Search

CN-122020582-A - Power data anomaly detection method, system, computer equipment and storage medium

CN122020582ACN 122020582 ACN122020582 ACN 122020582ACN-122020582-A

Abstract

The invention provides a power data anomaly detection method, a system, computer equipment and a storage medium, wherein a multisource data correlation matrix is constructed by dynamically correlating power data with non-power scene data, characteristic information corresponding to a specific area or equipment with anomaly labels is migrated to a similar scene according to the time sequence evolution characteristics of the power data and various scene influence variables, the cooperative training of a cross-area anomaly detection model is realized on the premise of not sharing original power data of each area, a coarse screening and fine inspection mode is adopted, a topological connection structure of a power system is combined, and the initial node position of anomaly triggering is reversely deduced and determined by tracking and tracing the propagation path of anomaly data. The method and the device can accurately distinguish real anomalies, provide directions for fault investigation, improve the detection adaptation capability of unlabeled data, reduce cloud transmission quantity and network delay, eliminate the need of manual frequent intervention, and improve the efficiency and meanwhile strength enemy labor intensity.

Inventors

  • YANG WEI
  • LAI GUOHUI
  • LIN DENGYONG
  • Peng Ningyun
  • ZHANG SHENGQIANG
  • Deng Shuiqun
  • XU ZHIFENG
  • HUANG ZHIHUI

Assignees

  • 惠州市焕能电气安装有限公司

Dates

Publication Date
20260512
Application Date
20251218

Claims (10)

  1. 1. A method for detecting anomalies in power data, the method comprising: S1, constructing a multi-source data association matrix by dynamically associating power data with non-power scene data, wherein the power data comprises a load change curve, equipment running state parameters and voltage and current time sequence monitoring data, the non-power scene data comprises meteorological environment data and traffic flow statistical data, and the dynamic association process comprises uniform alignment processing of time granularity and dynamic adaptation distribution of scene weights; s2, dynamically adjusting a baseline threshold interval in real time according to the time sequence evolution characteristics of the power data and various scene influence variables through a sliding window algorithm, wherein the baseline threshold interval can show nonlinear telescopic adaptation characteristics along with fluctuation and change of the scene variables; S3, migrating the feature information corresponding to the specific area or equipment with the abnormal labeling to a similar scene through a migration learning algorithm, carrying out elimination processing on the distribution difference between different data sources by combining a field self-adaptive model, introducing a field self-adaptive fusion coefficient, dynamically balancing the adaptation degree of the data features of a source field and a target field, and realizing the initial screening of the abnormal situation of unlabeled data; S4, adopting a coarse screening and fine inspection mode, firstly defining an abnormal candidate set through a coarse granularity algorithm, and then carrying out authenticity verification on abnormal conditions in the candidate set through a fine granularity algorithm; s5, combining a topological connection structure of the power system, tracking and tracing the propagation path of the abnormal data, and reversely deducing and determining the position of the initial node triggered by the abnormality.
  2. 2. The power data anomaly detection method according to claim 1, wherein the domain adaptive fusion coefficient introduced in S3 The expression is: Wherein the method comprises the steps of The method is a field self-adaptive fusion coefficient; Is a source domain data characteristic value; for the characteristic value of the data of the target domain, Is the number of feature dimensions; The pearson correlation coefficient of the source domain and the target domain; Setting migration trust coefficients through federation migration of cross-regional data features A value range [0.6,0.9], dynamically adjusting the weight of the migration of the source domain features to the target domain, The value is determined by the scene similarity and the data distribution consistency of the source domain and the target domain, and the expression is: Wherein the method comprises the steps of Is scene similarity; Data distribution consistency; Scene similarity The method is obtained by calculating the weighted sum of the weather type matching degree, the traffic flow grade matching degree and the topological structure similarity of the source domain and the target domain, and the expression is as follows: Wherein the method comprises the steps of For the degree of matching of the weather type, The traffic flow grade matching degree is; Is the similarity of topological structures; Data distribution consistency By passing through Calculating the data distribution difference between the source domain and the target domain by using the divergence, normalizing to the interval of [0,1], Wherein the method comprises the steps of Is the source domain With the target domain A kind of electronic device The degree of dispersion is determined by the degree of dispersion, Maximum for history A divergence value; Through the federal learning framework, on the premise of not sharing original power data of each region, collaborative training of a cross-region anomaly detection model is realized, and multi-region data features are aggregated through encryption gradient transmission.
  3. 3. The method for detecting power data anomalies according to claim 2, wherein the step of dynamically associating the power data with the non-power scene data in step S1, further comprising the steps of: S11, carrying out outlier rejection and normalization processing on the electric power data, and uniformly mapping to a [0,1] interval, wherein the expression is as follows: Wherein the method comprises the steps of Representing an original value of the power data; representing a reasonable minimum of the power data; representing a reasonable maximum of the power data; Representing the normalized result; representing scene impact coefficients; representing a scene adaptation factor; scene influence system The value rule of the system is weather data corresponding coefficient range [0.1,0.8], traffic flow data corresponding coefficient range [0.05,0.5], scene adaptation factor The value of (3) is dynamically adjusted according to the association accuracy of the historical data through the system, and the range is [0.8,1.2]; The scene adaptation factor The dynamic adjustment mechanism of (1) is that the system calculates the historical data association accuracy according to the preset time, if the accuracy is higher than the preset value, the system calculates the historical data association accuracy Increasing on the basis of the current value, if the accuracy is lower than the preset value Decreasing on the basis of the current value, if the accuracy is within the preset value range Remain unchanged; Carrying out format standardization processing on the non-electric scene data, converting meteorological data into an influence factor index, converting traffic flow data into an area activity intensity value, and uniformly adjusting the time granularity of the two types of data through a time stamp alignment algorithm; S12, calculating the Pearson correlation coefficient of the power data and the non-power data in the window range through a sliding time window, capturing the correlation change of the power data and the non-power data in the time sequence dimension, constructing a scene influence factor library, configuring initial correlation weights for different meteorological types and traffic states, and dynamically adjusting the weights according to data feedback information acquired in real time; increasing the weight of abrupt node in the power data time sequence through a time sequence attention mechanism, and strengthening the data association strength of an abnormal sensitive period through attention coefficients when constructing a multisource data association matrix, wherein the expression of the attention coefficients is as follows: Wherein the method comprises the steps of Is that Is a time attention coefficient of (a); Is a sensitive regulatory factor; Is that Time power data; Is the time sequence length; S13, taking a power data type as a row dimension of a matrix, taking a non-power data type as a column dimension, setting an initial value of a matrix element as a real-time correlation coefficient of a corresponding dimension, setting a scene weight layer on the basic matrix, multiplying a scene adaptation weight and the correlation coefficient to obtain a weighted correlation value, setting a power equipment topology correlation column, filling in a corresponding coefficient of the non-power data on the power data of different topology nodes, and forming a three-dimensional correlation matrix comprising a data type, a scene type and the topology nodes; And S14, triggering matrix updating operation according to a preset time interval, recalculating a correlation coefficient and scene weights through newly acquired data, replacing outdated elements in the matrix, starting an updating process when detecting that the data distribution has abrupt change, completing dynamic adjustment of the matrix, establishing an association validity assessment index system, and adjusting a correlation coefficient calculation model and a weight distribution rule by the system if the association accuracy of the matrix is lower than a preset threshold value, wherein the association validity assessment index system comprises an association accuracy, a time sequence synchronous error rate and a weight adaptation degree, the association accuracy is an effective association data occupation ratio in the matrix, the time sequence synchronous error rate is a statistic value of alignment deviation of data time stamps, and the weight adaptation degree is a matching degree of the scene weights and the association strength of actual data.
  4. 4. The method for detecting an abnormality of power data according to claim 3, wherein in S2, the dynamically adjusting the baseline threshold interval in real time by a sliding window algorithm specifically includes: s21, setting a first initial window, a second initial window and a third initial window in combination with the power scene, wherein the first initial window is provided with 15-20 data points, the second initial window is provided with 30-50 data points and the third initial window is provided with 60-90 data points; s22, calculating variances and trend slopes of data in the first initial window, the second initial window and the third initial window; Wherein the method comprises the steps of Representing dynamic fusion weights of windows; 、 And (d) sum Standard deviations of the first initial window, the second initial window and the third initial window data are respectively represented; 、 And Respectively representing window basic weights; Establishing a window adaptation evaluation model, distributing dynamic weights for a first initial window, a second initial window and a third initial window according to fluctuation and change of data, lifting the first window weight if the short-term fluctuation of the data is severe, lifting the second window weight if the data is stable in a middle period, lifting the third window weight if the long-term trend of the data is stable, and generating a calculation base number of an initial baseline threshold interval based on the data characteristics of the first initial window, the second initial window and the third initial window fused by weight weighting; S23, collecting physical constraint parameters of the operation of the power equipment in real time, wherein the physical constraint parameters comprise a rated voltage range, a rated current range and a maximum load threshold of the equipment, converting the rated voltage range, the rated current range and the maximum load threshold into corresponding threshold boundary values, fusing the threshold boundary values corresponding to the physical constraint parameters with a dynamically adjusted baseline threshold interval, taking an intersection of the threshold boundary values and the baseline threshold interval as a final baseline threshold interval, taking the threshold of the physical constraint parameters as the reference if the intersection of the threshold value and the threshold of the physical constraint parameters does not exist, and triggering a system to alarm and prompt threshold conflict.
  5. 5. The method for detecting abnormal power data according to claim 4, wherein S23 further comprises introducing a physical constraint threshold of the power equipment, clipping and calibrating the calculated threshold interval so that the threshold does not exceed the safe operation boundary of the equipment, monitoring the distribution density of the data in the threshold interval in real time, triggering iterative adjustment of window parameters if the normal data distribution duty ratio is low, calculating the difference between the occurrence time of the abnormal data and the response time of the threshold adjustment, and optimizing the window weight distribution rule if the difference exceeds a preset threshold; and the light-weight abnormal coarse screening algorithm is deployed at the power terminal equipment through an edge computing node deployment mechanism, so that the local identification and preliminary processing of abnormal data are realized.
  6. 6. The power data anomaly detection method according to claim 5 is characterized in that the window parameter iterative adjustment is specifically that a dynamic duty ratio threshold model is firstly constructed based on scene weights and topological association coefficients in a three-dimensional association matrix, a normal data distribution duty ratio threshold is calculated in real time through the model, the model takes a current scene influence coefficient and topological node association strength as input, a normal distribution duty ratio reference value in similar scenes in historical data is combined, a dynamic threshold is output through an adaptive weighting algorithm, if the actual normal data duty ratio is lower than the dynamic threshold, a hierarchical adjustment mechanism is started, the number of first initial window data points is reduced by 2-4 in a self-adaptive mode according to scene fluctuation strength, specifically, the upper limit is taken when short-term fluctuation is severe, the lower limit is taken when fluctuation is mild, the number of second initial window data points is dynamically increased by 1-3 according to the time sequence mutation slope of power data, specifically, the time mutation slope is increased when exceeding the threshold, the threshold is lower than the threshold, the number of third initial window data points is increased by 2-5, the time sequence attention coefficient is introduced after adjustment to conduct weighted correction on window data, the data weight of anomaly sensitivity time period is calculated, and the threshold is changed again, and the threshold is adapted to a base line change interval after the adjustment is carried out.
  7. 7. The method of claim 6, wherein the fast defining the anomaly candidate set by coarse-grained algorithm comprises setting a dynamic cluster radius reference value based on the space-time correlation characteristic of the power data, wherein the dynamic cluster radius reference value setting method comprises setting an initial reference value based on the average radius of normal data clusters in historical data, combining the power data types, dynamically fine-tuning according to real-time scene weights, adopting an improved density clustering algorithm, taking power data time sequence nodes in a three-dimensional correlation matrix as cores, dividing data with adjacent time dimensions and space dimensions belonging to the same topological correlation area into basic data clusters, calculating the core feature statistical value of each data cluster, including the data mean value, fluctuation variance and deviation degree from a base line threshold interval, and setting the deviation degree threshold as The said The expression of (2) is Wherein the method comprises the steps of Representing the comprehensive deviation degree of the data; Is the baseline threshold interval mean; the change slope of the time sequence section where the abnormal data are located; Maximum allowable slope for normal operation of the device; Is a slope influencing factor; Screening out data clusters with the deviation degree exceeding a set threshold, marking all data in the clusters as abnormal candidate data, integrating to form an abnormal candidate set, and simultaneously recording space-time coordinates and associated scene information of each candidate data; The fine-granularity algorithm performs authenticity verification on abnormal conditions in the candidate set, and comprises the steps of extracting physical operation constraint parameters of the power equipment corresponding to each piece of abnormal data in the candidate set, constructing an equipment constraint verification library, introducing a time sequence consistency verification model, comparing the change trend of adjacent time sequence data before and after the abnormal data, calculating the mutation slope of the abnormal data and the normal data of the preamble, and entering the next verification step if the slope exceeds the maximum allowable change rate in the normal operation state of the equipment; combining scene weight and topology association coefficient in the three-dimensional association matrix, analyzing whether abnormal data are associated with the current scene condition, judging as suspicious abnormality if abnormal performance is contrary to scene influence rules, calling a device operation mechanism model, simulating whether an operation state corresponding to the abnormal data accords with the physical characteristics of the power device, judging as real abnormality if the deviation exceeds a preset threshold, comprehensively grading candidate data after the steps, confirming as real abnormality if the grading meets the standard, marking as to-be-rechecked or false abnormality if the grading does not meet the standard, and finishing authenticity verification.
  8. 8. A power data anomaly detection system, the system comprising: s100, a multi-source data association module, which is used for carrying out outlier rejection, normalization and format standardization processing on power data and non-power scene data, unifying time granularity through a time stamp alignment algorithm, calculating a Pearson correlation coefficient, dynamically adjusting weights by combining a time sequence attention mechanism, constructing a three-dimensional association matrix containing data types, scene types and topological nodes, and completing matrix dynamic updating and effectiveness evaluation according to preset rules; S200, a baseline threshold dynamic adjustment module is used for setting initial windows with different data point numbers, calculating data variances and trend slopes in the windows, distributing dynamic weights, generating a final baseline threshold interval by fusing physical constraint parameters of the power equipment, deploying a lightweight coarse screening algorithm through edge calculation nodes, and iteratively adjusting window parameters according to data distribution conditions; s300, a cross-domain collaborative detection module is used for migrating marked abnormal characteristics to similar scenes through a migration learning algorithm, eliminating data source distribution differences by combining a field self-adaptive model, realizing cross-domain model collaborative training based on an encryption aggregation mechanism of a federal learning framework, and guaranteeing data privacy safety and node fault tolerance; S400, an abnormal precise screening module is used for defining an abnormal candidate set through coarse granularity screening of an improved density clustering algorithm, removing duplication, and carrying out authenticity verification and comprehensive scoring on candidate set data through a fine granularity algorithm of time sequence consistency verification, scene association analysis and equipment operation mechanism simulation; s500, an anomaly tracing and positioning module is used for tracking the propagation path of anomaly data by combining the topological connection structure of the power system, reversely deducing and determining the position of an initial node triggered by anomaly, and synchronously recording the space-time coordinates and associated scene information of the anomaly data.
  9. 9. A computer device comprising a memory, a processor and a computer program stored on the memory, the processor executing the computer program to implement the steps of the power data anomaly detection method of any one of claims 1 to 7.
  10. 10. A computer-readable storage medium, characterized in that a computer program is stored thereon, which, when being executed by a processor, implements the steps of the power data anomaly detection method of any one of claims 1 to 7.

Description

Power data anomaly detection method, system, computer equipment and storage medium Technical Field The present invention relates to the field of power data anomaly detection technologies, and in particular, to a power data anomaly detection method, a system, a computer device, and a storage medium. Background In the running process of the power system, accurate monitoring and abnormal identification of power data are important to ensuring stable and efficient running of the system, and the power data cover core indexes such as load change curves, equipment running state parameters, voltage and current time sequence monitoring data and the like, and the data quality directly influences key works such as power dispatching, fault investigation, equipment maintenance and the like; In the prior art, a sliding window algorithm is adopted in a threshold adjustment link of abnormal detection of electric data, on one hand, if sudden abnormal data is mixed in a window, a threshold calculation result is directly interfered, baseline deviation is caused, normal data is misjudged, on the other hand, the fixed window size cannot adapt to multi-scale fluctuation characteristics of the electric data, the adaptability to complex scenes of short-term sudden changes or long-term gradual changes is poor, meanwhile, the traditional method only depends on local characteristics of the data in the window, ignores the inherent time sequence relevance of the electric data and the influence of non-electric scene data such as meteorological environment and traffic flow, so that hysteresis exists in threshold adjustment, the abnormal characteristics of the data cannot be accurately captured, in addition, the traditional detection method lacks an effective cross-domain adaptation mechanism for the electric data of different areas and different devices, the abnormal screening accuracy of unlabeled data is low, when the cross-area detection is performed, the original data sharing is easy to cause privacy security problems, part of node faults can influence the training effect of an overall detection model, meanwhile, the traditional detection method is not careful enough to verify the authenticity of the abnormal data, the false redundant data and the abnormal data is easy to be truly difficult to accurately judge the abnormal nodes, and the abnormal node is difficult to accurately trace back to detect. Accordingly, those skilled in the art have been directed to providing a power data anomaly detection method, system, computer device and storage medium that can effectively solve the above-mentioned technical problems. Disclosure of Invention In view of the above-mentioned drawbacks of the prior art, the present invention is to provide a method, a system, a computer device and a storage medium for detecting an abnormality of power data, which are used for solving the problems existing in the prior art. In order to achieve the above object, the present invention provides a method for detecting abnormality of power data, the method comprising: S1, constructing a multi-source data association matrix by dynamically associating power data with non-power scene data, wherein the power data comprises a load change curve, equipment running state parameters and voltage and current time sequence monitoring data, the non-power scene data comprises meteorological environment data and traffic flow statistical data, and the dynamic association process comprises uniform alignment processing of time granularity and dynamic adaptation distribution of scene weights; s2, dynamically adjusting a baseline threshold interval in real time according to the time sequence evolution characteristics of the power data and various scene influence variables through a sliding window algorithm, wherein the baseline threshold interval can show nonlinear telescopic adaptation characteristics along with fluctuation and change of the scene variables; S3, migrating the characteristic information corresponding to the specific area or equipment with the abnormal labeling to the similar scene through a migration learning algorithm, carrying out elimination processing on the distribution difference between different data sources by combining a field self-adaptive model, introducing a field self-adaptive fusion coefficient, dynamically balancing the adaptation degree of the data characteristics of a source field and a target field, and realizing the initial screening of the abnormal condition of unlabeled data, wherein the field self-adaptive fusion coefficient introduced in the S3 is as follows The expression is: Wherein the method comprises the steps of The method is a field self-adaptive fusion coefficient; Is a source domain data characteristic value; for the characteristic value of the data of the target domain, Is the number of feature dimensions; The pearson correlation coefficient of the source domain and the target domain; setting migration trust coefficient A value range [