A method for detecting abnormal power consumption in a distribution network containing distributed power generation

Abstract

The present invention provides a method for detecting abnormal power consumption in a distribution network containing distributed power sources, comprising the following steps: data acquisition and data preprocessing; constructing an input feature set and converting it into a pixel image; establishing an abnormal power consumption feature extraction and detection model that combines a convolutional neural network, a gated recurrent unit, and a support vector machine; and training the detection model using training set data. Compared to traditional methods in which the input feature set only considers historical load data and temperature and meteorological data, the input feature set of the present invention can more comprehensively reflect the influencing factors of the net load containing distributed power sources, and is more conducive to analyzing the change pattern of the net load, thereby helping to improve the accuracy of abnormal power consumption detection. The present invention fully explores the local characteristics, global characteristics, and time-series change laws of the net load, and reveals the situation in which the net load is affected by meteorological factors, thereby more comprehensively grasping the change pattern of the net load and achieving higher-precision abnormal power consumption detection.

Description

Technical Field

[0001] The present invention relates to the technical field of abnormal power consumption detection in power grids, and in particular to a method for detecting abnormal power consumption in a power distribution network containing distributed power sources. Background Art

[0002] Abnormal electricity usage, such as electricity theft, is a major cause of power transmission and distribution losses, resulting in significant economic losses for power grid companies and a major challenge facing power companies worldwide. Accurate and efficient methods for detecting abnormal electricity usage can quickly and accurately identify electricity theft and mitigate the resulting economic losses, thus being of significant value.

[0003] Existing methods for detecting abnormal electricity usage can be categorized as system-state-based and data-driven. System-state-based methods utilize distribution network measurements such as voltage, current, and node power to estimate the state and identify abnormal electricity users. However, this method requires comprehensive network topology parameters, resulting in high costs and difficulty in practical implementation.

[0004] The data-driven abnormal electricity consumption detection method directly starts from the characteristics of the user's electricity consumption curve and the distribution of electricity, and uses clustering technology and classification technology to identify abnormal electricity users. It has the advantages of low cost, high efficiency, and easy implementation, and is a current research hotspot.

[0005] Existing data-driven abnormal power consumption detection methods usually include three steps:

[0006] (1) Constructing the input feature set, i.e. determining the input data of the detection model;

[0007] (2) Feature extraction, that is, further extracting important features that are beneficial to anomaly detection from the input data;

[0008] (3) Abnormal identification, that is, identifying normal and abnormal electricity usage behaviors based on the extracted important features.

[0009] When it comes to distribution networks with distributed power generation, existing abnormal power consumption detection methods have the following deficiencies in the above links:

[0010] When constructing input feature sets, existing methods do not consider the characteristics of loads containing distributed power sources to construct input features for anomaly detection models. For example, the patent "A Method and System for Detecting Power Consumption Anomalies Based on Deep Neural Networks" (CN202111237405.3) uses historical power consumption data and ambient temperature data as inputs for the anomaly detection model, while the patent "A Method and System for Analyzing Distribution Network Losses Based on Multi-Source Measurement Data" (CN202111433526.5) uses power consumption data, temperature and climate data, and substation information as inputs for the detection model. However, using only this data as an input feature set fails to reflect the characteristics of loads containing distributed power sources. This is because, when connected to distributed power sources, the load is affected by the coupling of multiple factors, including user electricity usage habits, wind turbine or photovoltaic panel rated capacity, wind speed, irradiance, temperature, and so on. Existing abnormal power consumption detection methods do not fully consider these factors, and therefore cannot accurately identify load variation patterns, making them unsuitable for abnormal power consumption detection in distribution networks containing distributed power sources.

[0011] In the feature extraction phase, existing methods for detecting abnormal power consumption use only use models suitable for conventional loads. There are no models specifically designed for distributed power generation (DGs). Consequently, existing methods for detecting abnormal power consumption have low accuracy when applied to distribution networks containing DGs.

[0012] With the access of a large number of distributed renewable energy sources to the load side of current distribution networks, the power load has become a "net load" consisting of actual power consumption minus the power generated by distributed power sources. The net load includes both the actual power consumption of users and the power generated by renewable energy. This means that the power load is affected not only by the user's power consumption behavior, but also by multiple meteorological factors such as wind speed, irradiance, and temperature. The load characteristics become more complex and more uncertain. Existing data-driven abnormal power consumption detection methods mainly target conventional loads with strong regularity and are difficult to apply to net loads with more complex characteristics including distributed power sources. Therefore, it is necessary to develop abnormal power consumption detection methods for distribution networks containing distributed power sources. Summary of the Invention

[0013] The purpose of the present invention is to provide a method for detecting abnormal power consumption in a distribution network containing distributed power sources, which can solve the problem of low detection accuracy of the abnormal power consumption detection method in the prior art when used in a distribution network containing distributed power sources.

[0014] The object of the present invention is achieved through the following technical solutions:

[0015] A method for detecting abnormal power consumption in a distribution network containing distributed power sources comprises the following steps:

[0016] Step S1, data acquisition and data preprocessing;

[0017] Step S2: construct an input feature set and convert it into a pixel image;

[0018] Step S3, establishing a CNN-GRU-SVM model for abnormal power consumption feature extraction and detection, which is a combination of a convolutional neural network (CNN), a gated recurrent unit (GRU), and a support vector machine (SVM);

[0019] Step S4: Use the training set data to train the detection model CNN-GRU-SVM described in step S3.

[0020] Furthermore, the step S1 includes:

[0021] Step S101: Acquire historical net load data, historical meteorological data, parameter data of installed distributed power sources, and user tag data of multiple users in the area to be detected;

[0022] Step S102: For missing data and abnormal mutation data, use the mean interpolation method to fill and replace; for duplicate data, delete the data in and out;

[0023] Step S103: Divide the preprocessed data into a training set and a test set in a ratio of 8:2.

[0024] Furthermore, the construction input feature set includes construction power load features, construction new energy related meteorological features, construction new energy equipment features and construction date features.

[0025] Furthermore, the power load characteristics are divided into load curve image characteristics and load statistical characteristics. The load curve image characteristics are obtained by converting the user net load matrix into a pixel image.

[0026] Furthermore, step S3 includes:

[0027] Step S301: For the load curve image features in the power load features, a combination of n convolutional layers and pooling layers is used to extract features to obtain deep features F1;

[0028] Step S302: For the load statistical features in the power load features, k GRU networks are used to extract features to obtain deep features F2;

[0029] Step S303: For new energy related meteorological features, a combination of m convolutional layers and pooling layers is used to extract features to obtain deep features F3;

[0030] Step S304: For new energy equipment features, a fully connected Dense network is used to extract input features to obtain deep features F4;

[0031] Step S305: For date features, a fully connected Dense network is used to extract the input features to obtain deep features F5;

[0032] Step S306: Input the deep features F1, F2, and F3 into the Dense network for further feature extraction to obtain a new deep feature F6;

[0033] Step S307: input the deep features F4, F5, and F6 into the Dense network for further feature extraction to obtain a new deep feature F7;

[0034] Step S308: Input the obtained new deep features F6 and F7 into the SVM classifier to obtain the recognition result of the user's electricity usage behavior.

[0035] Furthermore, the calculation formula of the convolution layer in step S301 is:

[0036] c i =f c (A i *w c,i +b c,i );

[0037] Among them, c i is the output of the i-th convolutional layer, i = 1, 2, ..., n; A i 、w c,i 、b c,i are the input, weight, and bias of the i-th convolutional layer respectively; f c is the activation function of the convolutional layer; * represents the dot product calculation, that is, the corresponding elements of the two matrices are multiplied and then accumulated.

[0038] Furthermore, the calculation formula of the pooling layer in step S301 is:

[0039] d i =w d,i *z i ;

[0040] Among them, d i is the output of the i-th pooling layer; w d,i is the weight of the i-th pooling layer; z i It is d i The corresponding pooling area.

[0041] Furthermore, the calculation formula of the GRU network in step S302 is as follows:

[0042]

[0043] Among them, r t、o t 、 h t are the outputs of the reset gate, update gate, candidate gate, and state gate of the t-th step input in the GRU unit; W r 、W o 、 They are the weights of reset gate, update gate and candidate gate respectively; x t is the input data of GRU; t is the output data of GRU; σ is the activation function.

[0044] Furthermore, in step S4, the test set or the data set to be tested is input into the trained detection model CNN-GRU-SVM to obtain the identification result of abnormal power consumption; and the identification result is evaluated using the precision rate and recall rate indicators.

[0045] Furthermore, the precision is defined as:

[0046]

[0047] The recall is defined as:

[0048]

[0049] Where TP is the number of samples identified as normal users that are actually normal users; FP is the number of samples identified as normal users that are actually abnormal users; and FN is the number of samples identified as abnormal users that are actually normal users.

[0050] The present invention constructs an input feature set that includes power load characteristics, new energy-related meteorological characteristics, new energy equipment characteristics, and date characteristics. In particular, constructing power load characteristics and new energy-related meteorological characteristics using pixel images facilitates the extraction of potential electricity usage information by the CNN network. Compared to traditional methods that only consider historical load data and temperature and meteorological data as input feature sets, the input feature set of the present invention can more comprehensively reflect the factors affecting the net load of distributed power sources, making it more conducive to analyzing net load variation patterns and thus helping to improve the accuracy of abnormal power consumption detection.

[0051] The present invention constructs an abnormal power consumption feature extraction and detection model (CNN-GRU-SVM) based on the characteristics of the net load containing distributed power sources. In particular, a multi-layer CNN network is used to perform deep feature extraction on the load curve image features and new energy-related meteorological features, deeply exploring the local and global variation patterns of net load and meteorology; a GRU network is used to perform deep feature extraction on the load statistical features, exploring the temporal variation patterns of net load; a Dense network is used to fuse the extracted net load features and meteorological features, which helps to explore the correlation between net load and meteorology and improve detection accuracy; a Dense network is used to further fuse the extracted deep-level features, and then input into the SVM classifier to achieve high-precision abnormal power consumption identification.

[0052] The CNN-GRU-SVM model proposed in this invention can fully explore the local characteristics, global characteristics, and time-series variation patterns of the net load, and reveal the influence of meteorological factors on the net load, thereby more comprehensively grasping the variation pattern of the net load and achieving high-precision abnormal power consumption detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 Schematic diagram of a pixel image converted from a user payload matrix according to the present invention;

[0054] Figure 2 Schematic diagram of pixel image converted from a matrix formed by meteorological data according to the present invention;

[0055] Figure 3 Schematic diagram of the detection model CNN-GRU-SVM of the present invention. DETAILED DESCRIPTION

[0056] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0057] The following describes the embodiments of the present disclosure through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present disclosure.

[0058] The present invention provides a method for detecting abnormal power consumption in a distribution network containing distributed power sources, comprising the following steps:

[0059] Step S1: data acquisition and data preprocessing.

[0060] Specifically, step S1 includes:

[0061] Step S101: Acquire historical net load data, historical meteorological data, parameter data of installed distributed power sources, and user tag data of multiple users in the area to be detected.

[0062] Step S102: For missing data and abnormal mutation data, the mean interpolation method is used to fill and replace them; for duplicate data, in and out deletion is performed.

[0063] Step S103: Divide the preprocessed data into a training set and a test set in a ratio of 8:2.

[0064] Step S2: construct an input feature set and convert it into a pixel image.

[0065] Furthermore, constructing the input feature set includes constructing power load features, constructing new energy related meteorological features, constructing new energy equipment features, and constructing date features. The details are as follows:

[0066] Furthermore, the power load characteristics are divided into load curve image characteristics and load statistical characteristics: the load curve image characteristics are obtained by converting the user net load matrix into a pixel image. Each column in the user net load matrix corresponds to the energy consumption at a certain time of the day, and each row corresponds to a different day. For example, the user's 2-week net load data constitutes a 14*24 matrix, where 14 is the number of days and 24 is the load at 24 o'clock every day. After converting the matrix into a pixel image, the following can be obtained: Figure 1 Picture shown.

[0067] The advantage of using pixel images to represent load curve features is that after converting load features into pixel images, the lines, shapes, and colors of the images contain information about user electricity usage behavior. This information can then be captured using convolutional neural networks (CNNs), which are known for their translation-invariant properties. Compared to numerical matrices, CNNs are more adept at extracting information from images, so converting load data into pixel images is more conducive to extracting load features.

[0068] Furthermore, the load statistical characteristics include: daily power consumption, daily maximum load, daily minimum load, and daily average load. The load statistical characteristics reflect the distribution of user net load.

[0069] Furthermore, the meteorological characteristics related to new energy include wind speed, solar irradiance, temperature, humidity and other meteorological data corresponding to the characteristic time of power load. The matrix formed by these meteorological data is converted into a pixel image. Taking the solar irradiance data as an example, after conversion into a pixel image, as shown in the attached figure Figure 2 shown.

[0070] Furthermore, new energy equipment characteristics include the maximum capacity, rated capacity, rated temperature, and fan cut-out wind speed of distributed power generation equipment. These parameters are closely related to the power generated by distributed power generation. Using these as input feature data helps analyze net load variations and improves the accuracy of abnormal power consumption detection.

[0071] Furthermore, date features include date, week, month, and quarter information. Consumer electricity usage and distributed generation characteristics vary across different dates, weeks, months, and quarters. Using these date features as input helps understand net load variations and improves the accuracy of abnormal power usage detection.

[0072] Step S3: Establish an abnormal power consumption feature extraction and detection model (CNN-GRU-SVM) that combines a convolutional neural network (CNN), a gated recurrent unit (GRU) and a support vector machine (SVM).

[0073] The established CNN-GRU-SVM model is shown in the attached Figure 3 As shown, step S3 includes:

[0074] Step S301: For the load curve image features in the power load features, a combination of n convolutional layers and pooling layers is used to extract features, and the deep features extracted from the load curve image by CNN are recorded as F1.

[0075] Among them, the calculation formula of the convolution layer is:

[0076] c i =f c (A i *w c,i +b c,i )(1)

[0077] (1) In the formula, c i is the output of the i-th convolutional layer, i = 1, 2, ..., n; A i 、w c,i 、b c,i are the input, weight, and bias of the i-th convolutional layer respectively; f cis the activation function of the convolutional layer; * represents the dot product calculation, that is, the corresponding elements of the two matrices are multiplied and then accumulated.

[0078] The calculation formula of the pooling layer is:

[0079] d i =w d,i *z i (2)

[0080] (2) In the formula, d i is the output of the i-th pooling layer; w d,i is the weight of the i-th pooling layer; z i It is d i The corresponding pooling area.

[0081] Step S302: For the load statistical features in the power load features, k GRU networks are used to extract features, and the deep features extracted by GRU from the load statistical features are recorded as F2.

[0082] Furthermore, the calculation formula of the GRU network is as follows:

[0083]

[0084] (3) In the formula, r t 、o t 、 h t are the outputs of the reset gate, update gate, candidate gate, and state gate of the t-th step input in the GRU unit; W r 、W o 、 They are the weights of reset gate, update gate and candidate gate respectively; x t is the input data of GRU; t is the output data of GRU; σ is the activation function.

[0085] Step S303: For new energy related meteorological features, a combination of m convolutional layers and pooling layers is used to extract features.

[0086] The deep features extracted by CNN from the meteorological features related to new energy are denoted as F3.

[0087] Step S304: For new energy equipment features, a fully connected Dense network is used to extract input features.

[0088] The deep features extracted by the Dense network from the new energy equipment features are denoted as F4.

[0089] Step S305: For date features, a fully connected Dense network is used to extract input features.

[0090] The deep features extracted by the Dense network from the date feature are denoted as F5.

[0091] Step S306: Input the deep features F1, F2 and F3 into the Dense network for further feature extraction to obtain a new deep feature, which is recorded as F6.

[0092] Step S307: Input the deep features F4, F5 and F6 into the Dense network for further feature extraction to obtain a new deep feature, which is recorded as F7.

[0093] Step S308: Input the extracted comprehensive features F6 and F7 into the SVM classifier to obtain the recognition result of the user's electricity usage behavior.

[0094] Step S4: Using the training set data, train the CNN-GRU-SVM model described in step S3. The hyperparameters n, k, and m mentioned in step 3, as well as the number of hidden layer neurons in each CNN network, GRU network, and Dense network, are optimized using a grid search method.

[0095] The test set or the dataset to be tested is fed into the trained model to obtain the abnormal power usage identification results. The identification results are evaluated using the precision and recall metrics.

[0096] The precision is defined as:

[0097]

[0098] Recall is defined as:

[0099]

[0100] Where TP is the number of samples identified as normal users that are actually normal users; FP is the number of samples identified as normal users that are actually abnormal users; and FN is the number of samples identified as abnormal users that are actually normal users.

[0101] A simulation analysis is performed using a certain net load data set as an example, where the total number of samples is 4619, the number of normal users is 2772, and the number of abnormal users is 1874. The abnormal power consumption detection is performed using the method of the present invention, and the results are shown in Table 1 below:

[0102] Table 1 Detection accuracy of the method of the present invention for a certain example

[0103]

[0104] The average precision is 0.87 and the average recall is 0.83. It can be seen that the method of the present invention has a high accuracy in detecting abnormal power consumption for net loads containing distributed power sources.

[0105] In the present invention, unless otherwise expressly specified or limited, a first feature "on" a second feature may mean that the first and second features are in direct contact, or that the first and second features are in indirect contact through an intermediary. "Multiple" means at least two, such as two or three, unless otherwise expressly specified or limited.

[0106] In the present invention, unless otherwise specified or limited, the terms "installed," "connected," "connect," "fixed," etc. should be understood in a broad sense. For example, they can refer to fixed connection, detachable connection, or integration; mechanical connection, electrical connection, or communication; direct connection or indirect connection through an intermediate medium; internal communication between two elements or interaction between two elements, unless otherwise specified. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0107] The above is only for explaining the embodiments of the present invention and is not intended to limit the present invention. For those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present invention without creative work should be included in the scope of protection of the present invention.

Claims

1. A method for detecting abnormal power consumption in a distribution network containing distributed power sources, characterized in that: The following steps are involved: Step S1, data acquisition and data preprocessing; Step S2: construct an input feature set and convert it into a pixel image; Step S3, establishing a detection model CNN-GRU-SVM composed of a convolutional neural network CNN, a gated recurrent unit GRU, and a support vector machine SVM combined with abnormal power consumption feature extraction; Step S4: Using the training set data, train the detection model CNN-GRU-SVM described in step S3; The step S3 comprises: Step S301: For the load curve image features in the power load features, a combination of n convolutional layers and pooling layers is used to extract features to obtain deep features F1; The calculation formula of the convolutional layer is: c i =f c (A i *w c,i +b c,i ); Among them, c i is the output of the i-th convolutional layer, i = 1, 2, ..., n; A i 、w c,i 、b c,i are the input, weight, and bias of the i-th convolutional layer respectively; f c is the activation function of the convolution layer; * represents the dot product calculation, that is, the corresponding elements of the two matrices are multiplied and then accumulated; The calculation formula of the pooling layer is: d i =w d,i *z i ; Among them, d i is the output of the i-th pooling layer; w d,i is the weight of the i-th pooling layer; z i It is d i The corresponding pooling area; Step S302: For the load statistical features in the power load features, k GRU networks are used to extract features to obtain deep features F2; The calculation formula of the GRU network is as follows: Among them, r t 、o t 、 h t are the outputs of the reset gate, update gate, candidate gate, and state gate of the t-th step input in the GRU unit; W r 、W o 、 They are the weights of reset gate, update gate and candidate gate respectively; x t is the input data of GRU; t is the output data of GRU; σ is the activation function; Step S303: For new energy related meteorological features, a combination of m convolutional layers and pooling layers is used to extract features to obtain deep features F3; Step S304: For new energy equipment features, a fully connected Dense network is used to extract input features to obtain deep features F4; Step S305: For date features, a fully connected Dense network is used to extract the input features to obtain deep features F5; Step S306: Input the deep features F1, F2, and F3 into the Dense network for further feature extraction to obtain a new deep feature F6; Step S307: input the deep features F4, F5, and F6 into the Dense network for further feature extraction to obtain a new deep feature F7; Step S308: Input the obtained new deep features F6 and F7 into the SVM classifier to obtain the recognition result of the user's electricity usage behavior.

2. The abnormal power consumption detection method for a distribution network containing distributed power sources according to claim 1, characterized in that: The step S1 comprises: Step S101: Acquire historical net load data, historical meteorological data, parameter data of installed distributed power sources, and user tag data of multiple users in the area to be detected; Step S102: For missing data and abnormal mutation data, use the mean interpolation method to fill and replace; for duplicate data, delete the data in and out; Step S103: Divide the preprocessed data into a training set and a test set in a ratio of 8:

2.

3. The abnormal power consumption detection method for a distribution network containing distributed power sources according to claim 1, characterized in that: The construction input feature set includes construction power load features, construction new energy related meteorological features, construction new energy equipment features and construction date features.

4. The abnormal power consumption detection method for a distribution network containing distributed power sources according to claim 3, characterized in that: The power load characteristics are divided into load curve image characteristics and load statistical characteristics. The load curve image characteristics are obtained by converting the user net load matrix into a pixel image.

5. The abnormal power consumption detection method for a distribution network containing distributed power sources according to claim 1, characterized in that: In step S4, the test set or the data set to be tested is input into the trained detection model CNN-GRU-SVM to obtain the identification result of abnormal power consumption; and the identification result is evaluated using the precision rate and recall rate indicators.

6. The abnormal power consumption detection method for a distribution network containing distributed power sources according to claim 5, characterized in that: The precision is defined as: The recall is defined as: Where TP is the number of samples identified as normal users that are actually normal users; FP is the number of samples identified as normal users that are actually abnormal users; and FN is the number of samples identified as abnormal users that are actually normal users.

Images