Transformer fault rate short-term prediction method based on AM-STGCN

Abstract

The invention discloses an AM-STGCN-based transformer fault rate short-term prediction method, and the method comprises the steps: carrying out the data processing of a transformer based on attribute analysis, and carrying out the feature representation of the transformer; then, on the basis of transformer feature model construction and finally, model training, the extracted attribute feature and structural feature representation are used as input of AM-STGCN to obtain the transformer predicted fault rate. According to the method provided by the invention, comprehensive monitoring, analysis and early warning of the health degree of the transformer can be realized, so that transformer shutdown, protection maloperation or operation refusal events caused by transformer faults are reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of transformer fault prediction, and in particular to a transformer fault rate short-term prediction method based on AM-STGCN. Background Art

[0002] As the core equipment for power conversion, the reliability of transformers is of vital importance. A lot of manpower and material resources are spent on transformer operation and inspection every year. The main reason is that transformers are highly independent and real-time individuals, and it is difficult to implement fault diagnosis and status assessment online. Transformers are affected by the environment such as temperature, humidity, and climate in different locations during the year-round non-stop operation. There are problems such as equipment aging and damage, node adhesion of relays, and unstable power supply voltage, which lead to abnormal exit of transformers. In severe cases, it will lead to malfunction and serious consequences.

[0003] At present, the main research is based on fault diagnosis and analysis after the transformer fault occurs. Therefore, if the transformer fault can be predicted before it occurs and measures can be taken immediately, it will be of great significance to the safe and stable operation of the power system. It is also necessary for the maintenance and overhaul strategy based on the reliability of the transformer. At the same time, it will help to improve the utilization rate of the device and reduce cost losses. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a short-term prediction method for transformer failure rate based on AM-STGCN, which effectively mines and analyzes the feature extraction in the transformer spatiotemporal series to ensure the prediction accuracy.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: A short-term prediction method for transformer failure rate based on AM-STGCN includes the following steps: Step 1: Obtain the fault rate feature representation of the station transformer through the graph convolution neural network GCN; including Step 1.1, extract the climate parameter data such as location temperature and humidity of the station transformer, use the MIC coefficient to pre-process the data, and then obtain the adjacency matrix K between each transformer to be predicted; Step 1.2, according to the spectral analysis method, that is, according to the convolution theorem on the graph, the graph convolution is defined from the spectral domain, and the Laplacian matrix L=DK of the defined graph A is obtained to obtain the regularized Laplacian matrix; Step 1.3, the positive definite matrix As a basis, the Fourier transform representation of the signal x in A is obtained; Step 2: Establish a gated recurrent unit network GRU model that integrates the Attention mechanism, including: Step 2.1: Calculate the hidden vector corresponding to each instantaneous quantity inside the transformer features using the gated recurrent unit network (GRU). ; Step 2.2: Weight and sum the attention probability distribution values assigned to each hidden state based on the Attention mechanism. ; Step 2.3: Calculate the attention layer state value at time t containing key information based on the hidden vector and the attention probability distribution value ; ; Step 2.4: Map the spatio-temporal features of the transformer and the failure rate to obtain a non-linear relationship, and the model construction is completed; Step 3: Model training and failure rate prediction, including: Step 3.1: Divide the failure rate features of the substation transformer obtained in Step 1 into a transformer training data set and a test data set, and input the transformer training data set into the structural model; Step 3.2: Complete the model iteration by judging whether the loss function no longer decreases; Step 3.3: Input the transformer test data set; Step 3.4: Output the failure rate prediction result.

[0006] The specific process of the above Step 1.1 is as follows: Extract the climate parameter data such as the position temperature and humidity of the substation transformer, and use the MIC coefficient to express the spatio-temporal correlation degree between transformers. The MIC coefficient is expressed as: ;(1) In the formula, ab < B represents the constraint condition of the quantity; B represents the 0.6th power of the total number of samples; I(X; Y) represents the mutual information; Define the feature map of multiple transformers and multiple elements as A = (V, E), and the set of nodes composed of N transformers , determined by the MIC coefficient between the historical data of transformers in different positions , and obtain the adjacency matrix K between transformers, which is expressed as: ;(2).

[0007] In the above Step 1.2, the regularized Laplacian matrix can be expressed as: ;(3) In the formula, represents the identity matrix; represents the degree matrix; ; It indicates that the L spectrum is transformed into the product of eigenvector and eigenvalue; .

[0008] The specific process of Step 1.3 above is: The Fourier transform of the signal x in A is expressed as: ; (4) After inverse Fourier transform of both ends of equation (4), we can get x: ; (5) For signals x and y, after Fourier transform and inverse transform, the graph convolution operator can be combined with the convolution theorem: ; (6) In the formula, represents the graph convolution operator; represents the Hadamard product, that is, the product of corresponding elements between vectors; By convolution kernel Instead, the Hadamard algorithm can be transformed into matrix multiplication: ; (7) In the formula, * represents the convolution operation; represents the learning parameters; Finally, the convolution hierarchical propagation formula of the local structural characteristic map A of the power station transformer is expressed as: ; (8) In the formula, represents node self-connection; X represents the feature matrix; Represents the learning parameters of the lth layer.

[0009] In the above Step 2.1, the gated recurrent unit network GRU is one of the recurrent neural network deformations. To update the gate, control the influence of the current state information of the transformer on the previous state; To reset the gate, control the integration degree of the current state information of the transformer with that of the previous moment, the formula is: ; (9) ; (10) ; (11) ; (12) In the formula, Represents the state of the graph convolution process; represents the hidden state at time t; Indicates the feature information at the current moment; Indicates the influence of the current state input on the hidden layer state hand; is the sigmoid activation function; , represents the gate structure weight matrix; , represents the gate structure bias vector; Represents the updated state weight matrix; Represents the updated state bias vector.

[0010] The above Step 2.2-Step 2.4 process is: set up for t The state value at each moment, the Attention mechanism is introduced into the GRU model, and the calculation formula is: ; (13) ; (14) ; (15) in represents the attention score; represents the attention probability distribution; u and is the weight coefficient; b is the bias coefficient.

[0011] The present invention provides a short-term prediction method for transformer failure rate based on AM-STGCN, and proposes a spatiotemporal graph convolutional neural network with an attention fusion mechanism, namely, the AM-STGCN fault prediction model predicts the failure rate of plant transformers. The Attention mechanism is integrated into the gated recurrent unit network to establish a model, and the data of the plant transformer is extracted through the graph convolution neural network GCN, and the model is trained and tested, which can significantly improve fault prediction under multiple environmental climatic conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0013] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and embodiments.

[0014] A short-term prediction method for transformer failure rate based on AM-STGCN is characterized by comprising the following steps: 1. Transformer Data Feature Extraction Hydroelectric power station transformers are affected by environmental factors such as location, temperature, humidity, and climate. To deeply explore the complex spatio-temporal characteristics coupled by transformers, the MIC coefficient is used to express the spatio-temporal correlation degree between transformers, and the MIC coefficient can be expressed as: ;(1) In the formula, ab < B represents the constraint condition of quantity; B represents the 0.6th power of the total number of samples; I(X;Y) represents mutual information.

[0015] Define the feature map of multiple transformers and multiple elements as A = (V, E), and the set of nodes consists of N transformers , determined by the MIC coefficient between the historical data of transformers at different locations , and the adjacency matrix K between transformers can be obtained as: ;(2).

[0016] II. Establishment of Transformer Prediction Model In actual engineering, power enterprises are more interested in the failure rate of the current or future transformer group, so as to realize the full-coverage failure rate monitoring, analysis, and early warning of substation transformers.

[0017] The installation locations of transformers are scattered and affected by environmental factors such as location, temperature, humidity, and climate. Existing prediction models do not establish relevant models based on the spatio-temporal characteristics of transformers, so as to effectively and accurately predict the failure rate of transformers with different spatio-temporal characteristics. Therefore, the focus of the present invention is to find a model that can deeply explore the internal correlation of transformers with different spatio-temporal characteristics and propose a prediction method suitable for short-term prediction.

[0018] Propose a spatio-temporal graph convolutional neural network fault prediction model with a fusion mechanism to predict the failure rate of substation transformers. The model is denoted as AM-STGCN, and the specific steps are as follows: (1) The graph convolutional neural network GCN realizes the local structural characteristics of substation transformers and representation: ① MIC data preprocessing, and then obtain the adjacency matrix K between transformers to be predicted; ② According to the spectral graph, the Laplacian matrix L of the defined graph A = D - K, and the regularized Laplacian matrix can be expressed as: ;(3) In the formula, represents the identity matrix; represents the degree matrix; ; represents the conversion of the L spectrum into the product of eigenvectors and eigenvalues; .

[0019] ③ is a positive definite matrix, and taking it as the basis, the Fourier transform of the signal x in A is expressed as: ; (4) After inverse Fourier transform of both ends of equation (4), we can get x: ; (5) For signals x and y, after Fourier transform and inverse transform, the graph convolution operator can be combined with the convolution theorem: ; (6) In the formula, represents the graph convolution operator; represents the Hadamard product, that is, the product of corresponding elements between vectors; By convolution kernel Instead, the Hadamard algorithm can be transformed into matrix multiplication: ; (7) In the formula, * represents the convolution operation; represents the learning parameters.

[0020] Finally, the convolution hierarchical propagation formula of the local structural characteristic map A of the power station transformer is expressed as: ; (8) In the formula, represents node self-connection; X represents the feature matrix; Represents the learning parameters of the lth layer.

[0021] (2) Gated Recurrent Unit Network (GRU) with Attention Mechanism ① GRU extracts the time series characteristics of the fault factor diagram of the power station transformer GRU is one of the recurrent neural network deformations. As the update gate, it can realize the "forget" and "select memory" functions to control the influence of the transformer's current state information on the previous moment; To reset the gate, control the integration degree of the current state information of the transformer with that of the previous moment, the formula is: ; (9) ; (10) ; (11) ; (12) In the formula, Represents the state of the graph convolution process; represents the hidden state at time t; Indicates the feature information at the current moment; Indicates the influence of the current state input on the hidden layer state hand; is the sigmoid activation function; , represents the gate structure weight matrix; , represents the gate structure bias vector; Represents the updated state weight matrix; Represents the updated state bias vector.

[0022] ②Attention mechanism enhances key information The Attention mechanism can reliably solve the long-range gradient vanishing problem of GRU gating characteristics, enhance the contribution of key information related to the transformer, deepen the model feature mining capability, and design for t The state value at a moment is calculated as follows: ; (13) ; (14) ; (15) in represents the attention score; represents the attention probability distribution; u and is the weight coefficient; b is the bias coefficient.

[0023] 3. Model training and failure rate prediction Input the transformer training data set into the structural model; complete the model iteration by judging whether the loss function no longer decreases; input the transformer test data set; output the failure rate prediction result.

[0024] The application client using this method is deployed in the intranet of the Three Gorges Power Plant and can be accessed through the subsystem under the web page. The server of this project is deployed in Area III. The failure rate of the relay protection device is obtained from the industrial Internet platform in Area IV. To ensure network security, a firewall needs to be installed between the two areas, and customized software needs to be deployed on the big data platform side of Area IV for data transmission.

Claims

1. A short-term prediction method for transformer failure rate based on AM-STGCN, characterized in that: It includes the following steps: Step 1. Obtain the failure rate feature representation of the substation transformer through the graph convolutional neural network GCN; It includes Step 1.

1. Extract the position temperature and humidity parameter data of the substation transformer, preprocess the data using the MIC coefficient, and then obtain the adjacency matrix K between each transformer to be predicted; Step 1.

2. Use the spectral graph analysis method, that is, define the graph convolution from the spectral domain according to the graph convolution theorem, and obtain the Laplacian matrix L = D - K of the defined graph A to get the regularized Laplacian matrix; Step 1.3, the positive definite matrix As a basis, the Fourier transform representation of the signal x in A is obtained; Step 2. Establish a gated recurrent unit network GRU model integrating the Attention mechanism, including: Step 2.

1. Use the gated recurrent unit network GRU to calculate the hidden vector corresponding to each instantaneous quantity inside the transformer feature. ; Step 2.2: Weighted sum of the attention probability distribution values ​​assigned to each hidden state based on the Attention mechanism ; Step 2.3, based on hidden vector and the attention probability distribution value Calculate the state value of the attention layer at time t ; Step 2.

4. Map the non-linear relationship between the spatio-temporal characteristics and the failure rate of the transformer, and complete the construction of the model; Step 3. Model training and failure rate prediction, including: Step 3.

1. Divide the failure rate feature representation of the substation transformer obtained in Step 1 into a transformer training data set and a test data set, and input the transformer training data set into the structure model; Step 3.

2. Complete the model iteration by judging whether the loss function no longer decreases; Step 3.

3. Input the transformer test data set; Step 3.

4. Output the failure rate prediction result.

2. According to claim 1, a transformer failure rate short-term prediction method based on AM-STGCN is characterized in that: The specific process of the above Step 1.1 is: Extract the position temperature and humidity parameter data of the substation transformer, and use the MIC coefficient to express the spatio-temporal correlation degree between each transformer. The MIC coefficient is expressed as: ； (1) In the formula, ab < B represents the constraint condition of the quantity; B represents the 0.6th power of the total number of samples; I(X;Y) represents the mutual information; Define the characteristic graph of multiple transformers and multiple elements as A=(V, E), and N transformers form a node set , determined by the MIC coefficient between the historical data of transformers at different locations , the adjacency matrix K between each transformer is obtained as: ； (2)。 3. According to claim 2, a transformer failure rate short-term prediction method based on AM-STGCN is characterized in that: In the above Step 1.2, the regularized Laplacian matrix can be expressed as: ；(3) In the formula, represents the identity matrix; represents the degree matrix; ; It indicates that the L spectrum is transformed into the product of eigenvector and eigenvalue; .

4. According to claim 3, a transformer failure rate short-term prediction method based on AM-STGCN is characterized in that: The specific process of the above Step 1.3 is: The Fourier transform of the signal x in A is expressed as: ；(4) After performing the inverse Fourier transform on both ends of Equation (4), x can be obtained: ；(5) For the signals x and y, after Fourier transform and inverse transform, combined with the convolution theorem, the graph convolution operator can be obtained: ； (6) In the formula, represents the graph convolution operator; represents the Hadamard product, that is, the product of corresponding elements between vectors; By convolution kernel Instead, the Hadamard algorithm can be transformed into matrix multiplication: ；(7) In the formula, * represents the convolution operation; represents the learning parameters; Finally, the formula for the convolutional hierarchical propagation of the local structure feature map A of the substation transformer is expressed as: ； (8) In the formula, represents node self-connection; X represents the feature matrix; Represents the learning parameters of the lth layer.

5. According to claim 4, a transformer failure rate short-term prediction method based on AM-STGCN is characterized in that: In the above Step 2.1, the gated recurrent unit network GRU is one of the recurrent neural network deformations. To update the gate, control the influence of the current state information of the transformer on the previous state; To reset the gate, control the integration degree of the current state information of the transformer with that of the previous moment, the formula is: ； (9) ；(10) ；(11) ； (12) In the formula, Represents the state of the graph convolution process; represents the hidden state at time t; Indicates the feature information at the current moment; Indicates the influence of the current state input on the hidden layer state hand; is the sigmoid activation function; , represents the gate structure weight matrix; , represents the gate structure bias vector; Represents the updated state weight matrix; Represents the updated state bias vector.

6. The short-term prediction method for transformer failure rate based on AM-STGCN according to claim 5 is characterized in that: The processes of the above Step 2.2 - Step 2.4 are: set up for t The state value at each moment, the Attention mechanism is introduced into the GRU model, and the calculation formula is: ；(13) ；(14) ；(15) in represents the attention score; represents the attention probability distribution; u and is the weight coefficient; b is the bias coefficient.

Images