Electric energy metering detection data quality evaluation method and system

Abstract

The invention provides an electric energy metering detection data quality evaluation method and system, and relates to the field of data processing. Data abnormity generated by a current transformer aiming at gradual change characteristics caused by contact oxidation is firstly eliminated through improved wavelet transform and feature extraction; the abnormal rate of the processed data is calculated according to a support vector machine model, the conditions of noise interference and abnormal fluctuation in actual current data are considered when a support vector machine is designed, and the contact resistance is slowly increased due to contact oxidation of the secondary winding of the current transformer; therefore, the problems that the gradual change abnormity is difficult to find in time and the electric charge metering deviation is easily caused by the existing fixed threshold value evaluation method are solved.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and more specifically, to a method and system for evaluating the quality of electric energy metering detection data. Background Art

[0002] With the development of digitalization and intelligence of power systems, the accuracy and reliability of data of electric energy metering, as the core link of power monitoring, billing and energy efficiency management, has attracted more and more attention. As a key device in the electric energy metering system, the secondary current data of the current transformer directly affects the accuracy of electric energy metering. However, in actual operation, the data collected on the secondary side of the current transformer may be affected by various factors such as equipment aging, environmental interference, electromagnetic noise, etc., resulting in data anomalies or data quality degradation. Such data anomalies may not only cause errors in electric energy metering, but may even bring safety hazards to the monitoring and dispatching of power grid operation. Therefore, how to conduct real-time and accurate evaluation of the quality of data collected on the secondary side of the current transformer has become an important research direction in the operation and management of power systems.

[0003] At present, for the monitoring and evaluation of current transformer data quality, traditional methods are usually based on simple threshold judgments or statistical analysis methods. For example, simple characteristic indicators such as maximum value, minimum value or mean value are used to judge whether the data is abnormal. Although these methods are simple to calculate, they often ignore the dynamic change characteristics and complex frequency domain characteristics of current data, and cannot effectively capture subtle anomalies in the data. In addition, some advanced methods implement data anomaly detection based on machine learning or statistical models, but these methods usually rely on a large amount of historical labeled data for training, which makes it difficult to apply in scenarios where there is a lack of labeled data. At the same time, they are not robust enough to specific noise interference, resulting in unstable detection results. Therefore, the existing technology still has great limitations in terms of accuracy and robustness in the evaluation of secondary current data quality of current transformers. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a method and system for evaluating the quality of electric energy metering detection data, which solves the problem that the contact resistance slowly increases due to contact oxidation of the secondary winding of the current transformer, causing the collected current data to gradually attenuate. The existing fixed threshold evaluation method is difficult to detect such gradual anomalies in time, which easily causes electricity metering deviations.

[0005] According to one aspect of the present invention, there is provided a method for evaluating the quality of electric energy metering detection data, comprising: The collected current transformer secondary side current data is divided into N data segments according to a 24-hour time window, and the N data segments are respectively subjected to wavelet transformation to obtain corresponding wavelet coefficient matrices; Extracting energy entropy eigenvalues ​​and standard deviation eigenvalues ​​of adjacent sampling points of each data segment based on the wavelet coefficient matrix, and constructing the energy entropy eigenvalues ​​and standard deviation eigenvalues ​​of adjacent sampling points into a eigenvector; Establishing a current data anomaly detection model, and calculating an abnormal probability value of each data segment through the current data anomaly detection model; The abnormal probability value is compared with a preset threshold value. When the abnormal probability value is greater than the preset threshold value, the corresponding data segment is marked as an abnormal data segment to obtain a quality assessment result of the secondary side current data of the current transformer.

[0006] Furthermore, in the process of wavelet transform, an asymmetric extension method is used to process the signal boundary, and when the signal is at the boundary position, the boundary data is extended by a forward prediction method; The asymmetric extension method includes: In the wavelet decomposition of each layer, a first-order linear prediction equation is used at the starting end to predict the extended value before the sampling point of the first preset interval. When an obvious unidirectional trend appears, the extended value will continue in the trend direction. A first-order linear prediction equation is also used at the ending end to predict the extended value after the sampling point of the first preset interval. If a continuous downward trend of the current value is detected, the extended value will maintain the downward trend.

[0007] Furthermore, the first-order linear prediction equation uses the arithmetic mean of the last four sampling points as the prediction reference value, captures the changing trend of the current by calculating the average rate of change between the first and fourth most recent points, and automatically adjusts the weight ratio of the base layer and the trend layer according to the severity of the current change to obtain the predicted value at time t.

[0008] Furthermore, the energy entropy characteristic value is extracted by weighted processing, including: Calculate the weight value of each coefficient point based on the time sequence; Multiply each normalized energy probability value by its corresponding time weight to obtain a weighted probability value; Calculate the natural logarithm of each weighted probability value; Multiply the weighted probability value by its natural logarithm; Sum all the product results and take the negative value to get the weighted energy entropy value.

[0009] Furthermore, the calculation of the weight value includes: Determine the current moment and set it as the reference point; According to the time distance from each coefficient point to the reference point, the corresponding weight value is calculated through the exponential decay function; When the following situations are detected, the attenuation coefficient needs to be adjusted dynamically: If the current values ​​in three consecutive time windows show a monotonically decreasing trend, the attenuation coefficient is reduced to the first preset multiple of the original value to enhance the tracking capability of the long-term trend; If the current value fluctuations in two adjacent time windows exceed a preset percentage, the attenuation coefficient is increased to a second preset multiple of the original value to enhance the responsiveness to rapid changes; If the current value in the current time window is within the normal fluctuation range, the attenuation coefficient remains unchanged.

[0010] Furthermore, the extraction of the standard deviation characteristic value of adjacent sampling points includes: Construct a sliding time window and calculate the standard deviation of the difference sequence of adjacent sampling points in the window; The trend discrimination condition is set. If five consecutive difference points are all negative and the amplitude gradually increases, it is considered that there is a decay trend caused by contact oxidation. The nonlinear weighted method is used to process the difference sequence to obtain the standard deviation of adjacent sampling points. If it is found that multiple consecutive windows all present similar attenuation characteristics, by recording and superimposing the attenuation characteristic intensity detected in each window, when the cumulative intensity exceeds the preset threshold, the weight of the difference sequence in the standard deviation calculation is further increased; The standard deviation values ​​of all windows are combined to obtain the standard deviation characteristic values ​​of adjacent sampling points.

[0011] Furthermore, the step of constructing the current data anomaly detection model includes: Construct a classification hyperplane based on the input feature vector; The Gaussian kernel function is selected as the core mapping function, and the optimal classification hyperplane parameters are obtained by maximizing the classification interval between normal samples and abnormal samples while minimizing the classification error; Based on the obtained support vectors and weight coefficients, a decision function for discriminating new samples is established.

[0012] Furthermore, when constructing the classification hyperplane, a fault-tolerant mechanism is used to divide normal data and abnormal data: When the fault-tolerant mechanism divides data, if a sample point crosses the boundary, a penalty is imposed according to the degree of the cross-boundary.

[0013] According to another aspect of the present invention, there is provided a system for evaluating the quality of electric energy metering detection data, comprising: A data processing module, used for dividing the collected current transformer secondary side current data into N data segments according to a 24-hour time window, and performing wavelet transformation on the N data segments to obtain corresponding wavelet coefficient matrices; A feature construction module, used for extracting energy entropy eigenvalues ​​and adjacent sampling point standard deviation eigenvalues ​​of each data segment based on the wavelet coefficient matrix, and constructing the energy entropy eigenvalues ​​and adjacent sampling point standard deviation eigenvalues ​​into a feature vector; A model building module, used to establish a current data anomaly detection model, and calculate an abnormal probability value of each data segment through the current data anomaly detection model; The abnormality judgment module is used to compare the abnormal probability value with a preset threshold value. When the abnormal probability value is greater than the preset threshold value, the corresponding data segment is marked as an abnormal data segment to obtain a quality assessment result of the secondary side current data of the current transformer.

[0014] Compared with the prior art, the present invention provides an electric energy metering detection data quality assessment method and system, which considers the data anomalies caused by the gradual characteristics of the current transformer due to contact oxidation, first eliminates them through improved wavelet transform and feature extraction, and then calculates the anomaly rate of the processed data according to the support vector machine model. In addition, when designing the support vector machine, the existence of noise interference and abnormal fluctuations in the actual current data is considered. Since the contact resistance due to the contact oxidation of the secondary winding of the current transformer increases slowly, the collected current data produces a progressive attenuation, which solves the problem that the existing fixed threshold assessment method is difficult to detect such gradual anomalies in time, which easily causes electricity metering deviations. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative work. In the drawings: Figure 1 is a flow chart of a method for evaluating quality of electric energy metering detection data according to an embodiment of the present invention; Figure 2 The present invention is a flowchart of extracting standard deviation characteristic values ​​of adjacent sampling points in a method for evaluating the quality of electric energy metering detection data according to an embodiment of the present invention. DETAILED DESCRIPTION

[0016] Below, the exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments of the present invention, and it should be understood that the present invention is not limited to the exemplary embodiments described here.

[0017] Figure 1 1 is a system block diagram of a method for evaluating the quality of electric energy metering detection data according to an embodiment of the present invention. Figure 1As shown, in the electric energy metering detection data quality assessment method, it includes: S1: Divide the collected current transformer secondary side current data into N data segments according to a 24-hour time window, and perform wavelet transform on the N data segments to obtain corresponding wavelet coefficient matrices; First, the current transformer secondary current data collection cycle is 5 minutes, so there are 288 sampling points in 24 hours. Considering the fluctuation law of commercial building power load, 24 hours is selected as the time window length to divide the continuously collected current data into multiple data segments. Specifically, each data segment contains 288 sampling points, and there is no overlap between adjacent data segments, thus obtaining N data segments.

[0018] For each data segment, the db4 wavelet basis function is used to perform a 4-layer discrete wavelet decomposition. Through wavelet transform, the current data can be decomposed into low-frequency approximate components and high-frequency detail components at different scales. The wavelet coefficient matrix obtained after the transformation contains 5 sub-bands: the 4th layer approximate coefficient (a4) and the 1st to 4th layer detail coefficients (d1-d4). Among them, a4 reflects the long-term change trend of the current data, which is particularly suitable for capturing the slow attenuation characteristics caused by contact oxidation; d1-d4 contains fluctuation information at different frequency scales, which helps to distinguish normal load fluctuations from abnormal fluctuations.

[0019] Specifically in mathematical expression: Let the current data sequence of the i-th data segment be Xi(t), t=1,2,...,288, and the wavelet coefficient matrix obtained after wavelet transform is Wi. Wi contains 18 a4 coefficients, 18 d4 coefficients, 36 d3 coefficients, 72 d2 coefficients and 144 d1 coefficients, which together constitute the complete feature representation of the data segment in the time-frequency domain.

[0020] In this embodiment, when the secondary current data of the current transformer is subjected to wavelet transformation, the db4 wavelet basis function is selected because it has good orthogonality and compact support, and can effectively capture the mutation characteristics of the current signal under the vanishing moment of the fourth-order derivative. In view of the gradual characteristics caused by contact oxidation, the traditional db4 wavelet decomposition method is improved: an asymmetric extension method is used to process the signal boundary. When the signal is at the boundary position, the boundary data is extended by a forward prediction method, avoiding the boundary distortion introduced by the traditional symmetric extension when processing the gradual signal.

[0021] For the current data sequence of 288 sampling points, when performing the first layer of wavelet decomposition, when the filter of the db4 wavelet basis function moves to the first 4 sampling points (points 1-4) at the beginning of the sequence and the last 4 sampling points (points 285-288) at the end, boundary extension processing is required. Traditional symmetric extension will form an artificial symmetric structure at these boundary positions, which is not conducive to capturing the gradual trend. Therefore, an asymmetric extension method is used: For the starting end: by analyzing the trend of sampling points 5-12, the first-order linear prediction equation is used to predict the extension value before points 1-4. When there is an obvious unidirectional trend, the extension value will continue in the direction of the trend, avoiding the trend reversal caused by symmetrical extension.

[0022] For the end: Based on the change trend of sampling points 281-284, the first-order linear prediction equation is also used to predict the extension value after points 285-288. If a continuous downward trend in the current value is detected, the extension value will maintain the downward trend, thereby more accurately reflecting the signal attenuation characteristics caused by contact oxidation.

[0023] In the subsequent second to fourth layer decomposition, the same asymmetric extension strategy is used each time the low-frequency coefficients are processed. For example, when processing the 72 a1 coefficients of the second layer, the first and last two coefficient points are asymmetric extended, and the extension direction is consistent with the currently observed local trend. This processing method runs through the entire multi-layer decomposition process, ensuring the consistency of boundary processing and the continuity of gradient features. This improved boundary processing method overcomes the limitations of traditional symmetric extension in processing gradient signals and improves the accuracy of wavelet transform in extracting slowly changing features in current data.

[0024] The first-order linear prediction equation prediction can be expressed by the following formula: ; in, for The predicted value at time, is the current sampling value at the historical moment, is the trend adjustment coefficient, and its value range is [0.5,1.5], is the sensitivity coefficient, and its value range is [0.1,1], is the average slope of the last four sampling points, is the base of natural logarithms, Indicates the current sampling value of the secondary side of the current transformer.

[0025] Among them, the determination logic of the trend adjustment coefficient and the sensitivity coefficient is as follows: Confirmation of the α parameter: First calculate the mean and standard deviation of the last four current sampling points. When the standard deviation is large, it means that the data fluctuates violently. At this time, a smaller α value (close to 0.5) is preferred to reduce the influence weight of the trend term and rely more on the moving average for prediction; when the standard deviation is small and the data shows a unidirectional change trend, a larger α value (close to 1.5) is selected to enhance the influence of the trend term so that the predicted value can better follow the trend change. Specifically, if the change direction of three consecutive sampling points is consistent and the standard deviation is less than the preset threshold (the standard deviation threshold can be determined based on the distribution of historical data, such as taking 20% ​​of the historical standard deviation), the α value is set to 1.5; if the standard deviation is greater than twice the preset threshold, the α value is set to 0.5; in other cases, the α value is determined by linear interpolation, and the following formula can be referred to: ; in, is the standard deviation of the last four sampling points, is the average of the last four sampling points.

[0026] Confirmation of the β parameter: First calculate the slope value at the current moment and compare it with the maximum slope value observed in history. When the current slope value is small (such as less than 10% of the maximum slope), select a smaller β value (close to 0.1) to reduce the impact of the exponential modulation term; when a significant slope change is observed (such as greater than 50% of the maximum slope), select a larger β value (close to 1) to enhance the impact of the exponential modulation term, so that the forecast is more sensitive to trend changes. Specifically, before each calculation, update the historical maximum slope value first, and then determine the β value based on the ratio of the current slope to the maximum slope: if the ratio is less than 0.1, β is 0.1; if the ratio is greater than 0.5, β is 1; when it is between the two, the β value is determined by linear mapping, which can be referred to the following formula: ; in, is the maximum slope value observed in history.

[0027] It is worth noting that although the above description outlines the general process of the wavelet transform multi-scale analysis method that uses an asymmetric extension method to process signal boundaries, the specific implementation details may vary depending on different application scenarios and their data characteristics. For example, in the analysis of electricity consumption data in a large shopping mall, due to the large proportion of air-conditioning load, the current data shows obvious seasonal variation characteristics. At this time, it is necessary to seasonally adjust the data before the first layer of wavelet decomposition. Specifically, by analyzing historical data from the same period, a seasonal index model is established, and the original current data is divided by the corresponding seasonal index to obtain a data sequence after eliminating the seasonal effect, and then wavelet decomposition is performed.

[0028] In another case, the current data of an office building showed a gradual decay of about 5% within 3 months due to the humid installation environment of the current transformer and the fast contact oxidation rate. When processing such data, the traditional symmetric continuation method will produce false trend reversals at the boundaries of the data segments, resulting in a decrease in the sensitivity of anomaly detection. The improved asymmetric continuation method accurately continued the decay trend through the prediction equation, allowing the anomaly detection model to detect this anomaly 15 days in advance, nearly 10 days earlier than the traditional method.

[0029] This multi-scale analysis method based on wavelet transform can effectively extract the time-varying features in the current data, laying the foundation for subsequent anomaly detection. Compared with traditional time domain analysis, this method has obvious advantages in processing non-stationary signals, especially for the current data quality assessment problem that needs to pay attention to both long-term trends and short-term fluctuations.

[0030] S2: extracting energy entropy eigenvalues ​​and standard deviation eigenvalues ​​of adjacent sampling points of each data segment based on the wavelet coefficient matrix, and constructing the energy entropy eigenvalues ​​and standard deviation eigenvalues ​​of adjacent sampling points into a eigenvector; The first step is to extract the energy entropy eigenvalue. For the wavelet coefficient matrix of each data segment, the energy entropy of the five sub-bands (a4 and d1-d4) needs to be calculated separately. When calculating, the wavelet coefficients of each sub-band are first squared to obtain the energy distribution, and then the energy value is normalized to the probability distribution. More specifically, the weighted processing method is as follows: The weight value of each coefficient point is calculated based on the time sequence. The weight calculation uses an exponential decay function. The specific process is: first determine the current time and set it as the reference point; then calculate the corresponding weight value through the exponential decay function based on the time distance from each coefficient point to the reference point. When the following situations are monitored, the attenuation coefficient needs to be dynamically adjusted: If the current values ​​in three consecutive time windows show a monotonically decreasing trend, the attenuation coefficient is reduced to the first preset multiple of the original value to enhance the tracking capability of the long-term trend; If the current value fluctuations in two adjacent time windows exceed a preset percentage, the attenuation coefficient is increased to a second preset multiple of the original value to enhance the responsiveness to rapid changes; If the current value in the current time window is within the normal fluctuation range, the attenuation coefficient remains unchanged.

[0031] Further, the calculation of weighted energy entropy: Multiply each normalized energy probability value by its corresponding time weight to obtain a weighted probability value; Calculate the natural logarithm of each weighted probability value; Multiply the weighted probability value by its natural logarithm; Sum all the product results and take the negative value to get the weighted energy entropy value of the subband.

[0032] The second is to extract the standard deviation characteristic values ​​of adjacent sampling points. Considering the gradual change characteristics caused by contact oxidation of the current transformer, the traditional standard deviation calculation method is improved, such as Figure 2 As shown in the figure: Instead of using the original sampling point values ​​directly, a difference sequence is constructed: when there are 12 sampling points in the window, 11 difference points are calculated. If the difference between two adjacent sampling points exceeds the preset threshold, it is considered that the difference point may be interfered with and the local median is used instead. When the window slides to the boundary of the data segment, adaptive boundary processing is used: if an obvious unidirectional change trend is detected, the boundary data is supplemented by trend extrapolation.

[0033] Oxidation trend detection and processing are added to the traditional standard deviation calculation method: when calculating the difference sequence within the window, set the trend judgment condition - if 5 consecutive difference points are negative and the amplitude gradually increases, it is considered that there may be an attenuation trend caused by contact oxidation, and the difference sequence is processed by a nonlinear weighted method: for the difference points that meet the attenuation characteristics, a larger weight coefficient is assigned to make the standard deviation calculation more sensitive to the attenuation trend. The specific weight coefficient is determined in the form of an exponential function: when the difference is negative and the amplitude gradually increases, the weight increases with the increase in the amplitude; when the difference is positive or the amplitude does not satisfy the increasing relationship, the basic weight is maintained.

[0034] If it is found that multiple consecutive windows show similar attenuation characteristics, the attenuation feature intensity detected in each window is recorded and superimposed. When the cumulative intensity exceeds the preset threshold, the weight of the difference sequence in the standard deviation calculation is further increased. This accumulation mechanism is particularly suitable for detecting slowly developing contact oxidation problems.

[0035] To prevent normal load fluctuations from being misjudged as contact oxidation characteristics, after the initial standard deviation value is calculated, it is compared with the fluctuation pattern of historical data in the same period. If the current standard deviation change trend deviates significantly from the historical fluctuation law and conforms to the characteristic performance of contact oxidation, it is considered that abnormal attenuation exists.

[0036] The standard deviation values ​​of all windows are combined: when the standard deviation values ​​of adjacent windows change suddenly, the influence of instantaneous fluctuations is eliminated through smoothing; when the standard deviation values ​​of multiple consecutive windows show a consistent change trend, this trend feature is retained. Through this multi-window feature combination, the final standard deviation feature value that can reflect the overall change trend is obtained.

[0037] This improved feature extraction method significantly improves the ability of feature vectors to characterize gradual anomalies in current data through mechanisms such as time weighting, frequency band weighting and adaptive normalization.

[0038] S3: using a support vector machine to train the feature vector to establish a current data anomaly detection model, and calculating an abnormal probability value of each data segment through the current data anomaly detection model; Support vector machines have excellent nonlinear feature mapping capabilities and strong generalization performance, and are particularly suitable for classifying normal and abnormal data in current transformers. When faced with gradual abnormal features in current data, support vector machines map low-dimensional features to high-dimensional feature space through kernel functions, which can effectively capture the nonlinear distribution characteristics of data. At the same time, based on the principle of structural risk minimization, support vector machines can avoid overfitting problems while ensuring classification accuracy.

[0039] The steps to build a current data anomaly detection model include: First, the optimal classification hyperplane is constructed. In the high-dimensional feature space, the classification hyperplane is constructed based on the input feature vector so that normal data and abnormal data can be separated to the greatest extent. After the feature vector is mapped to the high-dimensional space, the Gaussian kernel function is selected as the core mapping function to obtain the optimal classification hyperplane parameters by maximizing the classification interval between normal samples and abnormal samples while minimizing the classification error.

[0040] Considering the existence of noise interference and abnormal fluctuations in the actual current data, a fault-tolerant mechanism is introduced when constructing the classification hyperplane. More specifically, the core of the fault-tolerant mechanism is to allow some sample points to "cross the boundary". When the support vector machine divides normal data and abnormal data, ideally all normal samples are required to be on one side of the classification hyperplane, and all abnormal samples are required to be on the other side. However, there is often noise interference in the actual current data, which may cause some normal samples to cross the classification boundary, or abnormal samples to mix into the normal area. At this time, if strict classification is required, the generalization ability of the model will be reduced. Therefore, this embodiment introduces a fault-tolerant mechanism for fault-tolerant processing. Specifically, if a sample point crosses the boundary slightly (such as a normal sample slightly deviates from the normal area), a smaller penalty is given; if the sample point crosses the boundary seriously (such as a normal sample goes deep into the abnormal area), a larger penalty is imposed, and the penalty intensity is determined by the distance factor. When a sample point crosses the boundary, the distance to the classification boundary of the area to which it should belong is calculated: the larger the cross-boundary distance, the greater the difference between the characteristics of the sample point and its labeled category, and the penalty intensity is increased at this time; the smaller the cross-boundary distance, the sample point is only slightly deviated, and the penalty intensity is reduced at this time. Through this graded punishment strategy, the model can treat different degrees of boundary violations differently: maintaining a certain degree of tolerance for minor violations and imposing strict restrictions on serious violations.

[0041] Preferably, the fluctuations during normal operation are avoided from being misjudged as abnormal states by introducing a fault-tolerant mechanism. If the current transformer has a gradual abnormality caused by contact oxidation, the fault-tolerant mechanism dynamically adjusts the penalty value so that the model remains sensitive to slight changes in the early stages of abnormal development. When sudden current fluctuations occur, the fault-tolerant mechanism can quickly adjust the classification boundaries through a larger penalty value and capture abnormal features in a timely manner, so that neither misjudgment nor reduction of abnormal judgment can be achieved. Instead, the detection accuracy is improved while reducing the false positive rate. After verification under multiple working conditions, the fault-tolerant mechanism significantly reduces the false alarm rate while ensuring the detection accuracy: compared with the traditional fixed threshold detection method, the false alarm rate is reduced by about 40%; compared with the support vector machine model without the introduction of the fault-tolerant mechanism, the false alarm rate is reduced by about 25%. At the same time, the detection rate of gradual abnormalities is increased by about 30%.

[0042] Furthermore, based on the obtained support vectors and weight coefficients, a decision function for discriminating new samples is established. When a data point needs to be discriminated, the kernel function value between the point and all support vectors is calculated, and the weighted sum is performed in combination with the corresponding weight coefficients, and finally the final discrimination result is obtained by adjusting the bias term. It should be noted that the decision function directly reflects the effect of the fault-tolerant mechanism. When the appropriate penalty value and slack variable are determined by the fault-tolerant mechanism, these parameters will directly affect the selection of support vectors and the size of the weight coefficient. For those sample points that are out of bounds but allowed by the fault-tolerant mechanism, they may be selected as support vectors, and their corresponding weight coefficients will be adjusted accordingly according to the size of the penalty value: the larger the penalty value, the larger the weight coefficient of the support vector in the decision function, and the more significant the impact on the discrimination result of the new sample. When a new data point needs to be discriminated, the decision function calculates the kernel function value between the point and each support vector. Since the support vectors contain sample points that have been processed for fault tolerance, the weight coefficients carried by these support vectors reflect the discrimination tendency of the fault tolerance mechanism: if a support vector is an out-of-bounds point allowed due to fault tolerance, its weight coefficient will be reduced accordingly, thereby reducing its impact when discriminating new samples.

[0043] Furthermore, the decision function value is calculated according to the constructed and trained support vector machine model. First, the kernel function value between the feature vector and all support vectors is calculated, and then these kernel function values ​​are weighted summed with the corresponding weight coefficients, and finally the bias term is added to obtain the decision function value.

[0044] After the decision function value is obtained, it is mapped to a probability interval of 0 to 1 through the probability conversion function. If the decision function value is larger, it means that the data segment is closer to the abnormal state, and the probability value obtained by mapping is closer to 1; if the decision function value is smaller, it means that the data segment is closer to the normal state, and the probability value obtained by mapping is closer to 0.

[0045] After the initial probability value is obtained, the time series correlation of adjacent data segments is considered for correction. If the probability value of the current data segment is too different from that of the adjacent data segment, appropriate adjustments are made according to the time series smoothing criterion; if multiple consecutive data segments show a high abnormal probability, the abnormal probability value of the current data segment is increased through the cumulative effect.

[0046] The calculation of the kernel function value between the feature vector and all support vectors can be expressed by the following formula: ; Where D(x) is the decision function value, is the normalized feature vector of the input, is the number of support vectors, is the category label of the ith support vector (value +1 or -1), is the i-th support vector, is the width parameter of the Gaussian kernel function, is the time series correlation function of the samples, and b is the bias term.

[0047] S4: Compare the abnormal probability value with a preset threshold value. When the abnormal probability value is greater than the preset threshold value, mark the corresponding data segment as an abnormal data segment to obtain a quality assessment result of the secondary side current data of the current transformer.

[0048] According to the operating characteristics of the current transformer, three key threshold points are set: the mild abnormality threshold is set to 0.6, the moderate abnormality threshold is set to 0.75, and the severe abnormality threshold is set to 0.85. When the abnormal probability value is in different threshold intervals, it indicates different abnormality levels.

[0049] When the abnormal probability value of a data segment is detected to exceed the mild abnormality threshold (0.6) but lower than the moderate abnormality threshold (0.75), the system will continuously monitor the abnormal probability values ​​of the next five data segments. If the abnormal probability values ​​of these data segments all exceed the mild abnormality threshold, the current data segment will be marked as "mild abnormality"; if the abnormal probability values ​​of subsequent data segments fluctuate, no abnormal marking will be performed for the time being and observation will continue.

[0050] When the abnormal probability value of a data segment exceeds the moderate abnormality threshold (0.75) but is lower than the severe abnormality threshold (0.85), the system will shorten the observation period to 3 data segments. If the abnormal probability values ​​of these data segments continue to exceed the moderate abnormality threshold, the current data segment will be marked as "moderate abnormality" and an alarm signal will be issued; if the abnormal probability value drops significantly during the observation period, it will be downgraded to mild abnormality processing.

[0051] When the abnormal probability value of a data segment exceeds the severe abnormality threshold (0.85), the system will immediately mark the data segment as "severely abnormal" and trigger the emergency alarm mechanism. At the same time, the adjacent data segments are retrospectively analyzed to determine the starting time point and development process of the abnormality.

[0052] The influence of seasonal factors is also considered in the marking process. When making abnormal judgments during the high temperature period in summer, considering the impact of temperature on the performance of the current transformer, the thresholds of each level will be appropriately increased by 0.05; when making judgments in cold seasons, the thresholds of each level remain at the standard settings. After completing the abnormal marking, the system will generate a detailed quality assessment report. The report content includes: the time distribution of abnormal data segments, the evolution process of abnormal degree, the correlation analysis of adjacent data segments and other information.

[0053] Here, those skilled in the art can understand that the specific operations of each step in the above-mentioned electric energy metering detection data quality assessment system have been referred to above. Figure 1 and Figure 2 The method for evaluating the quality of electric energy metering detection data has been described in detail, and therefore, its repeated description will be omitted.

[0054] In summary, the method for evaluating the quality of electric energy metering detection data based on the embodiment of the present invention is explained, which considers the data anomalies caused by the gradual characteristics of the current transformer due to contact oxidation, first eliminates them through improved wavelet transform and feature extraction, and then calculates the abnormality rate of the processed data according to the support vector machine model, and considers the existence of noise interference and abnormal fluctuations in the actual current data when designing the support vector machine. In this way, due to the slow increase in contact resistance caused by contact oxidation of the secondary winding of the current transformer, the collected current data produces a progressive attenuation, which solves the problem that the existing fixed threshold evaluation method is difficult to detect such gradual anomalies in time, which easily causes the deviation of electricity metering.

Claims

1. A method for evaluating the quality of electric energy metering detection data, characterized in that: include: The collected current transformer secondary side current data is divided into N data segments according to a 24-hour time window, and the N data segments are respectively subjected to wavelet transformation to obtain corresponding wavelet coefficient matrices; Extracting energy entropy eigenvalues ​​and standard deviation eigenvalues ​​of adjacent sampling points of each data segment based on the wavelet coefficient matrix, and constructing the energy entropy eigenvalues ​​and standard deviation eigenvalues ​​of adjacent sampling points into a eigenvector; Establishing a current data anomaly detection model, and calculating an abnormal probability value of each data segment through the current data anomaly detection model; The abnormal probability value is compared with a preset threshold value. When the abnormal probability value is greater than the preset threshold value, the corresponding data segment is marked as an abnormal data segment to obtain a quality assessment result of the secondary side current data of the current transformer.

2. The method for evaluating the quality of electric energy metering detection data according to claim 1, characterized in that: In the process of wavelet transform, an asymmetric extension method is used to process the signal boundary. When the signal is at the boundary position, the boundary data is extended by a forward prediction method. The asymmetric extension method includes: In the wavelet decomposition of each layer, a first-order linear prediction equation is used at the starting end to predict the extended value before the sampling point of the first preset interval. When an obvious unidirectional trend appears, the extended value will continue in the trend direction. A first-order linear prediction equation is also used at the ending end to predict the extended value after the sampling point of the first preset interval. If a continuous downward trend of the current value is detected, the extended value will maintain the downward trend.

3. The method for evaluating the quality of electric energy metering detection data according to claim 2, characterized in that: The first-order linear prediction equation uses the arithmetic mean of the last four sampling points as the prediction reference value, captures the current change trend by calculating the average change rate between the first and fourth points, and automatically adjusts the weight ratio of the base layer and the trend layer according to the severity of the current change to obtain the predicted value at time t.

4. The method for evaluating the quality of electric energy metering detection data according to claim 1, characterized in that: The energy entropy characteristic value is extracted by weighted processing, including: Calculate the weight value of each coefficient point based on the time sequence; Multiply each normalized energy probability value by its corresponding time weight to obtain a weighted probability value; Calculate the natural logarithm of each weighted probability value; Multiply the weighted probability value by its natural logarithm; Sum all the product results and take the negative value to get the weighted energy entropy value.

5. The method for evaluating the quality of electric energy metering detection data according to claim 4, characterized in that: The calculation of the weight value includes: Determine the current moment and set it as the reference point; According to the time distance from each coefficient point to the reference point, the corresponding weight value is calculated through the exponential decay function; When the following situations are detected, the attenuation coefficient needs to be adjusted dynamically: If the current values ​​in three consecutive time windows show a monotonically decreasing trend, the attenuation coefficient is reduced to the first preset multiple of the original value to enhance the tracking capability of the long-term trend; If the current value fluctuations in two adjacent time windows exceed a preset percentage, the attenuation coefficient is increased to a second preset multiple of the original value to enhance the responsiveness to rapid changes; If the current value in the current time window is within the normal fluctuation range, the attenuation coefficient remains unchanged.

6. The method for evaluating the quality of electric energy metering detection data according to claim 1, characterized in that: The extraction of the standard deviation characteristic values ​​of adjacent sampling points includes: Construct a sliding time window and calculate the standard deviation of the difference sequence of adjacent sampling points in the window; The trend discrimination condition is set. If five consecutive difference points are all negative and the amplitude gradually increases, it is considered that there is a decay trend caused by contact oxidation. The nonlinear weighted method is used to process the difference sequence to obtain the standard deviation of adjacent sampling points. If it is found that multiple consecutive windows all present similar attenuation characteristics, by recording and superimposing the attenuation characteristic intensity detected in each window, when the cumulative intensity exceeds the preset threshold, the weight of the difference sequence in the standard deviation calculation is further increased; The standard deviation values ​​of all windows are combined to obtain the standard deviation characteristic values ​​of adjacent sampling points.

7. The method for evaluating the quality of electric energy metering detection data according to claim 1, characterized in that: The steps of constructing the current data anomaly detection model include: Construct a classification hyperplane based on the input feature vector; The Gaussian kernel function is selected as the core mapping function, and the optimal classification hyperplane parameters are obtained by maximizing the classification interval between normal samples and abnormal samples while minimizing the classification error; Based on the obtained support vectors and weight coefficients, a decision function for discriminating new samples is established.

8. The method for evaluating the quality of electric energy metering detection data according to claim 6, characterized in that: When constructing the classification hyperplane, a fault-tolerant mechanism is used to divide normal data and abnormal data: When the fault-tolerant mechanism divides data, if a sample point crosses the boundary, a penalty is imposed according to the degree of the cross-boundary.

9. An electric energy metering detection data quality assessment system, characterized in that: include: A data processing module, used for dividing the collected current transformer secondary side current data into N data segments according to a 24-hour time window, and performing wavelet transformation on the N data segments to obtain corresponding wavelet coefficient matrices; A feature construction module, used for extracting energy entropy eigenvalues ​​and adjacent sampling point standard deviation eigenvalues ​​of each data segment based on the wavelet coefficient matrix, and constructing the energy entropy eigenvalues ​​and adjacent sampling point standard deviation eigenvalues ​​into a feature vector; A model building module, used to establish a current data anomaly detection model, and calculate an abnormal probability value of each data segment through the current data anomaly detection model; The abnormality judgment module is used to compare the abnormal probability value with a preset threshold value. When the abnormal probability value is greater than the preset threshold value, the corresponding data segment is marked as an abnormal data segment to obtain a quality assessment result of the secondary side current data of the current transformer.

Images