. Extract semantic feature vectors from the text spatio-temporal data set through the Power BERT model, specifically including the semantic encoding of the fault location, severity, and maintenance suggestions of the defect description text; S214. Adopt wavelet transform to extract the frequency domain energy distribution features from the time series spatio-temporal data set, and combine the dynamic time warping algorithm to match the historical fault waveform pattern.

[0024] According to the above method for intelligent dynamic evaluation of substation equipment, the specific steps of S3 are as follows: S31. The basic rule layer adopts the fuzzy comprehensive evaluation method to calculate the initial health score based on the device status feature vector. The scoring formula is: ; where is the empirical weight, is the feature normalization function, is the initial score, where n is the total number of devices, and i represents the i-th device; It should also be noted that The weights of each state feature set by domain experts according to the device historical fault data and operation experience reflect the influence degree of different features on the health state; The original feature (such as temperature, vibration amplitude, etc.) is mapped to the [0,1] interval function, such as Min-Max normalization or Sigmoid function; Quantify the initial health state of the device by weighted sum of multi-dimensional state features, as the basis for subsequent causal adjustment; Traditional methods usually directly use raw data or simple statistical weights (such as the entropy weight method), while here expert knowledge is combined to avoid weight bias driven purely by data and improve the physical interpretability of the initial score;

[0025] S32. The causal intervention layer uses the PC causal discovery algorithm to construct a fault correlation network, identify key causal paths and calculate causal effect values, and adjust the initial score through a structural equation model: ; where is the causal path influence factor, is the causal effect coefficient, is the dynamic adjustment parameter, is the adjusted initial score, is the initial score; It should also be noted that is the contribution degree of the key causal path (such as "humidity → insulation deterioration → discharge") identified by the PC causal discovery algorithm to the fault, which is calculated by the conditional probability of the nodes in the path; is the regression coefficient of the causal path in the structural equation model, quantifying the strength of the causal relationship (such as the increase in the probability of insulation deterioration caused by a 1-unit increase in humidity); is a parameter that adapts to environmental mutations and the historical state of the device, controlling the influence weight of causal intervention on the total score, and taking a value of 0.97 in this embodiment; On the basis of the initial score, the causal effect is superimposed to correct the evaluation bias caused by ignoring the fault propagation path. Existing methods are mostly based on correlation analysis (such as the Pearson coefficient), while here the fault propagation mechanism is modeled through causal reasoning (PC algorithm + structural equation) to avoid the interference of spurious correlation and more accurately identify the true causal impact of key fault paths such as environmental factors on insulation aging; S33. Input the environmental factor label into the graph neural network to generate the graph embedding features of the interaction between the device and the environment, and fuse them with the causally adjusted score to generate the final evaluation state.

[0026] It should be noted that the device and environment nodes are constructed into a graph structure, and the interaction features between the device and the environment (such as local topological relationships) are extracted through a graph convolutional network.

[0027] The graph embedding features are concatenated with the causally adjusted score H and input into the fully connected layer to generate the final evaluation state. Traditional methods usually process environmental data separately (such as threshold alarms), while the GNN models the dynamic coupling relationship between the device and the environment; Capture the indirect impact of environmental mutations on the device state (such as salt spray corrosion conducting to device failures through multiple intermediate nodes); As an alternative embodiment: The specific steps of constructing the maintenance decision-making model in S4 include: S41. Establish a multi-agent game framework and define the utility functions of three parties: maintenance resources, failure risks, and operation and maintenance costs: ; Among them , , and are the utility weights for the three aspects of maintenance resources, failure risks, and operation and maintenance costs; is the failure risk utility function, is the operation and maintenance cost function, is the resource constraint function; It should be noted that the failure risk utility function is expressed in exponential form as , where γ is the risk sensitivity coefficient, which is used to control the attenuation rate of risk with the health score H; The larger the γ value, the faster the risk decreases under the same health score H, which is applicable to high-risk equipment; The exponential function is selected because it conforms to the concavity characteristic of the risk-averse utility function and can reflect the actual situation that "the worse the health state, the increasing marginal effect of risk". For example, the risk of insulation deterioration increases slowly in the initial stage and will explode exponentially in the later stage. The acquisition method of the γ parameter is determined by fitting historical failure data. For example, the maximum likelihood estimation method is used to optimize γ to make the risk curve predicted by the model as close as possible to the actual failure rate distribution; The operation and maintenance cost function is expressed in a linear relationship as Frequency + β Spare part price. This is because the maintenance frequency is positively correlated with the spare part consumption. For example, an increase in the frequency of transformer oil chromatographic analysis will directly lead to an increase in labor costs, and the spare part price will linearly affect the total cost due to market fluctuations; Among them, h is the frequency cost coefficient, which is calculated by multiplying the labor hour unit price by the time-consuming of a single maintenance; β is the spare part cost coefficient, which is the moving average of the average purchase price of spare parts; The resource constraint function adopts the form of a logistic function, that is ; Here, K represents the upper limit of resource capacity, such as the total number of deployable technical personnel; N is the current resource occupancy; N_c is the critical value of resource utilization rate, usually taking half of K; θ is the curve steepness parameter, which reflects the mutation threshold of resource tension; The parameter K is determined by the resource ledger of the operation and maintenance department, and θ is calibrated through equipment load pressure testing. For example, the inflection point of the change in the human resource response delay rate under different loads is simulated.

[0028] The solution of the utility weight is realized based on the Nash equilibrium; Construct a game model with agents in three aspects: risk control, cost management, and resource scheduling; The strategy space is , , and The value ranges of, and, and their sum is usually constrained to be 1. The solution goal of the Nash equilibrium is to minimize the utility conflicts among the three parties. The optimization algorithm adopted is Sequential Least Squares Programming (SLSQP), whose initial value is taken from the weighted mean fitted from historical operation and maintenance data, and then approaches the Pareto optimal solution set through iteration. In addition, a dynamic adjustment mechanism is designed. When environmental mutation detection is introduced (such as the monthly increase in salt fog concentration exceeding 15%), it will trigger the rebalancing of weights. And the game results of multiple sites are aggregated through the federated learning framework to update the global equilibrium parameters; S42. Solve the Pareto optimal solution set based on the NSGA-II multi-objective optimization algorithm to generate maintenance strategies that take into account risk reduction and resource constraints; S43. Automatically trigger a red warning for defects with high weights in the causal explanation, preferentially allocate maintenance resources, and generate a disposal work order within 48 hours.

[0029] It should be noted that non-dominated sorting is combined with crowding degree calculation to retain a diverse solution set.

[0030] Defects with β>0.7 in the causal path (such as the strong causal effect of insulation defects on discharge) automatically generate work orders; traditional methods rely on fixed-threshold alarms, while here it is dynamically triggered in combination with causal effects; Preferentially handle defects with high causal weights to reduce the risk of cascading failures; As an optional embodiment: The specific steps of S5 are as follows: S51. Adopt the federated learning framework, and each substation node uploads the local causal structure to the cloud platform; It should be noted that the specific steps are as follows: Define a causal structure similarity metric function: ; where is the set of causal edges of site ; Use the hierarchical clustering algorithm to divide sites with causal graph similarity > threshold into homogeneous groups, and use the weighted average method to fuse the causal graphs within the group. In this embodiment, the threshold value is taken as 0.8; Start an expert arbitration mechanism for conflicting causal paths (opposite directions), and determine the retained path in combination with physical mechanism analysis; where is the causal graph similarity, and calculate the proportion of common causal edges among sites (such as both sites having a "humidity → insulation deterioration" path); Hierarchical clustering groups similar sites, and within the group, a causal graph is fused (such as weighted average edge weights).

[0031] Traditional federated learning aggregates model parameters, while here, the causal structure is aggregated to extract global knowledge while preserving local causal specificity (such as the salt spray corrosion path of coastal sites); S52. Establish a dual verification mechanism: For historical data verification, the leave-one-out cross-validation method is used to calculate the F1 value of the model. For on-site verification, the equipment owner is required to conduct a secondary confirmation of the evaluation results. Results with a confidence level lower than 85% trigger model reconstruction; S53. Set a dynamic update period. Under normal operating conditions, the model is updated once every quarter. When an environmental mutation is detected, an emergency update is initiated.

[0032] As an optional embodiment: The specific steps of S51 are as follows: S511. Define a causal structure similarity metric function; S512. Use the hierarchical clustering algorithm to divide sites with similarity greater than the threshold into homogeneous groups, and within the group, the causal graph is fused using the weighted average method; S513. Initiate an expert arbitration mechanism for conflicting causal paths, and determine the retained path in combination with physical mechanism analysis.

[0033] As an optional embodiment: S5 further includes S54, specifically to establish a feedback closed-loop mechanism to correct the causal knowledge graph through real-time feedback of the maintenance effect at the edge; The specific implementation of the feedback closed-loop mechanism is as follows: S101. Deploy a lightweight causal inference engine at the edge to receive the device status data after maintenance in real time; S102. Calculate the maintenance intervention effect index: ; Where is the maintenance intervention effect index, is the real-time health score before maintenance, reflecting the current degradation state of the device, is the health score after maintenance, calculated from the status monitoring data, is the real-time health score before maintenance, reflecting the current degradation state of the device; By comparing the differences in the health status before and after maintenance with the reference value, the actual effect of the maintenance intervention is quantified. The numerator in the formula represents the "health potential" restored by the device, and the denominator represents the health improvement actually brought by the maintenance; S103. When it is greater than the threshold, it is determined as an inefficient maintenance, and the causal knowledge graph backtracking correction process is automatically triggered to update the weight parameters of the environment-fault association edge.

[0034] In this embodiment, the threshold is 0.5, indicating that less than 50% of the potential health losses have been restored after the overhaul, which may be caused by incorrect identification of causal paths (such as ignoring key environmental factors) or model parameter deviation, and it is necessary to trigger the backtracking repair of the causal knowledge graph; The above are only the preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions within the idea of the present invention belong to the protection scope of the present invention. It should be pointed out that for those of ordinary skill in the art, several improvements and refinements made without departing from the principle of the present invention should also be regarded as within the protection scope of this template.

Claims

1. A method for intelligent dynamic evaluation of substation equipment, characterized in that, It includes the following steps: S1. Collect multi-source heterogeneous data, construct a substation equipment integration model with substation equipment as the core, associate cross-system data through unified coding, achieve cross-system equipment matching, and add spatio-temporal dimension tags to each substation equipment data to obtain a spatio-temporal data set of substation equipment; S2. Extract features and tag the spatio-temporal data set of the collected substation equipment to obtain substation equipment operation status tags and environmental factor tags, and integrate them to obtain a complete substation equipment operation status tag and tag set; S3. Construct an evaluation model for substation equipment, input the substation equipment operation status tags into the substation equipment evaluation model to classify the status of different types of substation equipment; Establish a fault association network between substation equipment and the environment, and input it into the substation equipment evaluation model; Establish a dynamic scoring for the substation equipment evaluation model, specifically including a basic rule layer and a causal intervention layer. The substation equipment evaluation model outputs the evaluation status and causal explanations of different types of substation equipment according to the multi-source heterogeneous data and environmental association relationship in S1; S4. Receive the evaluation status and causal explanations of different types of substation equipment output by S3, screen the evaluation status of all substation equipment to obtain compliant substation equipment, construct a maintenance decision-making model, dynamically generate a maintenance strategy for substation equipment, and based on the causal explanation, give priority to handling high-weight defects; S5. Input the substation equipment evaluation model constructed in S3 into historical data for verification, set the period of dynamic calculation and evaluation, support the second confirmation by the owner of the substation equipment for the output result, and dynamically revise the model according to the verification result and actual situation.

2. The method for intelligent dynamic evaluation of substation equipment according to claim 1, characterized in that The specific working steps of S1 are as follows: S11. Collect data from multi-source heterogeneous data sources such as equipment ledgers, defect records, maintenance records, hidden danger records, test records, patrol records, centralized control measurement data, auxiliary oil chromatogram data, SF6 on-line monitoring data, and fault recorder data; S12. With substation equipment as the core, associate cross-system data through unified coding, match cross-system substation equipment, and output the globally unique code of each substation equipment; S13. Add spatio-temporal dimension tags to each substation equipment data, and output a fusion data set of substation equipment with spatio-temporal dimension tags.

3. The method for intelligent dynamic evaluation of substation equipment according to claim 2, characterized in that, The specific steps of S13 are as follows: Connect to multi-source heterogeneous data sources in the substation; Perform format conversion and standardization processing on the accessed data, associate the same equipment data in different data sources through the globally unique code of each substation equipment, establish a data quality evaluation index system, classify the multi-source heterogeneous data, and evaluate the accuracy, integrity, consistency, and timeliness of the data according to different classifications, identify and mark and remove low-quality data; Obtain the geographical location information of substation equipment, including longitude and latitude coordinates, and associate it with each equipment data record; Combine the terrain data and the geographical location of the substation to determine the altitude where the equipment is located, and add it as a tag to the data to reflect the influence of different altitudes on equipment performance and aging. According to the corrosivity data of the substation environment, specifically including salt spray concentration, industrial pollution degree, and humidity; Use the geographical location information, altitude, and corrosivity data of the substation equipment as the spatio-temporal dimension labels of the substation equipment, add spatio-temporal dimension labels to each substation equipment data, and obtain the spatio-temporal dataset of the substation equipment.

4. The method for intelligent dynamic evaluation of substation equipment according to claim 1, wherein The specific steps of S2 are as follows: S21. Connect to the spatio-temporal dataset output by S1, classify the spatio-temporal dataset of the substation equipment, and extract features from different types of spatio-temporal datasets according to the classification results to obtain the operation data features of the substation equipment; S22. According to the operation data features of the substation equipment, combined with the rated parameters and operation standards of the substation equipment, generate the operation status labels of the substation equipment; Collect the spatio-temporal dimension labels in the spatio-temporal dataset output by S1, extract the geographical location information, altitude, and corrosivity data of each substation equipment, and assign corresponding environmental factor labels to each substation equipment; S23. Integrate the extracted operation status labels and environmental factor labels of the substation equipment to construct a complete equipment status feature vector and label set.

5. The method for intelligent dynamic evaluation of substation equipment according to claim 4, characterized in that The specific steps of S21 are as follows: S211. Divide the spatio-temporal dataset into three categories: numerical type, text type, and time series type, and use standardization, word embedding, and time-frequency transformation methods for feature extraction respectively; S212. Calculate the sliding window statistics for the numerical spatio-temporal dataset, specifically including near mean, variance, extreme value, and change trend, and use the above sliding window statistics as the extracted features; S213. Extract semantic feature vectors from the text spatio-temporal dataset through the Power BERT model, specifically including semantic encoding of the fault location, severity, and maintenance suggestions of the defect description text; S214. Use wavelet transform to extract the frequency domain energy distribution features from the time series spatio-temporal dataset, and combine the dynamic time warping algorithm to match the historical fault waveform pattern.

6. The method for intelligent dynamic evaluation of substation equipment according to claim 1, characterized in that, The specific steps of S3 are as follows: S31. The basic rule layer uses the fuzzy comprehensive evaluation method to calculate the initial health score based on the equipment status feature vector, and the scoring formula is: ； wherein is the empirical weight, is the feature normalization function, is the initial score, where n is the total number of devices and i represents the i-th device; S32. The causal intervention layer uses the PC causal discovery algorithm to construct a fault correlation network, identify the key causal path and calculate the causal effect value, and adjust the initial score through the structural equation model: ； Among them is the causal path influence factor, is the causal effect coefficient, is the dynamic adjustment parameter, is the adjusted initial score, is the initial score; S33. Input the environmental factor labels into the graph neural network to generate graph embedding features of the interaction between the equipment and the environment, and fuse them with the causally adjusted score to generate the final evaluation status.

7. A method for intelligent dynamic evaluation of substation equipment according to claim 1, characterized in that, The specific steps of constructing the maintenance decision-making model in S4 include: S41. Establish a multi-agent game framework and define the utility functions of three parties: maintenance resources, fault risks, and operation and maintenance costs; ； Among them 、 、 and are the utility weights for the three aspects of maintenance resources, failure risks, and operation and maintenance costs; is the failure risk utility function, is the operation and maintenance cost function, is the resource constraint function; S42. Solve the Pareto optimal solution set based on the NSGA-II multi-objective optimization algorithm to generate a maintenance strategy that takes into account risk reduction and resource constraints; S43. Automatically trigger a red warning for high-weight defects in the causal explanation, prioritize the allocation of maintenance resources and generate a disposal work order within 48 hours.

8. A method for intelligent dynamic evaluation of substation equipment according to claim 1, characterized in that, The specific steps of S5 are as follows: S51. Adopt the federated learning framework, and each substation node uploads the local causal structure to the cloud platform, and aggregates and generates a global causal knowledge base through knowledge distillation technology; S52. Establish a dual verification mechanism: For historical data verification, the F1 value of the model is calculated using the leave-one-out cross-validation method. For on-site verification, the equipment owner is required to confirm the evaluation results again. Results with a confidence level lower than 85% trigger model reconstruction; S53. Set a dynamic update period. Under normal operating conditions, the model is updated quarterly. When an environmental mutation is detected, an emergency update is initiated; S54. Establish a feedback closed-loop mechanism to correct the causal knowledge graph by real-time feedback of the maintenance effect at the edge; 9. The method for intelligent dynamic evaluation of substation equipment according to claim 8, characterized in that, The specific steps of S51 are as follows: S511. Define a causal structure similarity metric function; S512. Use the hierarchical clustering algorithm to divide stations with similarity greater than the threshold into homogeneous groups, and use the weighted average method to fuse causal graphs within the groups; S513. Start an expert arbitration mechanism for conflicting causal paths, and determine the retained path by combining physical mechanism analysis; 10. A method for intelligent dynamic evaluation of substation equipment according to claim 8, characterized in that The specific implementation of the feedback closed-loop mechanism is as follows: S101. Deploy a lightweight causal inference engine at the edge to receive the equipment status data after maintenance in real time; S102. Calculate the maintenance intervention effect index: ； Among them is the maintenance intervention effect index is the real-time health score before maintenance, reflecting the current degradation state of the equipment is the health score after maintenance, calculated from the condition monitoring data is the real-time health score before maintenance, reflecting the current degradation state of the equipment; S103、 When it is greater than the threshold, it is determined as inefficient maintenance, and the causal knowledge graph backtracking and correction process is automatically triggered to update the weight parameters of the environment-fault association edge.

Images