A power monitoring event identification and intelligent analysis method based on a graph neural network

Abstract

The application discloses a power monitoring event identification and intelligent analysis method based on a graph neural network, which is used for inferring specific power grid events when a large number of discrete alarm signals occur in the power grid. The main steps of the method are as follows: firstly, collecting the discrete alarm signals of the power grid; then, extracting triplets to construct a knowledge graph through the steps of knowledge extraction, knowledge fusion and knowledge processing; on the basis, adopting a graph neural network method with a attention mechanism, feature extraction is performed on the fault events occurring at present, a hidden feature vector of an entity is calculated, the neighborhood information of a node is learned, a possible fault event is inferred, and an auxiliary solution is provided. The method has the following advantages: 1) can meet the needs of multiple business scenarios; 2) compared with a real-time processing mode, the efficiency is higher; and 3) the misjudgment and missed judgment of the monitoring screen are greatly reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power system, in particular to a power monitoring event identification and intelligent analysis method based on graph neural network. BACKGROUND

[0002] The power system is composed of power generation, power transformation, power transmission, power distribution, power consumption and dispatching. A large number of alarm signals occur in the power system every day, including accidents, abnormalities, overruns, position changes and notifications. At present, in the field of power monitoring technology, alarm window and system monitoring are the main methods used in monitoring work. However, these two methods cannot meet the current needs of power monitoring. The monitoring alarm window needs to process more than 1,700 monitoring alarms per day, and needs to analyze, judge and handle one alarm signal every 50 seconds. When faults occur frequently, the monitoring screen pressure increases dramatically, and it is easy to miss notifications and misjudge. System monitoring has the problems of insufficient identification of alarm signals and inaccurate analysis, which reduces the real-time processing efficiency of alarm signals.

[0003] Therefore, there is a need for a method that can improve the efficiency of analyzing and processing alarm signals, so as to meet the current needs of power monitoring, quickly analyze a large number of alarm signals, and improve the accuracy of judgment.

[0004] As a new research in the field of knowledge engineering, knowledge graph can realize the storage and accumulation of power knowledge when applied to the field of power system. The knowledge in the power field does not change over time, but gradually expands and improves. Due to the strict control of safety standards in the power industry, the reliability of expert experience is often better than various artificial intelligence algorithms. As an integrated corpus of expert experience, the retrieval and matching capabilities of knowledge graph perfectly meet the needs of safety and rapid response in the power field. Therefore, more and more attention is focused on the application of knowledge graph in different occasions in the power field. SUMMARY

[0005] In view of the defects in the prior art, the present application provides a power monitoring event identification and intelligent analysis method based on graph neural network, comprising the following steps:

[0006] Step 1, collecting alarm information when an alarm signal occurs in the power grid;

[0007] Step 2, constructing a power monitoring event knowledge graph by knowledge extraction, knowledge fusion and knowledge processing on the alarm information;

[0008] Step 3, the graph neural network reads the graph structure features of the power monitoring event knowledge graph and calculates the entity hidden feature vector;

[0009] Step 4, reasoning event fault type in knowledge graph based on graph neural network with attention mechanism, providing auxiliary solution.

[0010] Preferably, the alarm information when the power grid appears an alarm signal is collected. The alarm information of power monitoring indicates the state change of a certain device or action in the station, which is displayed in the alarm window in the text format and stored in the alarm record file, and the content includes voltage level, device name, device number and action behavior.

[0011] Preferably, the alarm information is processed through knowledge extraction, knowledge fusion and knowledge processing to construct the power monitoring event knowledge graph. Since the power monitoring text includes voltage level, device name, device number and action behavior, the mixed information description method of numbers and Chinese will affect the judgment of defect level. The conventional knowledge extraction model is applied to the power monitoring event, and the recognition accuracy of the extracted fault information is low, therefore, the influence of the text information around the number on the training of the number word vector is considered, and the original knowledge extraction model is optimized.

[0012] For the obtained power monitoring event, first, the named entity definition is performed, the power monitoring event entity to be extracted is defined as the device name, event content and event level; then, the original data is preprocessed to remove redundant, noise information and data labels, and the monitoring event text is expressed in the form of word vector, which is helpful for formal data training; then, the generated word vector is subjected to entity recognition training operation, and the application adds selectable position information on the basis of the BiLSTM-CRF model to obtain the text representation of the number in the defect text as the input of the model.

[0013] The application uses a bidirectional long short-term memory network BiLSTM to perform benchmarking, and can fully consider the connection between the context information of data and the bidirectional semantic dependency. The LSTM unit has three gate controllers, namely input gate, forgetting gate and output gate. Their forward propagation processes can be represented by formula (1) to formula (5) respectively.

[0014] i t =σ(W xi x t +W hi h t-1 +W ci c t-1 +b i ) (1)

[0015] f t =σ(W xf x t +W hf h t-1 +W cf c t-1+b f ) (2)

[0016] c t =f t c t-1 +i t tanh(W xc x t +W hc h t-1 +b c ) (3)

[0017] o t =σ(W xy x t +W hf h t-1 +W σ c t-1 +b f ) (4)

[0018] h t =o t tanh(c t ) (5)

[0019] Wherein, b represents the bias, t represents the current time, x t represents the input, h t is the state of the hidden unit, sigma is the s-shaped activation function, i, f, o, c are the input gate, forget gate, output gate and unit vector, and the weight matrix W represents the unit corresponding to the subscript connection weight matrix.

[0020] Conditional random field CRF is essentially a discriminative undirected graph. If the observed sequence X={x1,x2,...,x n}, the corresponding label sequence is Y={y1,y2,...,y n}, then the conditional random field constructed between the two conditional probabilities is P(X|Y).

[0021] Before the input of the BiLSTM-CRF model, the word vectors near the number are weighted according to the distance, and the number vector is revalued; the calculation formula is as follows:

[0022] V i '=α i-1 ×V i-1 +...+α i-k ×V i-k +V i +α i+1 ×V i+1 +...+α i+k ×V i+k (6)

[0023] wherein a = {a1, a2,..., aK} is the weight of the previous and subsequent k data as a word vector, and V = {V1, V2,..., V n} is multiplied by V = {V1, V2,..., V n} to obtain a newly defined digital vector; a i-1 represents the weight of the i-1th vector relative to the digital vector, V i represents the digital vector calculated by the optimization model.

[0024] Preferably, the graph neural network reads the graph structure features of the power monitoring event knowledge graph, and calculates the entity hidden feature vector. In order to learn the power monitoring alarm signal features using the graph neural network, a graph neural network with an attention mechanism is used to extract valuable nodes in the alarm signal, a semantic attention layer is set, the nodes in the power monitoring event knowledge graph constructed in step 2 are embedded into a vector space to form an embedding vector matrix, and by keeping the node embedding vectors with the same dimension during training, the attention layer is ensured to obtain sufficient expressive feature information. Through the self-attention mechanism, the attention score of each label node y L ∈Y and its K neighbors H L = {h1, h2,..., h K} is calculated. The self-attention score between two nodes h i and h j is:

[0025] u ij = selfAtt(Wh i , Wh j ) (7)

[0026] It can be shown from the formula that the feature learning of the current node considers the influence of all neighbor nodes.

[0027] wherein H L = {h1, h2,..., h K} is the K neighbor nodes, W is the neighborhood entity feature matrix obtained by linear transformation of the neighbor nodes, u ij is the self-attention score, and selfAtt is the feedforward neural network.

[0028] When learning, the graph neural network learns the topology of the graph by adding the features of all adjacent nodes to the node features. The node set is embedded as a feature matrix H L , the adjacency matrix A is used as the input of the graph neural network, and the update of the node features is generated by convolution operation, which is:

[0029]

[0030]

[0031] The convolution operation is the neural transformation from layer l to layer l+1 in a graph neural network.

[0032] Where A is the adjacency matrix, representing the relationships between vertices, with a size of n*n, where n is the number of nodes; H is the feature matrix, with a size of n*d, where d is the dimension of the features of nodes in layer l of the graph neural network; l is the number of layers in the graph neural network; D is the degree matrix of A, representing the number of edges associated with each node in the graph, with a size of n*n; W is the transformation weight matrix, with a size of n*d', where d' is the dimension of the features of nodes in layer l+1 of the graph neural network, and W ensures the effective transmission of messages during the transformation process; ρ is the nonlinear activation function.

[0033] During training, the update of each node's feature vector affects the feature learning of its neighbors. Therefore, each node's features incorporate the feature information from its neighbors. In the feature fusion process, redundant feature information needs to be removed, and features from different sources need to be fused. The formula for calculating the combined hidden feature vector is as follows:

[0034]

[0035] In the formula, f i ' represents the forward hidden feature vector, f i " represents the backward hidden feature vector, Representing a sparse relational matrix The hidden state h' obtained by convolution operation (l+1) Multiplication operation.

[0036] During training, since the learning of the hidden feature vectors of nodes is carried out in parallel, the update of each node's feature vector will affect the feature learning of its neighboring nodes. Therefore, the learning process of the hidden feature vectors of nodes is continuously iterated until the features of each node tend to stabilize. During this process, the entity feature vector is optimized by minimizing the cross loss function until the update converges to a certain value or the maximum number of iterations.

[0037] Prioritizing the use of graph neural networks with attention mechanisms to infer event fault types within the knowledge graph, an auxiliary solution is provided. Alarm signals are input into the power monitoring event knowledge graph, and an attention-based graph convolutional neural network (Att-GCN) is employed to calculate the correlation between the alarm signal entity and all its neighboring entities. This accurately infers the hidden feature vector of the alarm signal, which is then linked to predict and determine the fault type, matching the auxiliary solution accordingly.

[0038] The beneficial effects of this invention are:

[0039] (1) This invention utilizes graph neural networks and knowledge graphs to improve the efficiency of power monitoring staff, reduce misjudgments and omissions, and meet the needs of multiple business scenarios.

[0040] (2) This invention belongs to the deep learning method based on graph neural networks. It has adaptive learning ability, strong learning ability, and can update the database according to the input data, continuously improve and gradually improve the efficiency of solving problems. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating the steps in an embodiment of the present invention.

[0042] Figure 2 The diagram below illustrates the optimization model in step 2 of this embodiment, showing the principle framework for parsing the input text information. First, the segmented word sequence is converted into a continuous sequence of word vectors and number vectors. Second, the BiLSTM-CRF model is used to extract features. Finally, the output is mapped to a tag set, with each word's tag representing the recognition result.

[0043] Figure 3 This is a schematic diagram of GCN extracting node features in step 3 of this embodiment of the invention, illustrating the process of learning the knowledge graph using a graph neural network. The knowledge graph is used as input, and the graph neural network learns the features of each node and the features of its neighboring nodes, with the output being the features of the graph structure. Detailed Implementation

[0044] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0045] Step 1: Collect alarm information when the power grid generates an alarm signal.

[0046] The alarm information of power monitoring indicates the status change of a certain device or action in the station. It is displayed in the alarm window in text format and stored in the alarm log file. The content includes voltage level, device name, device number and action behavior.

[0047] Step 2: Construct a power monitoring event knowledge graph by extracting, fusing, and processing the alarm information.

[0048] Knowledge extraction is performed on the alarm signal information, extracting entities, attributes, and relationships between entities. Based on this, an ontology-based knowledge representation is formed, yielding new knowledge. Then, the acquired new knowledge is categorized using traditional knowledge fusion methods, classifying alarm information based on similar locations, lines, and causes, and these alarm messages are then merged. Ontology mapping and instance matching are employed to address the heterogeneity of knowledge, eliminating ambiguity and contradictions. The quality of the merged new knowledge is assessed to check for irrelevant alarm information being grouped together, thus quantifying the reliability of the knowledge. High-quality knowledge is added to the constructed knowledge graph. The graph construction process uses the Neo4j graph database to store and visualize the power system alarm signal knowledge graph, storing the triples extracted from the power system alarm signal information in the Neo4j graph database.

[0049] For the acquired power monitoring events, named entity definition is first performed. For power monitoring events, such as "2019-07-22 20:09:21 Jiangsu. Gaoqiao Substation / 110kV. Gaohua Line 919 switch tripping," the entities to be extracted are defined as equipment name, event content, and event level. Then, a large amount of raw event data is preprocessed to remove redundancy, noise, and data labels, representing the monitoring event text as word vectors, which facilitates formal data training. Next, entity recognition training is performed on the generated word vectors. This invention adds selectable location information to the BiLSTM-CRF model to obtain the text representation of numbers in the defective text, which serves as the model's input vector. Finally, the recalculated numerical vectors, along with the text vectors, are input into the BiLSTM and CRF layers for knowledge extraction, extracting high-quality power monitoring event knowledge. This special processing of numerical information during input improves the accuracy of the extraction results and enhances the quality of power monitoring event knowledge.

[0050] This invention uses a bidirectional long short-term memory (BiLSTM) network for benchmarking, which can fully consider the relationship between the contextual information of the data and the bidirectional semantic dependencies. The LSTM unit has three gate controllers: an input gate, a forget gate, and an output gate. Their forward propagation processes can be represented by equations (1) to (5), respectively:

[0051] i t =σ(W xi x t +W hi h t-1 +W ci c t-1 +b i (1)

[0052] f t =σ(W xf xt +W hf h t-1 +W cf c t-1 +b f (2)

[0053] c t =f t c t-1 +i t tanh(W xc x t +W hc h t-1 +b c (3)

[0054] o t =σ(W xy x t +W hf h t-1 +W σ c t-1 +b f (4)

[0055] h t =o t tanh(c t (5)

[0056] Where b represents the bias, t represents the current time, and x t Indicates input, h t Let σ represent the state of the hidden unit, σ be the sigmoid activation function, i, f, o, and c be the input gate, forget gate, output gate, and unit vector, and W be the weight matrix W representing the unit corresponding to the subscript connected to the weight matrix.

[0057] A Conditional Random Field (CRF) is essentially a discriminative undirected graph. If the observed sequence X = {x1, x2, ..., x...} n The corresponding label sequence is Y = {y1, y2, ..., y}. n If}, then the conditional random field constructed between the two conditional probabilities is P(X|Y).

[0058] As attached Figure 2 As shown, before inputting into the BiLSTM-CRF model, word vectors near the digits are weighted according to distance, and the digit vectors are re-weighted; the calculation formula is as follows:

[0059] V i '=α i-1 ×V i-1 +...+α i-k ×V i-k +V i +α i+1 ×Vi+1 +...+α i+k ×V i+k (6)

[0060] Where α={α1,α2,...,α n The first and last k data points in the vector are used as the weights of the word vectors, and are related to V = {V1, V2, ..., V}. n Multiplying them together yields a newly defined numerical vector; α i-1 V represents the weight of the (i-1)th vector relative to the numerical vector. i ' represents the numerical vector calculated by the optimization model.

[0061] Step 3: The graph neural network reads the graph structure features of the power monitoring event knowledge graph and calculates the entity hidden feature vector.

[0062] The graph structure features include node features and relation features. In the power system, node features mainly include: 110kV line switch opening, 110kV line switch reclosing success, 110kV line switch reclosing failure, protection output, line protection output, line fault accompanying signal, substation fault common accompanying signal, and total substation accident. For learning power monitoring alarm signal features using graph neural networks, a graph neural network with an attention mechanism is used to extract valuable nodes from the alarm signals. A semantic attention layer is set up, embedding the nodes in the power monitoring event knowledge graph constructed in step 2 into a vector space to form an embedding vector matrix. By maintaining the same dimension of the node embedding vectors during training, sufficient expressive feature information is obtained in this attention layer. The feature matrix of a node can be obtained by searching the vector space. The correlation coefficient between each node and its surrounding neighbor nodes is calculated through the attention mechanism. Taking a node in the knowledge graph as an example, it is used as the node to be learned. During learning, the nodes related to it, i.e., neighbor nodes, can be obtained from the relation matrix. The attention mechanism is an inner product operation that measures the degree of influence of a node on any of its neighbor nodes. L ∈Y and its K neighbors H L ={h1,h2,...,h K}, two nodes h i and h j The correlation coefficient between them is:

[0063] u ij =selfAtt(Wh i ,Wh j (7)

[0064] The formula shows that the feature learning of a node takes into account the influence of all its neighboring nodes.

[0065] Among them, HL ={h1,h2,...,h K Let} be K neighboring nodes, W be the neighborhood entity feature matrix obtained by linear transformation of the neighboring nodes, and u ij The score represents self-attention, and selfAtt is a feedforward neural network.

[0066] As attached Figure 3 As shown, during learning, the graph neural network learns the graph's topology by adding the features of all neighboring nodes to the node features. The node set is embedded as a feature matrix, and the adjacency matrix A is used as the input to the graph neural network. The node feature updates are generated by convolution operations, which are as follows:

[0067]

[0068]

[0069] The convolution operation is the neural transformation from layer l to layer l+1 in a graph neural network.

[0070] Where A is the adjacency matrix, representing the relationships between vertices, with a size of n*n, where n is the number of nodes; H is the feature matrix, with a size of n*d, where d is the dimension of the features of nodes in layer l of the graph neural network; l is the number of layers in the graph neural network; D is the degree matrix of A, representing the number of edges associated with each node in the graph, with a size of n*n; W is the transformation weight matrix, with a size of n*d', where d' is the dimension of the features of nodes in layer l+1 of the graph neural network, and W ensures the effective transmission of messages during the transformation process; ρ is the nonlinear activation function.

[0071] During training, the update of each node's feature vector affects the feature learning of its neighbors. Therefore, each node's features incorporate the feature information from its neighbors. In the feature fusion process, redundant feature information needs to be removed, and features from different sources need to be fused. The formula for calculating the combined hidden feature vector is as follows:

[0072]

[0073] In the formula, f i ' represents the forward hidden feature vector, f i " represents the backward hidden feature vector, Representation matrix with h' (l+1) Multiplication operation Let h' be the sparse relation matrix of the relation set of node i. (l+1) Let be the hidden state of the set of neighboring nodes of node i.

[0074] During training, the learning of hidden feature vectors for each node is performed in parallel. Each node's feature vector update affects the feature learning of its neighbors. Therefore, the learning process is iterative and continuous until the features of each node stabilize. During this process, the entity feature vector is optimized by minimizing the cross-loss function, and the entity features are evaluated until the loss function converges to a certain value or the maximum number of iterations, indicating that the node feature vector update is complete.

[0075] Step 4: Based on the graph neural network with attention mechanism, reason about the fault type of the event in the knowledge graph and provide an auxiliary solution.

[0076] First, the alarm signals are preprocessed. The entities and relation types in the signals are numbered and embedded into the vector space using embedding to generate an embedding vector matrix, thereby obtaining the feature matrix and relation sparse matrix of the nodes. Then, a graph convolutional neural network (Att-GCN) based on attention mechanism is used to calculate the correlation between the target node (alarm signal node) and all its neighbor nodes in the knowledge graph. The calculation is performed as described in step 3 (7) to obtain the degree of influence of the neighbor nodes on the target node. The degree of influence of the neighbor nodes on the target node is fused, and the fusion calculation formula is described in step 3 (10) to obtain the hidden state of the target node. Then, the graph neural network is used to learn the features of the neighbor nodes of the target node. The hidden feature vector of the target node is obtained by fusing the features of the neighbor nodes and the relation features. The feature vector of the target node is updated by fusing the features of the neighbor nodes and the relation sparse matrix information. The overall information is obtained by fusing the hidden feature vector of the target node. Finally, the best prediction result is obtained by link prediction on the power monitoring event knowledge graph, and an auxiliary scheme is given.

[0077] The above description is only a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. Any equivalent modifications or changes made by those skilled in the art based on the content disclosed in the present invention should be included within the scope of protection set forth in the claims.

Claims

1. A method for power monitoring event recognition and intelligent analysis based on graph neural networks, characterized in that: The method includes the following steps: Step 1: Collect alarm information when the power grid generates an alarm signal; Step 2: Construct a power monitoring event knowledge graph by extracting, fusing, and processing the alarm information. In step 2, an optimized model based on Bi-LSTM-CRF is used to extract knowledge from the unstructured data of alarm signals. Before inputting the model, word vectors near numbers in the event information are weighted according to distance, and the number vectors are re-assigned. The formula for calculating the number vectors of the optimized extraction model is as follows: in, Middle and front Each data point serves as a weight for the word vector, compared to the numerical vector. Multiplying them yields a newly defined numerical vector; Indicates the first The weights of each vector relative to the numerical vector. This represents the numerical vector calculated by the optimized extraction model; Step 3: Use a graph neural network with attention mechanism to read the graph structure features of the power monitoring event knowledge graph, calculate the attention scores between nodes to obtain node feature vectors, combine the entity features and relation features of nodes to obtain bidirectional hidden states, and calculate the entity hidden feature vectors. In step 3, a graph neural network with an attention mechanism is used to learn the features of the power monitoring event knowledge graph. Features of neighboring nodes are added to the features of the node itself. The attention score between each labeled node and its one-step neighbor nodes is calculated to obtain the node feature vector. The feature vectors of two nodes are then calculated. and The correlation formula between them is: in, for 1 neighboring node, This is the neighborhood entity feature matrix obtained by performing a linear transformation on the neighboring nodes. For self-attention, It is a feedforward neural network; Step 3 integrates the entity and relational features of the research nodes to mine the bidirectional hidden states of the research entities in the power monitoring event graph when they are the head and tail entities, respectively, and infers the hidden feature vectors of the research entities. The calculation formula is as follows: in, This represents the forward hidden feature vector. This represents the backward hidden feature vector. Representing a sparse relational matrix Hidden state obtained by convolution operation Multiplication operation; Step 4: Based on the graph neural network with attention mechanism, reason about the fault type of the event in the knowledge graph and provide an auxiliary solution.

2. The power monitoring event recognition and intelligent analysis method based on graph neural networks according to claim 1, characterized in that: The alarm signals collected in step 1 include accidents, anomalies, exceeding limits, changes in position, and notifications.

3. The power monitoring event recognition and intelligent analysis method based on graph neural networks according to claim 1, characterized in that: In step 2, knowledge is extracted from the entity names in the alarm information using named entity recognition technology. Based on term comparison in natural language processing, this knowledge is segmented. Through ontology mapping and instance matching, the alarm information is classified according to the same location, line, cause, etc., and these alarm information are fused. The fused knowledge is then checked for irrelevant alarm information being classified together through confidence evaluation, and the reliability of the knowledge is calculated to complete the knowledge processing. Based on common fault events in substations, the main protection action trajectory (SOE) is obtained. Then, the initially defined process is summarized and organized to construct a complete power monitoring event knowledge graph.

4. The power monitoring event identification and intelligent analysis method based on graph neural networks according to claim 1, characterized in that: Step 4 inputs the alarm signal into the power monitoring event knowledge graph, and uses an attention-based graph convolutional neural network Att-GCN to calculate the correlation between the alarm signal entity and all its neighboring entities in the knowledge graph, accurately infers the hidden feature vector of the alarm signal, and then links and predicts to determine the fault type and match auxiliary schemes.

Images