Intelligent identification method of illegal power consumption in distribution substation area considering the quality of multi-source measurement data

Abstract

The present invention discloses an intelligent identification method for illegal electricity consumption in distribution station areas considering the quality of multi-source measurement data, comprising the following steps: performing preliminary abnormal data point screening on the voltage time series data and power time series data of each node respectively to obtain abnormal time series data clusters; performing correlation analysis on the abnormal time series data clusters by association rule mining to obtain data quality abnormal data; repairing the marked data quality abnormal data based on a conditional diffusion model to obtain repaired voltage time series data and power time series data; performing multi-image fusion on the voltage time series graph and the power time series graph corresponding to the repaired voltage time series data and power time series data to obtain a node time series fusion image; inputting the node time series fusion image into a trained illegal electricity consumption identification model to perform image abnormal point detection to obtain abnormal pixel points, and mapping them back to the node voltage time series data and the node power time series data to achieve the traceability and positioning of illegal electricity consumption in the distribution station area.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, and in particular to an intelligent identification method for illegal electricity consumption in a distribution station area taking into account the quality of multi-source measurement data. Background Art

[0002] Identification of illegal electricity use in distribution stations is of great significance for improving the power supply quality of the power system, ensuring power supply safety, and improving the lean management level of the substation. In recent years, with the construction of transparent distribution stations, the real-time and accuracy of the identification of illegal electricity use by users in the substation area has gradually improved. However, due to the access of a large number of new sources and loads such as distributed photovoltaics, energy storage, and charging piles to the low-voltage distribution network, their random, intermittent, decentralized, and uncontrollable characteristics have made the complexity of substation operation and the difficulty of management and control exponentially increase, bringing more challenges to the identification of illegal electricity use by users in the substation area. On the one hand, as a bridge connecting a large number of user terminals, the distribution station area's measurement data is the core production factor for improving the distribution network's state perception ability and operation and maintenance management level. In actual engineering, due to factors such as equipment communication quality and data structure differences, the quality of data collected by the substation intelligent terminal is uneven, which greatly affects the accuracy of the identification of illegal electricity use in the distribution station area. On the other hand, with the access of a large number of new sources and loads to the distribution station area, the diversity and complexity of the measurement data in the substation area have increased, and the electrical coupling characteristics between nodes have changed, which has greatly reduced the efficiency of identifying illegal electricity use in the distribution station area.

[0003] At present, the detection and repair of abnormal data in distribution substations only focus on data quality abnormal data. Commonly used abnormal data detection methods include statistical analysis, density-based spatial data clustering, and isolation forest algorithms. Abnormal data repair methods include cubic Hermite interpolation, long short-term memory neural networks, and regression algorithms based on shallow neural networks. However, the above abnormal data detection and repair methods do not consider the confusion between abnormal data of illegal power consumption and abnormal data of data quality, which may lead to misjudgment of the actual operating status measurement data of the distribution substation, affecting the results of the state perception of the substation. Although the identification of illegal power consumption in distribution substations has also attracted the attention of many scholars, the commonly used illegal power consumption identification methods mainly use machine learning algorithms to analyze and mine the measurement data of the substation. For example, based on the abnormal measurement data marked in advance by humans, the illegal power consumption identification model is constructed using random forests, neural networks and other algorithms. However, due to the large number of new sources and loads connected to the distribution substation, the voltage, power and other electrical quantities in the substation fluctuate, and the diversity and complexity of the measurement data in the substation increase, which poses a huge challenge to the identification of illegal power consumption in the substation. Summary of the invention

[0004] In view of this, the present invention provides an intelligent identification method for illegal electricity consumption in a distribution station area taking into account the quality of multi-source measurement data, so as to at least solve the problem of inaccurate detection and repair of abnormal data in distribution station areas in the prior art.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

[0006] The intelligent identification method of illegal power consumption in distribution substation area considering the quality of multi-source measurement data includes the following steps:

[0007] S1. Voltage timing data of each node and power timing data Perform preliminary screening of abnormal data points to obtain abnormal time series data clusters, where the nodes include user nodes, photovoltaic nodes, and charging pile nodes in the substation area; is the length of a single node time series, is the number of topological nodes in the distribution area;

[0008] S2. Perform correlation analysis on the abnormal time series data cluster marked in S1 by association rule mining to obtain strongly correlated nodes in the abnormal time series data cluster. If only one voltage time series data or power time series data of a strongly correlated node is marked as abnormal data at the same time, the marked abnormal data is determined to be data quality abnormal data;

[0009] S3. Repair the marked data quality abnormality data based on the conditional diffusion model to obtain repaired voltage time series data and power time series data;

[0010] S4. Encode and convert the repaired voltage time series data and power time series data to obtain the voltage time series diagram and power time series diagram of each node, extract the features of the voltage time series diagram and the power time series diagram, and perform multi-image fusion to obtain the node time series fusion image ;

[0011] S5. Fusion of nodes in time series The trained illegal electricity consumption recognition model is input to detect image anomaly points, obtain abnormal pixels, and map them back to node voltage time series data and node power time series data to achieve the traceability and location of illegal electricity consumption in the distribution station area.

[0012] Preferably, the specific content of S1 includes:

[0013] S11. Obtain voltage timing data of each node Power Timing Data ;

[0014] S12. Detect repeated data in the voltage time series data and the power time series data by using a sliding window method, and mark data points with a repetition rate greater than a preset repetition threshold as abnormal data points;

[0015] S13. Perform outlier detection on the voltage time series data and the power time series data based on the local anomaly factor algorithm, and mark the data points whose anomaly factors are greater than the preset anomaly threshold as abnormal data points;

[0016] S14. Preliminarily mark all abnormal data points as abnormal time series data clusters.

[0017] Preferably, the specific content of S12 includes:

[0018] Choose a sliding window size of , sliding the window from the starting position of the voltage time series data or the power time series data, one step forward each time; when the number of occurrences of repeated data in the time series data selected by the sliding window exceeds the preset repetition threshold of the sliding window size, the first repeated data point in the selected time series is abnormal data;

[0019] The specific contents of S13 include:

[0020] according to and The data density is calculated and The local accessibility density of each data point in is further calculated to obtain the abnormal factor size of the data point. The abnormal factor size indicates the degree of outlier of each data point, and the data points with abnormal factors greater than the preset abnormal threshold are marked as abnormal data points. The local accessibility density and abnormal factor are:

[0021] ,

[0022] In the formula, Indicates the sampling time point; Indicates voltage and power timing data points The local accessibility density of For data points No. Neighborhood, representing the data point Distance The set of all time series data points within the distance; Indicates The number of data points in the neighborhood; For time series data points To the time series data point No. distance; is the node time series data point The size of the abnormal factor.

[0023] Preferably, the specific contents of association analysis through association rule mining include:

[0024] The correlation and confidence between all voltage time series data and power time series data are calculated. When the correlation and confidence of the voltage time series data and power time series data of two nodes meet the preset minimum threshold, it is considered that the time series of the two nodes are strongly correlated, otherwise the time series of the two nodes are considered to be unrelated.

[0025] Preferably, the specific contents of calculating the correlation degree and the confidence degree include:

[0026] The lengths are L The sliding window of the voltage time series data and power time series data to be analyzed for correlation is intercepted as M time series and M power subsequences, and linearly fit the voltage subsequence and the power subsequence respectively, so that the value interval of the slope of each subsequence is [-1,1] and divide it into various slope segments, and symbolically represent each slope segment;

[0027] Each slope of the voltage subsequence or power subsequence constitutes a complete event set;

[0028] All voltage timing data or power timing data The event database composed of subsequences of E , the number of each subsequence in the event library is recorded as ;

[0029] Will To indicate the k -1 node and k The correlation between the voltage timing data or power timing data between nodes, and Respectively k -1 node and k The voltage subsequence or power subsequence of each node, express In the database E The closer the correlation is to 1, the more k -1 node and k The higher the correlation degree of voltage timing data or power timing data of each node is:

[0030] ,

[0031] In the formula, For the k -1 node and k The voltage subsequence or power subsequence of each node is and The correlation between for and The confidence level between Represents a database E Also included Also included probability; express and The correlation between and The size of the correlation between express and The confidence level or and Confidence between The number of associated subsequences that satisfy the minimum support;

[0032] Get and .

[0033] Preferably, the specific content of S3 includes:

[0034] (1) Generate estimated time series data of data quality abnormality data by using the strongly correlated node power flow calculation, and define the remaining time series data after removing the time series fragment corresponding to the data quality abnormality data as the remaining monitorable time series data, and use the remaining monitorable time series data as the normal time series data; the normal time series data includes normal voltage time series data or normal power time series data;

[0035] (2) Add noise to the data with abnormal data quality to obtain the masked abnormal time series data corresponding to the current diffusion step, and calculate the estimated noise value corresponding to the current diffusion step based on the normal time series data and the estimated time series data;

[0036] (3) Calculate the estimated noise value The actual noise added corresponding to the current diffusion step The difference between Minimize

[0037] The optimization goal of repairing abnormal voltage time series data based on the conditional diffusion model is:

[0038] ,

[0039] In the formula, represents the denoising part of the model; For the diffusion step The masked abnormal voltage timing data obtained after adding noise; is Gaussian noise, It is the normal voltage timing data; To calculate the voltage timing data, estimate the noise value and ; It means taking the average of the probability distribution of normal voltage timing data and Gaussian noise;

[0040] The optimization goal of repairing abnormal power time series data based on the conditional diffusion model is:

[0041] ,

[0042] In the formula, For the diffusion step The masked abnormal power time series data obtained after adding noise; It is the normal power timing data; To calculate the power timing data, estimate the noise value and ; Represents the average of the probability distributions of normal power time series data and Gaussian noise.

[0043] Preferably, the specific content of S4 includes:

[0044] The repaired voltage timing data Power Timing Data Perform a cosine polar transformation, where the timestamps of the time series are converted into radii R , the cosine value is converted into the corresponding angle value , the transformation process is:

[0045] ,

[0046] In the formula, and Represents the repaired node k Voltage and power timing data points; and Representation Node k exist t The angle value of the voltage and power time series data points corresponding to the polar coordinates; Indicates the radius value in polar coordinates; Indicates the timestamp of the corresponding time series data, and divides the timestamp interval into Z The equal division regularizes the polar coordinate span, For timestamp A collection of;

[0047] Based on the Gram angular field imaging algorithm, the voltage time series data and power time series data after cosine value polar coordinate transformation are encoded into a two-dimensional image matrix and :

[0048] ,

[0049] ,

[0050] In the formula, The voltage timing data after repair The transpose of The voltage timing data after repair The transpose of It is a unit and Corresponding angle in polar coordinates;

[0051] Will and Converted into a two-dimensional image, the node voltage timing diagram is obtained Power Timing Diagram , where the tensor of each time series image is , G and F is the spatial resolution of the node timing graph, 1 means the node timing graph is a single-channel image;

[0052] Use convolutional neural network to extract the features of each time series graph :

[0053] ,

[0054] In the formula, The tensors are , concatenate all the feature graphs of the timing graph according to the channel to form a node timing feature matrix , the concatenated feature matrix The size is ;

[0055] Compute node timing feature matrix The mean , and subtract it from all node time series feature graphs to eliminate the impact of the mean:

[0056] ,

[0057] In the formula, Represents the decentralized node time series feature matrix; Represents the node time series feature matrix in k Node, j Elements on a feature;

[0058] calculate The covariance matrix of ,right Perform eigenvalue decomposition to obtain the corresponding eigenvalue With the eigenvector ;

[0059] According to the feature vector Get the projection matrix , use the projection matrix to transform the decentralized node time series feature matrix Project to feature space to get the projected feature matrix ;

[0060] For the fused projection feature matrix Perform inverse transformation to reconstruct it into a feature map , and the feature map Convert to the spatial representation of the image to get the final node temporal fusion image :

[0061] .

[0062] Preferably, the specific content of S5 includes:

[0063] Fusion of nodes into images The trained illegal electricity consumption recognition model is input to detect image anomalies, and the difference score of each pixel of the two images is calculated using the pixel-by-pixel difference method:

[0064] ,

[0065] Thresholds based on difference score statistics P for:

[0066] ,

[0067] In the formula, for The mean of for The standard deviation of is the adjustment factor;

[0068] The current difference score is greater than the threshold P The pixel points are marked as illegal electricity usage feature points, and are mapped back to the node voltage time series data and the node power time series data to achieve the traceability and positioning of illegal electricity usage in the distribution station area.

[0069] Preferably, the specific contents of training the illegal electricity usage identification model include:

[0070] The training model for illegal electricity consumption recognition includes a discriminator and a generator. A self-attention mechanism is added to the middle layer of the discriminator and the generator to improve the image generation capability. The training process includes the training of the discriminator and the training of the generator:

[0071] The training process of the discriminator includes:

[0072] Fuse the node time series of the current batch into an image Generate images with nodes At the same time, it is input into the discriminator, the corresponding time series feature graph is generated and a layer of convolution operation is performed on it, the discriminator loss is calculated and the parameters are updated:

[0073] ,

[0074] In the formula, Represents the prediction probability of the discriminator for the node temporal fusion feature map; Represents the predicted probability of the discriminator generating a time series feature graph for the node; It is a 0-1 variable, which is 1 when the node time series fusion feature graph is generated, and 0 when the node generates the time series feature graph;

[0075] The training process of the generator includes:

[0076] The obtained node time series fusion image Generate a time series feature vector through the encoder e , and then the time series feature vector e Input decoder to decode and get node generated image , generate the node image Input the encoder again to get the timing generation vector , calculate the overall loss of the generator and perform parameter updates.

[0077] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses an intelligent identification method for illegal electricity consumption in distribution station areas taking into account the quality of multi-source measurement data, which has the following beneficial effects:

[0078] Compared with the prior art, the present invention has higher data repair capabilities, more efficient illegal electricity use detection, more accurate traceability and positioning capabilities, and a wide range of application scenarios. By detecting and repairing data with abnormal data quality, the accuracy and reliability of the measurement data in the substation area are improved, thereby improving the accuracy of identifying illegal electricity use. The use of multi-source time series feature fusion image anomaly point detection can enable the time series data characteristics of each node to be fully reflected, better identify complex abnormal patterns and illegal behaviors, and improve the efficiency of identifying illegal electricity use. It can adapt to the identification of illegal electricity use in distribution stations with a high proportion of distributed energy access, accurately locate specific illegal electricity use nodes in the substation area, and significantly improve the efficiency, accuracy and practicality of identifying illegal electricity use. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0080] Figure 1 A flow chart of the intelligent identification method for illegal electricity consumption in distribution station areas taking into account the quality of multi-source measurement data provided by the present invention;

[0081] Figure 2 A diagram showing the results of identifying illegal electricity use in a distribution station area provided by an embodiment of the present invention;

[0082] Figure 3 A comparison chart of illegal electricity usage identification results using various algorithms provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0083] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0084] The present invention provides an intelligent identification method for illegal power consumption in distribution substations considering the quality of multi-source measurement data, such as Figure 1 As shown, the following steps are included:

[0085] S1. Voltage timing data of each node and power timing data Perform preliminary screening of abnormal data points to obtain abnormal time series data clusters, where the nodes include user nodes, photovoltaic nodes, and charging pile nodes in the substation area; is the length of a single node time series, is the number of topological nodes in the distribution area;

[0086] S2. Perform correlation analysis on the abnormal time series data cluster marked in S1 by association rule mining to obtain strongly correlated nodes in the abnormal time series data cluster. If only one voltage time series data or power time series data of a strongly correlated node is marked as abnormal data at the same time, the marked abnormal data is determined to be data quality abnormal data;

[0087] S3. Repair the marked data quality abnormality data based on the conditional diffusion model to obtain repaired voltage time series data and power time series data;

[0088] S4. Encode and convert the repaired voltage time series data and power time series data to obtain the voltage time series diagram and power time series diagram of each node, extract the features of the voltage time series diagram and the power time series diagram, and perform multi-image fusion to obtain the node time series fusion image ;

[0089] S5. Fusion of nodes in time series The trained illegal electricity consumption recognition model is input to detect image anomaly points, obtain abnormal pixels, and map them back to node voltage time series data and node power time series data to achieve the traceability and location of illegal electricity consumption in the distribution station area.

[0090] It should be noted that:

[0091] The measurement data of user nodes, photovoltaic nodes and charging pile nodes in low-voltage distribution substations can be divided into normal operation data, abnormal data of illegal electricity consumption and abnormal data quality data according to data characteristics. Both normal operation data and abnormal data of illegal electricity consumption can reflect the real operating status of the distribution substation and are the main basis for the status perception of the distribution substation. The abnormal data quality data is due to the changeable operating conditions and complex electromagnetic environment affecting the collection, transmission and storage of measurement data of the intelligent terminal in the substation area, resulting in duplicate data, missing data and erroneous data in the measurement data. In the process of identifying illegal electricity consumption in the distribution substation, abnormal data quality data and abnormal data of illegal electricity consumption are easily confused. Therefore, it is necessary to detect and repair the abnormal data quality data to ensure the integrity and accuracy of the data used for identifying illegal electricity consumption.

[0092] In order to further implement the above technical solution, the specific contents of S1 include:

[0093] S11. Obtain voltage timing data of each node Power Timing Data ;

[0094] S12. Detect repeated data in the voltage time series data and the power time series data by using a sliding window method, and mark data points with a repetition rate greater than a preset repetition threshold as abnormal data points;

[0095] S13. Perform outlier detection on the voltage time series data and the power time series data based on the local anomaly factor algorithm, and mark the data points whose anomaly factors are greater than the preset anomaly threshold as abnormal data points;

[0096] S14. Preliminarily mark all abnormal data points as abnormal time series data clusters.

[0097] In order to further implement the above technical solution, the specific contents of S12 include:

[0098] Choose a sliding window size of , sliding the window from the starting position of the voltage time series data or the power time series data, one step forward each time; when the number of occurrences of repeated data in the time series data selected by the sliding window exceeds the preset repetition threshold of the sliding window size, the first repeated data point in the selected time series is abnormal data;

[0099] The specific contents of S13 include:

[0100] according to and The data density is calculated and The local accessibility density of each data point in is further calculated to obtain the abnormal factor size of the data point. The abnormal factor size indicates the degree of outlier of each data point, and the data points with abnormal factors greater than the preset abnormal threshold are marked as abnormal data points. The local accessibility density and abnormal factor are:

[0101] ,

[0102] In the formula, Indicates the sampling time point; Indicates voltage and power timing data points The local accessibility density of For data points No. j Neighborhood, representing the data point Distance j The set of all time series data points within the distance; Indicates j The number of data points in the neighborhood; For time series data points To the time series data point No. j distance; is the node time series data point The size of the abnormal factor.

[0103] It should be noted that:

[0104] In the specific application process, after collecting the node voltage time series data and the node power time series data through the intelligent terminal, the data needs to be preprocessed, including normalization processing, that is, the average value of the voltage time series data and the power time series data of each node is converted to 0, and the standard deviation is converted to 1.

[0105] In the process of detecting duplicate data in time series data by the sliding window method, the preset duplication threshold is set to 70% in this embodiment, that is, when the number of occurrences of duplicate data in the time series window exceeds 70% of the window size, the first duplicate data point in the window is abnormal data, and all duplicate data points in the time series of the mark node are abnormal data.

[0106] In order to further implement the above technical solution, the specific contents of association analysis through association rule mining include:

[0107] The correlation and confidence between all voltage time series data and power time series data are calculated. When the correlation and confidence of the voltage time series data and power time series data of two nodes meet the preset minimum threshold, it is considered that the time series of the two nodes are strongly correlated, otherwise the time series of the two nodes are considered to be unrelated.

[0108] In order to further implement the above technical solution, the specific contents of calculating the correlation degree and the confidence degree include:

[0109] The lengths are L The sliding window of the voltage time series data and power time series data to be analyzed for correlation is intercepted as M time series and M power subsequences, and linearly fit the voltage subsequence and the power subsequence respectively, so that the value interval of the slope of each subsequence is [-1,1] and divide it into 5 segments to obtain the slope segmentation, and symbolically represent each slope segmentation;

[0110] Each slope of the voltage subsequence or power subsequence constitutes a complete event set, defining all voltage timing data Power Timing Data The event database composed of subsequences of E , the number of each subsequence in the event library is recorded as ;

[0111] definition To indicate the k -1 node and k The correlation between the voltage timing data of the nodes, and Respectively k -1 node and k A subsequence of the voltage timing data of each node, express In the database E The closer the correlation is to 1, the and The higher the correlation:

[0112] ,

[0113] In the formula, For the k -1 node and k The voltage subsequence of each node is and The correlation between express In the database E The number of occurrences in ; for and The confidence level between Represents a database E Also included Also included probability; Represents the node voltage sequence The node voltage sequence The size of the correlation between Represents the node voltage sequence The node voltage sequence Confidence between The number of associated subsequences that satisfy the minimum support;

[0114] Similarly, we can obtain the node k -1 Power timing data and nodes k The correlation between power time series data is , the confidence level is .

[0115] It should be noted that:

[0116] In this embodiment, the numerical interval [-1, 1] of the subsequence slope is divided into 5 segments for symbolic representation, as shown in Table 1:

[0117] ,

[0118] The level of confidence can reflect the credibility of the degree of correlation between the time series of two nodes.

[0119] In order to further implement the above technical solutions, the specific contents of S3 include:

[0120] (1) Generate estimated time series data of data quality abnormality data by using the strongly correlated node power flow calculation, and define the remaining time series data after removing the time series fragment corresponding to the data quality abnormality data as the remaining monitorable time series data, and use the remaining monitorable time series data as the normal time series data; the normal time series data includes normal voltage time series data or normal power time series data;

[0121] (2) Add noise to the data with abnormal data quality to obtain the masked abnormal time series data corresponding to the current diffusion step, and calculate the estimated noise value corresponding to the current diffusion step based on the normal time series data and the estimated time series data;

[0122] (3) Calculate the estimated noise value The actual noise added corresponding to the current diffusion step The difference between Minimize

[0123] The optimization goal of repairing abnormal voltage time series data based on the conditional diffusion model is:

[0124] ,

[0125] In the formula, represents the denoising part of the model; For the diffusion step The masked abnormal time series data obtained after adding noise; is Gaussian noise, It is the normal voltage timing data; To calculate the voltage timing data, estimate the noise value and ; It means taking the average of the probability distribution of normal voltage timing data and Gaussian noise;

[0126] The optimization goal of repairing abnormal power time series data based on the conditional diffusion model is:

[0127] ,

[0128] In the formula, For the diffusion step The masked abnormal power time series data obtained after adding noise; It is the normal power timing data; To calculate the power timing data, estimate the noise value and ; Represents the average of the probability distributions of normal power time series data and Gaussian noise.

[0129] It should be noted that:

[0130] The conditional diffusion model mainly consists of two TThe Markov chain of the forward diffusion process converts the abnormal time series data into noise, and the Markov chain of the reverse diffusion process gradually denoises the noise to generate normal operating time series data. In the present invention, the forward diffusion process based on the conditional diffusion model fills and repairs the data quality abnormal data marked in S2, and uses a bidirectional encoder representation model to perform mask training on the data quality abnormal data time series fragments to convert the abnormal time series data into noise; at the same time, the strong correlation time series of the data quality abnormal data is used to infer the normal operating time series data of the data quality abnormal time series fragments through the flow calculation, and based on the reverse diffusion process, the normal operating time series data fragments unmasked in the node time series and the normal time series data inferred by the strong correlation time series are used to guide the diffusion model to fill and repair the abnormal time series fragments.

[0131] The voltage timing data with masked abnormal timing data is defined as , the power timing data is In the formula, and Abnormal voltage time series data and abnormal power time series data representing the mask; and Indicates normal voltage timing data and normal power timing data;

[0132] Voltage timing estimation data generated by strong correlation node power flow calculation With normal voltage timing data The conditional diffusion model is guided to fill in the abnormal voltage time series data. The model training expression is as follows: ,

[0133] In the formula, Represents the transfer kernel of the reverse Markov chain. The training goal of the model is to use the reverse transfer kernel Data distribution close to real voltage timing data ; express and The joint probability distribution of .

[0134] In each iterative training of the denoising diffusion module, the abnormal data is first sampled from the abnormal voltage time series data , using Gaussian noise With diffusion step The abnormal voltage time series data sampled Masking is performed to normal voltage timing data Voltage timing estimation data The noise value added to guide the calculation of the current diffusion step , and according to the calculated noise value The actual noise value added Perform difference minimization gradient descent. The abnormal voltage time series data output in the previous round As the next round of output , and maintain normal voltage timing data Voltage timing estimation data The abnormal voltage time series data remains unchanged, and the abnormal voltage time series data is gradually repaired to efficiently fill the abnormal voltage time series fragments. The repair steps of the abnormal power time series data are the same as those of the abnormal voltage time series data.

[0135] In order to further implement the above technical solution, the specific contents of S4 include:

[0136] The repaired voltage timing data Power Timing Data Perform a cosine polar transformation, where the timestamps of the time series are converted into radii R , the cosine value is converted into the corresponding angle value , the transformation process is:

[0137] ,

[0138] In the formula, and Represents the repaired node k Voltage and power timing data points; and Representation Node k exist t The angle value of the voltage and power time series data points corresponding to the polar coordinates; Indicates the radius value in polar coordinates; Indicates the timestamp of the corresponding time series data, and divides the timestamp interval into Z The equal division regularizes the polar coordinate span, For timestamp A collection of;

[0139] Based on the Gram angular field imaging algorithm, the voltage time series data and power time series data after cosine value polar coordinate transformation are encoded into a two-dimensional image matrix and :

[0140] ,

[0141] ,

[0142] In the formula, The voltage timing data after repair The transpose of The voltage timing data after repair The transpose of It is a unit and Corresponding angle in polar coordinates;

[0143] Will and Converted into a two-dimensional image, the node voltage timing diagram is obtained Power Timing Diagram , where the tensor of each time series image is , G and F is the spatial resolution of the node timing graph, 1 means the node timing graph is a single-channel image;

[0144] Use convolutional neural network to extract the features of each time series graph :

[0145] ,

[0146] In the formula, The tensors are , concatenate all the feature graphs of the timing graph according to the channel to form a node timing feature matrix , the concatenated feature matrix The size is ;

[0147] Compute node timing feature matrix The mean , and subtract it from all node time series feature graphs to eliminate the impact of the mean:

[0148] ,

[0149] In the formula, Represents the decentralized node time series feature matrix; Represents the node time series feature matrix in k Node, j Elements on a feature;

[0150] calculate The covariance matrix of ,right Perform eigenvalue decomposition to obtain the corresponding eigenvalue With the eigenvector ;

[0151] According to the feature vector Get the projection matrix , use the projection matrix to transform the decentralized node time series feature matrix Project to feature space to get the projected feature matrix ;

[0152] For the fused projection feature matrix Perform inverse transformation to reconstruct it into a feature map , and the feature map Convert to the spatial representation of the image to get the final node temporal fusion image :

[0153] .

[0154] It should be noted that:

[0155] For any given node voltage and power timing data, there is only one corresponding polar coordinate value in the polar coordinate system. Transforming the timing data into polar coordinates can not only fully preserve the data feature information, but also accurately restore the node voltage and power timing data through inverse mapping.

[0156] Voltage timing matrix based on Gram angle field GAF ​​encoding and Power Timing Matrix The correlation between the timing information is retained and enhanced, and the angle and size information of the node timing data is included.

[0157] In order to further implement the above technical solution, the specific contents of S5 include:

[0158] Fusion of nodes in time series The trained illegal electricity consumption recognition model is input to detect image anomalies, and the difference score of each pixel of the two images is calculated using the pixel-by-pixel difference method:

[0159] ,

[0160] Thresholds based on difference score statistics P for:

[0161] ,

[0162] In the formula, for The mean of for The standard deviation of is the adjustment factor;

[0163] The current difference score is greater than the threshold P The pixel points are marked as illegal electricity usage feature points, and are mapped back to the node voltage time series data and the node power time series data to achieve the traceability and positioning of illegal electricity usage in the distribution station area.

[0164] In order to further implement the above technical solution, the specific contents of training the illegal electricity consumption recognition model include:

[0165] The training model for illegal electricity consumption recognition includes a discriminator and a generator. A self-attention mechanism is added to the middle layer of the discriminator and the generator to improve the image generation capability. The training process includes the training of the discriminator and the training of the generator:

[0166] The training process of the discriminator includes:

[0167] Fusion image of the nodes of the current batch in time series Generate images with nodes At the same time, it is input into the discriminator, the corresponding time series feature graph is generated and a layer of convolution operation is performed on it, the discriminator loss is calculated and the parameters are updated:

[0168] ,

[0169] In the formula, Represents the prediction probability of the discriminator for the node temporal fusion feature map; Represents the predicted probability of the discriminator generating a time series feature graph for the node; It is a 0-1 variable, which is 1 when the node time series fusion feature graph is generated, and 0 when the node generates the time series feature graph;

[0170] The training process of the generator includes:

[0171] The obtained node time series fusion image Generate a time series feature vector through the encoder e , and then the time series feature vector e Input decoder to decode and get node generated image , generate the node image Input the encoder again to get the timing generation vector , calculate the overall loss of the generator and perform parameter updates.

[0172] The present invention will be further described below based on specific examples:

[0173] The present invention is applied to a 187-node area, which includes 70 residential users, 110 electric vehicle charging piles and 7 distributed photovoltaics. The results of illegal electricity use identification in the area at a certain time are as follows: Figure 2 The comparison of illegal electricity use identification results using various algorithms is shown in the figure below. Figure 3 As shown, five running methods are considered for comparison:

[0174] 1) Method 1: the method proposed by the present invention;

[0175] 2) Method 2: Isolation forest algorithm-based illegal electricity usage identification method;

[0176] 3) Method 3: Identification method of illegal electricity consumption based on density space clustering algorithm;

[0177] 4) Method 4: Identification method of illegal electricity consumption based on optimal path forest clustering algorithm;

[0178] 5) Method 5: Illegal electricity usage identification method based on shallow neural network algorithm.

[0179] The results show that the recognition accuracy and recall rate of illegal electricity consumption by the proposed method are higher than those of other commonly used algorithms. The recognition accuracy can be increased by at least 5.88%, and the recognition recall rate can be increased by at least 5.41%. The calculation method of the recognition accuracy and the recognition recall rate is:

[0180] ,

[0181] In the formula, Indicates recognition accuracy; represents the recognition recall rate; is the number of nodes correctly identified as illegal electricity users; is the number of nodes correctly identified as operating normally; The number of nodes that are mistakenly identified as violating power usage when they are operating normally; The number of nodes with illegal power usage that are mistakenly identified as normal nodes.

[0182] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. An intelligent identification method for illegal power consumption in distribution substations considering the quality of multi-source measurement data, characterized in that: The following steps are involved: S1. Voltage timing data of each node and power timing data Perform preliminary screening of abnormal data points to obtain abnormal time series data clusters, where the nodes include user nodes, photovoltaic nodes, and charging pile nodes in the substation area; is the length of a single node time series, is the number of topological nodes in the distribution area; S2. Perform correlation analysis on the abnormal time series data cluster marked in S1 by association rule mining to obtain strongly correlated nodes in the abnormal time series data cluster. If only one voltage time series data or power time series data of a strongly correlated node is marked as abnormal data at the same time, the marked abnormal data is determined to be data quality abnormal data; S3. Repair the marked data quality abnormality data based on the conditional diffusion model to obtain repaired voltage time series data and power time series data; S4. Encode and convert the repaired voltage time series data and power time series data to obtain the voltage time series diagram and power time series diagram of each node, extract the features of the voltage time series diagram and the power time series diagram, and perform multi-image fusion to obtain the node time series fusion image ; S5. Fusion of nodes in time series The trained illegal electricity consumption recognition model is input to detect image anomaly points, obtain abnormal pixel points, and map them back to node voltage time series data and node power time series data to achieve the traceability and location of illegal electricity consumption in the distribution station area. The specific contents of S1 include: S11. Obtain voltage timing data of each node Power Timing Data ; S12. Detect repeated data in the voltage time series data and the power time series data by using a sliding window method, and mark data points with a repetition rate greater than a preset repetition threshold as abnormal data points; S13. Perform outlier detection on the voltage time series data and the power time series data based on the local anomaly factor algorithm, and mark the data points whose anomaly factors are greater than the preset anomaly threshold as abnormal data points; S14. Preliminarily mark all abnormal data points as abnormal time series data clusters.

2. The intelligent identification method for illegal power consumption in distribution substations considering the quality of multi-source measurement data according to claim 1 is characterized in that: The specific contents of S12 include: Choose a sliding window size of , sliding the window from the starting position of the voltage time series data or the power time series data, one step forward each time; when the number of occurrences of repeated data in the time series data selected by the sliding window exceeds the preset repetition threshold of the sliding window size, the first repeated data point in the selected time series is abnormal data; The specific contents of S13 include: according to and The data density is calculated and The local accessibility density of each data point in is further calculated to obtain the abnormal factor size of the data point. The abnormal factor size indicates the degree of outlier of each data point, and the data points with abnormal factors greater than the preset abnormal threshold are marked as abnormal data points. The local accessibility density and abnormal factor are: ， In the formula, Indicates the sampling time point; , Indicates voltage and power timing data points The local accessibility density of For data points No. Neighborhood, representing the data point Distance The set of all time series data points within the distance; Indicates The number of data points in the neighborhood; For time series data points To the time series data point No. distance; is the node time series data point The size of the abnormal factor.

3. The intelligent identification method for illegal power consumption in distribution substations considering the quality of multi-source measurement data according to claim 1 is characterized in that: The specific contents of association analysis through association rule mining include: The correlation and confidence between all voltage time series data and power time series data are calculated. When the correlation and confidence of the voltage time series data and power time series data of two nodes meet the preset minimum threshold, it is considered that the time series of the two nodes are strongly correlated, otherwise the time series of the two nodes are considered to be unrelated.

4. The intelligent identification method for illegal power consumption in distribution substations considering the quality of multi-source measurement data according to claim 3 is characterized in that: The specific contents of calculating the correlation and confidence include: The lengths are respectively L The sliding window of the voltage time series data and power time series data to be analyzed for correlation is intercepted as M time series and M power subsequences, and linearly fit the voltage subsequence and the power subsequence respectively, so that the value interval of the slope of each subsequence is [-1,1] and divide it into various slope segments, and symbolically represent each slope segment; Each slope of the voltage subsequence or power subsequence constitutes a complete event set; All voltage timing data or power timing data The event database composed of subsequences of E , the number of each subsequence in the event library is recorded as ; Will To indicate the k -1 node and k The correlation between the voltage timing data or power timing data between nodes, and Respectively k -1 node and k The voltage subsequence or power subsequence of each node, express In the database E The closer the correlation is to 1, the more k -1 node and k The higher the correlation degree of the voltage timing data or power timing data of each node is; ， In the formula, For the k -1 node and k The voltage subsequence or power subsequence of each node is and The correlation between for and The confidence level between Represents a database E Also included Also included probability; express and The size of the correlation or and The size of the correlation ; express and Confidence between or and Confidence between ; n The number of associated subsequences that satisfy the minimum support; Get , , and .

5. The intelligent identification method for illegal power consumption in distribution substations considering the quality of multi-source measurement data according to claim 1 is characterized in that: The specific contents of S3 include: (1) Generate estimated time series data of data quality abnormality data by using the strongly correlated node power flow calculation, and define the remaining time series data after removing the time series fragment corresponding to the data quality abnormality data as the remaining monitorable time series data, and use the remaining monitorable time series data as the normal time series data; the normal time series data includes normal voltage time series data or normal power time series data; (2) Add noise to the data with abnormal data quality to obtain the masked abnormal time series data corresponding to the current diffusion step, and calculate the estimated noise value corresponding to the current diffusion step based on the normal time series data and the estimated time series data; (3) Calculate the estimated noise value The actual noise added corresponding to the current diffusion step The difference between Minimize The optimization goal of repairing abnormal voltage time series data based on the conditional diffusion model is: ， In the formula, represents the denoising part of the model; For the diffusion step The masked abnormal voltage timing data obtained after adding noise; is Gaussian noise, It is the normal voltage timing data; To calculate the voltage timing data, estimate the noise value and ; It means taking the average of the probability distribution of normal voltage timing data and Gaussian noise; The optimization goal of repairing abnormal power time series data based on the conditional diffusion model is: ， In the formula, For the diffusion step The masked abnormal power time series data obtained after adding noise; It is the normal power timing data; To calculate the power timing data, estimate the noise value and ; Represents the average of the probability distributions of normal power time series data and Gaussian noise.

6. The intelligent identification method for illegal power consumption in distribution substations considering the quality of multi-source measurement data according to claim 1 is characterized in that: The specific contents of S4 include: The repaired voltage timing data Power Timing Data Perform a cosine polar transformation, where the timestamps of the time series are converted into radii R , the cosine value is converted into the corresponding angle value , the transformation process is: ， In the formula, and Indicates the node after repair k Voltage and power timing data points; and Representation Node k exist t The angle value of the voltage and power time series data points corresponding to the polar coordinates; Indicates the radius value in polar coordinates; Indicates the timestamp of the corresponding time series data, and divides the timestamp interval into Z The equal division regularizes the polar coordinate span, Timestamp A collection of; Based on the Gram angular field imaging algorithm, the voltage time series data and power time series data after cosine value polar coordinate transformation are encoded into a two-dimensional image matrix and : ， ， In the formula, The voltage timing data after repair The transpose of The voltage timing data after repair The transpose of It is a unit and Corresponding angle in polar coordinates; Will and Converted into a two-dimensional image, the node voltage timing diagram is obtained Power Timing Diagram , where the tensor of each time series image is , G and F is the spatial resolution of the node timing graph, 1 means the node timing graph is a single-channel image; Convolutional neural network is used to extract the features of each time series graph : ， In the formula, The tensors are , concatenate all the feature graphs of the timing graph according to the channel to form a node timing feature matrix , the concatenated feature matrix The size is ; Compute node timing feature matrix The mean , and subtract it from all node time series feature graphs to eliminate the impact of the mean: ， In the formula, Represents the decentralized node time series feature matrix; Represents the node time series feature matrix in k Node, j Elements on a feature; calculate The covariance matrix of ,right Perform eigenvalue decomposition to obtain the corresponding eigenvalue With the eigenvector ; According to the feature vector Get the projection matrix , use the projection matrix to transform the decentralized node time series feature matrix Project to feature space to get the projected feature matrix ; For the fused projection feature matrix Perform inverse transformation to reconstruct it into a feature map , and the feature map Convert to the spatial representation of the image to get the final node temporal fusion image : 。 7. The intelligent identification method for illegal power consumption in distribution substations considering the quality of multi-source measurement data according to claim 1 is characterized in that: The specific contents of S5 include: Fusion of nodes into images Input the trained illegal electricity consumption recognition model to detect image anomalies and output the image , use the pixel-by-pixel difference method to calculate the difference score of each pixel in the two images: ， Thresholds based on difference score statistics P for: ， In the formula, for The mean of for The standard deviation of is the adjustment factor; The current difference score is greater than the threshold P The pixel points are marked as illegal electricity usage feature points, and are mapped back to the node voltage time series data and the node power time series data to achieve the traceability and positioning of illegal electricity usage in the distribution station area.

8. The intelligent identification method for illegal power consumption in distribution substations considering the quality of multi-source measurement data according to claim 7 is characterized in that: The specific contents of training the illegal electricity usage identification model include: The training model for illegal electricity consumption recognition includes a discriminator and a generator. A self-attention mechanism is added to the middle layer of the discriminator and the generator to improve the image generation capability. The training process includes the training of the discriminator and the training of the generator: The training process of the discriminator includes: Fusion image of the nodes of the current batch in time series Generate images with nodes At the same time, it is input into the discriminator, the corresponding time series feature graph is generated and a layer of convolution operation is performed on it, the discriminator loss is calculated and the parameters are updated: ， In the formula, Represents the prediction probability of the discriminator for the node temporal fusion feature map; Represents the predicted probability of the discriminator generating a time series feature graph for the node; It is a 0-1 variable, which is 1 when the node time series fusion feature graph is generated, and 0 when the node generates the time series feature graph; The training process of the generator includes: The obtained node time series fusion image Generate a time series feature vector through the encoder e , and then the time series feature vector e Input decoder to decode and get node generated image , generate the node image Input the encoder again to get the timing generation vector , calculate the overall loss of the generator and perform parameter updates.

Images