Distribution network measurement data anomaly detection method based on neural network model

Abstract

The invention provides a distribution network measurement data anomaly detection method based on a neural network model. The method is oriented to an SCADA distribution network on-line topology model, a power grid on-line topology graph set is constructed through a cut set based on a graph theory, and an input form of a GCN graph convolutional network model expressed as a feature vector is formed; forming a training sample set and an input data set of the GCN network based on the distribution network topology cut set, and carrying out model training and testing by adopting a GCN graph convolutional network to determine equipment possibly containing error measurement; and outputting the abnormal nodes in sequence to form the input of the CNN equipment measurement error discrimination model, and outputting an abnormal type after discrimination is completed, thereby realizing the detection of the power distribution network measurement data abnormity. According to the method, high-quality clean data can be provided for establishment of a power distribution network abnormal data detection model, the model precision is high, the prediction effect is good, the detection delay is reduced, the robustness of the method is high, the accuracy of abnormal data detection is ensured, and the final detection precision is improved.

Description

Technical Field

[0001] The present invention relates to the field of photovoltaic power generation, and in particular to a method for detecting abnormalities in distribution network measurement data based on a neural network model. Background Art

[0002] The high proportion of new energy access has led to changes in the power system structure, and the abnormal identification of distribution network data has become more important. Abnormal data at the distribution network equipment level mainly comes from various parameters and indicators in the equipment operation status, power transmission and distribution process. The abnormality of these data may indicate equipment failure, energy efficiency problems or safety hazards. Abnormal data in the distribution network flow mainly involves the steady-state distribution of parameters such as voltage, current, and power. The abnormality of these parameters may indicate failures, energy efficiency problems or safety hazards in the distribution network. Abnormal data usually has significant abnormal values ​​or abnormal patterns, which are statistically or logically significantly different from normal data. There are many possible reasons for the occurrence of abnormal power quality data, such as abnormal network access, incorrect collection of equipment information, abnormalities in the operation of external computer hardware and software systems, and human errors. Therefore, how to effectively identify abnormal data during the operation of the distribution network has become one of the key issues that need to be solved.

[0003] Taking the distribution network as the main research object, some scholars started with analytical models and used mechanism deduction method to analyze the problem of abnormal data identification during the operation of the distribution network; in addition, experimental identification was also carried out through neural network methods. Summary of the invention

[0004] In order to solve the above problems, the present invention provides a distribution network measurement data anomaly detection method based on a neural network model. The method uses the SCADA system to collect distribution network data, and after preprocessing, the GCN-CNN model is used to analyze and extract features of the data, thereby realizing the identification of abnormal data in the distribution network.

[0005] Preferably, the method comprises:

[0006] S1: Build a distribution network online topology model for SCADA, and generate a grid topology cut set according to the equipment topology distance less than 3;

[0007] S2. Based on graph theory, the online topology atlas of the power grid is constructed through the above cut sets. The devices contained in different online topology graphs are divided according to the connection relationship, thereby obtaining the feature matrix X and adjacency matrix A of the graph structure as the input of the GCN graph convolutional network model;

[0008] S3. Build an abnormal device identification model, construct a training sample set and input data set of the GCN network based on the distribution network topology cut set, and use the GCN graph convolutional network to train and test the model to determine the abnormal nodes in each cut set, that is, the devices that may contain erroneous measurements;

[0009] S4: Construct a device measurement error discrimination model, sequentially output the abnormal nodes containing the output feature vector in step S3, constitute the input of the CNN device measurement error discrimination model, perform device measurement error discrimination, and realize the detection of distribution network measurement data anomalies.

[0010] Preferably, the step S1 comprises:

[0011] The extreme gradient boosting tree XGBoost is used to measure the importance of each variable in the SCADA system, where the number of feature splits in the tree model is used as a criterion for measuring the above importance, thereby determining the variables that can effectively characterize the power grid state, namely the active power, reactive power, current and voltage of the equipment in a certain period of time as the model input;

[0012] The SCADA system is used to collect data from the power distribution network, including the active power, reactive power, current and voltage of the equipment during a certain period of time, and the collected data is input into the database for data training.

[0013] Preferably, step S2 comprises:

[0014] Based on graph theory, the online power grid topology atlas is constructed through the above-mentioned power grid topology cut set. The devices contained in different online topology graphs are divided according to the connection relationship and expressed in the form of "node-edge". Each device can be regarded as a node of the graph, and the connection relationship is the edge. Each device has a corresponding feature vector, which is composed of the measurement data of the device in a certain period of time <active power, reactive power, current, voltage>. In this way, the feature matrix X and adjacency matrix A of the graph structure can be obtained as the input of the GCN graph convolutional network model.

[0015] Preferably, the GCN graph convolutional network model consists of an input layer, two graph convolutional layers, a pooling layer, two fully connected layers and an activation function; and step S3 includes:

[0016] S21: In the first graph convolution layer, the feature matrix X of the electric quantity data graph structure and the normalized adjacency matrix A" are convolved to obtain the output feature matrix

[0017]

[0018] A”=D -1 / 2 A'D -1 / 2

[0019] A′=A+IN

[0020] Where: W 1 G and b 1 G are the weight matrix and bias vector of the first graph convolutional layer respectively; H 1 G is the output vector of the first graph convolutional layer; ReLU represents the activation function; I N is the identity matrix of order N; E is the edge set, e ij represents the edge connecting nodes i and j, A=(a ij ) n×n is the adjacency matrix of the graph, if e ij ∈E, then a ij =1, otherwise, ai j =0; n represents the number of nodes; D is the degree matrix of the graph, which is a diagonal matrix with the following diagonal elements:

[0021]

[0022] S22: Output data obtained after the second graph convolution layer:

[0023]

[0024] Where: and They are the weight matrix and bias vector of the second graph convolutional layer respectively; is the output vector of the second graph convolutional layer;

[0025] S23: In the fully connected layer, output data The weight matrix W 3 G , bias vector And the activation function ReLU gets:

[0026]

[0027] S24: The output result of the last layer is:

[0028]

[0029] Where: and are the weight matrix and bias vector of the last layer respectively; Y G is the output feature vector of the GCN module; the output feature vector includes the measurement data of the abnormal device node.

[0030] Preferably, step S4 includes: the measurement data of the abnormal node are sampled in a short time window respectively, and then represented in a time series graphical form, expressed as a graph of 50 sampling points according to a time series line graph, each measurement graph is spliced ​​into a picture to be identified, the measurement data features in the picture are extracted to generate a feature matrix X', which is input into the CNN device measurement error discrimination model for discrimination, and the error type is output after the discrimination is completed, so as to realize the detection of abnormal measurement data of the distribution network.

[0031] Preferably, in step S4,

[0032] CNN convolutional neural network model, including input layer, convolution layer 1, convolution layer 2, pooling layer, fully connected layer and output layer;

[0033] In the first convolutional layer, the input data is the feature matrix X', and the output data is:

[0034]

[0035] Where: W 1 C and are the weight matrix and bias vector respectively.

[0036] Similar to the first convolutional layer, the output features after the second convolutional layer are:

[0037]

[0038] Where: and They are the weight matrix and bias vector of the second convolutional layer respectively.

[0039] Using maximum pooling in the pooling layer can reduce the parameters of CNN, and the output features are:

[0040]

[0041] In the last fully connected layer of CNN, the output is:

[0042]

[0043] Where: and are the weight matrix and bias vector of the fully connected layer respectively; Y C is the output feature vector of the CNN module; after extracting the information of the output abnormal data feature vector, it outputs <overload, heavy load, low voltage, white noise, data missing, three-phase imbalance, single-phase current large / small> to realize the detection of abnormal distribution network measurement data.

[0044] This method preprocesses the abnormal data collected by SCADA, and can provide high-quality clean data for the establishment of the distribution network abnormal data detection model; the GCN-CNN model is used to detect abnormal data in the power grid. Compared with other comparison models, it has higher accuracy and better prediction effect, reduces detection delay, enhances the robustness of the overall method, ensures the accuracy of abnormal data detection, and achieves the improvement of final detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation on the embodiments of the present invention;

[0046] Figure 1 A flowchart of a method according to the present invention;

[0047] Figure 2 FIG. 1 is a block diagram of a GCN-CCN model according to the present invention. DETAILED DESCRIPTION

[0048] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0049] The present invention provides a method for detecting abnormality in distribution network measurement data based on a neural network model. The method includes four parts, namely, data input and preprocessing, abnormal equipment identification model, abnormal measurement judgment model and output:

[0050] like Figure 1-2 As shown, the method specifically comprises the following steps:

[0051] S1: Build a distribution network online topology model for SCADA, and generate a grid topology cut set according to the equipment topology distance less than 3;

[0052] In order to improve the generalization ability of the model, the extreme gradient boosting tree XGBoost is used to measure the importance of each variable in the SCADA system, and several variables that can effectively characterize the power grid state are determined as model input.

[0053] The splitting process of XGBoost starts from the root node, splits it into leaf nodes, and builds a tree structure by repeating this process. By calculating the gain of the structure score, the best feature is selected as the split point. Therefore, the number of feature splits in the tree model can be regarded as a criterion for measuring feature importance. The formula for defining feature weights is:

[0054]

[0055] Where: W frepresents the weight of feature f in the model; T represents the total number of model decision trees; n i It represents the number of times feature f is split in the i-th decision tree. The sum of the splits of the feature in all decision trees of the model is the feature weight.

[0056] According to the number of feature splitting in the tree model as the criterion for measuring feature importance, the variables that can effectively characterize the state of the power grid, namely the active power, reactive power, current and voltage of the equipment in a certain period of time, are determined as model inputs.

[0057] The distribution network data collection process based on deep neural networks uses the SCADA system to collect data in the distribution network, including the active power, reactive power, current and voltage of the equipment in a certain period of time, and inputs the collected data into the database for data training.

[0058] S2. Based on graph theory, the online topology atlas of the power grid is constructed through the above-mentioned power grid topology cut set. First, each device can be regarded as a node of the graph; secondly, the devices contained in different online topology graphs are divided according to the connection relationship, which is expressed in the form of "node-edge". The device is the node and the connection relationship is the edge. Each device has a corresponding feature vector, which is composed of the measurement data of the device in a certain period of time <active power, reactive power, current, voltage>. In this way, the feature matrix X and adjacency matrix A of the graph structure can be obtained as the input of the GCN graph convolutional network model.

[0059] The graph structure G is represented by an ordered pair (V, E). V is called the node set, v i ∈V represents node i, ||V||=n represents the number of nodes; E is called the edge set, e ij represents the edge connecting nodes i and j.

[0060] The characteristic information of the graph structure G = (V, E) is expressed in the form of a Laplacian matrix, which is expressed as follows:

[0061] L=DA

[0062] The normalized Laplace equation is expressed as:

[0063] L sys =I N -D -1 / 2 AD -1 / 2

[0064] Where: L represents the Laplace matrix, L sys represents the normalized Laplace matrix, A=(a ij ) n×n is the adjacency matrix of the graph, if e ij ∈E, then a ij =1, otherwise, ai j =0;IN is the identity matrix of order N; D is the degree matrix of graph G, which is a diagonal matrix with the following diagonal elements:

[0065]

[0066] S3. Construct an abnormal device identification model, construct the training sample set and input data set of the GCN network based on the distribution network topology cut set, and use the GCN graph convolutional network for model training and testing to determine the abnormal nodes in each cut set, that is, the devices that may contain erroneous measurements.

[0067] The GCN graph convolutional network module is mainly composed of an input layer, a graph convolution layer, a pooling layer, a fully connected layer and an activation function. This embodiment takes a GCN model with two graph convolution layers and two fully connected layers as an example to illustrate the change process of data flow in the GCN module.

[0068] When extracting features, in order to consider the node's own information, the adjacency matrix A needs to be modified according to the following formula, thereby adding a self-connection to each node.

[0069] A′=A+I N

[0070] In the first graph convolution layer, the feature matrix X of the electric quantity data graph structure and the normalized adjacency matrix A" are convolved to obtain the output feature matrix

[0071]

[0072] A”=D -1 / 2 A'D -1 / 2

[0073] A′=A+I N

[0074] Where: W 1 G and b 1 G are the weight matrix and bias vector of the first graph convolutional layer respectively; H 1 G is the output vector of the first graph convolutional layer; ReLU represents the activation function; I N is the identity matrix of order N; E is the edge set, e ij represents the edge connecting nodes i and j, A=(a ij ) n×n is the adjacency matrix of the graph, if e ij ∈E, then a ij =1, otherwise, ai j=0; n represents the number of nodes; D is the degree matrix of the graph, which is a diagonal matrix with the following diagonal elements:

[0075]

[0076] Similarly, the output data obtained after the second graph convolution layer is:

[0077]

[0078] Where: and They are the weight matrix and bias vector of the second graph convolutional layer respectively; is the output vector of the second graph convolutional layer.

[0079] In the fully connected layer, the output data The weight matrix W 3 G , bias vector And the activation function ReLU gets:

[0080]

[0081] Similarly, the output of the last layer can be obtained as follows:

[0082]

[0083] Where: and are the weight matrix and bias vector of the last layer respectively; Y G is the output feature vector of the GCN module. The output feature vector includes the measurement data of the abnormal device node.

[0084] In the above data processing process, the feature matrix X and the normalized adjacency matrix A' are used as the input of the GCN module, which contains the periodicity and time correlation information of the power data. After being processed by the multi-layer graph convolution layer, the output feature vector Y G The periodic laws and time-correlation characteristics of these electrical quantity data can be extracted from them.

[0085] S4: Construct a device measurement error discrimination model, and convert the output feature vector Y contained in step S3 into GThe abnormal nodes are sequentially output to form the input of the CNN device measurement error discrimination model. The measurement data of the abnormal nodes are sampled in a short time window and then represented in a time series graphical form. The time series line graph is expressed as a graph of 50 sampling points. Each measurement graph is spliced ​​into a picture to be identified. The measurement data features in the picture are extracted to generate a feature matrix X, which is input into the CNN device measurement error discrimination model for discrimination. After the discrimination is completed, <overload, heavy load, low voltage, white noise, data missing, three-phase imbalance, single-phase current large / small...> is output to realize the detection of abnormal distribution network measurement data.

[0086] The specific steps of S4 are:

[0087] The CNN convolutional neural network model used in the present invention includes an input layer, a convolutional layer 1, a convolutional layer 2, a pooling layer, a fully connected layer and an output layer.

[0088] In the first convolutional layer, the input data is the feature matrix X', and the output data is:

[0089]

[0090] Where: W 1 C and are the weight matrix and bias vector respectively.

[0091] Similar to the first convolutional layer, the output features after the second convolutional layer are:

[0092]

[0093] Where: and They are the weight matrix and bias vector of the second convolutional layer respectively.

[0094] Using maximum pooling in the pooling layer can reduce the parameters of CNN, and the output features are:

[0095]

[0096] In the last fully connected layer of CNN, the output is:

[0097]

[0098] Where: and are the weight matrix and bias vector of the fully connected layer respectively; Y C is the output feature vector of the CNN module.

[0099] After extracting the information of the output abnormal data feature vector, the output is <overload, heavy load, low voltage, white noise, data missing, three-phase imbalance, single-phase current large / small...>, realizing the detection of abnormal distribution network measurement data.

[0100] In the process of designing a distribution network data anomaly recognition algorithm using a convolutional neural network, the difference features are classified according to the collected distribution network anomaly data samples, the characteristics of the distribution network anomaly data are identified, and cluster analysis is performed on the characteristics of each type of distribution network anomaly data to set the recognition threshold of the distribution network anomaly data.

[0101] Specifically, the online topology model of the distribution network for SCADA first generates a power grid topology cut set according to the device topology distance less than 3; then, based on graph theory, the online topology atlas of the power grid is constructed through the cut set, which is expressed in the form of "node-edge". The device is the node, the connection relationship is the edge, and each device has a corresponding feature vector. The feature vector is composed of the measurement data of the device in a certain period of time <active power, reactive power, current, voltage>, which is expressed as the input form of the GCN graph convolutional network model; the training sample set and input data set of the GCN network are constructed based on the distribution network topology cut set, and the GCN graph convolutional network is used for model training and Test to determine the abnormal nodes in each cut set, that is, the devices that may contain erroneous measurements; the abnormal nodes are output sequentially to form the input of the CNN device measurement error discrimination model, the measurement data of the abnormal nodes are sampled in a short time window, and then represented in a time series graphical form, expressed as a graph of 50 sampling points according to the time series line graph, and the various measurement graphs are spliced ​​into a picture to be identified, which is input into CNN for discrimination. After the discrimination is completed, <overload, heavy load, low voltage, white noise, data missing, three-phase imbalance, single-phase current large / small...> is output to realize the detection of abnormal distribution network measurement data.

[0102] This method preprocesses SCADA abnormal data and can provide high-quality clean data for the establishment of distribution network abnormal data detection model; combined with mechanism analysis to select the input and output parameters of the model, the GCN-CNN model is used to detect abnormal data of the power grid. Compared with other comparative models, it has higher accuracy and better prediction effect, reduces detection delay, enhances the robustness of the overall method, ensures the accuracy of abnormal data detection, and achieves the improvement of final detection accuracy.

[0103] The method proposed in the present invention is simple and easy to implement, has high reliability and strong adaptability, does not require any hardware equipment investment, and has good economy and practicality. According to different actual engineering needs, combined with actual data, the operation data information can be further explored to increase attention to abnormal data.

[0104] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When the use is implemented in whole or in part in the form of a computer program product, the computer program product includes one or more computer instructions. When the computer program instructions are loaded or executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL) or wireless (e.g., infrared, wireless, microwave, etc.) mode) to another website site, computer, server or data center. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state hard disk SolidState Disk (SSD)), etc.

[0105] The above are only preferred embodiments of the present invention, but do not limit the patent scope of the present invention. Although the present invention is described in detail with reference to the above embodiments, those skilled in the art can still modify the technical solutions recorded in the above embodiments, or replace some of the technical features with equivalent ones. Any equivalent structure made by using the contents of the present invention specification and drawings, directly or indirectly used in other related technical fields, is also within the scope of patent protection of the present invention.

Claims

1. A method for detecting abnormality in distribution network measurement data based on a neural network model, characterized in that: This method uses the SCADA system to collect distribution network data, preprocesses it, and then uses the GCN-CNN model to analyze and extract features of the data, thereby realizing the identification of abnormal data in the distribution network.

2. The method according to claim 1, characterized in that The method comprises: S1: Build a distribution network online topology model for SCADA, and generate a grid topology cut set according to the equipment topology distance less than 3; S2. Based on graph theory, the online topology atlas of the power grid is constructed through the above cut sets. The devices contained in different online topology graphs are divided according to the connection relationship, thereby obtaining the feature matrix X and adjacency matrix A of the graph structure as the input of the GCN graph convolutional network model; S3. Build an abnormal device identification model, construct a training sample set and input data set of the GCN network based on the distribution network topology cut set, and use the GCN graph convolutional network to train and test the model to determine the abnormal nodes in each cut set, that is, the devices that may contain erroneous measurements; S4: Construct a device measurement error discrimination model, sequentially output the abnormal nodes containing the output feature vector in step S3, constitute the input of the CNN device measurement error discrimination model, perform device measurement error discrimination, and realize the detection of distribution network measurement data anomalies.

3. The method according to claim 2, characterized in that The step S1 comprises: The extreme gradient boosting tree XGBoost is used to measure the importance of each variable in the SCADA system, where the number of feature splits in the tree model is used as a criterion for measuring the above importance, thereby determining the variables that can effectively characterize the power grid state, namely the active power, reactive power, current and voltage of the equipment in a certain period of time as the model input; The SCADA system is used to collect data from the power distribution network, including the active power, reactive power, current and voltage of the equipment during a certain period of time, and the collected data is input into the database for data training.

4. The method according to claim 3, characterized in that The step S2 comprises: Based on graph theory, the online power grid topology atlas is constructed through the above power grid topology cut set. The devices contained in different online topology graphs are divided according to the connection relationship and expressed in the form of "node-edge". Each device can be regarded as a node of the graph, and the connection relationship is the edge. Each device has a corresponding feature vector, which is composed of the measurement data of the device in a certain period of time <active power, reactive power, current, voltage>. In this way, the feature matrix X and adjacency matrix A of the graph structure can be obtained as the input of the GCN graph convolutional network model.

5. The method according to claim 4, characterized in that The GCN graph convolutional network model consists of an input layer, two graph convolutional layers, a pooling layer, two fully connected layers and an activation function; the step S3 includes: S21: In the first graph convolution layer, the feature matrix X of the electric quantity data graph structure and the normalized adjacency matrix A" are convolved to obtain the output feature matrix A”=D -1 / 2 A'D -1 / 2 A’=A+I N Where: W1 G and b1 G are the weight matrix and bias vector of the first graph convolutional layer respectively; H1 G is the output vector of the first graph convolutional layer; ReLU represents the activation function; I N is the identity matrix of order N; E is the edge set, e ij represents the edge connecting nodes i and j, A=(a ij ) n×n is the adjacency matrix of the graph, if e ij ∈E, then a ij =1, otherwise, ai j =0; n represents the number of nodes; D is the degree matrix of the graph, which is a diagonal matrix with the following diagonal elements: S22: Output data obtained after the second graph convolution layer: Where: and They are the weight matrix and bias vector of the second graph convolutional layer respectively; is the output vector of the second graph convolutional layer; S23: In the fully connected layer, output data The weight matrix Bias Vector And the activation function ReLU gets: S24: The output result of the last layer is: Where: and are the weight matrix and bias vector of the last layer respectively; Y G is the output feature vector of the GCN module; the output feature vector includes the measurement data of the abnormal device node.

6. The method according to claim 5, characterized in that The step S4 comprises: the measurement data of the abnormal node are sampled in a short time window respectively, and then represented in a time series graphical form, expressed as a graph of 50 sampling points according to a time series line graph, each measurement graph is spliced ​​into a picture to be identified, the measurement data features in the picture are extracted to generate a feature matrix X', which is input into the CNN device measurement error discrimination model for discrimination, and the error type is output after the discrimination is completed, so as to realize the detection of abnormal distribution network measurement data.

7. The method according to claim 6, characterized in that In the step S4, CNN convolutional neural network model, including input layer, convolution layer 1, convolution layer 2, pooling layer, fully connected layer and output layer; In the first convolutional layer, the input data is the feature matrix X', and the output data is: Where: and are the weight matrix and bias vector respectively. Similar to the first convolutional layer, the output features after the second convolutional layer are: Where: and They are the weight matrix and bias vector of the second convolutional layer respectively. Using maximum pooling in the pooling layer can reduce the parameters of CNN, and the output features are: In the last fully connected layer of CNN, the output is: Where: and are the weight matrix and bias vector of the fully connected layer respectively; Y C is the output feature vector of the CNN module; after extracting the information of the output abnormal data feature vector, it outputs <overload, heavy load, low voltage, white noise, data missing, three-phase imbalance, single-phase current large / small> to realize the detection of abnormal distribution network measurement data.

Images