CN117056714A - Method, system, equipment and storage medium for identifying PMU bad data of intelligent power distribution network - Google Patents

Abstract

The invention relates to the technical field of data processing, in particular to a method, a system, equipment and a storage medium for identifying bad PMU data of an intelligent power distribution network. The method comprises the following steps: acquiring a PMU measurement sequence, and acquiring a corresponding measurement reconstruction sequence based on the PMU measurement sequence; forming a PMU measurement sequence combination based on the PMU measurement sequence and the corresponding measurement reconstruction sequence, and generating corresponding two-dimensional image data; processing the two-dimensional image data by utilizing hybrid clustering to obtain a preliminary PMU measurement sequence point identification result; and performing integrated learning identification and result correction on the identification results of each point of the initial PMU measurement sequence to obtain the final identification results of each point of the PMU measurement sequence. The method based on the chart identifies the characteristics of normal and bad data, exerts the potential of dimension clustering and classifier, and improves the space-time correlation analysis efficiency and the identification sensitivity.

Description

Intelligent distribution network PMU bad data identification method, system, device and storage medium

Technical Field

The present invention relates to the field of data processing technology, and in particular to a method, system, device and storage medium for identifying bad data of a PMU in a smart distribution network.

Background Art

Hybrid clustering: refers to using multiple clustering methods to cluster the same data set.

Smart distribution network: Smart distribution network is one of the key links of smart grid. Usually, power networks of 10kV and below belong to the distribution network (some areas have 20kV). The distribution network is the part of the entire power system that is directly connected to the dispersed users. The smart distribution network system uses modern electronic technology, communication technology, computer and network technology to integrate the online and offline data of the distribution network, the distribution network data and user data, the grid structure and geographic graphics, and realize the intelligent monitoring, protection, control, power consumption and distribution management of the normal operation and accident of the distribution system.

PMU: Synchronous phasor measurement unit (PMU: phasor measurement unit) is a phasor measurement unit that uses the global positioning system (GPS) second pulse as a synchronous clock. It can be used in the fields of dynamic monitoring, system protection, system analysis and prediction of power systems. It is an important device to ensure the safe operation of power grids. PMU based on GPS clock can measure phasor data such as voltage phase and current phase at the hub of the power system, and transmit the data to the monitoring master station through the communication network. The monitoring master station measures the phase amplitude at different points.

Reconstructed data: refers to simulated measurement data generated by algorithms such as adversarial generative networks, which has the same distribution as the actual measurement data.

Bad data: In addition to normal data, abnormal data obtained by PMU measurement also includes missing data, outliers and event data, among which outliers and missing data are bad data caused by poor measurement quality. The purpose of bad data identification is to classify outliers and missing data as bad data, while event data is classified as normal data through accurate measurement.

Generative Adversarial Network: GAN for short. GAN is a deep generative model consisting of a discriminant module and a generator module. During the training process, the generator G inputs Gaussian noise of the same dimension as the target data, and the discriminator D inputs normal measurement information and pseudo data output by the generator. The two are trained alternately and iteratively to form a game confrontation. Finally, the generator and the discriminator reach a Nash equilibrium, at which time the generator outputs the reconstructed measurement data.

PMU data is the basis for monitoring, control and analysis of power systems. Therefore, the quality of PMU data is crucial and has a significant impact on the analysis results and even the safety of power system operation. However, PMU equipment has a complex structure and is easily affected by internal and external factors, which causes the time series of PMU measurements to contain bad or abnormal data. Therefore, it is necessary to identify bad data in PMU measurement data in power systems and strive to improve the quality of PMU measurement data.

Outliers in PMU data are data points that deviate significantly from the expected measurement. In the absence of system events, outliers are usually shaped like spikes that rise and fall suddenly. Event data is generated when system events such as switching events and sudden load changes occur. Typically, step event data is displayed because the PMU measurement value changes from the pre-event period to the post-event period. If the PMU measurement value has a small deviation before and after the transient period, the event data is spike-shaped, and the PMU measurement curve has similar sudden changes and outliers. Bad data identification methods based on checking the occurrence of sudden changes can easily identify outliers, such as data peaks under static operating conditions. However, the similar profiles of outliers and spike-shaped event data can lead to failure and inaccuracy in bad data identification.

Summary of the invention

In order to solve the technical problems existing in the above-mentioned prior art, the present invention provides a method, system, device and storage medium for identifying bad data of PMU in a smart distribution network.

To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

In a first aspect, in one embodiment provided by the present invention, a method for identifying bad data of a PMU in a smart distribution network is provided, the method comprising the following steps:

Acquire a PMU measurement sequence, and obtain a corresponding measurement reconstruction sequence based on the PMU measurement sequence;

Based on the PMU measurement sequence and the corresponding measurement reconstruction sequence, a PMU measurement sequence combination is formed, and corresponding two-dimensional image data is generated;

Processing the two-dimensional image data by hybrid clustering to obtain preliminary identification results of each point in the PMU measurement sequence;

The preliminary identification results of each point in the PMU measurement sequence are subjected to integrated learning identification and result correction to obtain the final identification results of each point in the PMU measurement sequence.

As a further solution of the present invention, before obtaining the PMU measurement sequence and obtaining the corresponding measurement reconstruction sequence based on the PMU measurement sequence, the following steps are also included:

A sample data set is obtained, and a GAN model is trained using the sample data set to obtain a trained GAN model; wherein the sample data set includes a sample PMU measurement sequence combination and corresponding sample two-dimensional image data.

As a further solution of the present invention, the acquiring of the PMU measurement sequence and obtaining the corresponding measurement reconstruction sequence based on the PMU measurement sequence includes:

A PMU measurement sequence is obtained, and a corresponding measurement reconstruction sequence is generated based on the PMU measurement sequence using a trained GAN model.

As a further solution of the present invention, the PMU measurement sequence is _Xi , the measurement reconstruction sequence is _Xj , the PMU measurement sequence combination is _Xij , and _Xij is calculated by the following formula:

As a further solution of the present invention, the two-dimensional image data is processed by hybrid clustering to obtain preliminary PMU measurement sequence point identification results, including:

The two-dimensional image data are processed in turn by using linear regression identifier, DBSCAN identifier and Gaussian mixture models (GMM) identifier to obtain preliminary identification results of each point in the PMU measurement sequence.

As a further solution of the present invention, integrated learning identification and result correction are performed on the preliminary identification results of each point in the PMU measurement sequence to obtain the final identification results of each point in the PMU measurement sequence, including:

Perform integrated learning and identification on the identification results of each point in the preliminary PMU measurement sequence to obtain the voting results;

The voting results are corrected to obtain the final identification results of each point in the PMU measurement sequence.

As a further solution of the present invention, the ensemble learning identification is a collection method using majority voting as a basic identifier.

In a second aspect, in another embodiment provided by the present invention, a smart distribution network PMU bad data identification system is provided, the system comprising: a measurement reconstruction sequence acquisition module, a construction sequence combination module, a preliminary identification module and a final identification module;

The measurement reconstruction sequence acquisition module is used to acquire a PMU measurement sequence and obtain a corresponding measurement reconstruction sequence based on the PMU measurement sequence;

The sequence combination construction module is used to construct a PMU measurement sequence combination based on the PMU measurement sequence and the corresponding measurement reconstruction sequence, and generate corresponding two-dimensional image data.

The preliminary identification module is used to process the two-dimensional image data by using hybrid clustering to obtain preliminary identification results of each point in the PMU measurement sequence;

The final identification module is used to perform integrated learning identification and result correction on the preliminary identification results of each point in the PMU measurement sequence to obtain the final identification results of each point in the PMU measurement sequence.

In a third aspect, in another embodiment provided by the present invention, a device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method for identifying bad data of PMU in a smart distribution network when loading and executing the computer program.

In a fourth aspect, in another embodiment provided by the present invention, a storage medium is provided, storing a computer program, and when the computer program is loaded and executed by a processor, the steps of the method for identifying bad data of PMU in the smart distribution network are implemented.

The technical solution provided by the present invention has the following beneficial effects:

The present invention provides a method, system, device and storage medium for identifying bad data of PMU in a smart distribution network. The present invention obtains a PMU measurement sequence, and obtains a corresponding measurement reconstruction sequence based on the PMU measurement sequence; based on the PMU measurement sequence and the corresponding measurement reconstruction sequence, a PMU measurement sequence combination is formed, and corresponding two-dimensional image data is generated; the two-dimensional image data is processed by hybrid clustering to obtain preliminary identification results of each point of the PMU measurement sequence; the preliminary identification results of each point of the PMU measurement sequence are integrated learning identification and result correction to obtain the final identification results of each point of the PMU measurement sequence; the present invention forms coordinates of the PMU measurement sequence and its reconstruction sequence, and draws a two-dimensional scatter plot. The method based on two-dimensional images improves the efficiency of spatiotemporal correlation analysis of the PMU measurement data time series and its reconstruction sequence, and specifically distinguishes the difference characteristics between normal and bad data and their reconstructed data. Hybrid clustering is used to divide normal and bad data, and the identification results are further corrected by integrated learning. The chart-based method identifies the characteristics of normal and bad data and exerts the potential of dimensional clustering and classifiers. Therefore, the efficiency of spatiotemporal correlation analysis and identification sensitivity are improved.

These and other aspects of the present invention will become more concise and understandable in the following description of the embodiments. It should be understood that the above general description and the following detailed description are only exemplary and explanatory and cannot limit the present invention.

BRIEF DESCRIPTION OF THE DRAWINGS

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 description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other embodiments can be obtained based on these drawings without paying creative work.

FIG1 is a flow chart of a method for identifying bad data of a PMU in a smart distribution network according to an embodiment of the present invention.

FIG2 is a specific flow chart of a method for identifying bad data of a PMU in a smart distribution network according to an embodiment of the present invention.

Figure 3 shows different types of PMU anomaly diagrams.

Figure 4 is a diagram of the GAN model network structure.

FIG5 shows a PMU measurement sequence and its measurement reconstruction sequence.

FIG6 is a scatter plot of the PMU measurement sequence and its measurement reconstruction sequence.

Figure 7 is a modified diagram of the safety zone.

Figure 8 shows the PMU measurement curve of normal data (correlation coefficient: 0.9669) under typical conditions.

Figure 9 shows the step event data under typical conditions and the outlier (correlation coefficient: 0.9665) PMU measurement curve.

Figure 10 shows the spike event data under typical conditions and the outlier (correlation coefficient: 0.9435) PMU measurement curve.

Figure 11 shows the PMU measurement curve with an outlier value (correlation coefficient: 0.6527) under typical conditions.

FIG12 shows the bad data identification process, results and effects of normal data under typical conditions.

FIG13 shows the bad data identification process, results and effects containing step event data and outliers under typical conditions.

FIG14 shows the bad data identification process, results and effects containing spike event data and outliers under typical conditions.

Figure 15 shows the bad data identification process, results and effects containing outliers under typical conditions.

FIG. 16 is a structural diagram of a smart distribution network PMU bad data identification system according to an embodiment of the present invention.

In the figure: measurement and reconstruction sequence acquisition module-100, construction sequence combination module-200, preliminary identification module-300, final identification module-400.

DETAILED DESCRIPTION

The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations/steps, nor must they be executed in the order described. For example, some operations/steps may also be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.

It should be understood that the terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

Here, the PMU sequentially acquires measurements from a single bus at a sampling interval of Δt. Within a time window of length T for observation, a PMU measurement time series of length n (=T/Δt) is captured as the basic analysis unit for bad PMU data identification. The sampled time series consists of several electrical quantities, such as the amplitude and angle of voltage and current, active and reactive power injection. Further analysis for bad data identification is performed on the PMU voltage amplitude time series, expressed as

x _i =[x _i,1 ,x _i,2 ,…,x _i,n ] (1)

Where i represents the bus number where the PMU is deployed.

In addition to normal data, abnormal data obtained by PMU measurement also includes missing data, outliers and event data, among which outliers and missing data are bad data caused by poor measurement quality. The purpose of bad data identification is to classify outliers and missing data as bad data, while event data is classified as normal data through accurate measurement.

Missing data is usually caused by missing faults in the data measurement, transmission and storage process, which results in the value of zero, empty or NaN (not a number) at the corresponding position in the historical data set. When the data is dropped, the missing data is easy to find in the visible graph of the PMU time series and can be easily eliminated by exact label matching. Therefore, this section pays little attention to the identification of missing data.

FIG3 shows an illustrative example of a single-bus PMU measurement time series with a 5-second window length, which demonstrates different types of PMU abnormal data, including missing data, outliers, and event data. The PMU device samples 50 data points per second with an interval of 20 milliseconds. Undoubtedly, the sharp data drop of 4 seconds is identified as missing data by its marker "0". Based on its unique graphical features, the step change of the PMU measurement value at approximately 4.60 seconds can be identified as step event data. However, the data spikes at approximately 0.76 seconds, 1.44 seconds, 2.26 seconds, and 3.24 seconds are confused by similar shapes and cannot be identified as outliers or spike event data. Since the system event is unknown, more information is needed as a criterion for identifying similar spike-like data.

Specifically, the embodiments of the present invention are further described below in conjunction with the accompanying drawings.

Please refer to FIG. 1 and FIG. 2 . FIG. 1 is a flow chart of a method for identifying bad data of a PMU in a smart distribution network provided by an embodiment of the present invention. As shown in FIG. 1 , the method for identifying bad data of a PMU in a smart distribution network includes steps S10 to S40 .

S10, obtaining a PMU measurement sequence, and obtaining a corresponding measurement reconstruction sequence based on the PMU measurement sequence, wherein the PMU measurement sequence is V _i , and the measurement reconstruction sequence is V _j . The PMU measurement sequence is a node voltage PMU measurement data time sequence.

In an embodiment of the present invention, before acquiring the PMU measurement sequence and obtaining the corresponding measurement reconstruction sequence based on the PMU measurement sequence, the following steps are also included:

A sample data set is obtained, and a GAN model is trained using the sample data set to obtain a trained GAN model; wherein the sample data set includes a sample PMU measurement sequence combination and corresponding sample two-dimensional image data.

In an embodiment of the present invention, the sample PMU measurement sequence combination includes a PMU measurement sequence and a sample measurement reconstruction sequence.

In an embodiment of the present invention, the sample PMU measurement sequence set is all node voltage measurement data including PMU measurements in the smart grid system at the same time.

In an embodiment of the present invention, the sample measurement reconstruction sequence set is historical moment measurement data.

The acquiring of the PMU measurement sequence and obtaining a corresponding measurement reconstruction sequence based on the PMU measurement sequence includes:

A PMU measurement sequence is obtained, and a corresponding measurement reconstruction sequence is generated based on the PMU measurement sequence using a trained GAN model.

The GAN model includes a generation model (G) and a discrimination model (D), and its network structure is shown in FIG4 .

Among them, the generative model (G) attempts to generate samples with the same probability distribution as the real data samples. The discriminative model (D) determines whether the input sample is a real sample based on the binary classifier. During the training process, the generative ability of the generative model (G) and the discriminative model (D) are improved. Finally, the discriminative model (D) does not distinguish between the generated data samples and the real data samples, and the game between the discriminative model (D) and the generative model (G) reaches a dynamic Nash equilibrium. The training process of GAN is to solve the following maximum and minimum binary game:

Where x represents the real data sample; z represents the random noise of the generative model; G(z) represents the generated sample considering the random noise z; D(x) and D(G(z)) represent the probabilities of identifying the real sample and the generated sample respectively; p _d (x) and p _z (z) are the probability distributions of the real sample and the random noise respectively; V(D,G) is the value function, which measures the difference between the probability distributions of the generated sample and the real sample. The parameters of D and G, denoted as θ _D and θ _G , are updated according to the gradient descent method, and their corresponding random gradients g _θD and g _θG are obtained according to (1-3) and (1-4) respectively.

It can be determined that the dynamic voltage distribution of the single node voltage PMU measurement sequence and its measurement reconstruction sequence enhances the potential of mining the spatiotemporal correlation of the local bus to identify bad data. As shown in Figure 5, after filling the data points by linear interpolation, the voltage amplitude curve is represented as the PMU measurement sequence _Vi . The voltage amplitude curve of the same period of the measurement reconstruction sequence is represented as _Vj , and the correlation coefficient with _Vi is 0.9408. Based on the difference analysis, the data spikes of _Vi at 1.44s, 2.26s, and 3.24s are likely to be outliers because a sudden change in voltage amplitude occurs on _Vi , while _Vj shows a smooth trend. The same is true for the data peak of _Vj at 3s. However, the data spikes of _Vi and _Vj at 0.76 seconds are preferably identified as event data, based on the condition that the two profiles experience similar sudden change trends. In reality, the comprehensive analysis of graphics and data trends is not standardized, resulting in manual misjudgment. Therefore, the qualitative analysis of bad data identification needs further research.

S20. Based on the PMU measurement sequence and the corresponding measurement reconstruction sequence, a PMU measurement sequence combination is constructed, and corresponding two-dimensional image data is generated.

In an embodiment of the present invention, the forming of a PMU measurement sequence combination based on the PMU measurement sequence and the corresponding measurement reconstruction sequence, and generating corresponding two-dimensional image data, includes:

The PMU measurement sequence of a single node and its measurement reconstruction sequence form coordinates, which are used as inputs of the bad data identification program, and two-dimensional image data is generated accordingly.

The PMU measurement sequence is _Xi , and the measurement reconstruction sequence is _Xj .

The PMU measurement sequence combination is _Xij , and _Xij is calculated by the following formula:

The two-dimensional coordinates are generated by combining the PMU measurement sequence _Xij , which is implemented by the following formula:

in, Each coordinate in the coordinates corresponds to a pair of PMU measurements of buses i and j at the same time period. In addition, Scatter plots of coordinates to show spatiotemporal distribution and correlation.

FIG6 is a scatter plot of an exemplary example, in which the middle data cluster framed by a dotted line is normal data, the data clusters on both sides are event data, and the scattered points outside the dotted line frame are outliers.

As shown in Figure 6, most of the data points are densely distributed on the diagonal due to the strong correlation of the two selected PMU measurement sequences. In order to tolerate the systematic errors of normal measurements and the changes in the system operating status, the "safe zone" for excluding bad data is designed as a tilted bar, which is visible in the area near the diagonal. Normal data and event data are expected to be located in the safe zone because the paired PMU measurements in the normal and event periods meet the strong correlation characteristics. Due to the severe deviation of the measurement data, the outliers with larger _Vi and smaller _Vj are expected to be sparsely located in the lower triangular area, while the outliers with smaller _Vi and larger _Vj are expected to be sparsely located in the upper triangular area.

The high resolution advantage of PMU leads to the clustered distribution of measurement data from the same type, which is fully utilized in the proposed method. In addition, it should be noted that if two PMU measurements encounter outliers at the same time, the corresponding data points may be located in the safe zone. However, since the occurrence of bad data on a single PMU is a low-probability event, the probability of the above situation is extremely small and is ignored in this section.

The clustering of data of the same type and the division of safe areas are not suitable for manual work in practice, so clustering and classifiers are used in this method, namely the basic bad PMU data identifier. The idea of bad data identification based on the two-dimensional spatiotemporal correlation analysis of the scatter plot measured by PMU realizes the potential of the currently used clusterers and classifiers, because their original application objects are two-dimensional images. Unlike traditional analysis, this method is effective in spatiotemporal analysis.

S30 , using hybrid clustering to process the two-dimensional image data to obtain preliminary identification results of each point in the PMU measurement sequence.

In an embodiment of the present invention, the two-dimensional image data is processed by hybrid clustering to obtain preliminary PMU measurement sequence point identification results, including:

The two-dimensional image data are processed in sequence using linear regression identifier, DBSCAN identifier and GMM identifier to obtain preliminary identification results of each point in the PMU measurement sequence.

The present invention utilizes the proposed graph-based method to identify the characteristics of normal and bad data and exerts the potential of dimensional clustering and classifiers. Therefore, the efficiency of spatiotemporal correlation analysis and the sensitivity of identification are improved.

The linear regression classifier (LR) is designed to minimize the residual sum of squares between the observed target y in the dataset and the target predicted by the linear approximation. In a two-dimensional model, the regression line is considered to be

Where y and x represent the approximate target vector and input vector, respectively, and a and b represent the slope and intercept of the regression line, respectively. To solve the classification problem, _xi and _xj are used as the target and input in the linear model, respectively, and are rewritten as

To measure the error between the approximated target and the observed target, the standard deviation σ is calculated as follows

The 3σ principle is applied to tolerate measurement errors and avoid misjudgment of normal data. Then, the safety area of Figure 6 can be modeled as:

The regression line and σ are calculated by the linear regression identifier within each preset time window of the PMU measurement. The purpose of the linear regression identifier is to classify diagonally distributed normal data containing strongly correlated event data. Then, data points that exceed the range constraints of the safe area are identified as outliers. Linear regression is a regression-based identifier, so the uncertainty of abnormal deviations leads to changes in the optimal linear division. Therefore, the safety zone divided by the 3σ principle is likely to contain some bad data with relatively small deviations and exclude very few normal data that deviate from the regression line under extreme system operating conditions. Misjudgment will reduce the accuracy of the linear regression identifier.

In the present invention, the DBSCAN identifier treats clusters as high-density areas separated by low-density areas, so the clusters found by DBSCAN can be of any shape. The DBSCAN identifier finds high-density core samples and expands clusters from them. Therefore, a cluster is a set of core samples and a set of non-core samples close to the core samples.

There are two parameters to define the data density in the DBSCAN algorithm, min_samples and eps. Formally, a core sample is defined as a sample for which there is a minimum number of samples in the dataset within a distance of min_samples, i.e., neighbors. A cluster is a set of core samples that can be constructed by recursively taking a core sample, finding all its core sample's neighbors, and so on. A cluster also has a set of non-core samples that are neighbors of the core sample but on the edge of the cluster. Therefore, non-core samples that are at least 1/4 of the distance from any core sample are considered outliers by the algorithm. While the parameter min_samples mainly controls the algorithm's tolerance to noise, the parameter eps is crucial for the appropriate choice of dataset and distance function.

The DBSCAN identifier is executed on the two-dimensional scattering data of the PMU measurement sequence combination _Xij , and the densely distributed data points are identified as normal data through clustering. Then, clusters with small data volume and data points with long distance are selected and classified as outliers.

The DBSCAN identifier is a density-based identifier, so the uncertainty of the system operating state leads to changes in the data distribution density. Both parameters of the DBSCAN identifier are set manually, which may lead to misjudgment when identifying bad PMU data. Therefore, the clusters gathered by the DBSCAN algorithm may exclude a small amount of sparsely distributed normal data, such as event data measured during transient processes where the system changes suddenly.

The Gaussian mixture model algorithm is a variant of the probabilistic k-Means method. The GMM-based identifier assumes that the data follows a Gaussian mixture distribution, in other words, the data can be considered to be generated from k Gaussian distributions. Each Gaussian distribution is called a component, and the linear superposition of these components forms the probability density function p(z) of the GMM:

Where z is a data sample of the data set D, ω _l represents the probability that the lth component produces x, g(z|μ _l ,σ _l ) is the probability density function of the lth component, where μ _l and σ _l are the center and covariance matrix of the lth Gaussian distribution.

The kth Gaussian distribution corresponds to k clusters, so the clustering based on GMM is basically the parameter estimation of ω _l , μ _l , σ _l , and the likelihood function is used as the evaluation function E in the GMM parameter estimation:

Where n _s is the sample size of the data set, and F is the logarithmic form of E for computational convenience. When E reaches a maximum value called maximum likelihood, a pair of GMM parameters is obtained, which determines the probability distribution with the maximum probability of generating a data sample in D. In order to maximize the evaluation function E, after initializing ω _l , μ _l , σ _l , an iteration-based method similar to k-Means is adopted as follows:

1) Estimate the probability that z _m belongs to cluster C _l , which is generated by the lth Gaussian distribution of z _m :

2) According to the conclusion in 1), that is, the lth Gaussian distribution generates p(z ₁ ∈ _{C l} )z ₁ , p(z ₂ ∈ _{C l} )z ₂ , … , p(z _n ∈ _{C l} )z _n , by estimating the lth Gaussian distribution parameters:

Convergence is considered to be reached when the calculated parameters remain constant across iterations.

In an embodiment of the present invention, the purpose of the GMM identifier is to identify the main clusters of normal data with a relatively loose error tolerance. In order to solve the bad data identification problem proposed by GMM, the total number of clusters is set to 2 to better divide the normal data and bad data into two clusters. It is believed that the bad data rate in the PMU measurement time series is relatively low. Therefore, if there is a significant gap in the data volume of the two clusters, the data points with a high data volume in the cluster are classified as normal data, while the data points with a low data volume in the cluster are classified as bad data. However, if the lower cluster data volume still accounts for a significant part of the total data volume, the data points in both clusters are classified as normal data, which means that the GMM identifier cannot classify the bad data class.

In the embodiment of the present invention, the GMM identifier identification method is based on the model, and the total number of clusters of the GMM identifier is also a hyperparameter that needs to be preset. Due to the relatively loose tolerance characteristics, the GMM identifier may dig out abnormal values that are difficult to identify by other identifiers. At the same time, the GMM identifier may also introduce some abnormal values in normal data clusters, thereby causing misjudgment.

S40, performing integrated learning identification and result correction on the preliminary identification results of each point in the PMU measurement sequence to obtain the final identification results of each point in the PMU measurement sequence.

In an embodiment of the present invention, integrated learning identification and result correction are performed on the preliminary identification results of each point in the PMU measurement sequence to obtain the final identification results of each point in the PMU measurement sequence, including:

Perform integrated learning and identification on the identification results of each point in the preliminary PMU measurement sequence to obtain the voting results;

The voting results are corrected to obtain the final identification results of each point in the PMU measurement sequence.

In an embodiment of the present invention, the ensemble learning identification is a collection method using majority voting as a basic identifier.

The present invention uses majority voting as an ensemble method of basic identifiers to verify normal data through consistent voting. Voting condition analysis is performed on statistical data to verify confirmable normal data. Then, the controversial identification results in majority voting are corrected based on the revised dividing boundary of normal and bad data distribution determined by previously verified normal data. Due to the graphical analysis, the proposed method is unsupervised; therefore, it is suitable for online bad PMU data identification. Simulation results show that the proposed method has superior performance compared to a single identifier and obtains accurate results under various conditions in a shorter computation time.

The present invention forms coordinates of the PMU measurement sequence and its measurement reconstruction sequence, and draws a two-dimensional scatter plot. The method based on two-dimensional images improves the efficiency of spatiotemporal correlation analysis of two PMU measurement sequences, and specifically distinguishes the characteristics of normal and bad data. Then, a variety of clustering/classifiers based on spatial models are used to perform mixed clustering on the two-dimensional scatter plot to divide normal and bad data, and the identification results are further corrected by ensemble learning. The present invention can effectively solve the problem of bad data identification of PMU measurement data in power grids and improve the accuracy of bad data identification.

Specifically, in some cases, each identifier still has limitations, so a single identifier cannot identify the complete type of bad PMU data. Therefore, it is of great significance to use various identifiers to make a comprehensive judgment on bad data. The hybrid bad data identification method is expected to classify the entire PMU measurement and improve the accuracy of bad data identification. This method is completely unsupervised and can be applied to online bad PMU data identification of two measurement time series at the same time.

In an embodiment of the present invention, in order to handle binary classification problems, the main idea of the voting classifier is to combine conceptually different classifiers and classify the class labels using majority vote (MV). Such a classifier is useful for a group of well-performing models in order to balance their respective weaknesses. In majority voting, the classification class label of a particular sample is the class label that represents the majority of class labels classified by each individual classifier.

For bad PMU data identification, the category of the data point is determined by the majority of identification results of the basic bad PMU data identifier based on majority voting. If a data point is identified as normal/bad data by all identifiers, the consistent result is identified as the final output label of the normal/bad data. However, if different results for a data point are obtained from the basic identifiers, this data point is likely to involve limited conditions of some identifiers. Further analysis using DBSCAN is required because simply using majority voting may introduce false positives. Please note that if the basic identifiers maintain consistent results for all PMU data points within the sampling time window, no further action will be taken to reduce the computational burden and duration.

In the embodiment of the present invention, a two-stage structure of bad PMU data identification result verification is constructed with majority voting as the first stage, wherein the second stage is intended to correct the results of controversial data points in majority voting. Based on graphical analysis and laboratory verification data, the statistical results of the complete majority voting conditions (sample size is 12,000) are listed in Table 1, where the occurrence time of each condition is listed in columns 4 and 5. Based on the statistical data, the possible results of the relevant conditions are listed in column 6. As shown in conditions 6 and 7, the final results may be different from the majority voting results and need to be further modified.

Table 1 Voting results

Note: N: normal data; B: bad data; S: rarely seen

According to the statistics in Table 1, the main problem of mitigating majority misclassification lies in the correct classification of conditions 5, 6, and 7. In addition, it is very likely that the data points that meet conditions 1, 2, 3, 4, 5, and 8 are correctly classified. Then, with the available information of a large amount of normal data, we intend to modify the safe zone determined by linear regression and enhance it to a "safe zone".

First, the normal data identified by conditions 1, 2, 3 and 4 in the previous process are selected. With the parameters relaxed, a database scan is performed to find several clusters that can completely contain the previously identified normal data. These data serve as the basis for modifying the edge of the safe area and describing the normal data layout in more detail. As shown in Figure 7, the previously identified normal data is aggregated by two clusters.

Second, the safety region around each cluster is updated according to the data distribution. For example, a line with a slope of a ₁ obtained by linear regression and _l1 determine the upper and lower limits of the safety zone around cluster 1. Line The analytical expressions for and l ₁ are determined by

where tol is an error tolerance coefficient with a relatively small positive value, and _C'1 is the original index set of _xi or the data point _xj in cluster 1. The same is true for 2l and 2l. l ₁ ， The boundaries with l ₂ are also determined by the normal data distribution of each cluster, measured by the vertical lines l' ₁ and l' ₂ with slope -1/a ₁

Third, the vertical lines l' ₁ , l' ₂ and line l ₁ ， l ₂ produces four intersection points, which are used as turning points for the upper and lower limits of the modified safety zone, as shown in Figure 7. Connect the intersection with l ₁₂ and complete the upper and lower limits. The analytical expression with l ₁₂ is determined by the coordinates of the four intersection points.

Finally, in order to correct the bad data identification results of controversial data points in majority voting, the second stage verification of bad data identification relies on the modified variable width safety region. The data points between the upper and lower limits (l ₁ -l ₁₂ -l ₂ ) are classified as normal data, while those points outside the range are classified as bad data. In addition, the data points fitting case 6 are identified as bad data by the linear regression and GMM identifiers, which means that these data points are far away from the regression line and do not belong to the main cluster of normal data. The misjudgment result of normal data from the database scanning identifier may be caused by the overlap of these bad data points that are closely located in the scatter plot, and are therefore identified as data clusters by the database scanning. Based on the above characteristics, the data points that meet case 6 should be classified as bad data, which is consistent with the result of the majority vote.

In the invented case study, an open access PMU dataset from the EPFL (Switzerland) campus was used to validate the proposed bad data identification method. The time window of each PMU measurement sample is 60 seconds and includes 3000 data points. The test platform is Python 3.7 based on Intel i7-9700@3.00GHz CPU and 32GB RAM.

Four typical conditions, including "normal data", "step event data and outliers", "spike event data and outliers" and "outliers", are exemplarily selected for case study using the method of the present invention. Figures 8-11 show two PMU measurement curves of voltage magnitudes V ₂ and V ₃ under typical conditions. In the form of a two-dimensional scatter plot, where the x-axis represents V ₂ and the y-axis represents V ₃ , Figures 12-15 show the original clustering and bad data identification results under typical conditions.

The original clustering results of each method are shown in the first row of the figure, and the calculation time of each step is listed in the lower right corner. In Figure 12-15{1,1} (1st row, 1st column), the scattered points represent the original samples of PMU data, and the broken lines represent the upper and lower limits calculated by linear regression. The broken lines in Figure 12-15{1,5} represent the upper and lower limits of the modified safety zone.

The second row of Figures 12-15 shows the bad data identification results and effects of each method. The circular points represent the correct identification results (normal data is identified as normal data, or bad data is identified as bad data), while the triangle points (normal data is identified as bad data) and the square points (bad data is identified as normal data) represent the wrong results. The identification accuracy of each method, as defined by , is also presented in the lower right corner.

Acc＝(n _all -n _fn -n _fb )/n _all ×100% (22)

Where n _all is the sample size of the data points in the test instance, n _fn is the number of wrongly identified normal data points, and n _fb is the number of wrongly identified bad data points. The index Acc (accuracy) reflects the proportion of correctly identified samples in the total number of samples.

(1) Normal data

The PMU measurement sequence and its reconstructed sequence are strongly correlated, with a correlation coefficient of 0.9669. According to Figure 8 and Figure 12{1,1}, the characteristics of the correlated normal data include similar contours and diagonally distributed two-dimensional scatter plots. The linear regression identifier misjudged some normal data with relatively large deviations, while the DBSCAN identifier misjudged some sparsely distributed normal data. Both clusters of the GMM identifier had a large amount of data and were merged into one cluster as normal data. However, the MV method corrected the identification errors of Figure 12{2,1} and {2,2}, and obtained 100% accuracy after the results were corrected.

(2) Step event data and outliers

According to Figure 9 and Figure 13{1,1}, the normal PMU measurement data points before and after the step event are still strongly correlated and divided into two clusters. In Figure 13{2,2}, the DBSCAN identifier misjudged some scattered normal data in the transient process, while the two clusters of GMM had a large amount of data and were merged into one cluster as normal data, so the outliers in Figure 13{2,3} were misjudged. In Figure 13{2,4}, the wrong judgment was corrected by MV. After the result correction, the accuracy of MV remained at 100%.

(3) Spike event data and outliers

According to Figure 10 and Figure 14{1,1}, the normal PMU measurement data points within the spike event are strongly correlated. In Figure 14{2,2}, the DBSCAN identifier misjudged some scattered normal data in the transient process, while in Figure 14{2,3}, a small amount of normal data in the transient process caused the GMM to misjudge. As shown in Figure 14{2,4}, MV corrected some identification errors, which was due to the fact that most identifiers misjudged some normal data. However, as shown in Figure 14{2,5}, the misjudged data points in MV were located in the corrected safe zone during the second stage correction process, and their identification results were corrected, so 100% accuracy was obtained after the result correction.

(4) Outliers

The presence of outliers in the PMU measurement data from spatially adjacent buses reduces the correlation coefficient to 0.6527. According to Figure 11 and Figure 15{1,1}, the outliers of the PMU measurement data points deviate significantly from the diagonal. The DBSCAN and GMM identifiers can correctly distinguish between normal and bad data. The linear regression identifier misjudged some relatively large deviations of normal data in Figure 15{2,1}, as shown in Figure 15{2,4}, which were corrected by MV. The corrected results maintained 100% accuracy.

Based on the above analysis, each single identifier has limitations in some cases, which leads to wrong judgments and low precision. The introduction of MV enhances the generalization ability of mixed clustering, thereby improving the accuracy of bad data identification. After introducing the result correction procedure, the proposed method overcomes the limitations of MV and compensates for the errors of MV in a cost calculation time of less than one second, thus obtaining the most accurate results.

A comprehensive study was conducted to verify the superiority and high performance of the proposed method in online bad data identification, and the results were evaluated by the index -. The index Fal (false positive rate) reflects the ratio of the number of wrongly identified normal data samples to the total number of samples, the index Mis (missing positive rate) reflects the ratio of the number of wrongly identified bad data samples to the total number of samples, and the index Pre (precision) reflects the ratio of the number of correctly identified bad data samples to the number of bad data samples. The performance test was conducted using six sets of one-hour 18,000 PMU data convections (36,000 PMU data) from the open PMU dataset, which have different bad data ratios and deviation ranges. The proposed method was performed on a moving time window of PMU data with a length of one minute (3,000 pairs, 6,000 data). It is worth noting that the bad data ratio determines the ratio of the number of bad data in a single PMU time series. Therefore, when it comes to two-dimensional analysis, the actual bad data ratio refers to the total number of bad data in two PMU time series divided by twice the length of a single time series.

Fal＝n _fn /n _all ×100% (23)

Mis＝n _fb /n _all ×100% (24)

Pre＝n _tb /(n _tb +n _fb )×100% (25)

Where: n _tn is the number of correctly identified normal data points, and n _tb is the number of correctly identified bad data points.

Table 2 Comprehensive results of bad data identification (EPFL data)

Note: The bold font method is based on the two-dimensional analysis proposed in this section, while the light font method is based on the one-dimensional analysis. The "Proposed" column is the test results corresponding to the method in this project.

Table 2 lists the numerical test results of the online bad data identification effect. Due to the inapplicability of linear regression and GMM in one-dimensional analysis, only the performance comparison between the two-dimensional and traditional one-dimensional DBCSCAN methods is shown. According to the improvement of bad data identification accuracy/precision and the reduction of bad data identification omission rate/error rate in Table 2, the two-dimensional DBSCAN method is better than the one-dimensional method because of the density-based features in the scatter plot. Therefore, the two-dimensional method improves the bad data identification performance by analyzing the spatiotemporal correlation of different measurements.

In addition, the performances of single model-based methods (LR/DBSCAN/GMM), ensemble-based methods (MV), and the proposed method are compared, all of which are based on two-dimensional analysis. Although the performance of linear regression identifiers and GMM identifiers is worse than that of DBSCAN identifier, their role in finding the wrong and missing identification points of DBSCAN identifier is verified by the performance of the proposed hybrid method. The ensemble-based MV method performs moderately but performs better than most single-base identifiers. In the tests on six datasets, the proposed hybrid method with a two-stage structure has the highest bad data identification accuracy, the lowest missing rate, and a low error rate. The excellent performance of the proposed method verifies the effectiveness of the proposed bad data identification result verification and correction process by reducing the wrong and missing identifications. The improvement in the identification accuracy of the proposed method means that the proposed method is able to identify certain situations that other methods cannot identify and can adapt to more complex bad data conditions. With the increase of the bad data ratio and the decrease of the maximum deviation range, the bad data identification accuracy of the proposed method also decreases due to the increase of the bad data identification difficulty. However, due to the high generalization ability and low sensitivity, the proposed method still has a stable performance, and the bad data identification accuracy is higher than 99.9% under the Pre indicator definition.

The average calculation time of the proposed method for online bad data identification of PMU time series in a single time window is 0.1612s, which is much smaller than the time window length and meets the requirements of online identification.

This project also tested the effectiveness of this method in identifying bad data based on actual measurement data of PMU in domestic distribution network. The data results are shown in Table 3.

Table 3 Comprehensive results of bad data identification (Lingang Demonstration Zone data)

The test results show that the method of the present invention can still identify bad data with a precision higher than 99.9% for the actual measurement data of PMU in the Lingang Demonstration Area under the definition of the Pre indicator.

It should be understood that, although described in a certain order, these steps are not necessarily performed in sequence in the above order. Unless there is clear explanation in this article, the execution of these steps does not have strict order restriction, and these steps can be performed in other orders. Moreover, a part of the steps of the present embodiment may include a plurality of steps or a plurality of stages, and these steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a part of the steps or stages in other steps or other steps.

In one embodiment, as shown in FIG. 3 , a smart distribution network PMU bad data identification system is also provided in an embodiment of the present invention, the system includes a measurement reconstruction sequence acquisition module 100 , a construction sequence combination module 200 , a preliminary identification module 300 and a final identification module 400 .

The measurement reconstruction sequence acquisition module 100 is used to acquire a PMU measurement sequence, and obtain a corresponding measurement reconstruction sequence based on the PMU measurement sequence.

The constructed sequence combination module 200 is used to construct a PMU measurement sequence combination based on the PMU measurement sequence and the corresponding measurement reconstruction sequence, and generate corresponding two-dimensional image data.

The preliminary identification module 300 is used to process the two-dimensional image data by using hybrid clustering to obtain preliminary identification results of each point in the PMU measurement sequence.

The final identification module 400 is used to perform integrated learning identification and result correction on the preliminary identification results of each point in the PMU measurement sequence to obtain the final identification results of each point in the PMU measurement sequence.

In one embodiment, referring to FIG. 5 , a device is also provided in an embodiment of the present invention, including a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus.

Memory, used to store computer programs;

The processor is used to execute the computer program stored in the memory to execute the method for identifying bad data of PMU in the smart distribution network. The processor implements the steps in the above method embodiment when executing the instructions.

The communication bus mentioned in the above terminal can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.

The communication interface is used for communication between the above terminal and other devices.

The memory may include a random access memory (RAM) or a non-volatile memory, such as at least one disk memory. Optionally, the memory may also be at least one storage device located away from the aforementioned processor.

The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

The device includes user equipment and network equipment. The user equipment includes but is not limited to computers, smart phones, PDAs, etc. The network equipment includes but is not limited to a single network server, a server group consisting of multiple network servers, or a cloud consisting of a large number of computers or network servers based on cloud computing (Cloud Computing), wherein cloud computing is a type of distributed computing, a super virtual computer consisting of a group of loosely coupled computer sets. The device can be operated alone to implement the present invention, or it can be connected to the network and implement the present invention through interactive operations with other devices in the network. The network where the device is located includes but is not limited to the Internet, wide area network, metropolitan area network, local area network, VPN network, etc.

It should be further understood that the term “and/or” used in the present specification and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

In one embodiment of the present invention, a storage medium is further provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above method embodiment are implemented.

A person skilled in the art can understand that all or part of the processes in the above-mentioned embodiment method can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the process of the embodiment of the above-mentioned method. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include at least one of non-volatile and volatile memory.

It should be understood that, as used herein, the singular form "a" or "an" is intended to include the plural form as well, unless the context clearly supports an exception. It should also be understood that, as used herein, "and/or" refers to any and all possible combinations of one or more of the items listed in association. The serial numbers of the embodiments disclosed in the above embodiments of the present invention are for description only and do not represent the advantages and disadvantages of the embodiments.

A person skilled in the art should understand that the discussion of any of the above embodiments is only exemplary and is not intended to imply that the scope of the disclosure of the embodiments of the present invention (including the claims) is limited to these examples; under the concept of the embodiments of the present invention, the technical features in the above embodiments or different embodiments can also be combined, and there are many other changes in different aspects of the embodiments of the present invention as above, which are not provided in detail for the sake of simplicity. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present invention should be included in the protection scope of the embodiments of the present invention.

Claims (10)

1. A method for identifying bad data of a PMU of an intelligent power distribution network is characterized by comprising the following steps:

acquiring a PMU measurement sequence, and acquiring a corresponding measurement reconstruction sequence based on the PMU measurement sequence;

forming a PMU measurement sequence combination based on the PMU measurement sequence and the corresponding measurement reconstruction sequence, and generating corresponding two-dimensional image data;

processing the two-dimensional image data by utilizing hybrid clustering to obtain a preliminary PMU measurement sequence point identification result;

and performing integrated learning identification and result correction on the identification results of each point of the initial PMU measurement sequence to obtain the final identification results of each point of the PMU measurement sequence.

2. The method for identifying PMU fault data of an intelligent power distribution network according to claim 1, wherein before obtaining the PMU measurement sequence and obtaining a corresponding measurement reconstruction sequence based on the PMU measurement sequence, the method further comprises the steps of:

acquiring a sample data set, and training a GAN model by using the sample data set to obtain a trained GAN model; wherein the sample data set includes a sample PMU measurement sequence combination and corresponding sample two-dimensional image data.

3. The method for identifying PMU fault data of an intelligent power distribution network according to claim 2, wherein said obtaining a PMU measurement sequence and obtaining a corresponding measurement reconstruction sequence based on said PMU measurement sequence comprises:

And acquiring a PMU measurement sequence, and generating a corresponding measurement reconstruction sequence by utilizing a trained GAN model based on the PMU measurement sequence.

4. The method for identifying PMU fault data of intelligent power distribution network according to claim 2, wherein said PMU measurement sequence is X _i The reconstructed sequence is measured as X _j， The PMU measurement sequence is combined into X _ij The X is _ij Calculated by the following formula:

5. the method for identifying PMU failure data of an intelligent power distribution network according to claim 1, wherein said processing said two-dimensional image data by using hybrid clustering to obtain an identification result of each point of a preliminary PMU measurement sequence comprises:

and sequentially processing the two-dimensional image data by using a linear regression identifier, a DBSCAN identifier and a Gaussian mixture model identifier to obtain the identification result of each point of the initial PMU measurement sequence.

6. The method for identifying PMU fault data of intelligent power distribution network according to claim 1, wherein performing integrated learning identification and result correction on each point identification result of the preliminary PMU measurement sequence to obtain a final PMU measurement sequence each point identification result comprises:

performing integrated learning identification on identification results of each point of the initial PMU measurement sequence to obtain a vote result;

And correcting the result of the ticket to obtain the final identification result of each point of the PMU measurement sequence.

7. The method for identifying PMU failure data of a smart distribution network according to claim 6, wherein said ensemble learning identification is a ensemble method using majority vote as a base identifier.

8. Intelligent power distribution network PMU bad data identification system, its characterized in that, this system includes: the system comprises a measurement reconstruction sequence acquisition module, a construction sequence combination module, a preliminary identification module and a final identification module;

the measurement reconstruction sequence acquisition module is used for acquiring a PMU measurement sequence and acquiring a corresponding measurement reconstruction sequence based on the PMU measurement sequence;

the construction sequence combination module is used for forming a PMU measurement sequence combination based on the PMU measurement sequence and the corresponding measurement reconstruction sequence and generating corresponding two-dimensional image data;

the primary identification module is used for processing the two-dimensional image data by utilizing hybrid clustering to obtain the identification result of each point of the primary PMU measurement sequence;

and the final identification module is used for carrying out integrated learning identification and result correction on the identification results of each point of the initial PMU measurement sequence to obtain the identification results of each point of the final PMU measurement sequence.

9. An apparatus comprising a memory storing a computer program and a processor implementing the steps of the smart distribution network PMU bad data identification method according to any one of claims 1-7 when the computer program is loaded and executed by the processor.

10. A storage medium storing a computer program which, when loaded and executed by a processor, implements the steps of the smart distribution network PMU bad data identification method according to any one of claims 1-7.