Method and device for identifying abnormal line loss data when new energy sources are connected to the power grid

Abstract

This invention discloses a method and apparatus for identifying abnormal line loss data when new energy sources are connected to the power grid. The method includes: acquiring a raw statistical line loss data sequence; clustering the raw statistical line loss data sequence to obtain multiple line loss data clusters; calculating anomaly characteristic indicators for each line loss data cluster; inputting the anomaly characteristic indicators of the line loss data clusters into a preset model for analysis to obtain the location of abnormal data in the raw line loss data sequence corresponding to the anomaly characteristic indicators of the line loss data clusters. The preset model is trained using multiple sets of data, each set of data including: anomaly characteristic indicators of the sample statistical line loss data sequence and a label used to identify the location of abnormal data in the sample statistical line loss data sequence corresponding to the anomaly characteristic indicators of the sample statistical line loss data sequence. This method helps to locate abnormal data and improve the utilization rate of statistical line loss data.

Description

Technical Field

[0001] This invention relates to the field of eigenvalue calculation, and more specifically, to a method and apparatus for identifying abnormal line loss data when new energy sources are connected to the power grid. Background Technology

[0002] Statistical line loss data is a crucial foundation and basis for calculating line loss rates, which in turn are a key indicator reflecting the level of power grid production, operation, and management. In recent years, to further reduce line loss rates, the construction of integrated power and line loss management systems has been vigorously promoted. However, due to incomplete meter coverage, meter malfunctions, or data jumps, serious quality issues such as missing statistical line loss data and data mutations persist. Especially with the rapid development of new energy sources, the random fluctuation characteristics of these sources have profoundly changed the patterns of power grid line losses, posing new challenges to the identification of abnormal statistical line loss data.

[0003] Therefore, identifying statistical line loss anomalies in grid connections with a high proportion of renewable energy sources has become one of the urgent problems to be solved. Summary of the Invention

[0004] The main objective of this invention is to provide a method and apparatus for identifying abnormal line loss data when new energy sources are connected to the power grid, so as to solve the problem in the prior art that it is difficult to identify abnormal statistical line loss data when a high proportion of new energy sources are connected to the power grid.

[0005] To achieve the above objectives, according to one aspect of the present invention, a method for identifying abnormal line loss data of new energy grid connection is provided, comprising: acquiring an original statistical line loss data sequence; clustering the original statistical line loss data sequence to obtain multiple line loss data clusters; calculating an abnormal data feature index for each line loss data cluster; inputting the abnormal data feature index of the line loss data cluster into a preset model for analysis to obtain the position of abnormal data in the original line loss data sequence corresponding to the abnormal data feature index of the line loss data cluster, wherein the preset model is obtained by training multiple sets of data, and each set of data includes: an abnormal data feature index of the sample statistical line loss data sequence and a label for identifying the position of abnormal data in the sample statistical line loss data sequence corresponding to the abnormal data feature index of the sample statistical line loss data sequence.

[0006] Optionally, the original statistical line loss data sequence is clustered to obtain multiple line loss data clusters, including: using variational mode decomposition to granulate the multidimensional information of the original statistical line loss data sequence to obtain a multidimensional information granulated input data sequence; and using fuzzy C-means clustering algorithm to cluster the multidimensional information granulated input data sequence to obtain multiple line loss data clusters.

[0007] Optionally, the abnormal data characteristic index of each line loss data cluster in multiple line loss data clusters is calculated, including: determining the degree of abnormal data of the baseline component of the original statistical line loss data in each line loss data cluster; and obtaining the distribution stability factor characteristic index based on the degree of abnormal data of the baseline component, wherein the distribution stability factor characteristic index is used as the abnormal data characteristic index of the baseline component data in the original statistical line loss data.

[0008] Optionally, the abnormal data feature index of each line loss data cluster in multiple line loss data clusters is calculated, including: performing mode decomposition on each line loss data cluster to obtain a fluctuation component data sequence, the fluctuation component data sequence including multiple fluctuation component data points; performing autoregressive analysis on the fluctuation component data sequence using an ARIMA model to obtain multiple regression curves corresponding to the multiple fluctuation component data points; determining a first regression curve from the multiple regression curves, wherein the multiple regression curves other than the first regression curve are second regression curves, the difference between the first regression curve and the multiple second regression curves is greater than a preset threshold, and the difference between any two of the multiple second regression curves is less than or equal to the preset threshold; obtaining a local outlier feature index based on the difference between the first regression curve and the multiple second regression curves, wherein the local outlier feature index is used as the abnormal data feature index of the fluctuation component data sequence.

[0009] Optionally, the abnormal data feature index of each line loss data cluster in multiple line loss data clusters is calculated, including: performing mode decomposition on each line loss data cluster to obtain a periodic component data sequence, the periodic component data sequence including multiple periodic component data points; arranging the multiple periodic component data points into a periodic mode matrix; performing singular value decomposition on the periodic mode matrix to obtain singular values; determining the matrix low-rank degree of the periodic mode matrix based on the singular values; and obtaining the low-rank factor feature index according to the matrix low-rank degree, wherein the low-rank factor feature index serves as the abnormal data feature index of the periodic component data sequence.

[0010] Optionally, the abnormal data feature index of each line loss data cluster in multiple line loss data clusters is calculated, including: performing mode decomposition on each line loss data cluster to obtain a statistical line loss uncertain component data sequence; obtaining a theoretical line loss data sequence corresponding to the original statistical line loss data sequence; performing mode decomposition on the theoretical line loss data sequence to obtain a theoretical line loss uncertain component data sequence; calculating a difference matrix based on the statistical line loss uncertain component data sequence and the theoretical line loss uncertain component data sequence; ranking the difference matrix to obtain the positive rank sum and negative rank sum of the difference matrix; and obtaining the rank sum factor feature index based on the difference degree of the positive rank sum and negative rank sum, with the rank sum factor feature index serving as the abnormal data feature index of the statistical line loss uncertain component data sequence.

[0011] Optionally, the identification method further includes: constructing a deep convolutional neural network, which includes an input layer, multiple convolutional layers, multiple pooling layers, multiple fully connected layers, and an output layer connected in sequence; training the deep convolutional neural network to obtain a preset model, wherein the training includes forward propagation training and backward propagation training of parameters. Forward propagation training is used to input the abnormal data feature indicators of the sample statistical line loss data sequence into the deep convolutional neural network to obtain the actual position of the abnormal data in the sample statistical line loss data sequence. Backward propagation training is used to calculate the error between the actual position and the position of the abnormal data in the sample statistical line loss data sequence. The error is used to input into the deep convolutional neural network to update the parameters using the gradient descent method in both forward propagation training and backward propagation training.

[0012] According to another aspect of the present invention, a device for identifying abnormal line loss data of new energy grid access is provided, comprising: an acquisition module for acquiring an original statistical line loss data sequence; a clustering module for clustering the original statistical line loss data sequence to obtain multiple line loss data clusters; a calculation module for calculating an abnormal data feature index of each line loss data cluster; and an analysis module for inputting the abnormal data feature index of the line loss data cluster into a preset model for analysis to obtain the position of abnormal data in the original line loss data sequence corresponding to the abnormal data feature index of the line loss data cluster, wherein the preset model is obtained by training with multiple sets of data, and each set of data includes: an abnormal data feature index of the sample statistical line loss data sequence and a label for identifying the position of abnormal data in the sample statistical line loss data sequence corresponding to the abnormal data feature index of the sample statistical line loss data sequence.

[0013] According to another aspect of the present invention, an electronic device is provided, comprising: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to execute instructions to implement the identification method as described above.

[0014] According to another aspect of the present invention, a computer-readable storage medium is provided that, when the instructions in the computer-readable storage medium are executed by a processor of an electronic device, enables the electronic device to perform the identification method as described above.

[0015] The present invention provides a method for identifying abnormal line loss data when new energy is connected to the power grid. The method first obtains the original statistical line loss data sequence; then clusters the original statistical line loss data sequence to obtain multiple line loss data clusters; next, it calculates the abnormal data feature index of each line loss data cluster; and then inputs the abnormal data feature index of the line loss data cluster into a preset model for analysis to obtain the location of the abnormal data in the original line loss data sequence corresponding to the abnormal data feature index of the line loss data cluster. The preset model is trained using multiple sets of data, and each set of data includes: the abnormal data feature index of the sample statistical line loss data sequence and a label used to identify the location of the abnormal data in the sample statistical line loss data sequence corresponding to the abnormal data feature index of the sample statistical line loss data sequence. Identifying abnormal statistical line loss data from areas with a high proportion of renewable energy connected to the grid helps locate and repair these abnormal data, improving the utilization rate of statistical line loss data. Simultaneously, it eliminates interference from abnormal data, effectively addressing the problem of difficulty in determining the accuracy of statistical line loss data in high-loss areas caused by abnormal data. Furthermore, accurately locating high-loss areas helps in developing reasonable loss reduction strategies and optimizing the economic operation of the power grid. Attached Figure Description

[0016] The accompanying drawings, which form part of this specification, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0017] Figure 1 This is a flowchart illustrating a method for identifying abnormal line loss data when new energy is connected to the power grid, according to Embodiment 1 of the present invention.

[0018] Figure 2 This is a clustering flowchart shown in Embodiment 1 of the present invention;

[0019] Figure 3 This is a flowchart illustrating the feature extraction process for anomalous data of fluctuation components according to Embodiment 1 of the present invention;

[0020] Figure 4 This is a flowchart illustrating the feature extraction of periodic component anomaly data according to Embodiment 1 of the present invention;

[0021] Figure 5 This is a flowchart illustrating the feature extraction process for uncertain component anomaly data according to Embodiment 1 of the present invention;

[0022] Figure 6 This is a schematic diagram of the probability density distribution function of the β threshold when H0 is true, as shown in Embodiment 1 of the present invention;

[0023] Figure 7This is a schematic diagram of the probability density distribution function of the β threshold when H0 is false, as shown in Embodiment 1 of the present invention;

[0024] Figure 8 This is a schematic diagram of the probability density distribution function of the α∩β threshold of the positive unilaterality index according to Embodiment 1 of the present invention;

[0025] Figure 9 This is a schematic diagram of the probability density distribution function of the α∩β threshold of the negative unilaterality index according to Embodiment 1 of the present invention;

[0026] Figure 10 This is a flowchart illustrating the identification of statistical line loss anomaly data combinations according to Embodiment 1 of the present invention;

[0027] Figure 11 This is a flowchart illustrating the parameter training process based on a deep convolutional neural network according to Embodiment 1 of the present invention;

[0028] Figure 12 This is a flowchart of the identification method according to Embodiment 1 of the present invention;

[0029] Figure 13 This is a block diagram of the device for the data identification method according to Embodiment 2 of the present invention;

[0030] Figure 14 This is a device block diagram of a terminal according to Embodiment 3 of the present invention. Detailed Implementation

[0031] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0032] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0033] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of the invention described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0034] Example 1

[0035] According to an embodiment of the present invention, an embodiment of a method for identifying abnormal line loss data when new energy sources are connected to the power grid is provided. Figure 1 This is a flowchart of a method for identifying abnormal line loss data when new energy sources are connected to the power grid according to Embodiment 1 of the present invention, as follows: Figure 1 As shown, the method includes the following steps:

[0036] Step S102: Obtain the original statistical line loss data sequence;

[0037] Step S104: Cluster the original statistical line loss data sequence to obtain multiple line loss data clusters;

[0038] Step S106: Calculate the abnormal data characteristic indicators for each of the multiple line loss data clusters;

[0039] Step S108: Input the abnormal data feature index of the line loss data cluster into the preset model for analysis to obtain the position of abnormal data in the original line loss data sequence corresponding to the abnormal data feature index of the line loss data cluster. The preset model is trained by multiple sets of data. Each set of data includes: the abnormal data feature index of the sample statistical line loss data sequence and a label used to identify the position of abnormal data in the sample statistical line loss data sequence corresponding to the abnormal data feature index of the sample statistical line loss data sequence.

[0040] With the integration of a high proportion of renewable energy sources, the patterns of statistical line loss data change significantly, mainly in the following ways: 1. Wind and solar power exhibit strong volatility, with line loss showing a quadratic relationship with power output, and the fluctuations becoming more pronounced; 2. Solar power exhibits typical periodicity, significantly increasing the periodic component in line loss; the higher the proportion of solar power, the greater the periodic component in line loss; 3. Wind and solar power are inherently uncertain, increasing the uncertain component in line loss. All three factors alter the distribution patterns of statistical line loss, reducing data stability and consequently hindering the identification of anomalies in traditional statistical line loss data. The quality of statistical line loss data needs further improvement. Studying the impact of high-proportion renewable energy integration on the identification of anomalies in statistical line loss data is fundamental to achieving effective identification of anomalies in statistical line loss data for grids with high-proportion renewable energy integration.

[0041] The method described in this embodiment identifies abnormal statistical line loss data for areas with a high proportion of renewable energy connected to the grid, which helps to locate and repair abnormal data and improve the utilization rate of statistical line loss data. At the same time, it can eliminate the interference of abnormal data and effectively solve the problem that it is difficult to determine whether the statistical line loss in high-loss areas is real due to abnormal data. In addition, accurately locating high-loss areas helps to formulate reasonable loss reduction strategies and optimize the economic operation mode of the power grid.

[0042] In step S102 above, the original statistical line loss data sequence is obtained. The original statistical line loss data sequence to be identified includes statistical line loss data of high-proportion new energy sources such as wind power and photovoltaic power connected to the grid in different time periods.

[0043] In step S104 above, the obtained raw statistical line loss data sequence is clustered to obtain multiple line loss data clusters.

[0044] Because the output ratios of new energy sources such as wind power and solar power vary across different time periods, the composition of the original statistical line loss data also differs. Processing all time periods using the same standard could lead to significant identification errors. Therefore, this embodiment clusters the original statistical line loss data, grouping similar data segments into one category for classification and identification.

[0045] Considering the varying impacts of different wind and solar power integration ratios on statistical line loss patterns, one optional implementation involves clustering the original statistical line loss data sequence to obtain multiple line loss data clusters. This includes: using variational mode decomposition to granulate the multidimensional information of the original statistical line loss data sequence, resulting in a multidimensional granulated input data sequence; and using fuzzy C-means clustering to cluster the multidimensional granulated input data sequence, resulting in multiple line loss data clusters. The statistical line loss anomaly data identification model based on granulated fuzzy C-means clustering achieves higher accuracy in classification identification compared to the traditional unified identification model.

[0046] This embodiment employs an improved information-granular fuzzy C-means clustering method, which is essentially similar to the traditional FCM clustering method. The main difference is that the clustering input is changed from data points to information granules.

[0047] The aforementioned Fuzzy C-Means (FCM) clustering algorithm performs soft clustering by assigning each element a membership degree to each cluster, resulting in more flexible clustering results. Its principle is as follows:

[0048]

[0049]

[0050] Equation (1-1) is the value function, where C represents the number of clusters, N represents the number of samples, and m is the number of clusters; u ij Indicates sample x i The membership degree belonging to class j is the variable to be determined; c j It is the center of the j-cluster and is also the variable to be determined; ||*|| represents the Euclidean distance.

[0051] The number of clusters is determined using a partitioning coefficient index, and the calculation formula is as follows:

[0052]

[0053] The optimal number of clusters m is when the partitioning coefficient is minimized.

[0054] This embodiment uses variational mode decomposition to construct multidimensional data from one-dimensional line loss data, using this as input to better reflect the intrinsic characteristics of line loss. Specifically:

[0055] Let the one-dimensional statistical line loss data sequence be: L=(L1,L2,…,L n ), where n is the number of line loss data to be detected. Variational Mode Decomposition (VDM) is used to perform mode decomposition on the data sequence L, resulting in:

[0056] L = L u1 +L u2 +…+L uk (1-4)

[0057] Where u represents the membership degree and k represents different modes. Each data to be tested can be divided into membership degrees, and each membership degree also has a corresponding mode.

[0058] Then any statistical line loss data L i ,have They can all be uniquely characterized by the following data sequences:

[0059]

[0060] Then the original statistical line loss data can be uniquely characterized as Therefore, change the input to L. u Clustering can be performed.

[0061] Suppose the multidimensional statistical line loss data sequence is as follows:

[0062]

[0063] Let p represent the number of time windows after information granulation. When p = 1, it means the entire dataset is one information granule; when p = n, it means each data point is one information granule. The purpose of constructing information granules in this embodiment is mainly to prevent different data points in the same period from being divided into different clusters. Data points in the same period can be "bundled" together and clustered as a whole in the form of information granules. For example, each information granule of statistical line loss data contains 96 data points. The information granule input format is as follows:

[0064]

[0065] Specifically, such as Figure 2 As shown, the clustering process of the fuzzy C-means clustering algorithm used in this embodiment may include:

[0066] (1) Preprocessing of statistical line loss data:

[0067] Obtain the raw statistical line loss data sequence;

[0068] Based on the VMD method, the original one-dimensional data is decomposed into multiple intrinsic components, that is, each statistical line loss data corresponds to multiple component data, one-to-one. Based on this, the original statistical line loss sequence is transformed into a multi-dimensional data sequence according to Equation (1-5).

[0069] Using the period as the optimal time window (T=96), multidimensional information granulation is performed according to equation (1-7) to obtain the multidimensional information granulation input data sequence.

[0070] (2) Fuzzy C-means clustering:

[0071] Initialize the number of clusters, cluster centers, and membership degrees, i.e., the clustering parameters m and c mentioned above. j and u ij And the number of iterations l;

[0072] According to equations (1-1) and (1-2), J m A fuzzy C-means clustering model is established with the goal of minimizing the minimum value.

[0073] Solving the model for cluster centers and membership degrees yields the cluster center and membership degree parameters, which in turn updates the clustering parameter c.j and u ij ;

[0074] Calculate the partitioning coefficient PC according to equation (1-3), and determine whether the current number of clusters m is less than the iteration number l. If so, let m = m + 1, m < l, return to step (2) to continue the calculation and proceed to the next step; otherwise, end the iteration.

[0075] (3) Output the clustering result corresponding to the smallest PC as the final clustering result.

[0076] In step S108 above, the abnormal data feature index of each line loss data cluster in multiple line loss data clusters is calculated. Adaptive variational mode decomposition can be used to decompose the original line loss data sequence to obtain different mode line loss components, and then multidimensional information granular fuzzy C-means clustering method is used for clustering.

[0077] For each cluster, the baseline component characteristic index, fluctuation component characteristic index, periodic component characteristic index, and uncertain component characteristic index can be calculated separately.

[0078] To obtain the aforementioned baseline component characteristic index, in one optional implementation, the abnormal data characteristic index of each line loss data cluster in multiple line loss data clusters is calculated, including: determining the degree of abnormal data of the baseline component of the original statistical line loss data in each line loss data cluster; and obtaining the distribution stability factor characteristic index based on the degree of abnormal data of the baseline component, wherein the distribution stability factor characteristic index serves as the abnormal data characteristic index of the baseline component data in the original statistical line loss data.

[0079] For the baseline component data of statistical line loss, its patterns and properties are basically the same as those of traditional line loss, and its distribution function is also the same as that of traditional line loss, which is relatively stable and the distribution parameters do not change much. Therefore, the degree of abnormality of the baseline component data is characterized by the degree of change in the distribution parameters, and its expression is as follows:

[0080]

[0081] In the formula, μ i μ represents the mean parameter of the distribution function of sample i. j Let σ represent the mean parameter of the distributed function for sample j. The degree of variability is the difference between sample i and sample j, σ. i σ represents the standard deviation parameter of sample i. j The standard deviation parameter of sample j represents the degree of variation of the difference between sample i and sample j. The distribution stability factor is a characteristic index of the baseline component data; the larger the value of this index, the greater the possibility that the sample contains outlier data.

[0082] To obtain the aforementioned volatility component characteristic indicators, in one optional implementation, an anomaly data characteristic indicator for each of the multiple line loss data clusters is calculated, including: performing mode decomposition on each line loss data cluster to obtain a volatility component data sequence, the volatility component data sequence including multiple volatility component data points; performing autoregressive analysis on the volatility component data sequence using an ARIMA model to obtain multiple regression curves corresponding to the multiple volatility component data points; determining a first regression curve from the multiple regression curves, wherein the multiple regression curves other than the first regression curve are second regression curves, the difference between the first regression curve and the multiple second regression curves is greater than a preset threshold, and the difference between any two of the multiple second regression curves is less than or equal to the preset threshold; obtaining a local outlier characteristic indicator based on the difference between the first regression curve and the multiple second regression curves, wherein the local outlier characteristic indicator serves as the anomaly data characteristic indicator for the volatility component data sequence.

[0083] Each data point in the statistical line loss fluctuation component corresponds to a regression curve. If a data point is an outlier, its corresponding regression curve will differ significantly from the regression curves corresponding to other data points. Therefore, the degree of difference between the regression curves corresponding to a data point can be used to determine whether the data point is outlier. Based on this, a feature extraction model for outlier data in the statistical line loss fluctuation component is proposed as follows: Figure 3 As shown.

[0084] Specifically, the process for extracting the features of fluctuating component anomaly data may include:

[0085] (1) Obtain the original statistical line loss data sequence;

[0086] (2) Adaptive variational mode decomposition method is used to perform mode decomposition on the original statistical line loss data to obtain the fluctuation component data sequence;

[0087] (3) The ARIMA regression model is used to regress the fluctuation component data series, including:

[0088] Determining the order of the difference: The main purpose of the difference is to ensure the stationarity of the input ARMA model data. The successive difference calculation method is adopted. When the difference data is stationary, the order of the ARIMA model is determined to be the corresponding order.

[0089] Regression type determination: Determine the regression type based on the tailing and truncation properties of the data shown in the ARIMA model identification criteria in the table below;

[0090]

[0091]

[0092] Note: Tail indicates that the data sequence never tends to 0; truncation indicates that the data sequence rapidly tends to 0 after reaching a value greater than k.

[0093] Regression parameter estimation: The maximum likelihood method was used to estimate the ARIMA regression coefficients;

[0094] Model order determination and optimization: Based on the BIC criterion, optimize the p and q orders of the ARIMA regression model with the goal of minimizing the BIC value of the data.

[0095] Regression parameter update: After updating the p and q orders, the model parameters are estimated again.

[0096] (4) Eliminate the line loss data points one by one to form the N-1 fluctuation component data sequence, return to step (3) to perform regression, until all data have been traversed;

[0097] (5) Calculate the degree of change of each regression curve and use it as the abnormal data characteristic index of the fluctuation component data sequence, that is, the abnormal data characteristic of the statistical line loss fluctuation component.

[0098] Specifically, in step (5) above, the difference between each regression curve can be calculated by taking the average value of the local outlier factors at each point in the sample, which can be used as the abnormal data characteristic of the statistical line loss fluctuation component. The expression is as follows:

[0099]

[0100] In the formula, n is the number of samples of local outliers, k represents the data points, and LOF is the local outlier characteristic index. The larger the value of this index, the greater the possibility that the sample contains outliers.

[0101] Statistical line loss fluctuation component data has good ARIMA regression characteristics. When there are outliers in the statistical line loss, the data regression characteristics change significantly. Based on this, the concept of local outlier factor is introduced to characterize the degree of difference in the regression curve.

[0102] To obtain the aforementioned periodic component characteristic indicators, in one optional implementation, an anomaly data characteristic indicator for each of the multiple line loss data clusters is calculated, including: performing mode decomposition on each line loss data cluster to obtain a periodic component data sequence, the periodic component data sequence including multiple periodic component data points; arranging the multiple periodic component data points into a periodic mode matrix; performing singular value decomposition on the periodic mode matrix to obtain singular values; determining the matrix low-rank degree of the periodic mode matrix based on the singular values; and obtaining a low-rank factor characteristic indicator based on the matrix low-rank degree, wherein the low-rank factor characteristic indicator serves as the anomaly data characteristic indicator for the periodic component data sequence.

[0103] The low-rank properties of periodic component matrices composed of normal and abnormal data differ significantly. Based on this, a feature extraction model for periodic component abnormal data is proposed, such as... Figure 4 As shown.

[0104] Specifically, the feature extraction process for periodic component anomaly data may include:

[0105] (1) Obtain the original statistical line loss data sequence;

[0106] (2) Adaptive variational mode decomposition method is used to perform mode decomposition on the original statistical line loss data to obtain the statistical line loss periodic component data sequence;

[0107] (3) According to equation (1-9), the periodic component data are arranged into a two-dimensional periodic mode matrix, as follows:

[0108]

[0109] In the formula, L ij (i = 1, 2, ..., k; j = 1, 2, ..., T) represents the statistical line loss period component data for the i-th period and the j-th period. The line loss data is generally collected on a 15-minute time scale. Therefore, the period T = 96.

[0110] (4) Perform singular value decomposition on the periodic mode matrix, calculate the low rank of the matrix, and use it as an abnormal data feature index of the periodic component data sequence, that is, the abnormal data feature of the periodic component of statistical line loss.

[0111] Specifically, in step (4) above, since the statistical line loss periodic component data has typical low-rank characteristics, the low-rank characteristics disappear when there are abnormal data in the statistical line loss. Based on this, abnormal data features can be extracted based on the low-rank statistical line loss periodic component data. The extracted feature indicators are:

[0112]

[0113] In the formula, R l This is a low-rank factor characteristic index; the larger the index value, the lower the probability that the sample contains outliers; N(D) h ) is a set D h The capacity, D h s represents the set of singular values ​​that are much larger than the other singular values; s represents the singular values ​​of the line loss matrix, and r represents the rank sum.

[0114] To obtain the characteristic index of uncertain components, in one optional implementation, an anomaly data characteristic index is calculated for each of the multiple line loss data clusters. This includes: performing mode decomposition on each line loss data cluster to obtain a statistical line loss uncertain component data sequence; obtaining a theoretical line loss data sequence corresponding to the original statistical line loss data sequence; performing mode decomposition on the theoretical line loss data sequence to obtain a theoretical line loss uncertain component data sequence; calculating a difference matrix based on the statistical line loss uncertain component data sequence and the theoretical line loss uncertain component data sequence; ranking the difference matrix to obtain the positive rank sum and negative rank sum; and obtaining the rank sum factor characteristic index based on the difference between the positive rank sum and the negative rank sum. The rank sum factor characteristic index serves as the anomaly data characteristic index for the statistical line loss uncertain component data sequence.

[0115] Since uncertain component data are untraceable, it is difficult to identify anomalies based on the data itself. Therefore, comparative analysis with other data sequences is necessary to uncover patterns in the anomalies. Theoretical line loss and statistical line loss are two different methods of measuring line loss; in the absence of anomalies, they should be approximately equal. Based on this idea, this embodiment introduces the "rank sum" analysis theory, proposing the approximately equal "rank sum" characteristic of uncertain component data, and uses this as a basis for anomaly identification.

[0116] Abnormal data feature extraction patterns, such as Figure 5 As shown, the feature extraction process may include:

[0117] (1) Obtain the original statistical line loss data sequence and the corresponding theoretical line loss data sequence;

[0118] (2) Adaptive variational mode decomposition method is used to perform mode decomposition on the original statistical line loss data sequence and the theoretical line loss data sequence to obtain the uncertain component data sequence of statistical line loss and theoretical line loss;

[0119] (3) Compare the uncertain component data sequences of statistical line loss and theoretical line loss, and denote the theoretical line loss dataset as L. c The statistical line loss dataset is L s The capacity of each is n, that is:

[0120]

[0121] Remember L s With L c The average value of the elemental differences is E(L) s -L c ), let L s -L c -E(L s -L c The resulting set of differential sequences is:

[0122] ΔL={ΔL1,ΔL2,…,ΔL n} (1-12)

[0123] If the data is correct, the elements in ΔL should be uniformly and randomly distributed in the neighborhood of 0. Sort the elements in ΔL by their absolute values ​​from smallest to largest to obtain new numbers. Let ΔL be... i The corresponding new number is SN(ΔL) i If ), then the rank is:

[0124] R(ΔL i )=sgn(ΔL i )SN(ΔL i (1-13)

[0125] According to sgn(ΔL) i The sign of ) will determine the rank of each element, sgn(ΔL). i ) is divided into positive rank sets W + and negative rank set W - ,have:

[0126]

[0127] Among them, W + The sum of the elements is called a positive "rank sum" R. + W - The sum of the elements is called the negative "rank sum" R. - The calculation formula is:

[0128]

[0129] Calculate the difference matrix between the theoretical line loss and the statistical line loss uncertainty components according to equation (1-12);

[0130] (4) Rank the data according to Equation (1-13) and calculate the positive and negative "rank sums" of the difference matrix according to Equation (1-15). Calculate the degree of difference between the positive and negative "rank sums" and use them as the abnormal data characteristic index of the statistical line loss uncertain component data sequence, that is, the abnormal data characteristic of the statistical line loss uncertain component.

[0131] Specifically, in step (4) above, the statistical line loss uncertain component data has the property of approximately equal rank sums. When there are abnormal data in the statistical line loss, the property of approximately equal rank sums disappears. Therefore, abnormal data features can be extracted based on the statistical line loss uncertain component data with approximately equal rank sums. A positive rank sum R is selected. + The reciprocal of the distance from the median, D(R) + The extracted feature indexes characterize the degree of abnormality in uncertain components:

[0132]

[0133] In the formula, D(R) + The constant weight "rank sum" factor is a characteristic index. The higher the value of this index, the less likely the sample contains outlier data.

[0134] This embodiment proposes different anomaly data feature extraction modes based on the different characteristics of various statistical line loss components. For fluctuating component data, an ARIMA regression model is introduced, revealing significant differences in the regression characteristics of normal and anomaly data. Based on this, an anomaly data feature extraction mode for fluctuating component data in statistical line loss is proposed. For periodic component data, a low-rank matrix is ​​constructed according to the approximate period, and it is proven that normal periodic component data exhibits low-rank characteristics, which weaken under anomaly data. Based on this, an anomaly data feature extraction mode for periodic component data in statistical line loss is proposed. For uncertain component data, since it has no inherent pattern, theoretical line loss is introduced for comparative analysis to uncover the approximately equal "rank sum" characteristic of uncertain component data. Based on this, an anomaly data extraction mode for uncertain component data in statistical line loss is proposed. Furthermore, considering the different impacts of varying wind and solar power access ratios on the statistical line loss pattern, an anomaly data identification mode for statistical line loss based on information granular fuzzy C-means clustering is proposed. Compared to the traditional unified identification mode, this mode offers higher accuracy in classification and identification.

[0135] For baseline component data, this embodiment proposes a distribution stability factor as an indicator of outlier data characteristics; the larger the indicator, the greater the likelihood of data anomalies. For fluctuation component data, the concept of outlier factor is introduced, using the degree of outlierness of the regression parameter point set to characterize the degree of change in the regression characteristics of the fluctuation component. A local outlier factor indicator is proposed; the larger the indicator value, the greater the likelihood of data anomalies. For periodic component data, a low-rank factor indicator is proposed; the larger the indicator value, the smaller the likelihood of data anomalies. For uncertain component data, it is deduced that uncertain component data has an approximately equal "rank sum" characteristic, and based on this, a "rank sum" factor is proposed; the larger the indicator value, the smaller the likelihood of data anomalies.

[0136] In order to reduce both the false positive rate and the false negative rate to the ideal level and improve the accuracy of identifying abnormal line loss data in high-proportion renewable energy grid access statistics, this embodiment also proposes an improved α∩β threshold calculation method.

[0137] The basic principle of threshold calculation is that both the false positive rate (the probability of identifying normal data as abnormal data) and the false negative rate (the probability of identifying abnormal data as normal data) of anomaly identification are below the required level. In this embodiment, the improved α∩β threshold calculation method can be used to propose four types of indicators for four different statistical line loss components: distribution stability factor characteristic indicators, local outlier factor characteristic indicators, low-rank factor characteristic indicators, and rank-sum factor characteristic indicators. Among these, the distribution stability factor characteristic indicators and outlier factor characteristic indicators are positive unilateral indicators, meaning that when the indicator value is greater than a certain threshold, the data is abnormal; the low-rank factor characteristic indicators and the "rank-sum" factor characteristic indicators are negative unilateral indicators, meaning that when the indicator value is less than a certain threshold, the data is abnormal.

[0138] Specifically, the α threshold only considers the probability of a Type I error, that is, the probability α of misclassifying data as anomalous when there is no anomalous data in the data sequence. Once this misclassification probability is determined, the threshold x can be determined based on the probability density function of the feature index x. t .

[0139] We establish two hypotheses: H0: x ≤ x t H1: x > x t .

[0140] Suppose that when H0 is true, the probability density function of x is: When there are no outliers in the sample, the probability density distribution function of x is: The probability density function is calculated based on historical data, and its schematic diagram is shown below. Figure 6 As shown.

[0141] For the positive unilaterality index, the misjudgment probability α is:

[0142]

[0143] Therefore, the threshold α x can be obtained. tα for:

[0144] Where c represents the upper limit of the threshold determined by the probability density function of the feature index x;

[0145] For the negative unilaterality index, the misjudgment probability α is:

[0146]

[0147] Therefore, the threshold α x can be obtained. tα for:

[0148]

[0149] Since the above α threshold only considers the probability of misjudgment, there may be cases where abnormal data cannot be identified. Therefore, a second type of error probability - the probability of missed judgment β can be further introduced.

[0150] When H0 is false, the probability density function of x is: That is, when H1 is true, the probability density distribution function of x when there is outlier data in the sample is: Its probability density distribution function is as follows Figure 7 As shown.

[0151] For the positive unilaterality index, the probability of missed detection β is:

[0152]

[0153] Therefore, the β threshold x can be obtained. tβ for:

[0154]

[0155] For a negative unilateral index, the probability of missed detection β is:

[0156]

[0157] Therefore, the β threshold x can be obtained. tβ for:

[0158]

[0159] The α∩β threshold is determined by comprehensively considering both the α threshold and the β threshold, as follows: Figure 8 and Figure 9 As shown.

[0160] Depend on Figure 8 and Figure 9 As can be seen, whether it is a positive or negative unilateral index, there are three possible relationships between the α threshold and the β threshold. The results of α∩β threshold are different in different cases, which will be explained below.

[0161] Positive unilaterality index:

[0162] x tα <x tβ In this case, by Figure 8 (a) It can be seen that when the α∩β threshold is large, approaching the α threshold will reduce the false negative probability β; when the α∩β threshold is small, approaching the β threshold will also reduce the false positive probability α. As long as the α∩β threshold is between the α threshold and the β threshold, the requirement is met. Therefore, the α∩β threshold x t for:

[0163] x t ∈[x tα ,x tβ(1-17)

[0164] x tα >x tβ In this case, according to Figure 8 (b) It can be seen that when the α∩β threshold is large, biased towards the α threshold, the false positive rate of abnormal data decreases, but the false negative rate increases; when the α∩β threshold is small, biased towards the β threshold, the false negative rate of abnormal data decreases, but the false positive rate increases. If the α∩β threshold is less than the α threshold or greater than the β threshold, it will cause a significant increase in the false negative rate or the false positive rate, which is seriously unacceptable. Therefore, no matter what value the α∩β threshold takes, it is difficult to simultaneously meet the requirements of false negative rate and false negative rate. At this time, the α∩β threshold needs to be compromised, and the weighted average of the α threshold and the β threshold should be selected:

[0165]

[0166] When the actual requirement is to retain normal data as much as possible, that is, to have a smaller false positive probability α, the α∩β threshold in equation (1-18) is more biased towards the α threshold; conversely, if the actual requirement is to remove abnormal data as much as possible, that is, to have a smaller false negative probability β, the α∩β threshold in equation (1-18) is more biased towards the β threshold.

[0167] x tα =x tβ In this case, both the α threshold and the β threshold exactly meet the requirements, and can be set to the α threshold (or the β threshold, which are equal). If the α∩β threshold is less than this value, the false negative rate will be unacceptable; if the α∩β threshold is greater than this value, the false positive rate will be unacceptable. Therefore, the α∩β threshold x t for:

[0168] x t =x tα =x tβ (1-19)

[0169] (2) Negative unilaterality index:

[0170] x tα <x tβ In this case, by Figure 9 (a) It can be seen that, exactly the opposite of the positive unilaterality index, it is difficult to simultaneously meet the requirements of false positive rate and false negative rate regardless of the value of the α∩β threshold. In this case, the α∩β threshold needs to be compromised, and the weighted average of the α threshold and the β threshold should be selected.

[0171] x tα >x tβ In this case, by Figure 9 (b) It can be seen that as long as the α∩β threshold is between the α threshold and the β threshold, the requirement is met. Therefore, the α∩β threshold is any value between the two.

[0172] xtα =x tβ In this case, similar to the positive unilaterality index, either the α threshold or the β threshold can be used.

[0173] As the above analysis shows, the most crucial step in calculating the α∩β threshold is determining the two distribution functions: and Once these two distribution functions are determined, the α∩β threshold can be calculated according to equations (1-17) to (1-19). For the same region, since the essential characteristics of data anomalies caused by factors such as metering and communication are consistent, it can be considered that the data anomalies within the same region are... and It is definite, and therefore the α∩β threshold of the same region is also definite and will not change with the change of the sample to be identified.

[0174] The improved method for calculating the α∩β threshold for abnormal data described above can reduce both the false positive rate and the false negative rate of abnormal data to an ideal level, providing technical support for step S108.

[0175] In step S108, statistical line loss anomaly data is identified based on a deep convolutional neural network of full sample logic states. By modifying the input matrix to contain only anomaly data information, interference from irrelevant information can be avoided, thereby improving the comprehensive identification accuracy of statistical line loss anomaly data.

[0176] Specifically, the statistical line loss anomaly data combination identification process provided in this embodiment Figure 10 As shown, it includes:

[0177] Obtain the raw statistical line loss data sequence to be identified;

[0178] Adaptive variational mode decomposition is used to decompose the original line loss data sequence to obtain data sequences of different components of statistical line loss;

[0179] Since the different access ratios of wind and solar power may lead to different thresholds for abnormal data features, a multidimensional information granular fuzzy C-means clustering method is used to cluster the data, resulting in clusters 1 to 4: low line loss, low wind and low solar power, low wind and high solar power, high wind and low solar power, and high wind and high solar power.

[0180] For each cluster, the baseline component feature index, fluctuation component feature index, periodic component feature index and uncertain component feature index are calculated respectively, and the corresponding learners 1 to 4 are trained to identify abnormal data.

[0181] The final result is the identification of statistical line loss anomalies, that is, the location of the anomalies in the original line loss data sequence.

[0182] In an optional implementation, the identification method in this embodiment further includes: constructing a deep convolutional neural network, the deep convolutional neural network including an input layer, multiple convolutional layers, multiple pooling layers, multiple fully connected layers, and an output layer connected in sequence; training the deep convolutional neural network to obtain a preset model, wherein the training includes forward propagation training and backward propagation training of parameters, the forward propagation training is used to input the abnormal data feature indicators of the sample statistical line loss data sequence into the deep convolutional neural network to obtain the actual position of the abnormal data in the sample statistical line loss data sequence, the error backward propagation training is used to calculate the error between the actual position and the position of the abnormal data in the sample statistical line loss data sequence, the error is used to input into the deep convolutional neural network to update the parameters using the gradient descent method in the forward propagation training and backward propagation training.

[0183] Specifically, the process for identifying statistical line loss anomalies based on deep convolutional neural networks may include:

[0184] (1) Data preprocessing:

[0185] After obtaining a large number of historical data samples, the samples first need to be preprocessed, including:

[0186] Mean removal: Data centering of the sample data forms a dataset centered at the origin, which facilitates subsequent network learning;

[0187] Normalization: To remove errors caused by differences in the order of magnitude of different component data, sample data are normalized to the range of 0 to 1;

[0188] PCA / Whitening: Primarily used to eliminate the influence of statistical line loss residual component data (Gaussian white noise).

[0189] (2) Model training:

[0190] The training parameters mainly include: the weight vectors and bias values ​​of the fully connected layer and pooling layer, and the convolution kernel and bias value of the convolutional layer. The basic idea of ​​training is to update the parameters according to the negative gradient direction of the corresponding parameters according to the loss function until the accuracy requirements are met. The training process is explained in detail below.

[0191] (3) Parameter initialization:

[0192] The purpose of parameter initialization is to provide good initial values ​​for training deep convolutional neural networks. Currently, the most widely used method is random initialization, which means that the initial values ​​of the training parameters satisfy a certain probability distribution, such as uniform distribution or Gaussian distribution.

[0193] However, the parameters of the aforementioned distribution function are difficult to determine. Inappropriate distribution parameters may lead to excessive deviation in the initial values ​​of the neural network, thereby causing gradient vanishing. Research has shown that maintaining consistent input and output variance for each layer is a crucial means of preventing gradient vanishing. Based on this, this embodiment derives suitable initialization parameters according to the characteristics of the proposed Tanh-ReLU function. The specific process is as follows:

[0194] The variance of the (l+1)th layer of a deep convolutional neural network is:

[0195]

[0196] In the formula, n (l) w represents the number of neurons. (l) x (l) These are the kernel parameters and inputs for the l-th layer, respectively.

[0197] By ensuring that the variances of the input and output of this layer are consistent, we can obtain:

[0198]

[0199] The cumulative variance after k layers of forward propagation in the neural network is:

[0200]

[0201] The partial derivative of the error after backpropagation at layer k is:

[0202]

[0203]

[0204]

[0205]

[0206] Taking the harmonic mean of the two as the final variance, we get:

[0207]

[0208]

[0209] Using a uniform distribution for initialization, let the initialization range be [-a, a], then the variance is:

[0210]

[0211]

[0212] Therefore, the initial parameters are to follow a uniform distribution within the following interval:

[0213]

[0214] Each training parameter can be obtained by random sampling according to equation (1-20).

[0215] (4) Parameter training:

[0216] The training process of a convolutional neural network includes multiple forward and backward propagation processes. The forward propagation process calculates the parameters from the input layer to the output layer according to the previously updated parameters. The backward propagation process updates the parameters using gradient descent based on the error between the calculated results and the actual results, until the training is completed.

[0217] A deep convolutional neural network typically consists of an input layer, multiple convolutional and pooling layers, a fully connected layer, and an output layer.

[0218] The parameters to be trained mainly include: the weight vectors and bias values ​​of the fully connected layer and pooling layer, the convolution kernel and additive bias value of the convolutional layer. The basic idea of ​​training is to update the parameters according to the negative gradient direction of the corresponding parameters in accordance with the loss function until the accuracy requirements are met.

[0219] The input layer consists of multiple matrices, corresponding to the statistical line loss component matrices. Each convolutional layer contains multiple convolutional kernels, which perform convolution calculations with the input matrices to obtain multiple feature maps for that layer. Then, pooling layers are used for dimensionality reduction to extract the features most likely reflecting abnormal data. The deep convolutional neural network contains multiple convolutional and pooling layers. The result of the first pooling layer can be input to the next convolutional layer for further feature extraction. The abnormal features from the last pooling layer are used as the basis for the output layer to output the identification result.

[0220] The input layer is the image's basic color matrix, which in this embodiment corresponds to the matrix of each line loss component.

[0221] For the forward propagation process during the training of a convolutional neural network:

[0222] A convolutional layer mainly consists of two parts. The first part convolves the convolution kernel with the feature matrix output from the previous layer (or the input matrix if it's the first layer) to obtain the linear features of that layer. The second part adds a bias to the convolution output and then performs a non-linear transformation through an activation function to obtain the non-linear feature map of that layer. The calculation formula is as follows:

[0223]

[0224] In the formula, M represents the feature map; C represents the convolution kernel; b represents the additive bias value, both of which are parameters to be trained; the superscript L indicates the layer number in the neural network; and the subscript i indicates the feature map number. Generally, each feature map corresponds to one convolution kernel; N L-1is the set of feature maps; * is the convolution operator; f(x) is the activation function, which transforms the linear result of convolution into a non-linear result; j represents the j-th layer of the neural network.

[0225] The main function of pooling layers is to reduce the dimensionality of the convolution output and further extract the main features.

[0226] Currently, there are two main pooling methods: average pooling and max pooling. The calculation formula for max pooling is as follows:

[0227]

[0228] In the formula, is the multiplicative bias; down(x) is the downsampling function.

[0229] After the input image is propagated alternately through multiple convolutional and pooling layers, effective features for identifying statistical line loss anomalies are obtained. The fully connected layer classifies based on these features, and the calculation process is as follows:

[0230] z=f(ωx+b) (1-23)

[0231] In the formula: ω is the weight of the fully connected layer, which is also the parameter to be trained.

[0232] For the backpropagation process during the training of a convolutional neural network:

[0233] The calculation of fully connected layer parameter updates includes:

[0234] The output function of the l-th layer in a fully connected layer is:

[0235] x l =f(u l ),u l =ω l x l-1 +b l ,

[0236] Where b is the additive bias value and u is the output function variable.

[0237] Before updating the parameters using gradient descent, we first need to define the error loss function. The error loss for the nth training example is:

[0238]

[0239] In the formula, E n For error; ...

[0240] The parameters of a fully connected layer mainly include two types: weight vector w and bias value b. The training processes for the two are slightly different, and will be explained separately:

[0241] ① Update of weight vector w:

[0242] Suppose the network has L layers, then for the Lth layer:

[0243]

[0244] In the formula,

[0245]

[0246] For the weights of the l-th layer (l≠L), calculate the partial derivatives:

[0247]

[0248] Therefore, the update amount of the weights in the fully connected layer is:

[0249]

[0250] In the formula, η is the weight update coefficient.

[0251] ② Update of bias value b:

[0252] The partial derivative of the loss function with respect to the bias is:

[0253]

[0254] Where t is the calculated true value of the sample, defined for ease of expression:

[0255]

[0256] Therefore, the update amount of the bias value of the fully connected layer is:

[0257]

[0258] Pooling layer parameter update calculations include:

[0259] Theoretically, given the output error term, the parameter update can be obtained by taking the partial derivative with respect to the parameters. However, during backpropagation, the output error term changes significantly after passing through intermediate layers, leading to bias when using the output layer loss function. Therefore, for pooling and convolutional layers, it is necessary to first determine the error term propagated from the loss function to this layer before further calculating the partial derivative. For pooling layers, assuming the current pooling layer is layer l, its output value is calculated as follows:

[0260]

[0261] in, For multiplicative bias, For additive bias values, down(x) j ) represents the downsampling of neuron j.

[0262] Without loss of generality, to calculate the error of the j-th neuron, we first need to determine the neurons connected to node j in the (l+1)-th layer. Assuming the set of neurons connected to node j in the (l+1)-th layer is M, the error term for the l-th layer is:

[0263]

[0264] Where, k ij The kernel parameters of the i-th layer connected to node j.

[0265] Then, by taking the partial derivative, the update amount of the pooling layer weights and biases can be obtained.

[0266] The parameters of the pooling layer mainly include the weight vector w and the bias value b. The training processes for the two are slightly different, and will be explained separately:

[0267] ① Update of weight vector w:

[0268] The partial derivative of the loss function with respect to the weight β is:

[0269]

[0270] In the formula, u is the output function variable, and v is the optimization variable matrix.

[0271] The update formula for the weight vector w is the same as that in equation (1-24), except that the expression for the error term is changed to equation (1-26).

[0272] ② Update of bias value b:

[0273] The partial derivative of the loss function with respect to the bias value b is:

[0274]

[0275] The update formula for the bias value b is the same as that in equation (1-25), except that the expression for the error term is changed to equation (1-27).

[0276] Convolutional layer parameter update calculations include:

[0277] During backpropagation, the convolutional layer follows the fully connected and pooling layers. The error of the output layer changes after being accumulated through multiple layers. To ensure training efficiency and accuracy, it is necessary to first calculate the accumulated error term of the convolutional layer and then solve for the updated parameters. Let the convolutional layer to be analyzed be the l-th layer, and its output be:

[0278]

[0279] Where, k ij M represents the kernel parameters of the i-th layer connected to node j. j Let be the set of neurons connected to node j in the (l+1)th layer. This is an additive bias value.

[0280] Without loss of generality, the error of the (l+1)th layer is δ l+1 Then the error term for node j in the l-th layer is:

[0281]

[0282] Considering the weights from layer l to the next layer, the above error term is updated as follows:

[0283]

[0284] In the formula, ups(x) represents upsampling x, and its specific operation is related to the pooling method of the next layer (pooling layer) during the forward propagation process.

[0285] This embodiment uses the max pooling method. Therefore, when performing the ups(x) operation, it first determines whether the current position is at the maximum position in the pooling process. If so, there is no error accumulation, and the error of this layer is the same as that of the next layer; otherwise, it is 0. The expression is:

[0286]

[0287] In the formula, p max Number the location of the maximum value in the pooling region.

[0288] The parameters of a convolutional layer mainly include two types: kernel parameters k and bias values ​​b. The training processes for these two are slightly different, and will be explained separately:

[0289] ① Update of kernel parameter k:

[0290] The partial derivative of the loss function with respect to the weight w is:

[0291]

[0292] The update formula for the weight vector w is the same as that in equation (1-24), except that the expression for the error term is changed to equation (1-28).

[0293] ② Update of bias value b:

[0294] The partial derivative of the loss function with respect to the bias value b is:

[0295]

[0296] The update formula for the bias value b is the same as that in equation (1-25), except that the expression for the error term is changed to equation (1-29).

[0297] In summary, the parameter training process based on deep convolutional neural networks is as follows: Figure 11 As shown, the details are as follows:

[0298] (1) Initialize training parameters: Initialize each training parameter according to equation (1-20);

[0299] (2) Let the training sample number i = 1;

[0300] (3) Input sample data to begin training. The training process for each sample involves both forward propagation and backward propagation, such as... Figure 11 As shown, where:

[0301] ① Forward propagation process: Calculate the outputs of the convolutional layer, pooling layer, and fully connected layer according to equations (1-21), (1-22), and (1-23), respectively, and calculate the error function E(x). i ); where Equation (1-21) is used to calculate the convolutional layer, Equation (1-22) is used to calculate the pooling layer. The main function of the pooling layer is to reduce the dimensionality of the convolutional output and further extract the main features. Equation (1-23) is used to calculate the fully connected layer. After the input image is propagated through multiple convolutional and pooling layers, effective features for identifying statistical line loss anomalies are obtained. The fully connected layer classifies based on these features; if E(x i ) < e set If the training of the i-th sample ends, proceed to step (4); if E(x) i )≥e set Then it enters the reverse propagation process.

[0302] ② Backpropagation process: Update the fully connected layer parameters according to equations (1-24) and (1-25), update the pooling layer parameters according to equations (1-26) and (1-27), update the convolutional layer parameters according to equations (1-28) and (1-29), and then proceed to the forward propagation process to continue calculating the error function E(x). i ).

[0303] (4) Determine whether all samples have been trained. If training is complete, output the final training parameters. Otherwise, let i = i + 1, i ≤ N (N represents the total number of training sample datasets), return to step (2), and continue training with the next sample until all samples have been trained.

[0304] This embodiment leverages the advantage of deep convolutional neural networks in image recognition for convenient data location, proposing a method for identifying abnormal statistical line loss data in high-proportion renewable energy grid connections based on a full-sample logical state deep convolutional neural network. Compared to traditional machine learning methods such as BP neural networks and radial basis function neural networks, which take four modality matrices of the original statistical line loss data as input and output a 0-1 matrix indicating whether the data is abnormal, this method is highly convenient for identifying abnormal data. Furthermore, to avoid interference from irrelevant information in the original statistical line loss data, an improved full-sample logical state deep convolutional neural network is proposed. Based on the feature index thresholds in Chapter 4, the 0-1 logical states of the baseline component, fluctuation component, periodic component, and uncertain component data are determined and used as input, ensuring that the data contains only abnormal data information. This further improves the efficiency and accuracy of identifying abnormal statistical line loss data.

[0305] For example, the method for identifying abnormal line loss data in the statistical data of new energy grid connection provided in this embodiment, such as... Figure 12 As shown, it mainly consists of four stages: clustering of raw statistical line loss data, calculation of abnormal data feature indicators, training of deep convolutional neural network identification model, and identification of statistical line loss anomalies. The specific steps are as follows:

[0306] (1) Obtain the original statistical line loss data sequence;

[0307] (2) In the first stage, the multidimensional information of the original statistical line loss data sequence is granulated by fuzzy C-means clustering to obtain clusters 1 to 4.

[0308] (3) Second stage, calculation of abnormal data characteristic indicators. For each cluster, the baseline component characteristic indicator - distribution stability factor is calculated according to equation (1-8), the fluctuation component characteristic indicator - local outlier factor is calculated according to equation (1-9), the periodic component characteristic indicator - low rank factor is calculated according to equation (1-11), the uncertain component characteristic indicator - "rank sum" factor is calculated according to equation (1-16), and the threshold is calculated.

[0309] (4) The third stage is the training of the identification model parameters of the deep convolutional neural network. First, the statistical line loss matrix is ​​converted into a logic state information matrix according to the feature index threshold, obtaining logic state samples 1 to 4. Then, according to... Figure 11 The process for model training is shown below and will not be elaborated further.

[0310] (5) Fourth stage, identification of statistical line loss anomalies. After performing modal decomposition, multidimensional information granular fuzzy C-means clustering, calculation of anomaly data feature indicators and determination of logical state on the original statistical line loss data sequence to be identified, the data is substituted into the corresponding identification model (trained neural networks 1-4) to obtain the final anomaly data identification result.

[0311] The quality of statistical line loss data is fundamental to accurate line loss analysis. However, current integrated power generation and line loss management systems still exhibit a certain proportion of abnormal data, especially with a high proportion of renewable energy connected to the grid. The strong random fluctuations and uncertainties of renewable energy significantly alter the statistical line loss variation patterns of the power grid, rendering traditional methods for identifying abnormal statistical line loss data inapplicable. This embodiment, by understanding the impact mechanism of renewable energy access on the statistical line loss variation patterns, proposes a new model for identifying abnormal statistical line loss data in grids with a high proportion of renewable energy access, explores new characteristics of abnormal statistical line loss data in such grids, and proposes a new method for identifying abnormal statistical line loss data.

[0312] This embodiment proposes for the first time a multi-component model of new energy based on variational mode decomposition, analyzing the influence mechanism of each component on the variation law of statistical line loss. This not only helps in identifying abnormal data but also lays a theoretical foundation for future energy conservation and loss reduction, and has broad applicability. Furthermore, based on the characteristics of different components of statistical line loss data, it mines characteristic indicators of abnormal data for different components, making it more targeted and enabling effective identification of abnormal statistical line loss data for new energy grid connection. In addition, considering the specific characteristics of the problem of identifying abnormal statistical line loss data, it introduces and improves deep convolutional neural networks from the field of image recognition, proposing a full-sample logical state deep convolutional neural network, which significantly improves both identification efficiency and accuracy compared to other machine learning methods.

[0313] Furthermore, by identifying abnormal statistical line loss data of high-proportion renewable energy grid connections, this embodiment helps to locate and repair abnormal data, thereby improving the utilization rate of statistical line loss data. At the same time, it can eliminate the interference of abnormal data and effectively solve the problem of difficulty in determining the authenticity of statistical line loss in high-loss areas caused by abnormal data. In addition, accurately locating high-loss areas helps to rationally formulate loss reduction strategies and optimize the economic operation mode of the power grid.

[0314] Example 2

[0315] According to embodiments of the present invention, an apparatus is also provided for implementing the above-described method for identifying abnormal line loss data when new energy sources are connected to the power grid. Figure 13 This is a structural block diagram of a testing device for an application programming interface according to Embodiment 2 of the present invention, as shown below. Figure 13 As shown, the device includes: an acquisition module 202, a clustering module 204, a calculation module 206, and an analysis module 208. The device will be described in detail below.

[0316] The acquisition module 202 is used to acquire the raw statistical line loss data sequence;

[0317] Clustering module 204 is used to cluster the original statistical line loss data sequence to obtain multiple line loss data clusters;

[0318] Calculation module 206 is used to calculate the abnormal data characteristic indicators of each line loss data cluster in multiple line loss data clusters;

[0319] The analysis module 208 is used to input the abnormal data feature indicators of the line loss data cluster into the preset model for analysis, and obtain the position of the abnormal data in the original line loss data sequence corresponding to the abnormal data feature indicators of the line loss data cluster. The preset model is trained by multiple sets of data. Each set of data includes: the abnormal data feature indicators of the sample statistical line loss data sequence and a label used to identify the position of the abnormal data in the sample statistical line loss data sequence corresponding to the abnormal data feature indicators of the sample statistical line loss data sequence.

[0320] It should be noted that the above-mentioned acquisition module 202, clustering module 204, calculation module 206 and analysis module 208 correspond to steps S102 to S108 in Embodiment 1. The multiple modules and the corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in Embodiment 1.

[0321] Example 3

[0322] Embodiments of the present invention can provide an electronic device, which can be a terminal or a server. In this embodiment, the electronic device, as a terminal, can be any computer terminal device in a group of computer terminals. Optionally, in this embodiment, the terminal can also be a mobile terminal or other terminal device.

[0323] Optionally, in this embodiment, the terminal may be located in at least one of a plurality of network devices in a computer network.

[0324] Optionally, Figure 14 This is a structural block diagram of a terminal according to an exemplary embodiment. For example... Figure 14 As shown, the terminal may include: one or more (only one is shown in the figure) processors 31 and a memory 32 for storing processor-executable instructions; wherein the processor is configured to execute instructions to implement the above-mentioned method for identifying abnormal line loss data of new energy access to the power grid.

[0325] The memory can be used to store software programs and modules, such as the program instructions / modules corresponding to the method and device for identifying abnormal power grid line loss data in this embodiment of the invention. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby realizing the aforementioned method for identifying abnormal power grid line loss data. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to a computer terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0326] The processor can access information and applications stored in the memory via a transmission device to perform the following steps: acquiring the original statistical line loss data sequence; clustering the original statistical line loss data sequence to obtain multiple line loss data clusters; calculating the abnormal data feature index of each line loss data cluster; inputting the abnormal data feature index of the line loss data cluster into a preset model for analysis to obtain the location of abnormal data in the original line loss data sequence corresponding to the abnormal data feature index of the line loss data cluster. The preset model is trained using multiple sets of data, and each set of data includes: the abnormal data feature index of the sample statistical line loss data sequence and a label used to identify the location of abnormal data in the sample statistical line loss data sequence corresponding to the abnormal data feature index of the sample statistical line loss data sequence.

[0327] Optionally, the processor may also execute program code that performs the following steps: clustering the original statistical line loss data sequence to obtain multiple line loss data clusters, including: using variational mode decomposition to granulate the multidimensional information of the original statistical line loss data sequence to obtain a multidimensional information granulated input data sequence; and using fuzzy C-means clustering algorithm to cluster the multidimensional information granulated input data sequence to obtain multiple line loss data clusters.

[0328] Optionally, the processor may also execute program code that performs the following steps: calculating the abnormal data characteristic index of each of the multiple line loss data clusters, including: determining the degree of abnormal data of the baseline component of the original statistical line loss data; and obtaining the distribution stability factor characteristic index based on the degree of abnormal data of the baseline component, wherein the distribution stability factor characteristic index serves as the abnormal data characteristic index of the baseline component data in the original statistical line loss data.

[0329] Optionally, the processor may also execute program code for the following steps: calculating the abnormal data feature index of each of the multiple line loss data clusters, including: performing mode decomposition on each line loss data cluster to obtain a fluctuation component data sequence, the fluctuation component data sequence including multiple fluctuation component data points; performing autoregressive analysis on the fluctuation component data sequence using an ARIMA model to obtain multiple regression curves corresponding to the multiple fluctuation component data points; determining a first regression curve from the multiple regression curves, wherein the multiple regression curves other than the first regression curve are second regression curves, the difference between the first regression curve and the multiple second regression curves is greater than a preset threshold, and the difference between any two of the multiple second regression curves is less than or equal to the preset threshold; obtaining a local outlier feature index based on the difference between the first regression curve and the multiple second regression curves, wherein the local outlier feature index serves as the abnormal data feature index of the fluctuation component data sequence.

[0330] Optionally, the processor may also execute program code for the following steps: calculating the abnormal data feature index of each of the multiple line loss data clusters, including: performing mode decomposition on each line loss data cluster to obtain a periodic component data sequence, the periodic component data sequence including multiple periodic component data points; arranging the multiple periodic component data points into a periodic mode matrix; performing singular value decomposition on the periodic mode matrix to obtain singular values; determining the matrix low-rank degree of the periodic mode matrix based on the singular values; and obtaining the low-rank factor feature index according to the matrix low-rank degree, wherein the low-rank factor feature index serves as the abnormal data feature index of the periodic component data sequence.

[0331] Optionally, the processor may also execute program code for the following steps: calculating the abnormal data feature index of each of the multiple line loss data clusters, including: performing mode decomposition on each line loss data cluster to obtain a periodic component data sequence, the periodic component data sequence including multiple periodic component data points; arranging the multiple periodic component data points into a periodic mode matrix; performing singular value decomposition on the periodic mode matrix to obtain singular values; determining the matrix low-rank degree of the periodic mode matrix based on the singular values; and obtaining the low-rank factor feature index according to the matrix low-rank degree, wherein the low-rank factor feature index serves as the abnormal data feature index of the periodic component data sequence.

[0332] Optionally, the processor may also execute program code for the following steps: constructing a deep convolutional neural network, the deep convolutional neural network comprising an input layer, multiple convolutional layers, multiple pooling layers, multiple fully connected layers, and an output layer connected in sequence; training the deep convolutional neural network to obtain a preset model, wherein the training includes forward propagation training and backward propagation training of parameters, the forward propagation training being used to input the abnormal data feature indicators of the sample statistical line loss data sequence into the deep convolutional neural network to obtain the actual location of the abnormal data in the sample statistical line loss data sequence, the error backward propagation training being used to calculate the error between the actual location and the location of the abnormal data in the sample statistical line loss data sequence, the error being used to input into the deep convolutional neural network to update the parameters using the gradient descent method in the forward propagation training and backward propagation training.

[0333] Those skilled in the art will understand that Figure 14 The structure shown is for illustrative purposes only. For example, the terminal mentioned above can also be a smartphone (such as an Android phone, an iOS phone, etc.), a tablet computer, a mobile internet device (MID), a PAD, and other terminal devices. Figure 14 This does not limit the structure of the aforementioned electronic device. For example, it may also include devices that are more... Figure 14 The more or fewer components shown (such as network interfaces, display devices, etc.), or having the same Figure 14 The different configurations shown.

[0334] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0335] Example 4

[0336] In an exemplary embodiment, a computer-readable storage medium including instructions is also provided. When the instructions in the computer-readable storage medium are executed by a processor of a terminal, the terminal is able to execute the aforementioned method for identifying abnormal line loss data of new energy grid access. Optionally, the computer-readable storage medium may be a non-transitory computer-readable storage medium, such as a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device.

[0337] Optionally, in this embodiment, the computer-readable storage medium described above can be used to store the program code executed by the method for identifying abnormal line loss data of new energy access to the power grid provided in Embodiment 1.

[0338] Optionally, in this embodiment, the computer-readable storage medium may be located in any computer terminal in a group of computer terminals in a computer network, or in any mobile terminal in a group of mobile terminals.

[0339] Optionally, in this embodiment, the computer-readable storage medium is configured to store program code for performing the following steps: acquiring the original statistical line loss data sequence; clustering the original statistical line loss data sequence to obtain multiple line loss data clusters; calculating the abnormal data feature index of each line loss data cluster; inputting the abnormal data feature index of the line loss data cluster into a preset model for analysis to obtain the position of abnormal data in the original line loss data sequence corresponding to the abnormal data feature index of the line loss data cluster, wherein the preset model is obtained by training multiple sets of data, and each set of data includes: the abnormal data feature index of the sample statistical line loss data sequence and a label for identifying the position of abnormal data in the sample statistical line loss data sequence corresponding to the abnormal data feature index of the sample statistical line loss data sequence.

[0340] Optionally, in this embodiment, the computer-readable storage medium is configured to store program code for performing the following steps: clustering the original statistical line loss data sequence to obtain multiple line loss data clusters, including: using variational mode decomposition to granulate the multidimensional information of the original statistical line loss data sequence to obtain a multidimensional information granulated input data sequence; and using fuzzy C-means clustering algorithm to cluster the multidimensional information granulated input data sequence to obtain multiple line loss data clusters.

[0341] Optionally, in this embodiment, the computer-readable storage medium is configured to store program code for performing the following steps: calculating the abnormal data feature index of each of the multiple line loss data clusters, including: determining the degree of abnormal data of the baseline component of the original statistical line loss data; obtaining the distribution stability factor feature index based on the degree of abnormal data of the baseline component, wherein the distribution stability factor feature index serves as the abnormal data feature index of the baseline component data in the original statistical line loss data.

[0342] Optionally, in this embodiment, the computer-readable storage medium is configured to store program code for performing the following steps: calculating the abnormal data feature index of each line loss data cluster in multiple line loss data clusters, including: performing mode decomposition on each line loss data cluster to obtain a fluctuation component data sequence, the fluctuation component data sequence including multiple fluctuation component data points; performing autoregressive analysis on the fluctuation component data sequence using an ARIMA model to obtain multiple regression curves corresponding to the multiple fluctuation component data points; determining a first regression curve from the multiple regression curves, wherein the multiple regression curves other than the first regression curve are second regression curves, the difference between the first regression curve and the multiple second regression curves is greater than a preset threshold, and the difference between any two of the multiple second regression curves is less than or equal to the preset threshold; obtaining a local outlier factor feature index based on the difference between the first regression curve and the multiple second regression curves, wherein the local outlier factor feature index is used as the abnormal data feature index of the fluctuation component data sequence.

[0343] Optionally, in this embodiment, the computer-readable storage medium is configured to store program code for performing the following steps: calculating the abnormal data feature index of each of the multiple line loss data clusters, including: performing mode decomposition on each line loss data cluster to obtain a periodic component data sequence, the periodic component data sequence including multiple periodic component data points; arranging the multiple periodic component data points into a periodic mode matrix; performing singular value decomposition on the periodic mode matrix to obtain singular values; determining the matrix low-rank degree of the periodic mode matrix based on the singular values; and obtaining a low-rank factor feature index based on the matrix low-rank degree, wherein the low-rank factor feature index serves as the abnormal data feature index of the periodic component data sequence.

[0344] Optionally, in this embodiment, the computer-readable storage medium is configured to store program code for performing the following steps: calculating the abnormal data feature index of each of the multiple line loss data clusters, including: performing mode decomposition on each line loss data cluster to obtain a statistical line loss uncertain component data sequence; obtaining a theoretical line loss data sequence corresponding to the original statistical line loss data sequence; performing mode decomposition on the theoretical line loss data sequence to obtain a theoretical line loss uncertain component data sequence; calculating a difference matrix based on the statistical line loss uncertain component data sequence and the theoretical line loss uncertain component data sequence; ranking the difference matrix to obtain the positive rank sum and negative rank sum of the difference matrix; and obtaining the rank sum factor feature index based on the difference degree of the positive rank sum and negative rank sum, wherein the rank sum factor feature index serves as the abnormal data feature index of the statistical line loss uncertain component data sequence.

[0345] Optionally, in this embodiment, the computer-readable storage medium is configured to store program code for performing the following steps: constructing a deep convolutional neural network, the deep convolutional neural network including an input layer, multiple convolutional layers, multiple pooling layers, multiple fully connected layers, and an output layer connected in sequence; training the deep convolutional neural network to obtain a preset model, wherein the training includes forward propagation training and backward propagation training of parameters, the forward propagation training is used to input the abnormal data feature indicators of the sample statistical line loss data sequence into the deep convolutional neural network to obtain the actual position of the abnormal data in the sample statistical line loss data sequence, the error backward propagation training is used to calculate the error between the actual position and the position of the abnormal data in the sample statistical line loss data sequence, the error is used to input into the deep convolutional neural network to update the parameters using the gradient descent method in the forward propagation training and backward propagation training.

[0346] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0347] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0348] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of units or modules may be electrical or other forms.

[0349] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0350] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0351] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0352] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying abnormal data of line loss of a new energy access power grid, characterized in that, The method comprises the following steps: obtaining an original statistical line loss data sequence; clustering the original statistical line loss data sequence to obtain a plurality of line loss data clusters; calculating an abnormal data feature index of each line loss data cluster in the plurality of line loss data clusters; inputting the abnormal data feature index of the line loss data cluster into a preset model for analysis to obtain the position of the abnormal data in the original statistical line loss data sequence corresponding to the abnormal data feature index of the line loss data cluster, wherein the preset model is obtained by training a plurality of groups of data, and each group of data in the plurality of groups of data comprises an abnormal data feature index of a sample statistical line loss data sequence and a label for identifying the position of the abnormal data in the sample statistical line loss data sequence corresponding to the abnormal data feature index of the sample statistical line loss data sequence, the calculation of the abnormal data feature index of each line loss data cluster in the plurality of line loss data clusters comprises: modal decomposition of the each line loss data cluster to obtain a fluctuation component data sequence, the fluctuation component data sequence comprising a plurality of fluctuation component data points; autoregressive analysis of the fluctuation component data sequence using an ARIMA model to obtain a plurality of regression curves corresponding to the plurality of fluctuation component data points; determining a first regression curve from the plurality of regression curves, wherein the plurality of regression curves other than the first regression curve are second regression curves, the difference between the first regression curve and the plurality of second regression curves is greater than a preset threshold, and the difference between any two second regression curves in the plurality of second regression curves is less than or equal to the preset threshold; obtaining a local outlier factor feature index according to the difference between the first regression curve and the plurality of second regression curves, wherein the local outlier factor feature index is taken as the abnormal data feature index of the fluctuation component data sequence; alternatively, the calculation of the abnormal data feature index of each line loss data cluster in the plurality of line loss data clusters comprises: determining a reference component abnormal data degree of the original statistical line loss data in the each line loss data cluster; obtaining a distribution stability factor feature index according to the reference component abnormal data degree, wherein the distribution stability factor feature index is taken as the abnormal data feature index of the reference component data in the original statistical line loss data; alternatively, the calculation of the abnormal data feature index of each line loss data cluster in the plurality of line loss data clusters comprises: modal decomposition of the each line loss data cluster to obtain a periodic component data sequence, the periodic component data sequence comprising a plurality of periodic component data points; arranging the plurality of periodic component data points into a periodic modal matrix; singular value decomposition of the periodic modal matrix to obtain a singular value; determining a matrix low rank degree of the periodic modal matrix based on the singular value; obtaining a low rank factor feature index according to the matrix low rank degree, wherein the low rank factor feature index is taken as the abnormal data feature index of the periodic component data sequence. Or, the computing the abnormal data feature index of each line loss data cluster in the plurality of line loss data clusters comprises: performing modal decomposition on the each line loss data cluster to obtain a statistical line loss uncertainty component data sequence; obtaining a theoretical line loss data sequence corresponding to the original statistical line loss data sequence; performing modal decomposition on the theoretical line loss data sequence to obtain a theoretical line loss uncertainty component data sequence; calculating a difference matrix according to the statistical line loss uncertainty component data sequence and the theoretical line loss uncertainty component data sequence; performing rank coding on the difference matrix to obtain a positive rank sum and a negative rank sum of the difference matrix; obtaining a rank sum factor feature index according to a difference degree of the positive rank sum and the negative rank sum, the rank sum factor feature index serving as the abnormal data feature index of the statistical line loss uncertainty component data sequence.

2. The recognition method according to claim 1, characterized in that, The clustering the original statistical line loss data sequence to obtain a plurality of line loss data clusters comprises: The original statistical line loss data sequence is granulated by using a variational modal decomposition method to obtain a multi-dimensional information granulation input data sequence; The multi-dimensional information granulation input data sequence is clustered by using a fuzzy C-means clustering algorithm to obtain the plurality of line loss data clusters.

3. The recognition method of claim 1, wherein, Further comprising: A deep convolutional neural network is constructed, and the deep convolutional neural network comprises an input layer, a plurality of convolutional layers, a plurality of pooling layers, a plurality of fully connected layers and an output layer which are sequentially and signal-connected; The deep convolutional neural network is trained to obtain the preset model, wherein the training comprises forward propagation training and back propagation training of parameters, the forward propagation training is used to input the abnormal data feature index of the sample statistical line loss data sequence into the deep convolutional neural network to obtain an actual position of abnormal data in the sample statistical line loss data sequence, and the back propagation training is used to calculate an error between the actual position and a position of abnormal data in the sample statistical line loss data sequence, the error is used to input the deep convolutional neural network, and the parameters are updated by using a gradient descent method in the forward propagation training and the back propagation training.

4. A new energy access power grid line loss abnormal data identification device, characterized in that, Comprise: The acquisition module is configured to acquire an original statistical line loss data sequence; The clustering module is configured to cluster the original statistical line loss data sequence to obtain a plurality of line loss data clusters; The computing module is configured to calculate an abnormal data feature index of each line loss data cluster in the plurality of line loss data clusters; The analysis module is configured to input the abnormal data feature index of the line loss data cluster into a preset model for analysis to obtain a position of abnormal data in the original statistical line loss data sequence corresponding to the abnormal data feature index of the line loss data cluster, wherein the preset model is obtained by training a plurality of groups of data, each group of data in the plurality of groups of data comprises an abnormal data feature index of a sample statistical line loss data sequence and a label for identifying a position of abnormal data in the sample statistical line loss data sequence corresponding to the abnormal data feature index of the sample statistical line loss data sequence. The computing module is further configured to perform modal decomposition on each line loss data cluster to obtain a fluctuation component data sequence, the fluctuation component data sequence comprising a plurality of fluctuation component data points; perform autoregressive analysis on the fluctuation component data sequence using an ARIMA model to obtain a plurality of regression curves corresponding to the plurality of fluctuation component data points; determine a first regression curve from the plurality of regression curves, wherein the plurality of regression curves other than the first regression curve are second regression curves, the difference degree of the first regression curve and the plurality of second regression curves is greater than a preset threshold, and the difference degree of any two second regression curves in the plurality of second regression curves is less than or equal to the preset threshold; and obtain a local outlier factor feature index according to the difference degree of the first regression curve and the plurality of second regression curves, wherein the local outlier factor feature index is taken as an abnormal data feature index of the fluctuation component data sequence. Alternatively, the computing module is configured to determine a reference component abnormal data degree of original statistical line loss data in each line loss data cluster; and obtain a distribution stability factor feature index according to the reference component abnormal data degree, wherein the distribution stability factor feature index is taken as an abnormal data feature index of reference component data in the original statistical line loss data. Alternatively, the computing module is configured to perform modal decomposition on each line loss data cluster to obtain a periodic component data sequence, the periodic component data sequence comprising a plurality of periodic component data points; arrange the plurality of periodic component data points into a periodic modal matrix; perform singular value decomposition on the periodic modal matrix to obtain singular values; determine a matrix low-rank degree of the periodic modal matrix based on the singular values; and obtain a low-rank factor feature index according to the matrix low-rank degree, wherein the low-rank factor feature index is taken as an abnormal data feature index of the periodic component data sequence. Alternatively, the computing module is configured to perform modal decomposition on each line loss data cluster to obtain a statistical line loss uncertainty component data sequence; obtain a theoretical line loss data sequence corresponding to the original statistical line loss data sequence; perform modal decomposition on the theoretical line loss data sequence to obtain a theoretical line loss uncertainty component data sequence; calculate a difference matrix according to the statistical line loss uncertainty component data sequence and the theoretical line loss uncertainty component data sequence; perform rank coding on the difference matrix to obtain a positive rank sum and a negative rank sum of the difference matrix; and obtain a rank sum factor feature index according to the difference degree of the positive rank sum and the negative rank sum, the rank sum factor feature index being taken as an abnormal data feature index of the statistical line loss uncertainty component data sequence.

5. An electronic device, comprising: comprise: a processor; a memory for storing instructions executable by the processor; wherein the processor is configured to execute the instructions to implement the identification method of any one of claims 1 to 3.

6. A computer-readable storage medium, characterized in that, When the instructions in the computer-readable storage medium are executed by the processor of the electronic device, the electronic device is enabled to perform the identification method of any one of claims 1 to 3.

Images