A low-voltage power distribution network state sensing and abnormality warning method

Abstract

The application is a kind of low-voltage power distribution network state perception and abnormal alarm method, belonging to the field of intelligent power distribution network, in order to overcome the defects of the existing power distribution network state perception method, a kind of low-voltage power distribution network state perception and abnormal alarm method, specifically: using the topology related operation data in power distribution network to construct graph feature data; learn the mapping function relationship between power and voltage in physical power flow model, build power flow fitting model based on graph feature propagation network; by inputting the historical data of power distribution network in the power flow fitting model, complete the power flow fitting perception learning; the real-time data of low-voltage power distribution network fusion terminal is taken as the model input, the voltage amplitude and phase angle of all operating nodes of power distribution network are calculated, the nodes with deviation value exceeding threshold value are screened out, and abnormal alarm signal is sent. The graph feature propagation model based on graph feature propagation network is constructed, the abnormal nodes are screened out, and the security and stability operation and coordinated control of power distribution network are provided.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of intelligent power distribution network, and particularly relates to a low-voltage power distribution network state sensing and abnormal alarm method. BACKGROUND

[0002] With the large-scale access of new energy, distributed power supply, micro-grid and the like to the power distribution network, and in order to alleviate the negative impact of the uncertainty of renewable energy on the power distribution network, the types of power equipment for improving power flow distribution are increasing. A large number of new energy and power equipment will make the power flow model of the power distribution network tend to be complex, and the existing low-voltage power distribution network state sensing technology is difficult to model, and it is difficult to meet the timely sensing demand of the system state.

[0003] The existing power distribution network state sensing method is mainly divided into static and dynamic two types, the static state estimation method is mainly based on the least square method state estimation, and the dynamic state estimation method is mainly based on the Kalman filter algorithm, however, the above methods are severely dependent on complete power distribution network topology parameters, and the low-voltage power distribution network has a serious account missing problem, and is accompanied by the continuous access of new energy, and the above method cannot be directly applied to the state sensing of the low-voltage power distribution network. Therefore, it is urgent to develop a state sensing tool to realize the panoramic state sensing of the power distribution network by using the measurement information of the power distribution network. SUMMARY

[0004] In order to overcome the defects of the existing power distribution network state sensing method, the technical problem to be solved by the application is to provide a low-voltage power distribution network state sensing and abnormal alarm method, which aggregates node information by using a graph feature propagation network, and performs power flow mapping calculation, so as to provide protection for the safe and stable operation and coordinated control of the power distribution network.

[0005] The technical scheme adopted by the application is as follows: a low-voltage power distribution network state sensing and abnormal alarm method, comprising:

[0006] S1, constructing graph feature data by using topology-related operation data in the power distribution network;

[0007] S2, learning the mapping function relationship between power and voltage in the physical power flow model, and constructing a power flow fitting model based on a graph feature propagation network;

[0008] S3, completing power flow fitting sensing learning by inputting the historical data of the power distribution network in the power flow fitting model;

[0009] S4, taking the real-time data of the low-voltage power distribution network fusion terminal as model input, calculating the voltage amplitude and phase angle of all operating nodes of the power distribution network, comparing the calculation result with the measurement result, screening out the nodes with a deviation value exceeding a threshold value, and issuing an abnormal alarm signal.

[0010] The application constructs a graph feature transmission model based on a graph feature transmission network, performs power flow mapping calculation, filters out abnormal nodes, and provides protection for safe and stable operation and coordinated control of a power distribution network.

[0011] Further, the specific process of step S1 is: establishing a graph feature data G(M, A) based on an n-node system feature matrix M={m1, m2,..., mn} and an adjacency matrix A representing topological connection, each node feature m n} and an adjacency matrix A representing topological connection, each node feature m i contains input features and output features , that is where i is one of the n nodes, the value range is 1-n, are the active power and reactive power of the generator at node i, are the active power and reactive power of the load at node i, U i and θ i are the voltage amplitude and phase angle corresponding to node i;

[0012] The definition of element a ij in the adjacency matrix A is as follows:

[0013]

[0014] where j is one of the n nodes, the value range is 1-n, and n is an integer. It should be noted that when i and j have the same value, it represents autocorrelation, which is not a direct connection.

[0015] Further, in step S2, the graph feature network constructed includes three layers of input layer, hidden layer and output layer;

[0016] The input layer is composed of the node adjacency matrix A and the corresponding node feature matrix M, and is used for convolution operation to aggregate node power feature information; the hidden layer is used for further convolution operation of the aggregated node power feature information obtained by the input layer to transmit graph feature information; the output layer includes a convolution layer and a fully connected layer, which are used for convolution operation again and map the output result to output features . In this way, the transmission of graph feature information in the graph feature transmission network is completed.

[0017] Further, in step S2, the graph feature transmission rule adopted by the graph feature network is:

[0018]

[0019] In the formula: h (l) represents the input features of the lth layer, w (l)is the weight of the l-th layer, and σ is the activation function, is the normalized graph feature Laplacian matrix, f is the graph feature transfer function; l is selected from the input layer, the hidden layer and the output layer.

[0020] Further, the values of the adjacency matrix A at the diagonal position are all 0, so that the node does not have self-connection (itself has an edge with itself); in fact, the information of itself is also very important when extracting features; in order to solve this problem, the adjacency matrix A can be added to the identity matrix I, so that each node has a self-connection, and the expression of the adjacency matrix A' is:

[0021] A' = A + I

[0022] Since the adjacency matrix A is not normalized, after matrix transformation for multiple layers, it will be more and more different from the amplitude of the input feature, which is not conducive to the training of the neural network; in order to solve this problem, normalization processing can be performed:

[0023]

[0024] In the formula, D is a diagonal matrix, and the values at the diagonal position are also the degrees of the corresponding nodes, which are defined as the sum of the weights of all edges from the i-th node.

[0025] Further, the expression of the convolution operation is:

[0026]

[0027] In the formula, is the feature vector of the i-th node of the l+1-th layer, is the feature vector of the m-th node of the l-th layer, so it can be seen that the feature vector of node i at the l+1-th layer is only related to the feature vectors of its neighbors at the l-th layer; m is the number of adjacent nodes of node i, and k is the number of adjacent nodes, is the weight of the m-th node of the l-th layer corresponding to node i; σ is the activation function, which is used to establish the nonlinear function of the graph feature network. The principle of graph convolution operation in the graph feature transfer model is that the result of node convolution is the weighted sum of the feature values of adjacent nodes, which can establish the relationship between the feature values of the previous layer nodes and the corresponding elements of the next layer.

[0028] The voltage of a node in a power system has a nonlinear relationship with the voltage of adjacent branches, input power and load, so the calculation of node voltage can be realized by nonlinear superposition of adjacent node features, and the data-driven power flow calculation task can be completed.

[0029] Further, the network loss function is constructed, which is used to train the parameters of the graph feature network, and the specific expression is as follows:

[0030]

[0031] In the formula: W is the network training parameter, H is the network output feature, and L is the loss function.

[0032] Furthermore, the specific process of step 3 is as follows: using historical data of the distribution network as training data samples, calculate the node voltage amplitude and phase angle, complete the power flow fitting and perception learning of the graph feature transfer model, and in the model training, the optimization function selection adopts the adaptive moment estimation (ADAM) optimization algorithm.

[0033] Furthermore, in step S4, when the standard error of the power system measurement value is τ, nodes with deviation values ​​exceeding the given threshold 6τ will issue an abnormal alarm signal.

[0034] The beneficial effects of this invention are as follows: This invention proposes a method for low-voltage distribution network status perception and anomaly alarm. By using graph features to transmit network aggregation node information and performing power flow mapping calculation, it effectively avoids the defects of traditional status estimation methods that rely on network topology parameters. Combined with the computing power of the fusion terminal of low-voltage distribution area, it realizes the real-time perception of low-voltage distribution network node voltage and the real-time alarm function for abnormal status, providing strong support for the panoramic perception of low-voltage distribution area. Attached Figure Description

[0035] Figure 1 Flowchart for an embodiment;

[0036] Figure 2 This is a topology diagram of a 33-node distribution network. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be explained and described below with reference to the accompanying drawings. However, the following embodiments are only preferred embodiments of the present invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments in the implementation methods without creative effort are all within the protection scope of the present invention.

[0038] The low-voltage distribution network status perception and anomaly alarm method in this embodiment, such as Figure 1 As shown, it includes:

[0039] S1, constructing graph feature data using operational data of topological associations in the distribution network; specifically:

[0040] Establish a feature matrix M = {m1, m2, ..., mn} for a system with n nodes. n} and the graph feature data G(M, A) based on the adjacency matrix A representing topological connections, with each node having feature m i All of them contain input features and output features i.e. where i is one of the n nodes, ranging from 1 to n, Pi and Qi are the active power and reactive power of the generator at node i, respectively, Pi and Qi are the active power and reactive power of the load at node i, respectively, U i and θ i are the voltage amplitude and phase angle of node i, respectively;

[0041] The definition of element a ij in the adjacency matrix A is as follows:

[0042]

[0043] where j is one of the n nodes, ranging from 1 to n; i and j are integers; it should be noted that when i and j take the same value, it represents autocorrelation, which is not directly connected;

[0044] S2, learning the mapping function relationship between power and voltage in the physical power flow model, constructing a graph feature propagation model based on graph feature propagation network;

[0045] The constructed graph feature network includes three layers of input layer, hidden layer and output layer;

[0046] The input layer is composed of node adjacency matrix A and corresponding node feature matrix M, which is used for convolution operation to aggregate node power feature information; the hidden layer is used for further convolution operation of the aggregated node power feature information obtained by the input layer to propagate graph feature information; the output layer includes a convolution layer and a fully connected layer, which is used for convolution operation again and maps the output result to output feature ;

[0047] The graph feature propagation rule adopted by the graph feature network is:

[0048]

[0049] In the formula: h (l) represents the input feature of the lth layer, w (l) represents the weight of the lth layer, σ is the activation function, is the normalized graph feature Laplacian matrix, f is the graph feature propagation function; l is selected from the input layer, the hidden layer and the output layer;

[0050] The expression of convolution operation is:

[0051]

[0052] wherein, represents the feature vector of the i-th node of the (l+1)-th layer, represents the feature vector of the m-th node of the l-th layer; m is the number of adjacent nodes of node i, and k is the number of adjacent nodes, is the weight of the m-th node of the l-th layer corresponding to node i; σ is an activation function, which is used to establish the nonlinear function of the graph feature network;

[0053] There is no self-connection (there is an edge between itself and itself), and the values of the adjacency matrix A at the diagonal positions are all 0; in fact, the information of itself is also very important when extracting features.

[0054] To solve this problem, the adjacency matrix A can be added to the identity matrix I, thereby adding a self-connection to each node, and the expression of the formalized adjacency matrix A' is:

[0055] A' = A + I

[0056] wherein, I is an identity matrix, and A' is a formalized adjacency matrix;

[0057] Since the adjacency matrix A is not normalized, after a plurality of matrix transformations, it will be more and more different from the amplitude of the input feature, which is not conducive to the training of the neural network. In order to solve this problem, the adjacency matrix A can be normalized:

[0058]

[0059] wherein, D is a diagonal matrix, and the values at the diagonal positions are the degrees of the corresponding nodes, the degree of the corresponding node being defined as the sum of the weights of all edges from the i-th node;

[0060] A network loss function is constructed to minimize the node voltage prediction error, which is used to train network parameters and realize power flow mapping training, and the specific expression is as follows:

[0061]

[0062] wherein: W is a network training parameter, H is a network output feature, and L is a loss function;

[0063] S3, taking the historical data of the distribution network as the training data sample, calculating the node voltage amplitude and phase angle, completing the power flow fitting perception learning of the graph feature transmission model, and in the model training, the optimization function selection adopts an adaptive moment estimation (ADAM) optimization algorithm;

[0064] S4 uses real-time data from the low-voltage distribution network fusion terminal as model input to calculate the voltage amplitude and phase angle of all operating nodes in the distribution network. The calculated results are compared with the measured results. When the standard error of the power system measurements is τ, nodes with deviations exceeding a given threshold of 6τ are selected, and an abnormal alarm signal is issued and uploaded to the cloud center for processing. The standard error τ is related to the equipment.

[0065] To verify the effectiveness of the state awareness capability of the proposed method, an IEEE 33 system connected to a new energy source (such as...) was used. Figure 2 Simulation analysis was conducted. For the IEEE 33-bus system, wind farms with a rated capacity of 10MW were added to nodes 5 and 12, and photovoltaic (PV) power plants with a rated capacity of 16MW were introduced to nodes 10, 15, and 27. The conventional generators at node 2 were replaced with hydroelectric generators with a rated capacity of 100MW. Historical data from 24 points per day for 18 months were used for the wind, PV, and hydroelectric generators. The historical data reflects the power output at various points in the past, resulting in diverse power flow data samples.

[0066] z-score standardization reflects the severity of data deviation from the mean and highlights the characteristics of data with large deviations from the mean. The z-score standardization formula is:

[0067]

[0068] To compare the differences between the graph feature transfer network model, conventional deep learning models, and traditional classical machine learning methods, the input active and reactive power data were transformed into structured data and used as input for convolutional neural networks, multilayer perceptrons, and deep belief networks to fit the AC power flow of an IEEE 33-node system. The results were then compared with the power flow fitting results of the proposed method. The inputs to the CNN, MLP, and DBN models were the active and reactive power of the node generators and the active and reactive power of the loads. The outputs were the phase angle and magnitude of the node voltages.

[0069] The power flow mapping results are shown in Table 1.

[0070] Table 1

[0071]

[0072] To verify the effectiveness of the proposed method for anomaly alerting, 20% of the dataset was manually modified to represent anomaly samples in the simulation data. The evaluation system used was precision, recall, and the F1 score, defined as follows:

[0073]

[0074]

[0075]

[0076]

[0077] In the formula: a s and γ us respectively are the number of normal samples correctly or incorrectly predicted; a us and γ s respectively are the number of abnormal samples correctly or incorrectly predicted; A CC , η SR , η UR , O mean respectively are the accuracy, recall, precision, F1 index. The abnormal alarm results are as shown in Table 2.

[0078] Table 2

[0079] Method Accuracy / % Recall / % F1-index / % GCN (method of the invention) 84.00 90.40 91.23 CNN 82.73 89.00 90.45 MLP 75.00 76.60 84.86 DBN 80.61 89.48 89.26

[0080] Finally, it can be seen that the accuracy and effect of the application on state perception and abnormal alarm are improved compared with similar models.

[0081] The above is only a specific embodiment of the application, but the protection scope of the application is not limited to this. Those skilled in the art should understand that the application includes but is not limited to the contents described in the drawings and the above specific embodiments. Any modification that does not deviate from the functional and structural principles of the application will be included in the scope of the claims.

Claims

1. A method for low-voltage distribution network status sensing and anomaly alarm, characterized in that, include: S1, construct graph feature data using operational data of topological associations in the distribution network; Step S1 specifically involves: establishing a system based on... n The feature matrix of a node system and the adjacency matrix representing topological connections Based on graph feature data Features of each node m i All of them contain input features and output features ,Right now ,in i express n One of the nodes, with a value range of 1. -n , , , These are nodes i The active and reactive power of the generator. , These are nodes i Active and reactive power under load, , , It is a node i The corresponding voltage amplitude and phase angle; Adjacency matrix Middle elements a ij The definition is as follows: in, j express n A node with a value range of 1. -n； S2, learn the mapping function relationship between power and voltage in the physical power flow model, and construct a power flow fitting model based on a graph feature transfer network; the constructed graph feature network includes a three-layer structure of input layer, hidden layer and output layer; The input layer consists of a node adjacency matrix. A and the corresponding node feature matrix M The input layer consists of a convolutional layer for aggregating node power features; a hidden layer for further convolutional operations on the aggregated node power features obtained from the input layer to transfer graph feature information; and an output layer containing a convolutional layer and a fully connected layer for further convolutional operations, mapping the output to output features. superior; The graph feature transfer rule adopted by the graph feature network is as follows: In the formula: Indicates the first l The input features of the layer Indicates the first l Layer weights It is an activation function. It is the normalized graph characteristic Laplacian matrix. f For graph feature transfer function; l Selected from the input layer, hidden layer, and output layer; S3, by inputting historical data of the distribution network into the power flow fitting model, completes the power flow fitting perception learning; S4 takes the real-time data of the low-voltage distribution network fusion terminal as the model input, calculates the voltage amplitude and phase angle of all operating nodes in the distribution network, compares the calculation results with the measurement results, filters out nodes whose deviation values ​​exceed the threshold, and issues an abnormal alarm signal.

2. The low-voltage distribution network status sensing and anomaly alarm method according to claim 1, characterized in that, Formulated adjacency matrix The expression is: In the formula, I It is the identity matrix. For the formalized adjacency matrix; Normalized graph feature Laplacian matrix Satisfy the following expression: In the formula, D Given a diagonal matrix, the values ​​at the diagonal positions are the degrees of the corresponding nodes, where the degree of a node is defined as... i The sum of the weights of all edges originating from a node.

3. The low-voltage distribution network status sensing and anomaly alarm method according to claim 1, characterized in that, The expression for convolution is: In the formula, Indicates the first l +1 floor i The feature vector of each node Indicates the first l The first layer m The feature vectors of each node; i>m is related to the node i The numbers of the adjacent nodes, k It is the number of neighboring nodes. It is the first l The first layer m Each node corresponds to a node. i The weights; It is an activation function used to establish a non-linear function for graph feature networks.

4. The low-voltage distribution network status sensing and anomaly alarm method according to claim 3, characterized in that, Construct the network loss function, used to train the parameters of the graph feature network, with the following specific expression: In the formula: These are network training parameters. It is a network output characteristic. L This is the loss function.

5. The low-voltage distribution network status sensing and anomaly alarm method according to claim 1, characterized in that, Step 3 involves the following steps: using historical data from the distribution network as training data samples, calculating node voltage magnitudes and phase angles, and completing the power flow fitting and perceptual learning of the graph feature transfer model. During model training, the optimization function is selected using... ADAM Optimize the algorithm.

6. The low-voltage distribution network status sensing and anomaly alarm method according to claim 1, characterized in that, In step S4, the standard error of the power system measurements is: τ When the deviation value exceeds the given threshold of 6... τ The node sends an abnormal alarm signal.

Images