Household transformer data quality evaluation method and system based on data-model dual drive

Abstract

The invention provides a user transformer data quality evaluation method and system based on data-model dual drive, and relates to the technical field of power data anomaly detection. Comprising the following steps: dividing user change abnormal data into four classes, including missing data, sudden change data, noise data and user change relation abnormal data; identifying the missing data; by establishing a voltage waveform correlation and power constraint model, user-transformer relation abnormal data is detected; the abrupt change data and the noise data are detected through an improved Transform model; and determining the proportion of the four types of user transformer abnormal data, and performing quality evaluation on the user transformer data by adopting a network analytic hierarchy process. According to the method, data trend features and context features of a time sequence can be captured, and accurate detection of abnormal data is realized; an index evaluation system is constructed from three aspects of data accuracy, consistency and integrity, an index evaluation system network structure is established, and finally a comprehensive score of user transformer data quality is obtained.

Description

Technical Field

[0001] The present invention belongs to the technical field of power data anomaly detection, and in particular relates to a method and system for evaluating the quality of household transformer data based on data-model dual drive. Background Art

[0002] With the development of digital industrialization and industrial digitization, the data center, as a key strategy for digital transformation, has become an important measure for digital construction. As a data-driven core hub, the data center brings together massive data generated in various fields and links, realizes the sharing, integration and efficient use of information resources, and provides solid support for the digital transformation of enterprises.

[0003] At present, the data center has gathered data from various professional aspects of power grid production and operation, and has initially realized the collection and perception of distribution transformers and user-related data. However, in recent years, with the continuous development of urbanization, the residential load has continued to increase, and the number of new distribution transformers and feeders has also continued to increase. Many problems have been exposed in the management of distribution network data, such as some users defaulting on electricity bills, illegally connecting lines, and abnormal measurement data. These problems will cause the power information collection system to be unable to accurately determine which transformer the power meter belongs to, thereby affecting the completion of the line loss index, causing ambiguity in customer file information, and increasing the operating risks of power supply companies.

[0004] At present, the detection and management of abnormal data of household transformers mainly consider the use of interpolation processing or more traditional machine learning methods, such as using density clustering, BIRCH clustering and other algorithms to reduce feature dimensionality, and using decision trees, random forests, support vector machines and other methods to classify data. This type of algorithm is difficult to capture the long-term and short-term temporal and spatial geographical correlations of power data, which is large in scale, diverse in sources, complex in correlation, and prominent in heterogeneity. The accuracy of judging abnormal data is low, and the data analysis effect needs to be further improved.

[0005] At the same time, the lack of a targeted and comprehensive household change data quality assessment system makes it difficult to accurately perceive the reliability and availability of the data. Summary of the invention

[0006] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides a household change data quality assessment method and system based on data-model dual drive, which is classified according to data characteristics, and the abnormal data of household change relationship is identified by establishing voltage waveform correlation and power constraint model; the generator and discriminator of the generative adversarial network are introduced into the Transformer model, and the mutation data and noise data are judged according to the abnormal data scores calculated by the generator and the discriminator, which can better capture the data trend characteristics and the contextual characteristics of the time series, and realize the accurate detection of abnormal data.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0008] The first aspect of the present invention provides a method for evaluating the quality of household change data based on data-model dual drive.

[0009] The household change data quality assessment method based on data-model dual drive includes the following steps:

[0010] The abnormal household change data are divided into four categories, including missing data, mutation data, noise data and abnormal household change relationship data;

[0011] Identify missing data;

[0012] By establishing voltage waveform correlation and power constraint models, abnormal data of household transformer relationships can be detected;

[0013] Through the improved Transformer model, mutation data and noise data are detected;

[0014] The proportion of the above four types of abnormal household change data was determined, and the network hierarchy analysis method was used to evaluate the quality of household change data.

[0015] The second aspect of the present invention provides a user change data quality assessment system based on data-model dual drive.

[0016] The household change data quality assessment system based on data-model dual drive includes:

[0017] The data type classification module is configured to: classify the household change abnormal data into four categories, including missing data, mutation data, noise data and household change relationship abnormal data;

[0018] The missing data identification module is configured to: identify missing data;

[0019] The abnormal data detection module for household-to-substation relationship is configured to: detect abnormal data for household-to-substation relationship by establishing a voltage waveform correlation and power constraint model;

[0020] The mutation data and noise data detection module is configured to: detect mutation data and noise data through an improved Transformer model;

[0021] The quality assessment module is configured to: determine the proportion of the above four types of household change abnormal data, and use the network hierarchy analysis method to conduct quality assessment on the household change data.

[0022] The third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the user change data quality assessment method based on data-model dual drive as described in the first aspect of the present invention.

[0023] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the user change data quality assessment method based on data-model dual drive as described in the first aspect of the present invention are implemented.

[0024] One or more of the above technical solutions have the following beneficial effects:

[0025] The present invention provides a household change data quality assessment method and system based on data-model dual drive, innovatively divides household change abnormal data into four categories: missing data, household change relationship abnormal data, mutation data, and noise data, and classifies and distinguishes them according to the characteristics of different abnormal data types; the present invention proposes a comprehensive and accurate household change data quality assessment method, constructs an indicator evaluation system from three aspects of data accuracy, consistency, and integrity, and establishes an indicator evaluation system network structure, and finally obtains a comprehensive score for household change data quality.

[0026] The present invention proposes a voltage correlation and power correlation judgment method for the discrimination of abnormal data on household-to-household relationships. When the calculation result exceeds the threshold, the household-to-household relationship of the data is considered abnormal. For the discrimination of mutation point and noise point data, the Transformer model is combined with the GAN model, and the abnormal data scores obtained by the generator and the discriminator are used for judgment. The above method can better adapt to large-scale, diverse sources, complex correlations, and prominent heterogeneity of power data, better capture its long-term and short-term time and spatial and geographical correlations, and improve the accuracy of abnormal data judgment.

[0027] The generator and discriminator of Generative Adversarial Network (GAN) are introduced into the Transformer model. The real household change data is input into the Transformer encoder of the TGAN-AD model generator to learn the hidden representation of the normal time series; the hidden representation and hidden space of the normal time series are input into the Transformer decoder of the generator to generate fake household change data; the fake household change data and the real household change data are input into the Transformer encoder of the TGAN-AD model discriminator to obtain the hidden representation of the fake household change data; the hidden representation of the fake household change data and the real household change data are input into the Transformer decoder of the discriminator, and the abnormal data detection score is calculated based on the reconstruction loss of the generator and the discriminant loss of the discriminator to determine whether the real household change data is a mutation point or noise point data. Data processing based on the above model can better capture the data trend characteristics and the contextual characteristics of the time series, and finally judge the mutation data and noise data according to the abnormal data detection scores calculated by the generator and the discriminator, so as to achieve more accurate detection.

[0028] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0030] Figure 1 This is a flow chart of the method of the first embodiment.

[0031] Figure 2 Flowchart for TGAN-AD model training.

[0032] Figure 3 Schematic diagram of using the trained TGAN-AD model to identify outliers.

[0033] Figure 4 This is a diagram of the data quality assessment system.

[0034] Figure 5 This is the flow chart of network analysis method. DETAILED DESCRIPTION

[0035] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0036] It should be noted that the terms used herein are for describing specific embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.

[0037] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0038] Embodiment 1

[0039] As mentioned above, for the detection and management of abnormal household change data, the current main considerations are interpolation processing or more traditional machine learning methods, such as using density clustering, BIRCH clustering and other algorithms to reduce feature dimensions, and using decision trees, random forests, support vector machines and other methods to classify data.

[0040] The above algorithms are difficult to capture the long-term and short-term temporal and spatial correlations of large-scale, diverse-source, complex-correlated, and heterogeneous power data. The accuracy of abnormal data judgment is low, and the data analysis effect needs to be further improved. At the same time, there is a lack of a targeted and comprehensive household data quality assessment system, which makes it difficult to accurately perceive the reliability and availability of data.

[0041] Therefore, improving the governance level of abnormal data of household transformers and achieving accurate quality assessment of transformer and household related data are crucial to ensuring the quality of power data and improving the efficiency of power supply companies.

[0042] Based on this, this embodiment provides a method for evaluating the quality of household change data based on data-model dual drive.

[0043] like Figure 1 As shown, the household change data quality assessment method based on data-model dual drive is characterized by comprising the following steps:

[0044] The abnormal household change data are divided into four categories, including missing data, mutation data, noise data and abnormal household change relationship data;

[0045] Identify missing data;

[0046] By establishing voltage waveform correlation and power constraint models, abnormal data of household transformer relationships can be detected;

[0047] Through the improved Transformer model, mutation data and noise data are detected;

[0048] The proportion of the above four types of abnormal household change data was determined, and the network hierarchy analysis method was used to evaluate the quality of household change data.

[0049] In some embodiments, by establishing a voltage waveform correlation and power constraint model, abnormal data of household transformer relationship is detected, specifically including:

[0050] Obtain the voltage of the transformer and the voltage of each user connected to the transformer, calculate the Pearson correlation coefficient between the two, determine whether the Pearson correlation coefficient is greater than a set threshold, and then determine whether the relationship between the transformer and the user is correct;

[0051] Assuming that the sum of the power of the same-phase users is basically balanced with the active power of the head end, list the equations and solve them to find the active power coefficient a of each user. i ; If the user's active power coefficient converges to around 1, it means that the user belongs to the current area, otherwise it does not belong to the current area.

[0052] In some embodiments, mutation data and noise data are detected by using an improved Transformer model, specifically including:

[0053] Obtain real household transformer data of distribution network time series;

[0054] Introduce the generator and discriminator of the generative adversarial network into the Transformer model to build the TGAN-AD model;

[0055] The real household change data is input into the Transformer encoder of the TGAN-AD model generator to learn the hidden representation of the normal time series; the hidden representation and hidden space of the normal time series are input into the Transformer decoder of the generator to generate fake household change data;

[0056] The fake household change data and the real household change data are input into the Transformer encoder of the TGAN-AD model discriminator to obtain the hidden representation of the fake household change data; the hidden representation of the fake household change data and the real household change data are input into the Transformer decoder of the discriminator, and the abnormal data detection score is calculated based on the reconstruction loss of the generator and the discriminant loss of the discriminator to determine whether the real household change data is a mutation point or noise point data.

[0057] In some embodiments, the method further includes training the TGAN-AD model, including:

[0058] Divide the training data into subsequences using sliding windows;

[0059] The training data and latent space divided into subsequences are input into the TGAN-AD model. The TGAN-AD model continuously trains the generator and discriminator in the minimax game. The specific formula is:

[0060]

[0061] After multiple iterations, the fake samples generated by the TGAN-AD generator can deceive the discriminator, indicating that the training is complete.

[0062] In some embodiments, the abnormal data detection score is calculated based on the reconstruction loss of the generator and the discriminant loss of the discriminator, specifically:

[0063] ADS=αG loss +(1-α)D loss ;

[0064] Among them, α represents the weight of reconstruction loss, 1-α represents the weight of discrimination loss, ADS is the abnormal data detection score; G loss ,D loss They represent reconstruction loss and discrimination loss respectively; if ADS>τ, the data label is equal to 1, indicating that the data is abnormal, that is, mutation point data or noise data; if ADS≤τ, the data label is equal to 0, indicating that the data is normal; τ represents the threshold of abnormal data detection score;

[0065] in:

[0066] D loss =D Trans (X test );

[0067] X test represents the test time series; D Trans Indicates the discriminator trained with the Transformer model, according to X test Output discriminative loss;

[0068]

[0069] Z test Denotes the test set X test The random latent variable G is mapped; Trans Represents the generator trained with the Transformer model, whose input is X test and Z rest ; Represents X test The specific value of the i-th feature at the j-th time point in ; c is the feature dimension; t w Represents partition X test The length of the time window; i represents the i-th feature, and j represents the j-th time point.

[0070] In some embodiments, before calculating the reconstruction loss, it also includes finding the optimal Z test , including:

[0071] The TGAN-AD model generator uses the covariance as a reference to update the similarity between the generated sequence and the test sequence;

[0072] Then the gradient descent algorithm is used to find the optimal generation sequence.

[0073] In some embodiments, the proportion of the above four types of abnormal household change data is determined, and the quality of the household change data is evaluated using the network analysis method, which specifically includes:

[0074] Construct a data quality assessment system, including three first-level indicators and corresponding second-level indicators. The first-level indicators are data integrity, data consistency and data accuracy, and the corresponding second-level indicators are the proportion of missing data, the proportion of abnormal data with household change relationships, and the proportion of mutation point and noise point data;

[0075] Based on the data quality assessment system, establish the network structure of the indicator evaluation system;

[0076] The relative importance of each secondary indicator is calculated using the 1-9 scale method to form a judgment matrix;

[0077] Calculate the maximum eigenvalue of the judgment matrix and normalize the eigenvector;

[0078] The consistency test of the judgment matrix is ​​performed based on the maximum characteristic root of the judgment matrix;

[0079] The judgment matrix that has passed the consistency test is weighted to obtain the weight W of different secondary indicators under their corresponding primary indicators;

[0080] Assign points to each secondary indicator;

[0081] By multiplying the score of each secondary indicator by the corresponding weight and then summing them up, we can get a comprehensive score for the quality of household change data.

[0082] In general, the method of the embodiment of the present invention:

[0083] Firstly, the abnormal data of household change is classified into four categories: missing data, mutation data, noise data, and abnormal data of household change relationship, and then classified and identified;

[0084] The abnormal data of household transformer relationship is judged by establishing the voltage waveform correlation and power constraint model;

[0085] At the same time, since it is difficult to quantitatively analyze the difference between mutation data and noise data and normal data, an improved Transformer model is used to identify them. The structure of the Generative Adversarial Network (GAN) generator and discriminator is introduced into the Transformer model to better capture the data trend characteristics and the context characteristics of the time series. Finally, the mutation data and noise data are judged based on the abnormal data scores calculated by the generator and the discriminator.

[0086] After the identification of the four types of abnormal data, corresponding indicator calculation methods were formulated according to the proportion of the four types of abnormal data, and the expert indicator weight calculation was completed based on the network analysis method, and finally a comprehensive score of household change data quality considering multiple indicators was obtained.

[0087] The technical solution of the embodiment of the present invention will be explained in detail below with reference to the accompanying drawings:

[0088] 1. Classification and identification of abnormal data

[0089] Abnormal household change data exists in normal household change data in various forms. Their locations are basically random. According to the characteristics of different abnormal data, they are divided into the following four categories:

[0090] (1) Missing data: One or more missing data points suddenly appear on the voltage or power curve;

[0091] (2) Abnormal data on the relationship between the distribution transformer and the user: abnormal data on the relationship between the distribution transformer and the user caused by an error in the connection relationship;

[0092] (3) Sudden change point data: one or more upward or downward peaks suddenly appear on the voltage or power curve, which is quite different from the overall trend of the data curve;

[0093] (4) Noise point data: abnormal data caused by excessive noise during measurement or transmission. It usually has a jagged shape on the data curve and is easily generated in extreme weather and human operational errors.

[0094] Among the above abnormal data types, it is relatively easy to determine missing data, which can be achieved by using the screening and positioning function of the Excel spreadsheet. This embodiment focuses on the identification methods of the other three types of data.

[0095] 1.1 Identification of abnormal data of household change relationship based on waveform correlation analysis

[0096] Assume that there are n transformers and user nodes in each group, that is, one distribution transformer node is connected to n-1 user nodes, and the voltage time series data of n nodes are U1, U2, …, Un , where U i = {u i,1 ,u i,2 ,…,u i,N}, N represents the number of sampling time points, u i，1 ,u i，2 ,…,u i，N Represents the voltage sampling sequence of node i.

[0097] 1.1.1 Voltage correlation judgment

[0098] From the perspective of electrical correlation, under the same feeder, the voltages of distribution transformers and users have high similarity in value and waveform. Therefore, they can be judged by the waveform correlation index. The basic principle is to use the ratio of the covariance and standard deviation between the two types of data as the correlation coefficient, that is, the Pearson correlation coefficient. The calculation formula is as follows:

[0099]

[0100] In the above formula, u j，i represents the i-th sampling point of the j-th node; u k，i represents the i-th sampling point of the k-th node; and Represent the average values ​​of the voltage sequence respectively. Assuming that the reference phases of the voltages are the same, the correlation coefficient ρ j.k The value range is between 0 and 1. The closer it is to 1, the stronger the voltage correlation is, and the closer it is to 0, the worse the voltage correlation is. Generally, when the correlation coefficient is greater than or equal to 0.9, it means that the relationship between the distribution transformer and the user of the voltage data is correct, otherwise it means that the relationship between the user and the transformer is wrong.

[0101] 1.1.2 Power correlation judgment

[0102] Assuming that the sum of the power of the same-phase users is basically balanced with the active power of the head end, the formula is:

[0103]

[0104] In the above formula, P sum,t represents the total active power on the low-voltage side of the substation at time t, a i represents the active power coefficient of user i, P i,t represents the active power of user i at time t, e i It represents the power deviation of user i due to the measurement error of the smart meter and line loss.

[0105] Considering that the measurement error of the total meter in the substation is lower than that of the user's meter, it can be assumed that the measurement error of the total meter in the substation is approximately equal to 0 during calculation. Since equation (2) can be listed for different time periods, (2) is actually a set of equations. By solving this set of equations, the active power coefficient a of each user can be obtained. i If the active power coefficient of the user converges to 1, it means that the user belongs to the area. Through research on literature, the generally more reliable value is |a i -1|≤0.2, when it is greater than 0.2 in the current period, it means that the change of household relationship is wrong.

[0106] 1.2 Identification of mutation point and noise point data based on improved Transformer model

[0107] Because the difference between the size of noise point data and mutation point data and normal household data is usually difficult to quantitatively analyze. For example, for voltage data, when the line encounters a short circuit fault, its size will suddenly change, forming a mutation point. Although this belongs to normal household data, it is often difficult to accurately distinguish it from the mutation point formed by measurement error; similarly, for noise data, its typical feature is that the curve shape is mostly jagged and has obvious burrs, but normal load fluctuations can also cause this phenomenon. Therefore, a neural network black box model is used for discrimination. In order to better capture the trend characteristics of abnormal data points, the Transformer model is improved by introducing GAN to improve the accuracy of detection. Assume that there is a complete and accurate data set X for training and testing models.

[0108] 1.2.1 Model architecture and training process

[0109] Since the common GAN model and Transformer model structure have been mentioned in many literatures, this embodiment will not introduce them separately. The model architecture adopted in this embodiment first adopts the Transformer model as the generator and discriminator of GAN to capture the time correlation and other hidden behaviors in the multivariate time series.

[0110] like Figure 2As shown, the real time series X and the random hidden variable Z are both input into the generator G to generate a fake time series G(X,Z): X,Z→G(X,Z), and then the fake time series and the real time series are input into the discriminator (D). The discriminator distinguishes between real and fake time series data by learning parameters. The parameters of G and D are updated according to whether the output of D is true or false. G generates fake samples close to the existing normal sample X, and D can improve its discriminability. Through iterative training, the discriminative ability of D can accurately distinguish between normal sequences X and fake sequences G(X,Z), and then G can capture the hidden temporal correlations in the normal time series X and generate a fake time series G(X,Z) that can deceive D. The model training architecture is as follows Figure 2 shown.

[0111] After training, all parameters of G and D are fixed. Test time series X test is encoded into a latent space. The optimal Z is trained using gradient descent test , input into G to reconstruct the test sample: X test ,Z test →G(X test ,Z test ). The reconstruction loss Gaussian is given by X test With G(X test ,Z test ) is used to calculate the deviation of . At the same time, the test time series X test Input the trained D and calculate the discriminant loss D loss Finally, two losses (G loss ,D loss ) Calculate the anomaly score of the test data. Then determine whether the test data is in an abnormal state based on the anomaly score.

[0112] like Figure 2 The figure shows the training flow chart of the TGAN-AD model, where the dotted line means the feedback update parameter, which is the model training diagram; Figure 3 Schematic diagram of using the trained TGAN-AD model to identify outliers.

[0113] The generator and discriminator of the model are trained in an adversarial manner, and the core part is the encoder and decoder modules based on Transformer.

[0114] The specific training process of the generator is as follows: the dataset X used for training in the real sequence data X train Input into the Transformer encoder to learn the hidden representation of normal time series and help the generator generate data close to real-time time series data. trainThe hidden representation of is input to the Transformer decoder. To generate rich similar samples, the latent space Z is the input to the Transformer decoder. The fake samples are based on X train The hidden representation of Z is the same as that generated by Transformer. In short, the goal of the generator is to continuously train Transformer to generate time series data that is close to real-time sequence data, making it difficult for the discriminator to distinguish the generated data from the training data.

[0115] In the discriminator, samples from training data or generated data are input into the Transformer encoder to obtain the hidden representation of the sample; then the sample and its representation are sent to the decoder. The class distribution of the sample is output by the Transformer decoder. The goal of the discriminator is to accurately distinguish the generated fake data from the training data, so that the model can be well trained through the minimum maximum game.

[0116] 1.2.2 Abnormal Data Detection

[0117] The training data set is The test time series is Among them, A and B are the maximum lengths of timestamps, C is the feature dimension, and in order to better capture the hidden information before and after the time series, a sliding window is used to divide the multivariate time series into subsequences. train ={x1,x2,x3,…x a},a=(At w )+1,t w Indicates the length of the time window, X test ={x1,x2,x3,…,x b}, b=(Bt w )+1. Z is a random hidden space, and similarly, Z={z1,z2,z3,…z n}. X and Z are put into the improved Transformer, and the model continuously trains the generator and discriminator in the minimax game. The specific formula is shown below:

[0118]

[0119] In the above formula, V(G, D) represents the loss function of the model; D(G(X, Z)) represents the probability that the discriminator determines that the generated sample is true; D(x) represents the probability that the discriminator determines that the real sample is true; x represents the training data set; z represents the element in the random hidden space; p z (x, y) represents the noise distribution of the generator input; R x (x, z) represents the actual distribution of the data we want to generate.

[0120] After multiple iterations, the fake samples generated by the TGAN-AD generator can deceive the discriminator, indicating that the training is complete. Then the reconstruction loss and discrimination loss of the test sample are calculated, and the abnormal data score is obtained by using the training and recorded as ADS (Abnormal data score).

[0121] Among them, for the discriminant loss, the test time series is input into the trained discriminator. The discriminator has the ability to identify whether the input sample is abnormal data. The improved Transformer model directly outputs the discriminant loss. Intuitively, the discriminant loss represents the probability that the input sample is abnormal. The formula is expressed as formula (5).

[0122] D loss =D Trans (X test )(5)

[0123] For the reconstruction loss, we first search the latent space for the optimal Z for the test dataset. test , the Z test Can generate G Trans The most similar generated sequence in G Trans (X test ,Z test ). The improved Transformer model updates G based on the covariance Trans (X test ,Z test ), that is, the similarity between the generated sequence and the test sequence, and then the gradient descent algorithm is used to find the optimal sequence, as shown in formula (6)(7).

[0124]

[0125] in, represents the average value of x; Represents G Trans (X test ,Z test )’s average value.

[0126] In finding the optimal Z test After that, the reconstruction loss G can be calculated loss , as shown in formula (8).

[0127]

[0128] Since the generator learns the distribution pattern of normal data, the degree of dissimilarity between the two samples can be calculated by comparing the original sample and the reconstructed sample. In other words, the difference between abnormal data and normal data can be obtained.

[0129] The final anomaly detection is performed by Dloss and G loss The two losses are used to make a decision, and the abnormal data detection score ADS is calculated based on the two losses. The specific formula is (9).

[0130] ADS=αG loss +(1-α)D loss (9)

[0131] In the above formula, α represents the weight of reconstruction loss, and 1-α represents the weight of discrimination loss. The weight is mainly set based on experience. Finally, the ADS is marked X according to the obtained test If ADS>τ, the data label is equal to 1, indicating that the data is abnormal, that is, mutation point data or noise data; if ADS≤τ, the data label is equal to 0, indicating that the data is normal.

[0132] 2. Construction of household change data quality assessment system

[0133] After analyzing missing data points, abnormal data points of household change relationships, mutation data points and noise data points, a data quality evaluation system is constructed based on the proportion of various types of abnormal data in the total data. The evaluation system mainly includes three first-level indicators, namely data integrity, data consistency and data accuracy, which are further divided into specific second-level evaluation indicators. Among them, data integrity is mainly reflected by the proportion of missing data, data consistency is mainly reflected by the proportion of abnormal data of household change relationships, and data accuracy is mainly reflected by the proportion of mutation point and noise point data. The architecture of the evaluation system is as follows: Figure 2 As shown, what is special is that the first-level indicators and second-level indicators in the indicator system correspond one to one.

[0134] 2.1 Calculation of indicator weights based on analytic network hierarchy process

[0135] The Analytic Network Process (ANP) organically combines qualitative and quantitative methods, and can handle many practical problems that cannot be handled by traditional quantitative techniques. This embodiment uses the network analysis method to calculate the expert's indicator weights. The operation process is as follows: Figure 5 shown.

[0136] like Figure 4 The figure shows the established data quality assessment system. The network structure of ANP consists of the upper control layer and the lower network layer. The control layer is the overall data quality assessment indicator. The network layer is divided into three first-level indicators and their subordinate second-level indicators. Each indicator affects each other. According to the network structure relationship, an unweighted initial super matrix can be constructed. The relative importance of each second-level indicator is calculated using the 1-9 scaling method to form a judgment matrix A. The maximum eigenvalue root is calculated and the eigenvector is normalized, as shown in Table 1.

[0137]

[0138]

[0139] Table 1 Unweighted initial super matrix

[0140] In the above table, a1, a2, and a3 represent three secondary indicators. pq Indicates the importance of index p relative to index q, where a pq =1 / a qp , a pq The size of is determined in the same way as in the AHP method, and is determined subjectively by experts based on the scale's judgment criteria table.

[0141] Since the chaotic judgment matrix cannot be used as a tool for index quantification, it is necessary to perform consistency check on it. Formula (10) is introduced as the consistency index:

[0142]

[0143] In the above formula, λ max is the maximum eigenvalue of the judgment matrix, n is the order of the judgment matrix, and if the ratio CR of the consistency index CI to the average consistency index RI of the order satisfies formula (11), the consistency test is passed.

[0144]

[0145] After obtaining the judgment matrix, Super decision 3.0 software is used to calculate the weights and obtain the weights W of different secondary indicators under their respective primary indicators.

[0146] 2.2 Scoring table for different index values

[0147] Since the values ​​of the three secondary indicators of missing data percentage, abnormal data percentage of household change abnormal relationship, and mutation point and noise point data percentage are all between 0 and 1, and the scores obtained by the general evaluation system are all in percentage, it is necessary to assign points to the three indicator values ​​so that they become values ​​between 0 and 100, and the larger the value, the better the data quality. The reference scoring table compiled based on the literature review is shown in Table 2:

[0148]

[0149]

[0150] By multiplying the score of each indicator by the corresponding weight and then finding the sum, the final household change data quality score can be obtained.

[0151] The embodiment of the present invention proposes a household change data anomaly detection and quality assessment method based on model-data dual drive. The present invention first classifies the household change abnormal data into four categories: missing data, household change relationship abnormal data, mutation data, and noise data, and classifies and distinguishes them according to the data characteristics. The household change relationship abnormal data is identified by establishing a voltage waveform correlation and power constraint model; at the same time, since it is difficult to quantitatively analyze the difference between mutation data and noise data and normal data, they are identified by an improved Transformer model. The generator and discriminator of the Generative Adversarial Network (GAN) are introduced into the Transformer model to better capture the data trend characteristics and the context characteristics of the time series, and finally the mutation data and noise data are judged according to the abnormal data scores calculated by the generator and the discriminator. After the identification of the four types of abnormal data is completed by combining the establishment of a model with a neural network, the corresponding indicator calculation methods are respectively formulated according to the proportion of the four types of abnormal data, the expert indicator weights are determined based on the network hierarchy analysis method, and the influence of each indicator is determined by the weighted super matrix, and finally the comprehensive score of the household change data quality considering multiple indicators is obtained.

[0152] Embodiment 2

[0153] This embodiment discloses a household change data quality assessment system based on data-model dual drive.

[0154] The household change data quality assessment system based on data-model dual drive includes:

[0155] The data type classification module is configured to: classify the household change abnormal data into four categories, including missing data, mutation data, noise data and household change relationship abnormal data;

[0156] The missing data identification module is configured to: identify missing data;

[0157] The abnormal data detection module for household-to-substation relationship is configured to: detect abnormal data for household-to-substation relationship by establishing a voltage waveform correlation and power constraint model;

[0158] The mutation data and noise data detection module is configured to: detect mutation data and noise data through an improved Transformer model;

[0159] The quality assessment module is configured to: determine the proportion of the above four types of household change abnormal data, and use the network hierarchy analysis method to conduct quality assessment on the household change data.

[0160] In some embodiments, the household change relationship abnormal data detection module is specifically configured as follows:

[0161] For abnormal data of household transformer relationships, the voltage waveform correlation and power constraint model are established for identification, including:

[0162] Obtain the voltage of the transformer and the voltage of each user connected to the transformer, calculate the Pearson correlation coefficient between the two, determine whether the Pearson correlation coefficient is greater than a set threshold, and then determine whether the relationship between the transformer and the user is correct;

[0163] Assuming that the sum of the power of the same-phase users is basically balanced with the active power of the head end, list the equations and solve them to find the active power coefficient a of each user. i ; If the user's active power coefficient converges to around 1, it means that the user belongs to the current area, otherwise it does not belong to the current area.

[0164] In some embodiments, the mutation data and noise data detection module is specifically configured as follows:

[0165] Obtain real household transformer data of distribution network time series;

[0166] Introduce the generator and discriminator of the generative adversarial network into the Transformer model to build the TGAN-AD model;

[0167] The real household change data is input into the Transformer encoder of the TGAN-AD model generator to learn the hidden representation of the normal time series; the hidden representation and hidden space of the normal time series are input into the Transformer decoder of the generator to generate fake household change data;

[0168] The fake household change data and the real household change data are input into the Transformer encoder of the TGAN-AD model discriminator to obtain the hidden representation of the fake household change data; the hidden representation of the fake household change data and the real household change data are input into the Transformer decoder of the discriminator, and the abnormal data detection score is calculated based on the reconstruction loss of the generator and the discriminant loss of the discriminator to determine whether the real household change data is a mutation point or noise point data.

[0169] In some embodiments, the quality assessment module is specifically configured to:

[0170] After determining the proportion of abnormal data of the four types of household change, the network analysis method is used to conduct a comprehensive evaluation of the quality of household change data, including:

[0171] Construct a data quality assessment system, including three first-level indicators and corresponding second-level indicators. The first-level indicators are data integrity, data consistency and data accuracy, and the corresponding second-level indicators are the proportion of missing data, the proportion of abnormal data on household change relationships, and the proportion of mutation point and noise point data;

[0172] Based on the data quality assessment system, establish the network structure of the indicator evaluation system;

[0173] The relative importance of each secondary indicator is calculated using the 1-9 scale method to form a judgment matrix;

[0174] Calculate the maximum eigenvalue of the judgment matrix and normalize the eigenvector;

[0175] The consistency test of the judgment matrix is ​​performed based on the maximum characteristic root of the judgment matrix;

[0176] The judgment matrix that has passed the consistency test is weighted to obtain the weight W of different secondary indicators under their corresponding primary indicators;

[0177] Assign points to each secondary indicator;

[0178] By multiplying the score of each secondary indicator by the corresponding weight and then summing them up, we can get a comprehensive score for the quality of household change data.

[0179] More specifically, the above classification module is specifically used to perform:

[0180] 1. Classification and identification of abnormal data

[0181] Abnormal household change data exists in normal household change data in various forms. Their locations are basically random. According to the characteristics of different abnormal data, they are divided into the following four categories:

[0182] (1) Missing data: One or more missing data points suddenly appear on the voltage or power curve;

[0183] (2) Abnormal data on the relationship between the distribution transformer and the user: abnormal data on the relationship between the distribution transformer and the user caused by an error in the connection relationship;

[0184] (3) Sudden change point data: one or more upward or downward peaks suddenly appear on the voltage or power curve, which is quite different from the overall trend of the data curve;

[0185] (4) Noise point data: abnormal data caused by excessive noise during measurement or transmission. It usually has a jagged shape on the data curve and is easily generated in extreme weather and human operational errors.

[0186] Among the above abnormal data types, it is relatively easy to determine missing data, which can be achieved by using the screening and positioning function of the Excel spreadsheet. This embodiment focuses on the identification methods of the other three types of data.

[0187] The household change relationship abnormal data identification module is specifically configured as follows:

[0188] 1.1 Identification of abnormal data of household change relationship based on waveform correlation analysis

[0189] Assume that there are n transformers and user nodes in each group, that is, one distribution transformer node is connected to n-1 user nodes, and the voltage time series data of n nodes are U1, U2, …, U n , where U i = {u i,1 ,u i,2 ,…,u i,N}, N represents the number of sampling time points, u i，1 ,u i，2 ,…,u i，N Represents the voltage sampling sequence of node i.

[0190] 1.1.1 Voltage correlation judgment

[0191] From the perspective of electrical correlation, under the same feeder, the voltages of distribution transformers and users have high similarity in value and waveform. Therefore, they can be judged by the waveform correlation index. The basic principle is to use the ratio of the covariance and standard deviation between the two types of data as the correlation coefficient, that is, the Pearson correlation coefficient. The calculation formula is as follows:

[0192]

[0193] In the above formula, u j，i represents the i-th sampling point of the j-th node; u k，i represents the i-th sampling point of the k-th node; and represents the average value of the voltage sequence. Assuming that the reference phase of the voltage is the same, the correlation coefficient ρ j.k The value range is between 0 and 1. The closer it is to 1, the stronger the voltage correlation is, and the closer it is to 0, the worse the voltage correlation is. Generally, when the correlation coefficient is greater than or equal to 0.9, it means that the relationship between the distribution transformer and the user of the voltage data is correct, otherwise it means that the relationship between the user and the transformer is wrong.

[0194] 1.1.2 Power correlation judgment

[0195] Assuming that the sum of the power of the same-phase users is basically balanced with the active power of the head end, the formula is:

[0196]

[0197] In the above formula, P sum,t represents the total active power on the low-voltage side of the substation at time t, a i represents the active power coefficient of user i, P i,t represents the active power of user i at time t, e iIt represents the power deviation of user i due to the measurement error of the smart meter and line loss.

[0198] Considering that the measurement error of the total meter in the substation is lower than that of the user's meter, it can be assumed that the measurement error of the total meter in the substation is approximately equal to 0 during calculation. Since equation (2) can be listed for different time periods, (2) is actually a set of equations. By solving this set of equations, the active power coefficient a of each user can be obtained. i If the active power coefficient of the user converges to 1, it means that the user belongs to the area. Through research on literature, the generally more reliable value is |a i -1|≤0.2, when it is greater than 0.2 in the current period, it means that the change of household relationship is wrong.

[0199] The abnormal data identification module is specifically configured as follows:

[0200] 1.2 Identification of mutation point and noise point data based on improved Transformer model

[0201] Because the difference between the size of noise point data and mutation point data and normal household data is usually difficult to quantitatively analyze. For example, for voltage data, when the line encounters a short circuit fault, its size will suddenly change, forming a mutation point. Although this belongs to normal household data, it is often difficult to accurately distinguish it from the mutation point formed by measurement error; similarly, for noise data, its typical feature is that the curve shape is mostly jagged and has obvious burrs, but normal load fluctuations can also cause this phenomenon. Therefore, a neural network black box model is used for discrimination. In order to better capture the trend characteristics of abnormal data points, the Transformer model is improved by introducing GAN to improve the accuracy of detection. Assume that there is a complete and accurate data set X for training and testing models.

[0202] 1.2.1 Model architecture and training process

[0203] Since the common GAN model and Transformer model structure have been mentioned in many literatures, this embodiment will not introduce them separately. The model architecture adopted in this embodiment first adopts the Transformer model as the generator and discriminator of GAN to capture the time correlation and other hidden behaviors in the multivariate time series.

[0204] like Figure 2As shown, the real time series X and the random hidden variable Z are both input into the generator G to generate a fake time series G(X,Z): X,Z→G(X,Z), and then the fake time series and the real time series are input into the discriminator (D). The discriminator distinguishes between real and fake time series data by learning parameters. The parameters of G and D are updated according to the output of D. G generates fake samples close to the existing normal samples X, and D can improve its discriminability. Through iterative training, the discriminative ability of D can accurately distinguish between normal sequences X and fake sequences G(X,Z), and then G can capture the hidden temporal correlations in the normal time series X and generate a fake time series G(X,Z) that can deceive D. The model training architecture is as follows Figure 2 shown.

[0205] After training, all parameters of G and D are fixed. Test time series X test is encoded into a latent space. The optimal Z is trained using gradient descent test , input into G to reconstruct the test sample: X test ,Z test →G(X test ,Z test ). The reconstruction loss Gaussian is given by X test With G(X test ,Z test ) is used to calculate the deviation of . At the same time, the test time series X test Input the trained D and calculate the discriminant loss D loss Finally, two losses (G loss ,D loss ) calculates the anomaly score of the test data. Then, the anomaly score is used to determine whether the test data is in an abnormal state, as shown in the schematic diagram. Figure 2 shown.

[0206] The generator and discriminator of the model are trained in an adversarial manner, and the core part is the encoder and decoder modules based on Transformer.

[0207] The specific training process of the generator is as follows: the dataset X used for training in the real sequence data X train Input into the Transformer encoder to learn the hidden representation of normal time series and help the generator generate data close to real-time time series data. train The hidden representation of is input to the Transformer decoder. To generate rich similar samples, the latent space Z is the input to the Transformer decoder. The fake samples are based on X trainThe hidden representation of Z is the same as that generated by Transformer. In short, the goal of the generator is to continuously train Transformer to generate time series data that is close to real-time sequence data, making it difficult for the discriminator to distinguish the generated data from the training data.

[0208] In the discriminator D, samples from training data or generated data are input into the Transformer encoder to obtain the hidden representation of the sample; then the sample and its representation are sent to the decoder. The class distribution of the sample is output by the Transformer decoder. The goal of the discriminator is to accurately distinguish the generated fake data from the training data, so that the model can be well trained through the minimum maximum game.

[0209] 1.2.2 Abnormal Data Detection

[0210] The training data set is The test time series is Among them, A and B are the maximum lengths of timestamps, C is the feature dimension, and in order to better capture the hidden information before and after the time series, a sliding window is used to divide the multivariate time series into subsequences. train ={x1,x2,x3,…x a},a=(At w )+1,t w Indicates the length of the time window, X test ={x1,x2,x3,…,x b}, b=(Bt w )+1. Z is a random hidden space, and similarly, Z={z1,z2,z3,…z n}. X and Z are put into the improved Transformer, and the model continuously trains the generator and discriminator in the minimax game. The specific formula is shown below:

[0211]

[0212] After multiple iterations, the fake samples generated by the TGAN-AD generator can deceive the discriminator, indicating that the training is complete. Then the reconstruction loss and discrimination loss of the test sample are calculated, and the abnormal data score is obtained by using the training and recorded as ADS (Abnormal data score).

[0213] Among them, for the discriminant loss, the test time series is input into the trained discriminator. The discriminator has the ability to identify whether the input sample is abnormal data. The improved Transformer model directly outputs the discriminant loss. Intuitively, the discriminant loss represents the probability that the input sample is abnormal. The formula is expressed as formula (5).

[0214] D loss =D Trans (X test ) (5)

[0215] For the reconstruction loss, we first search the latent space for the optimal Z for the test dataset. test , the Z test Can generate G Trans The most similar generated sequence in G Trans (X test ,Z test ). The improved Transformer model updates G based on the covariance Trans (X test ,Z test ), that is, the similarity between the generated sequence and the test sequence, and then the gradient descent algorithm is used to find the optimal sequence, as shown in formula (6)(7).

[0216]

[0217] in, Represents the average value of the training data set x; Represents G Trans (X test ,Z test )’s average value.

[0218] In finding the optimal Z test After that, the reconstruction loss G can be calculated loss , as shown in formula (8).

[0219]

[0220] Since the generator learns the distribution pattern of normal data, the degree of dissimilarity between the two samples can be calculated by comparing the original sample and the reconstructed sample. In other words, the difference between abnormal data and normal data can be obtained.

[0221] The final anomaly detection is performed by D loss and G loss The two losses are used to make a decision, and the abnormal data detection score ADS is calculated based on the two losses. The specific formula is (9).

[0222] ADS=αG loss +(1-α)D loss (9)

[0223] In the above formula, α represents the weight of reconstruction loss, and 1-α represents the weight of discrimination loss. The weight is mainly set based on experience. Finally, the ADS is marked X according to the obtained testIf ADS>τ, the data label is equal to 1, indicating that the data is abnormal, that is, mutation point data or noise data; if ADS≤τ, the data label is equal to 0, indicating that the data is normal.

[0224] Embodiment 3

[0225] The purpose of this embodiment is to provide a computer-readable storage medium.

[0226] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the user change data quality assessment method based on data-model dual drive as described in Example 1 of the present disclosure.

[0227] Embodiment 4

[0228] The purpose of this embodiment is to provide an electronic device.

[0229] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps in the user change data quality assessment method based on data-model dual drive as described in Example 1 of the present disclosure are implemented.

[0230] The steps involved in the apparatuses of the above embodiments 2, 3 and 4 correspond to the method embodiment 1, and the specific implementation methods can refer to the relevant description part of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0231] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0232] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A household change data quality assessment method based on data-model dual drive, characterized in that: The following steps are involved: The abnormal household change data are divided into four categories, including missing data, mutation data, noise data and abnormal household change relationship data; Identify missing data; By establishing voltage waveform correlation and power constraint models, abnormal data of household transformer relationships can be detected; Through the improved Transformer model, mutation data and noise data are detected; The proportion of the above four types of abnormal household change data was determined, and the network hierarchy analysis method was used to evaluate the quality of household change data.

2. The household change data quality assessment method based on data-model dual drive according to claim 1 is characterized in that: By establishing the voltage waveform correlation and power constraint model, the abnormal data of household transformer relationship is detected, including: Obtain the voltage of the transformer and the voltage of each user connected to the transformer, calculate the Pearson correlation coefficient between the two, determine whether the Pearson correlation coefficient is greater than a set threshold, and then determine whether the relationship between the transformer and the user is correct; Assuming that the sum of the power of the same-phase users is basically balanced with the active power of the head end, list the equations and solve them to find the active power coefficient a of each user. i ; If the user's active power coefficient converges to around 1, it means that the user belongs to the current area, otherwise it does not belong to the current area.

3. The household change data quality assessment method based on data-model dual drive according to claim 1 is characterized in that: Through the improved Transformer model, mutation data and noise data are detected, including: Obtain real household transformer data of distribution network time series; Introduce the generator and discriminator of the generative adversarial network into the Transformer model to build the TGAN-AD model; The real household change data is input into the Transformer encoder of the TGAN-AD model generator to learn the hidden representation of the normal time series; the hidden representation and hidden space of the normal time series are input into the Transformer decoder of the generator to generate fake household change data; The fake household change data and the real household change data are input into the Transformer encoder of the TGAN-AD model discriminator to obtain the hidden representation of the fake household change data; the hidden representation of the fake household change data and the real household change data are input into the Transformer decoder of the discriminator, and the abnormal data detection score is calculated based on the reconstruction loss of the generator and the discriminant loss of the discriminator to determine whether the real household change data is a mutation point or noise point data.

4. The household change data quality assessment method based on data-model dual drive according to claim 3 is characterized in that: It also includes training the TGAN-AD model, including: Divide the training data into subsequences using sliding windows; The training data and latent space divided into subsequences are input into the TGAN-AD model. The TGAN-AD model continuously trains the generator and discriminator in the minimax game. The specific formula is: Where V(G, D) represents the loss function of the model; D(G(X, Z)) represents the probability that the discriminator determines that the generated sample is true; D(x) represents the probability that the discriminator determines that the real sample is true; x represents the training data set; z represents the element in the random hidden space; p z (x, z) represents the noise distribution of the generator input; p x (x, z) represents the actual distribution of the data that we want to generate; After multiple iterations, the fake samples generated by the TGAN-AD generator can deceive the discriminator, indicating that the training is complete.

5. The household change data quality assessment method based on data-model dual drive according to claim 2 is characterized in that: The abnormal data detection score is calculated based on the reconstruction loss of the generator and the discriminant loss of the discriminator, specifically: ADS=αG loss +(1-a)D loss ; Among them, α represents the weight of reconstruction loss, 1-α represents the weight of discrimination loss, ADS is the abnormal data detection score; G loss ,D loss They represent reconstruction loss and discrimination loss respectively; if ADS>τ, the data label is equal to 1, indicating that the data is abnormal, that is, mutation point data or noise data; if ADS≤τ, the data label is equal to 0, indicating that the data is normal; τ represents the threshold of abnormal data detection score; in: D loss =D Trans (X test ); X test represents the test time series; D Trans Indicates the discriminator trained with the Transformer model, according to X test Output discriminative loss; Z test Denotes the test set X test The random latent variable G is mapped; Trans Represents the generator trained with the Transformer model, whose input is X test and Z rest ; Represents X test The specific value of the i-th feature at the j-th time point in ; c is the feature dimension; t w Represents partition X test The length of the time window; i represents the i-th feature, and j represents the j-th time point.

6. The household change data quality assessment method based on data-model dual drive according to claim 5 is characterized in that: Before calculating the reconstruction loss, it also includes finding the optimal Z test , specifically including: Using the TGAN-AD model generator covariance as a reference, update the similarity between the generated sequence and the test sequence; Then the gradient descent algorithm is used to find the optimal generation sequence.

7. The household change data quality assessment method based on data-model dual drive according to claim 1 is characterized in that: Determine the proportion of the above four types of abnormal household change data, and use the network analysis method to evaluate the quality of household change data, including: Construct a data quality assessment system, including three first-level indicators and corresponding second-level indicators. The first-level indicators are data integrity, data consistency and data accuracy, and the corresponding second-level indicators are the proportion of missing data, the proportion of abnormal data with household change relationships, and the proportion of mutation point and noise point data; Based on the data quality assessment system, establish the network structure of the indicator evaluation system; The relative importance of each secondary indicator is calculated using the 1-9 scale method to form a judgment matrix; Calculate the maximum eigenvalue of the judgment matrix and normalize the eigenvector; The consistency test of the judgment matrix is ​​performed based on the maximum characteristic root of the judgment matrix; The judgment matrix that has passed the consistency test is weighted to obtain the weight W of different secondary indicators under their corresponding primary indicators; Assign points to each secondary indicator; By multiplying the score of each secondary indicator by the corresponding weight and then summing them up, we can get a comprehensive score for the quality of household change data.

8. The household change data quality assessment system based on data-model dual drive is characterized by: include: The data type classification module is configured to: classify the household change abnormal data into four categories, including missing data, mutation data, noise data and household change relationship abnormal data; The missing data identification module is configured to: identify missing data; The abnormal data detection module for household-to-substation relationship is configured to: detect abnormal data for household-to-substation relationship by establishing a voltage waveform correlation and power constraint model; The mutation data and noise data detection module is configured to: detect mutation data and noise data through an improved Transformer model; The quality assessment module is configured to: determine the proportion of the above four types of household change abnormal data, and use the network hierarchy analysis method to conduct quality assessment on the household change data.

9. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps in the user change data quality assessment method based on data-model dual drive as described in any one of claims 1-7 are implemented.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the user change data quality assessment method based on data-model dual drive as described in any one of claims 1-7 are implemented.

Images