A substation health management and control method fusing multi-dimensional information

Abstract

The application belongs to the technical field of substation health management and control, and provides a substation health management and control method fusing multi-dimensional information, comprising the following steps: obtaining historical operation data of equipment in a substation at all monitoring points, analyzing the correlation between monitoring points of each equipment, extracting a high-correlation monitoring point combination, and identifying an event in a historical fault event that is masked by the high-correlation monitoring points as an abnormal masking event. The application optimizes the problem of abnormal signal masking in high-dimensional monitoring data of a substation by fusing multi-dimensional information and adopting hierarchical correlation analysis and an improved PCA algorithm, accurately identifies abnormal masking events and targets dimension reduction, retains key information and eliminates redundant data, improves the accuracy of abnormal detection, reduces potential fault omissions, and provides a reliable basis for equipment health state evaluation.

Description

Technical Field

[0001] This invention belongs to the field of substation health management technology, specifically a substation health management method that integrates multi-dimensional information. Background Technology

[0002] Substation health management is a management model that integrates multi-source monitoring data, such as equipment status parameters, environmental sensor data, and operation logs, and uses artificial intelligence and data analysis technology to conduct real-time assessment and trend prediction of the operating status of power equipment (such as transformers, switchgear, and protection devices) in substations. Its core lies in building equipment health models, identifying potential fault symptoms, and realizing the transformation from passive emergency repair to proactive prevention, ultimately achieving the goals of improving equipment reliability, reducing operation and maintenance costs, and ensuring the safe and stable operation of the power grid.

[0003] In high-dimensional data spaces, anomalies may be difficult to detect due to the "curse of dimensionality" effect. This is because multi-dimensional monitoring systems may generate massive amounts of data. For example, each transformer generates thousands of records of temperature, gas concentration, vibration frequency, etc. every day. If most of these are normal data, abnormal signals may be submerged in redundant data, forming a masking effect of abnormal data. This leads to a sharp decline in the performance of anomaly detection algorithms, a significant increase in false detection rate and false negative rate, and makes it impossible to identify potential faults in substation equipment in a timely and accurate manner, seriously threatening the safe and stable operation of the power system.

[0004] Therefore, the present invention provides a substation health management method that integrates multi-dimensional information. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0006] The technical solution adopted by this invention to solve its technical problem is: a substation health management method integrating multi-dimensional information, comprising the following steps:

[0007] The system acquires historical operating data of equipment in the substation from all monitoring points, analyzes the correlation between monitoring points of each device, extracts highly correlated monitoring point combinations, and identifies anomaly detection events in historical fault events that are masked by highly correlated monitoring points as anomaly masking events.

[0008] The number of abnormal masking events in each device is counted, and their proportion in the number of historical fault events is calculated. Devices with high proportions are extracted as devices to be dimensionality reduced. An improved PCA algorithm is used to perform data dimensionality reduction and generate a set of principal components of device state features.

[0009] Obtain the principal component set after dimensionality reduction, analyze the discreteness and directional consistency of the coefficient sequence corresponding to each principal component, identify whether there is a fuzzy phenomenon in the physical meaning of the principal components, and if so, combine the knowledge graph of substation equipment fault mechanism to verify the matching degree between the fuzzy principal components and the known fault modes.

[0010] Extract principal components with high matching degree, perform principal component optimization and reconstruction, and update the optimized principal components to the substation equipment health status assessment model for dynamic adaptation of control strategies.

[0011] The beneficial effects of this invention are as follows:

[0012] This invention optimizes the problem of abnormal signal masking in high-dimensional monitoring data of substations by integrating multi-dimensional information and using hierarchical correlation analysis and improved PCA algorithm. It accurately identifies abnormal masking events and performs targeted dimensionality reduction, retaining key information while eliminating redundant data, thereby improving the accuracy of anomaly detection, reducing potential missed faults, and providing a reliable basis for equipment health status assessment.

[0013] This invention enables control strategies to accurately adapt to the real-time status of equipment through principal component physical meaning identification, optimized reconstruction, and dynamic updating of the health model. This transforms the system from passive emergency repair to proactive prevention, significantly improving the reliability of substation equipment, reducing operation and maintenance costs, and effectively ensuring the safe and stable operation of the power grid. Attached Figure Description

[0014] The invention will now be further described with reference to the accompanying drawings.

[0015] Figure 1 This is a flowchart of the steps of a substation health management method that integrates multi-dimensional information according to the present invention;

[0016] Figure 2 This is a flowchart of obtaining a highly correlated combination of monitoring points in a substation health management method that integrates multi-dimensional information according to the present invention;

[0017] Figure 3 This is an architecture diagram of a substation health management system that integrates multi-dimensional information according to the present invention. Detailed Implementation

[0018] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0019] Example 1

[0020] Please see Figure 1 and Figure 2 As shown in the embodiment of the present invention, a substation health management method integrating multi-dimensional information includes the following steps:

[0021] Step S10: Obtain historical operating data of equipment in the substation at all monitoring points, analyze the correlation between monitoring points of each device, extract highly correlated monitoring point combinations, and identify anomaly detection events in historical fault events that are masked by highly correlated monitoring points as anomaly masking events.

[0022] It should be noted that the definition of an anomaly masking event is as follows: In the historical fault events recorded in the above operation log, if an anomaly occurs at a certain monitoring point, but the data of other monitoring points in the high correlation combination are normal, the traditional anomaly detection algorithm fails to identify the anomaly.

[0023] In some embodiments, historical operating data of each device in the substation is acquired at all monitoring points.

[0024] The equipment in the substation includes, but is not limited to: transformers, switchgear, and protection devices.

[0025] Specifically, historical operating data includes, but is not limited to: equipment status parameters, environmental sensor data, and operating log records.

[0026] Furthermore, the equipment status parameters should include at least: temperature, gas concentration, and vibration frequency; environmental sensor data should include at least: humidity, air pressure, and dust concentration; and the operation log records should include at least: operation time, load changes, and maintenance records.

[0027] The acquired historical operation data is cleaned to remove invalid values, missing values, and obvious false alarms, and a standardized historical operation dataset is constructed. The historical operation data is then stored according to equipment type and time series.

[0028] For example, taking a transformer, a typical piece of equipment in a substation, as an example: collect historical operating data from all its monitoring points.

[0029] Equipment status parameters: daily recording of winding temperature, once per hour, range -20℃ to 120℃; dissolved gas concentration in oil, methane and acetylene, unit μL / L, measured at 8:00 AM daily; core vibration frequency, 50Hz-200Hz, collected once every 30 minutes.

[0030] Environmental sensor data: ambient humidity around the transformer, 30%-90%RH, recorded hourly; air pressure, 950hPa-1050hPa, 3 times daily; dust concentration, 0.01mg / m³-1mg / m³, collected every 2 hours.

[0031] Operation log records: the time of load adjustment operation twice a month; load change curve, the process record of increasing from 50% rated load to 80%; insulation resistance test records during quarterly maintenance.

[0032] For any given device, all monitoring points are paired. For any given pair of monitoring points, a hierarchical correlation analysis strategy is used to extract highly correlated monitoring point pairs. The extraction process is as follows:

[0033] First, the linear correlation of all monitoring point combinations is analyzed. Specifically, the Pearson correlation coefficient between the historical operating data of any monitoring point combination is calculated. Combinations with a Pearson correlation coefficient greater than the first correlation screening value are extracted as strongly linearly correlated combinations, and the remaining monitoring point combinations are regarded as weakly linearly correlated combinations.

[0034] The first relevance screening value is set by those skilled in the art, and is generally set to 0.8.

[0035] Next, the nonlinear correlation relationship is analyzed for all linear weakly correlated combinations. Specifically, the Spearman correlation coefficient between the historical running data of any linear weakly correlated combination is calculated, and combinations with Spearman correlation coefficients greater than the second correlation screening value are extracted as nonlinear strongly correlated combinations.

[0036] The second relevance screening value is set by those skilled in the art, and is generally set to 0.6.

[0037] By employing the aforementioned correlation analysis strategy, the limitations of single linear or nonlinear methods can be optimized. Linear analysis prioritizes capturing common linear correlations in equipment, such as direct proportional relationships between electrical parameters. Nonlinear analysis supplements this by capturing complex correlations, such as the nonlinear relationship between temperature and insulation material aging, thereby improving the comprehensiveness of correlation combinations. This is particularly important in substation health management scenarios, as equipment failure modes may involve both linear and nonlinear characteristics. Hierarchical analysis can more accurately locate potential fault correlations.

[0038] The advantages of using a hierarchical correlation analysis strategy are as follows: First, substation monitoring data usually contains massive amounts of time-series data from multiple devices and monitoring points. By employing a strategy of first linear filtering and then nonlinear supplementation, the calculation process can be optimized. In substation scenarios, early signs of equipment failure are often mainly linearly correlated, and hierarchical analysis can prioritize capturing key linear information.

[0039] Secondly, the judgment of equipment failure is highly dependent on the physical mechanism. It is necessary to clearly distinguish between linear and nonlinear dominant characteristics to facilitate subsequent fault location. Hierarchical analysis improves the interpretability of the results through step-by-step analysis.

[0040] Third, it covers a combination of highly correlated linear and nonlinear monitoring points, thereby accurately locating the associated structures that may mask anomalies, providing a more complete set of masking sources for the identification of anomaly masking events, and improving the accuracy of identification.

[0041] Combinations of strongly linearly correlated and strongly nonlinearly correlated data points are used as combinations of highly correlated monitoring points to identify anomaly masking events. The identification process is as follows:

[0042] Retrospective analysis was performed on the combined data of highly correlated monitoring points, and the threshold method (i.e., based on the abnormal threshold of the corresponding data) was used to identify whether there were abnormal signals related to the fault within a preset period before the fault occurred.

[0043] For example, the transformer winding temperature showed that the acetylene concentration in the oil exceeded the threshold of 15 μL / L 12 hours before the fault.

[0044] If there are abnormal signals related to the fault, compare the abnormal detection records. If a clear abnormal signal is identified, but the abnormal signal was not identified by the detection system at the time, and the characteristics of the abnormal signal are covered by the normal data of other monitoring points in the highly correlated combination, then it is preliminarily determined that the historical fault event is being masked.

[0045] For example, when the winding temperature is abnormal, the acetylene concentration in the oil and the ambient humidity are both within the normal range, causing the abnormal signal to be overwhelmed by redundant normal data. Therefore, it is preliminarily determined that the historical fault event is being masked.

[0046] By combining the knowledge graph of substation equipment fault mechanisms, such as winding overheating usually accompanied by abnormal temperature rise and excessive acetylene concentration, the physical correlation between the identified abnormal signals and the historical fault events can be verified.

[0047] If the characteristics of the abnormal signal (such as the rate of temperature increase and the gas concentration threshold) are highly matched with the clear fault precursors in the fault mechanism, and the possibility that the abnormal signal is an interference signal (such as a temporary sensor failure) is ruled out, then the historical fault event is determined to be an abnormal masking event.

[0048] This step involves acquiring historical operating data of substation equipment, using a hierarchical correlation analysis strategy to extract combinations of highly correlated monitoring points, and identifying anomaly masking events. This provides a foundation for subsequent targeted processing. Specifically, by analyzing the linear and nonlinear correlations of equipment monitoring points, it identifies historical fault events where highly correlated monitoring point combinations masked abnormal signals, thus optimizing the anomaly detection failure problem caused by the curse of dimensionality. This provides a clear basis for screening equipment to be downgraded and lays a data foundation for subsequent fault mechanism analysis and model optimization.

[0049] Step S20: Count the number of abnormal masking events in each device and calculate its proportion in the number of historical fault events. Extract the devices with high proportions as devices to be dimensionality reduced. Use the improved PCA algorithm to perform data dimensionality reduction and generate a set of principal components of device state features.

[0050] In some embodiments, for each device, the number of abnormal masking events is obtained, and the ratio of the number of abnormal masking events to the total number of historical fault events is calculated to obtain the percentage of abnormal masking events.

[0051] Devices with an excessive percentage of abnormal masking events are identified as high-percentage devices and designated as devices requiring dimensionality reduction.

[0052] Exceeding the standard means that the proportion of abnormal masking events is greater than the limit for the proportion of abnormal masking events, which is set by those skilled in the art based on experience and industry characteristics.

[0053] For the dimensionality reduction processing equipment, data dimensionality reduction processing is performed, and the process is as follows:

[0054] Extract the combined data of highly correlated monitoring points of the equipment to be downgraded.

[0055] For example, if the device to be processed is a transformer, and its highly correlated monitoring point combination is: winding temperature - acetylene concentration in oil - ambient humidity - core vibration frequency, then the complete time series data matrix of this combination within the historical period is extracted.

[0056] The combined data of highly correlated monitoring points were standardized using the Z-score standardization formula to eliminate the dimensional differences of different parameters.

[0057] The improved PCA algorithm reduces dimensionality by introducing the Kaiser-Guttman criterion and weighted optimization based on variance contribution rate. The specific process is as follows:

[0058] Among them, the improved PCA algorithm is a dimensionality reduction method that optimizes the processing flow based on the traditional PCA algorithm to address the characteristics of high-dimensional monitoring data in substations. It adds two key steps: screening principal components using the Kaiser-Guttman criterion and weighting principal components by variance contribution rate. This achieves the goal of eliminating redundant principal components and strengthening the weight of key information, generating a set of state feature principal components that is more suitable for the needs of substation equipment fault identification.

[0059] Specifically, the process of screening principal components using the Kaiser-Guttman criterion is as follows:

[0060] The core theory behind the Kaiser-Guttman criterion, where eigenvalues ​​> 1, is that the information entropy of principal components is higher than that of a single original monitoring variable. In substation scenarios, the information entropy of original monitoring variables (such as winding temperature and acetylene concentration in oil) reflects their ability to independently carry the equipment status. Principal components with eigenvalues ​​> 1 integrate the information of multiple related variables (such as the coupling relationship between temperature and gas concentration) through linear combination. Their information carrying efficiency is significantly higher than that of a single variable, which can effectively avoid noise interference from single variables.

[0061] The covariance matrix is ​​calculated on the standardized data matrix to reflect the degree of linear correlation between parameters at each monitoring point, such as the covariance between winding temperature and acetylene concentration in oil.

[0062] The covariance matrix is ​​decomposed into eigenvalues ​​(reflecting the variance contribution of the principal components) and corresponding eigenvectors (the coefficients of the principal components).

[0063] Among them, the eigenvalue is a quantitative indicator of the principal component's ability to contribute to the variance. The larger the eigenvalue, the more variance of the original data the principal component can explain, and the richer the information on the equipment's operating status it contains.

[0064] Principal components with eigenvalues ​​greater than 1 are retained. According to the Kaiser-Guttman criterion, principal components with eigenvalues ​​greater than 1 contain more information than a single original variable, which can effectively reduce redundancy.

[0065] For example, if the eigenvalues ​​are calculated to be [3.2, 1.8, 0.9, 0.5], then the first two principal components are retained.

[0066] Specifically, the process of introducing variance contribution rate weighting to enhance the impact is as follows:

[0067] The selected principal components are weighted according to their variance contribution rate, based on the proportion of the principal component eigenvalues ​​to the sum of the total eigenvalues, thereby strengthening the influence of principal components with high contribution rates.

[0068] The calculation formula is: ;

[0069] Among them, the variance contribution rate reflects the proportion of the principal component's explanation of the original data information. By assigning weights to the selected principal components through the variance contribution rate, the principal components with a high explanation proportion (principal components that are more strongly associated with equipment failure) can play a greater role in the final feature set.

[0070] For example, if the variance contribution rates of the first two principal components are 55% and 25% respectively, then they are assigned weights of 0.55 and 0.25 respectively after weighting.

[0071] The selected and weighted principal components are combined to form the set of principal components representing the state characteristics of the device to be dimensionality reduced.

[0072] For example, taking a substation "transformer" as an example, the dimensionality reduction process of the improved PCA algorithm is as follows:

[0073] Extract combined data from highly relevant monitoring points of the transformer, such as one-year historical time series data of winding temperature, acetylene concentration in oil, ambient humidity, and core vibration frequency;

[0074] Data is processed using the Z-score standardization formula to eliminate dimensions;

[0075] Calculate the covariance matrix to obtain the eigenvalues ​​[3.2, 1.8, 0.9, 0.5] and their corresponding eigenvectors;

[0076] Retain the first two principal components with eigenvalues ​​> 1;

[0077] The contribution rates of the first two principal components are calculated to be 64% and 36%, respectively, and weights of 0.64 and 0.36 are assigned.

[0078] Generate a weighted set of principal components (e.g., "0.64×PC1+0.36×PC2"), which will be used for subsequent principal component physical meaning analysis and health assessment model input.

[0079] This step identifies high-risk devices requiring dimensionality reduction by statistically analyzing the proportion of anomaly masking events. An improved PCA algorithm is then used for data dimensionality reduction to address the core issue of anomaly signals being overwhelmed by redundant data. By introducing the Kaiser-Guttman criterion and variance contribution rate weighting, redundant data is removed while retaining key information, generating a set of principal components for device status features. This simplifies the data dimensions and strengthens the core features crucial for fault early warning, providing high-quality feature input for subsequent analysis of the physical meaning of principal components and model optimization.

[0080] After performing data dimensionality reduction, more than 90% of the key information in the original data was retained, while redundant data from highly correlated monitoring points were removed, thus solving the problem of abnormal signals being masked.

[0081] Since dimensionality reduction is the key to cracking the obscuration of abnormal data, the principal components obtained after dimensionality reduction through principal component analysis may have ambiguous physical meanings and be difficult to interpret directly. This can lead to an inability to accurately identify the potential root causes of equipment failures, significantly reducing the accuracy of health management models in predicting early hidden faults. Consequently, this increases the risk of sudden failures in substation equipment and affects the safe and stable operation of the power grid. Therefore, after data dimensionality reduction, identifying and optimizing the physical meaning of the principal components can make fault detection and early warning more timely and accurate.

[0082] Step S30: Obtain the set of principal components after dimensionality reduction, analyze the discreteness and directional consistency of the coefficient sequence corresponding to each principal component, identify whether there is a fuzzy phenomenon in the physical meaning of the principal components, and if so, combine the knowledge graph of substation equipment fault mechanism to verify the matching degree between the fuzzy principal components and the known fault modes.

[0083] In some embodiments, the principal component set after dimensionality reduction is obtained.

[0084] For example, assuming the device to be dimensionality-reduced is a transformer, its highly correlated monitoring point combination is: winding temperature (T), acetylene concentration in oil (C), ambient humidity (H), and core vibration frequency (F). After dimensionality reduction using the improved PCA algorithm, two principal components, PC1 and PC2, are generated. The specific calculation method is as follows: PC1 = 0.8T + 0.7C - 0.1H + 0.2F; PC2 = 0.3T - 0.2C + 0.6H + 0.5F.

[0085] For the coefficient sequence corresponding to the principal component, calculate the mean and standard deviation of the coefficient sequence, and use the mean and standard deviation to calculate the coefficient of variation, which characterizes the degree of dispersion.

[0086] If the coefficient of variation is greater than the coefficient of variation threshold, it indicates that the coefficient sequence is highly discrete and the contribution of each parameter to the principal component varies greatly, which may lead to unclear physical meaning. If the coefficient of variation is less than or equal to the coefficient of variation threshold, it indicates that the coefficient sequence is not highly discrete.

[0087] The higher the degree of dispersion (coefficient of variation), the greater the difference in weights of the original parameters in the principal components, which will destroy the synergistic relationship between parameters and thus lead to ambiguity in physical meaning.

[0088] The number of positive and negative coefficients in the coefficient sequence is counted separately, and the ratio of positive coefficients to the total number of coefficients is calculated to obtain the proportion of positive coefficients.

[0089] If the proportion of positive coefficients is between 40% and 60% (mixed positive and negative), then the consistency of the direction is low; otherwise, the consistency of the direction is high.

[0090] Low directional consistency means that the parameters have contradictory effects on the principal components, such as some parameters promoting positive effects while others inhibiting them negatively, which is difficult to explain with a single physical mechanism.

[0091] If a principal component simultaneously satisfies the conditions that its coefficient of variation is greater than the coefficient of variation threshold and its directional consistency is low, it is determined to be a principal component with fuzzy physical meaning.

[0092] Conversely, if the discreteness is low and the directional consistency is high, it is determined to be a principal component with clear physical meaning.

[0093] For example, PC2 = 0.3T - 0.2C + 0.6H + 0.5, with a coefficient sequence of [0.3, -0.2, 0.6, 0.5], a mean of 0.3, a standard deviation of approximately 0.34, a coefficient of variation (1.13) > the coefficient of variation threshold (0.5), high dispersion, positive numbers (0.3, 0.6, 0.5) account for 3 / 4, negative numbers (-0.2) account for 1 / 4, positive and negative numbers are mixed, and the consistency of direction is low.

[0094] Extract fuzzy principal components with physical meaning, and verify their matching degree with fault modes through a knowledge graph of substation equipment fault mechanisms (which includes feature parameter association rules of known fault modes).

[0095] Among them, the knowledge graph of substation equipment failure mechanism is well known to those skilled in the art.

[0096] Extract the core parameter weights of the fuzzy principal components (such as H and F, which have higher weights in PC2), compare them with the feature parameters of each fault mode in the knowledge graph, and calculate the matching degree.

[0097] The formula for calculating the matching degree is: ;

[0098] The number of overlapping features is the number of parameters contained in the principal component coefficient sequence that overlap with the typical feature parameters of a certain fault mode in the knowledge graph.

[0099] It reflects the correlation between principal components and a certain failure mode at the parameter level. The greater the overlap, the more comprehensively the principal components cover the core parameters of the failure mode, providing a premise for subsequent directional consistency analysis and matching degree calculation, and reducing misjudgments of correlation caused by missing parameters.

[0100] The number of directional matching features is: among the overlapping features, the number of principal component coefficients whose positive and negative directions are consistent with the positive and negative directions of the fault mode features in the knowledge graph.

[0101] Further verification of the consistency between principal components and fault modes in terms of physical logic shows that the more the directions match, the more it indicates that the principal components not only cover parameters but also reflect the intrinsic mechanism of faults, thus improving the reliability of correlation judgment.

[0102] The total number of features is the total number of typical feature parameters of this fault mode in the knowledge graph.

[0103] For example, the coefficient sequence of PC2 is: 0.3T-0.2C+0.6H+0.5F; + indicates that the parameter is positively correlated with the principal component, and - indicates that it is negatively correlated.

[0104] In the knowledge graph of substation equipment failure mechanism, the typical characteristics of insulation moisture failure are: H↑ (increased ambient humidity), F↑ (increased core vibration frequency), T slightly↓ (slightly decreased winding temperature), C stable (no significant change in acetylene concentration in oil) (↑ indicates that the characteristic parameters are positively correlated with the failure, ↓ indicates that they are negatively correlated).

[0105] The parameters involved in PC2 are: T (winding temperature), C (acetylene concentration in oil), H (ambient humidity), and F (core vibration frequency).

[0106] Typical characteristic parameters of insulation moisture fault: H, F, T, C; the two parameters completely overlap, so the number of overlapping features = 4.

[0107] T (winding temperature): The PC2 coefficient is +0.3 (positive correlation). In the knowledge graph, T decreases slightly (negative correlation), so the directions do not match.

[0108] C (acetylene concentration in oil): The PC2 coefficient is -0.2 (negative correlation). In the knowledge graph, C is stable (without a clear direction, considered uncorrelated), so the directions do not match.

[0109] H (ambient humidity): The PC2 coefficient is +0.6 (positive correlation). If H↑ (positive correlation) in the knowledge graph, then the direction is consistent.

[0110] F (core vibration frequency): PC2 coefficient is +0.5 (positive correlation). If F↑ (positive correlation) in the knowledge graph, then the direction matches.

[0111] In summary, the parameters for directional matching are H and F, therefore the number of directional matching features = 2 parameters with complete overlap;

[0112] The calculated match rate is 75%.

[0113] If the matching degree is greater than the matching degree threshold (60%), then the fuzzy principal component is considered to have a significant relationship with a certain fault mode.

[0114] If the matching degree is less than or equal to the matching degree threshold (60%), it is considered that the fuzzy principal component does not have a significant relationship with a certain fault mode.

[0115] This step identifies the physical meaning of the principal components after dimensionality reduction. By analyzing the discreteness and directional consistency of the coefficient sequence, it determines whether there is any ambiguity in the physical meaning of the principal components. It also verifies the matching degree between the ambiguous principal components and known fault modes by combining the fault mechanism knowledge graph. This optimizes the problem of unclear physical meaning of principal components after dimensionality reduction in traditional PCA. By associating with fault mechanisms, it clarifies the correlation between ambiguous principal components and specific fault modes, providing a scientific basis for subsequent principal component optimization and reconstruction, and ensuring that the features after dimensionality reduction can accurately reflect the root cause of equipment failure.

[0116] Step S40: Extract principal components with high matching degree, perform principal component optimization and reconstruction, and update the optimized principal components to the substation equipment health status assessment model for dynamic adaptation of control strategies.

[0117] In some embodiments, for highly matched principal components, optimization and reconstruction measures based on the substation equipment fault mechanism knowledge graph include, but are not limited to:

[0118] Coefficient weight adjustment: Strengthen the weight of parameters in the principal components that are related to the core features of the failure mode, and weaken the weight of irrelevant or contradictory parameters.

[0119] For example: The original coefficient of PC2 is 0.3T-0.2C+0.6H+0.5F. Combined with the core characteristics of insulation moisture fault (H↑, F↑), it is adjusted to 0.1T-0.05C+0.7H+0.6F, which increases the weight of H and F and reduces the interference of T and C.

[0120] Redundant parameter removal: If a parameter in the principal component has no physical relationship with the fault mode, the coefficient of the parameter can be directly removed to simplify the principal component structure. For example, C in PC2 is not related to insulation moisture.

[0121] Feature integration: If a highly matched principal component has some related features with other principal components, a more comprehensive feature principal component can be formed by weighted fusion (such as assigning weights according to the variance contribution rate). For example, the "heat-gas coupling" of PC1 and the "moisture-vibration" of PC2 have cross parameters.

[0122] The optimized principal component set is input into the substation equipment health status assessment model to replace the original principal component parameters.

[0123] The model is retrained based on the new principal components, and the feature weights and fault warning thresholds of the model are updated using the sliding window method (such as real-time monitoring data from the past 3 months) to ensure that the model is adapted to the current operating status of the equipment.

[0124] The size of the sliding window is set by those skilled in the art based on the magnitude of data volatility and experience; for example, if the equipment data volatility is low (such as the gas concentration in transformer oil), the window is set to 3 months; if the data volatility is high (such as the temperature of switchgear contacts), the window is set to 1 month.

[0125] Among these, training and updating the model is a technique well known to those skilled in the art.

[0126] For example: Incorporate the optimized PC2 (0.1T-0.05C+0.7H+0.6F) into the transformer health assessment model, retrain the early warning logic of "insulation moisture fault", so that the model can provide early warning 48 hours in advance through small changes in H and F.

[0127] Based on the principal component features output by the updated model, the real-time health status of the equipment is reflected, and the control strategy is dynamically adjusted.

[0128] For example, if the principal component shows abnormal heat-gas coupling characteristics (such as a sudden increase in PC1 value), the monitoring frequency of winding temperature and gas concentration in oil is increased (from once per hour to once every 30 minutes), and an overheating risk inspection plan is triggered.

[0129] If the principal component shows abnormal moisture-vibration characteristics (such as a sudden increase in PC2 value), adjust the power of the environmental dehumidification equipment and shorten the insulation resistance test cycle.

[0130] This step achieves dynamic adaptation of control strategies by optimizing and reconstructing high-matching principal components and updating the health assessment model. By adjusting the principal component coefficient weights and eliminating redundant parameters, the correlation between principal components and fault modes is enhanced. Combined with real-time data to dynamically update the model, health management shifts from passive response to proactive prevention. Based on the optimized model output, strategies such as monitoring frequency and inspection plans are adjusted to improve the accuracy of equipment fault early warning, reduce operation and maintenance costs, and ensure the safe and stable operation of the power grid.

[0131] Example 2

[0132] Based on the same inventive concept as the substation health management method integrating multi-dimensional information in the foregoing embodiments, such as Figure 3 As shown, this application provides a substation health management and control system that integrates multi-dimensional information, wherein the system specifically includes:

[0133] Masking event identification module: acquires historical operating data of equipment in the substation at all monitoring points, analyzes the correlation between monitoring points of each device, extracts highly correlated monitoring point combinations, and identifies anomaly detection events in historical fault events that are masked by highly correlated monitoring points as anomaly masking events.

[0134] Historical operating data of equipment such as transformers and switchgear in the substation at various monitoring points are acquired, cleaned and standardized, and stored according to equipment type and time series. A hierarchical correlation analysis strategy is adopted. First, the Pearson correlation coefficient is used to screen the combination of strongly linearly correlated monitoring points. Then, the Spearman correlation coefficient is used to extract the combination of strongly nonlinearly correlated points from the remaining combinations and integrate them into a combination of highly correlated monitoring points. Historical fault events are traced back. If the abnormal signal of a certain monitoring point in the highly correlated combination is masked by other normal data and is not identified by the system at that time, its correlation with the fault is verified by combining the fault mechanism knowledge graph and it is determined to be an abnormal masking event.

[0135] Data dimensionality reduction module: Counts the number of abnormal masking events in each device and calculates its proportion in the number of historical fault events. Extracts devices with high proportions as devices to be dimensionality reduced. Uses an improved PCA algorithm to perform data dimensionality reduction and generate a set of principal components of device state features.

[0136] The number of abnormal masking events for each device was counted, and their proportion in historical failure events was calculated. Devices with a proportion exceeding the standard were listed as devices to be dimensionality reduced. Their highly correlated monitoring point combination data were extracted, and after Z-score standardization, an improved PCA algorithm with Kaiser-Guttman criterion and variance contribution rate weighting was used for dimensionality reduction. Finally, a set of principal components of device state features was generated to eliminate redundant data and optimize the problem of abnormal signal masking.

[0137] Dimensionality Reduction Effective Identification Module: Obtain the set of principal components after dimensionality reduction, analyze the discreteness and directional consistency of the coefficient sequence corresponding to each principal component, identify whether there is a fuzzy phenomenon in the physical meaning of the principal components, and if so, combine it with the knowledge graph of substation equipment fault mechanism to verify the matching degree between the fuzzy principal components and the known fault modes.

[0138] The principal component set after dimensionality reduction is obtained. By calculating the coefficient of variation and the proportion of positive coefficients in the coefficient sequence, the principal components with fuzzy physical meaning are identified. Combined with the equipment failure mechanism knowledge graph, the fuzzy principal components are compared with known failure modes. By calculating the number of overlapping features and the number of directional matching features, the matching degree is obtained, and the correlation between them and failure modes is verified.

[0139] Dimensionality reduction and optimization execution module: Extract principal components with high matching degree, perform principal component optimization and reconstruction, and update the optimized principal components to the substation equipment health status assessment model for dynamic adaptation of control strategies.

[0140] Extract principal components with high matching degree, optimize and reconstruct them in combination with fault mechanism, and verify the effectiveness of the optimized principal components through historical data. Update the optimized principal components to the equipment health status assessment model, retrain the model to adapt to the current operating status, and dynamically adjust the control strategy based on the real-time principal component features output by the model to achieve accurate adaptation of health control.

[0141] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A substation health management method integrating multi-dimensional information, characterized in that: Includes the following steps: Historical operating data of equipment in the substation is acquired from all monitoring points. The correlation between monitoring points of each device is analyzed, and highly correlated monitoring point combinations are extracted. Fault events of monitoring points that were detected as abnormal but were masked by highly correlated monitoring points in historical fault events are identified as anomaly masking events. This includes: performing retrospective analysis on the data of highly correlated monitoring point combinations to identify whether there are fault-related abnormal signals within a preset time period before the fault; if there are fault-related abnormal signals, the anomaly detection records are compared. If a clear abnormal signal is identified, but the abnormal signal was not identified by the detection system at the time, and the characteristics of the abnormal signal are covered by normal data from other monitoring points in the highly correlated combination, then it is preliminarily determined that the historical fault event is masked. By combining the knowledge graph of substation equipment fault mechanisms, the physical correlation between the identified abnormal signals and the historical fault events is verified: If the characteristics of the abnormal signal highly match the clearly defined fault precursors in the fault mechanism, and the possibility that the abnormal signal is an interference signal is ruled out, then the historical fault event is determined to be an abnormal masking event. The number of abnormal masking events in each device is counted, and its proportion in the number of historical fault events is calculated. Devices with high proportions are extracted as devices to be dimensionality reduced. An improved PCA algorithm is used to perform data dimensionality reduction and generate a set of principal components of device state features. Obtain the principal component set after dimensionality reduction, analyze the discreteness and directional consistency of the coefficient sequence corresponding to each principal component, identify whether there is a fuzzy phenomenon in the physical meaning of the principal components, and if so, combine the knowledge graph of substation equipment fault mechanism to verify the matching degree between the fuzzy principal components and the known fault modes. Extract principal components with high matching degree, perform principal component optimization and reconstruction, and update the optimized principal components to the substation equipment health status assessment model for dynamic adaptation of control strategies.

2. The substation health management method integrating multi-dimensional information according to claim 1, characterized in that: The extraction of highly correlated monitoring point combinations includes: For any device, all monitoring points are paired up. For any pair of monitoring points, a hierarchical correlation analysis strategy is used to extract the pairs of monitoring points with correlation. The stratified correlation analysis strategy is used to first extract strongly linearly correlated combinations, and then extract strongly nonlinearly correlated combinations. Combinations of strongly linearly correlated points and combinations of strongly nonlinearly correlated points are used as combinations of highly correlated monitoring points.

3. The substation health management method integrating multi-dimensional information according to claim 2, characterized in that: Using a hierarchical correlation analysis strategy, the combination of monitoring points for correlation is extracted, including: Calculate the Pearson correlation coefficient between historical data of any combination of monitoring points, extract combinations with Pearson correlation coefficients greater than the first correlation screening value as strongly linearly correlated combinations, and use the remaining combinations of monitoring points as weakly linearly correlated combinations. Calculate the Spearman correlation coefficient between historical data of any linear weakly correlated combination, and extract combinations with Spearman correlation coefficients greater than the second correlation screening value as nonlinear strongly correlated combinations.

4. The substation health management method integrating multi-dimensional information according to claim 1, characterized in that: Calculate its proportion in the number of historical failure events, extract the devices with high proportions as devices to be downgraded, including: For each device, the number of abnormal masking events is obtained, and the ratio of the number of abnormal masking events to the total number of historical fault events is calculated to obtain the percentage of abnormal masking events. Devices with an excessive proportion of abnormal masking events are identified as devices requiring dimensionality reduction processing.

5. The substation health management method integrating multi-dimensional information according to claim 1, characterized in that: An improved PCA algorithm is used to perform data dimensionality reduction, generating a set of principal components of equipment state features, including: Extract the high-correlation monitoring point combination data of the equipment to be dimensionality reduced, and standardize the high-correlation monitoring point combination data using the Z-score standardization formula. The improved PCA algorithm generates a set of principal components of equipment state features by introducing the Kaiser-Guttman criterion and using variance contribution rate weighted optimization for dimensionality reduction.

6. The substation health management method integrating multi-dimensional information according to claim 1, characterized in that: The identification of whether there is ambiguity in the physical meaning of principal components includes: For the coefficient sequence corresponding to the principal component, calculate the mean and standard deviation of the coefficient sequence, and then calculate the coefficient of variation using the mean and standard deviation; The number of positive and negative coefficients in the coefficient sequence is counted separately, and the ratio of positive coefficients to the total number of coefficients is calculated to obtain the proportion of positive coefficients. If the proportion of positive coefficients is between 40% and 60%, then the consistency of the direction is considered low. If a principal component simultaneously satisfies the conditions that its coefficient of variation is greater than the coefficient of variation threshold and its directional consistency is low, it is determined to be a principal component with fuzzy physical meaning.

7. The substation health management method integrating multi-dimensional information according to claim 1, characterized in that: The verification of the matching degree between the fuzzy principal components and the known fault modes includes: Extract the core parameter weights of fuzzy principal components; The core parameter weights are compared with the feature parameters of each fault mode in the knowledge graph. The matching degree is calculated by identifying the number of overlapping features and the number of directional matching features. If the matching degree is greater than the matching degree threshold, then the fuzzy principal component is considered to have a significant relationship with a certain fault mode.

8. The substation health management method integrating multi-dimensional information according to claim 7, characterized in that: The step of calculating the matching degree by identifying the number of overlapping features and the number of directional matching features includes: Obtain the number of overlapping features and the number of features with matching directions, and sum them up; The obtained value is then multiplied by twice the total number of features, and the percentage is calculated to obtain the matching degree.

9. A substation health management method integrating multi-dimensional information according to claim 8, characterized in that: The number of overlapping features refers to the number of parameters contained in the principal component coefficient sequence that overlap with the typical feature parameters of a certain fault mode in the knowledge graph. The number of directional matching features refers to the number of overlapping features in which the positive and negative directions of the principal component coefficients are consistent with the positive and negative directions of the fault mode features in the knowledge graph. The total number of features refers to the total number of typical feature parameters of the fault modes in the knowledge graph.

Images