A transformer status monitoring method and system
Through the combination of multi-class state data and neural network models, a transformer state monitoring system is built, which solves the subjectivity and environmental noise interference problems of traditional monitoring solutions, and achieves higher monitoring accuracy and efficiency.
Patent Information
- Application Number
- CN202411124086.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-16
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2044-08-16
AI Technical Summary
Traditional transformer monitoring solutions rely on manual judgment and are susceptible to subjective factors and have poor accuracy. They are susceptible to environmental noise and other factors when monitoring through transformer sound signal, resulting in a decrease in monitoring interpretability and accuracy.
Multiple state data (dissolved gas in oil, local discharge, iron core grounding current, oil temperature and oil pressure) are used for monitoring, and dimensionality reduction is reduced through principal component analysis, combined with recurrent neural network and convolutional neural network to build a state monitoring model to perform Global and local feature extraction and feature fusion to realize automated monitoring.
No manual participation is required, and subjective factors are avoided, which improves the accuracy and interpretability of transformer status monitoring, reduces unnecessary calculations and data processing, and improves monitoring efficiency and accuracy.
Smart Images

Figure CN118655398B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of data processing, and in particular relates to a transformer state monitoring method and system. Background Art
[0002] The transformer is the core equipment in the power grid, and its operating status is directly related to the safe operation of the equipment and the system. How to ensure the long-term and stable operation of the transformer equipment has always been the core content of the operation and management of the power grid equipment.
[0003] The traditional transformer monitoring solution is to check the oil temperature and oil level of the transformer oil in the transformer, as well as the gas composition in the transformer, and to judge whether the transformer is faulty by manual means. This is time-consuming and labor-intensive, and is easily affected by subjective factors. The accuracy of transformer monitoring is poor.
[0004] With the rapid development of technologies such as pattern recognition, machine vision, and deep learning, as well as the urgent need for transformer condition monitoring, more and more modern technologies are being applied to transformer condition monitoring.
[0005] At present, there are automated solutions for monitoring transformer status based on transformer sound signals, which can automatically monitor transformer status. However, status monitoring through transformer sound signals is easily interfered by environmental noise, and sound signals are also affected by factors such as load changes, temperature, and humidity, resulting in poor interpretability of status monitoring and reducing the accuracy of transformer status monitoring. Summary of the invention
[0006] In order to solve the technical problems that the traditional transformer monitoring solution relies on manual means to judge whether the transformer has a fault, which is time-consuming and labor-intensive, and is easily affected by subjective factors, the accuracy of transformer monitoring is poor, and status monitoring through transformer sound signals is easily interfered by environmental noise, and the sound signal is also affected by factors such as load changes, temperature, and humidity, resulting in poor interpretability of status monitoring and reducing the accuracy of transformer status monitoring, the present invention provides a transformer status monitoring method and system.
[0007] First aspect
[0008] The present invention provides a transformer state monitoring method, comprising:
[0009] S1: continuously acquiring transformer status data, wherein the transformer status data includes multiple types of status data: dissolved gas in oil, partial discharge, core grounding current, oil temperature and oil pressure;
[0010] S2: performing dimension reduction processing on the transformer state data through principal component analysis;
[0011] S3: constructing a transformer state monitoring model, and monitoring the state of the transformer through the transformer state monitoring model; S3 specifically includes:
[0012] S301: extracting global features of the transformer state data through a recurrent neural network;
[0013] S302: Preliminarily monitor the state of the transformer according to the fusion feature, and determine whether the preliminary monitoring result of the transformer is faulty. If so, proceed to the next step; otherwise, return to S1;
[0014] S303: extracting local features of the transformer state data through a convolutional neural network;
[0015] S304: Fusing the local feature with the global feature to obtain a fused feature;
[0016] S305: Based on the fusion feature, the transformer status is carefully monitored to determine whether the preliminary monitoring result of the transformer has a fault. If so, an alarm is issued; otherwise, the process returns to S1.
[0017] Second aspect
[0018] The present invention provides a transformer state monitoring system, comprising a processor and a memory for storing processor executable instructions; the processor is configured to call the instructions stored in the memory to execute the transformer state monitoring method in the first aspect.
[0019] Compared with the prior art, the present invention has at least the following beneficial technical effects:
[0020] (1) In the present invention, based on multiple types of transformer status data such as dissolved gas in oil, partial discharge, core grounding current, oil temperature and oil pressure, the transformer status is automatically and comprehensively monitored through the transformer status monitoring model, which can more comprehensively evaluate the status of the transformer without human intervention, saving time and effort, avoiding the influence of subjective factors, and not being disturbed by environmental noise. The status monitoring has strong interpretability, thereby improving the accuracy of transformer monitoring.
[0021] (2) In the present invention, a preliminary monitoring is first performed through a recurrent neural network. When it is determined that the preliminary monitoring result of the transformer is faulty, feature fusion is further performed with a convolutional neural network to more accurately monitor the transformer status. If the preliminary monitoring result shows that the transformer status is normal, time and computing resources can be saved, and there is no need to perform more complex feature extraction and analysis, which can reduce unnecessary calculations and data processing.
[0022] (3) In the present invention, the recurrent neural network can capture the long-term dependencies in the time series, while the convolutional neural network can capture the local features. Through feature fusion, the data can be analyzed more comprehensively, thereby improving the accuracy of transformer monitoring. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] The preferred implementation modes will be described below in a clear and understandable manner with reference to the accompanying drawings to further illustrate the above-mentioned characteristics, technical features, advantages and implementation methods of the present invention.
[0024] Figure 1 It is a flow chart of a transformer status monitoring method provided by the present invention.
[0025] Figure 2 It is a structural schematic diagram of a transformer state monitoring system provided by the present invention. DETAILED DESCRIPTION
[0026] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the specific implementation of the present invention will be described below with reference to the accompanying drawings. Figure 1 , showing a flow chart of a transformer state monitoring method provided by the present invention.
[0027] The present invention provides a transformer state monitoring method, comprising:
[0028] S1: Continuously obtain transformer status data.
[0029] Among them, the transformer status data includes multiple types of status data: dissolved gas in oil, partial discharge, core grounding current, oil temperature and oil pressure.
[0030] Among them, dissolved gas in transformer insulating oil is a useful monitoring indicator. When a fault or partial discharge occurs inside the transformer, gas is released. Monitoring the gas content in the oil can help detect whether there is a potential problem with the transformer. Different types of gases can be associated with specific fault types, so their monitoring can provide clues about the type of fault.
[0031] Partial discharge is a discharge phenomenon that usually occurs in small areas of the insulation system. This phenomenon may be an early sign of internal transformer faults, such as local insulation damage. Monitoring partial discharge can help detect these problems early so that maintenance measures can be taken to avoid more serious failures.
[0032] Among them, core ground current is a current in the transformer core, which is usually generated when the core is faulty or grounded. Monitoring the core ground current can help detect the condition of the core and whether there is a grounding problem.
[0033] Among them, the oil temperature of the transformer is an important monitoring parameter, because the increase in temperature may cause aging and damage to the insulation material. Monitoring the oil temperature can help ensure that the transformer operates within a safe temperature range and prevent overheating from causing failures.
[0034] Oil pressure refers to the pressure of the transformer's insulating oil. Oil pressure monitoring can help detect if there is an oil leak or oil leakage problem, as a drop in oil pressure can be a sign of these problems.
[0035] S2: Perform dimension reduction processing on transformer status data through principal component analysis.
[0036] Among them, Principal Component Analysis (PCA) is a commonly used multivariate statistical analysis method for dimensionality reduction and data dimensionality reduction processing. Its main goal is to reduce the dimension of the data while retaining as much information in the original data as possible.
[0037] In a possible implementation, S2 specifically includes sub-steps S201 to S207:
[0038] S201: Combine transformer status data into a transformer status matrix A .
[0039] S202: Standardize the transformer state matrix to obtain a standardized matrix C :
[0040]
[0041] in, c ik Represents the first i Class k Status data, a ik Represents the first i Class k Status data, Indicates i The average value of class state data, s i Indicates i The standard deviation of the class status data.
[0042] In the present invention, different categories of status data are placed on the same scale, eliminating unit differences between different data, which helps to better compare and analyze data.
[0043] S203: According to the standardized matrix, solve the correlation coefficients between various types of state data to obtain the correlation coefficient matrix R :
[0044]
[0045] in, r ij Indicates i Class status data and j The correlation coefficient of class status data, c ik Represents the first i Class k Status data, Represents the first i The average value of class state data, c jk Represents the first j Class k Status data, Represents the first j The average value of class state data, , K Indicates the total number of data.
[0046] In the present invention, calculating the correlation coefficient matrix helps to understand the relationship between data of different states. The correlation coefficient matrix can reveal the linear relationship between data and help identify the interdependence between data.
[0047] S204: Solve the eigenvalues and eigenvectors of the correlation coefficient matrix:
[0048]
[0049] in, R represents the correlation coefficient matrix, l Represents the discriminant matrix A The characteristic value of p Represents the discriminant matrix A The characteristic vector of I Represents the unit matrix, and arranges the eigenvalues in descending order as , n Represents the total number of eigenvalues.
[0050] S205: Calculate the variance contribution of each principal component:
[0051]
[0052] in, α i Indicates i The variance contribution of the principal components is l i It means that the eigenvalues are arranged in descending order. ifeature values.
[0053] S206: Calculate the cumulative contribution rate of the principal component:
[0054]
[0055] in, β j Before j The cumulative contribution rate of the principal components, l j It means that the eigenvalues are arranged in descending order. j feature values.
[0056] S207: When the cumulative contribution rate reaches the preset contribution rate, the principal component is retained and other components are discarded.
[0057] In the present invention, principal component analysis can help simplify data, improve computational efficiency, remove noise, visualize data, and retain key information. In transformer status data monitoring, it can help improve the efficiency of data processing and analysis, and help identify key information and potential problems in transformer status.
[0058] S3: Construct a transformer status monitoring model and monitor the status of the transformer through the transformer status monitoring model. S3 specifically includes:
[0059] S301: Extracting global features of transformer status data through a recurrent neural network.
[0060] Among them, Recurrent Neural Network (RNN) is a deep learning model that is mainly used to process sequence data and time series data. RNN has a recursive structure that allows information to be passed in the network, enabling it to capture the time dependency and order information in the data.
[0061] In a possible implementation, the recurrent neural network includes: an input layer, a state layer, and an attention layer. S301 specifically includes:
[0062] S3011: In the input layer, input transformer status data sequence ,in, C i Indicates i Transformer status data, m Indicates the total number of transformer status data, .
[0063] S3012: In the state layer, calculate i The hidden state of the transformer state data:
[0064]
[0065] in, h i Indicates i The hidden state of the transformer state data, Indicates i The previous state of the forward loop of the transformer state data, Indicates i The previous state of the backward loop of transformer state data, Indicates i -1 The previous state of the forward loop of the transformer state data, Indicates i +1 The previous state of the forward loop of the transformer state data, GRU() represents the nonlinear calculation after the recurrent neural network, u i express The weight coefficient of v i express The weight coefficient of b i Indicates i The bias term of the transformer status data.
[0066] S3013: In the attention layer, weights are assigned to each transformer state data, and the global features of the transformer state data are accumulated:
[0067]
[0068] in, H 1 represents the global characteristics of transformer status data, c i Indicates i The weight of each transformer status data, h i Indicates i The hidden state of the transformer state data.
[0069] In the present invention, the sequence modeling capability and attention mechanism of the recurrent neural network (RNN) are fully utilized to extract the global features of the transformer status data more comprehensively and accurately. This helps to improve the effect of transformer status monitoring and enhance the ability to identify and predict potential problems.
[0070] S302: Perform preliminary monitoring of the transformer status according to the global characteristics to determine whether there is a fault in the preliminary monitoring result of the transformer. If so, proceed to the next step; otherwise, return to S1.
[0071] In a possible implementation, S302 specifically includes:
[0072] S3021: Output the preliminary monitoring value of the transformer status according to the global characteristics of the transformer status data:
[0073]
[0074] in, S 1 represents the initial monitoring value of the transformer state, Softmax() represents the Softmax activation function, W 1 represents the weight coefficient of preliminary monitoring, H 1 represents the global characteristics of transformer status data, B 1 indicates the bias term for preliminary monitoring.
[0075] S3022: When the initial monitoring value of the transformer state is greater than the first preset value, it is determined that the transformer is faulty and the process proceeds to the next step; otherwise, the process returns to S1.
[0076] In the present invention, in transformer status monitoring, by automating preliminary monitoring and detecting faults in advance, the efficiency and accuracy of monitoring are improved, the manual workload is reduced, and it is helpful to maintain the stability and safety of power grid equipment. If the preliminary monitoring results show that the transformer status is normal, time and computing resources can be saved, and more complex feature extraction and analysis are not required, which can reduce unnecessary calculations and data processing.
[0077] S303: Extracting local features of transformer status data through a convolutional neural network.
[0078] Among them, Convolutional Neural Network (CNN) is a deep learning neural network architecture, which is mainly used to process and analyze data with grid structure.
[0079] In a possible implementation, the convolutional neural network includes: an input layer, a convolutional layer, a pooling layer, and a fully connected layer. S303 specifically includes sub-steps S3031 to S3034:
[0080] S3031: In the input layer, input transformer status data.
[0081] S3032: In the convolutional layer, extract the data features of the transformer status data:
[0082]
[0083] in, Indicates the current convolutional layer j The output of each channel, Represents the previous convolutional layerj In the channel i The output of the convolution kernel, Represents the convolution kernel weight of the current convolution layer, represents the bias term of the current convolutional layer, represents the selected input feature map, f c () represents the convolutional layer activation function.
[0084] S3033: In the pooling layer, the features extracted by the convolutional layer are compressed by dimensionality reduction:
[0085]
[0086] in, Indicates the current pooling layer j The output of each channel, f p () represents the pooling layer activation function, Represents the multiplication bias of the current pooling layer, f down () represents the downsampling function, Indicates the previous pooling layer j The output of each channel, Represents the additive bias of the current pooling layer.
[0087] S3034: In the fully connected layer, the output of the pooling layer is summarized to obtain the local features of the transformer state data.
[0088] In the present invention, the use of convolutional neural network (CNN) for local feature extraction helps to improve the ability to analyze and understand transformer status data, automate feature extraction, reduce dimensionality and compress data, and improve the efficiency and accuracy of the model. This helps to better monitor the transformer status and detect problems in advance.
[0089] S304: Fusing local features with global features to obtain fused features.
[0090] In a possible implementation, S304 specifically includes: performing feature fusion on the local features and the global features according to the following formula:
[0091]
[0092] in, H Represents the fusion features of transformer status data, H 1 represents the global characteristics of transformer status data, m represents the weight of the global feature, H 2 represents the local characteristics of transformer status data.
[0093] In the present invention, global features and local features are fused together, allowing the model to comprehensively consider the information of the data from different perspectives. Global features can provide overall status information, while local features can provide detailed local information. By fusing this information, the transformer status can be understood more comprehensively, thereby improving the accuracy of transformer status monitoring.
[0094] S305: Based on the fusion features, the transformer status is carefully monitored to determine whether the initial monitoring result of the transformer indicates a fault. If so, an alarm is issued; otherwise, the process returns to S1.
[0095] In a possible implementation, S305 specifically includes sub-steps S3051 to S3052:
[0096] S3051: Output the detailed monitoring value of the transformer status according to the fusion characteristics of the transformer status data:
[0097]
[0098] in, S 2 represents the fine monitoring value of the transformer state, Softmax() represents the Softmax activation function, W 2 represents the weight coefficient of fine monitoring, H Represents the fusion features of transformer status data, B 2 represents the bias term for fine monitoring.
[0099] S3052: When the fine monitoring value of the transformer status is greater than the second preset value, it is determined that the transformer is faulty and an alarm is issued; otherwise, the process returns to S1.
[0100] In the present invention, the state of the transformer is finely monitored based on the fusion features, which can more accurately evaluate the state of the transformer without human intervention, saving time and effort, avoiding the influence of subjective factors, and not being disturbed by environmental noise. The state monitoring has strong interpretability and improves the accuracy of transformer monitoring.
[0101] In a possible implementation, the training method of the transformer condition monitoring model includes:
[0102] With the highest monitoring accuracy as the goal, the objective function of the transformer condition monitoring model is constructed:
[0103]
[0104] in, F represents the objective function, max() represents the maximum value function, Accuracy It represents the detection accuracy, 1 Indicates the number of samples classified correctly, Total represents the total number of samples, i represents the set of model parameters, , a total of 11 model parameters, i 1 represents the number of layers of the convolutional neural network. i 2 represents the convolution kernel size of the convolutional neural network. i 3 represents the number of convolution kernels of the convolutional neural network. i 4 represents the number of layers of the recurrent neural network. i 5 represents the hidden layer dimension of the recurrent neural network, i 6 represents the discard ratio of the discard layer of the recurrent neural network. m represents the weight of the global feature, W 1 represents the weight coefficient of preliminary monitoring, W 1 represents the weight coefficient of preliminary monitoring, B 1 represents the bias term for preliminary monitoring, W 2 represents the weight coefficient of fine monitoring, B 2 represents the bias term for fine monitoring.
[0105] The transformer condition monitoring model is trained with the goal of maximizing the function value of the objective function.
[0106] In the present invention, the training method with monitoring accuracy as the highest goal helps to build a transformer state monitoring model with excellent performance, improve accuracy, adaptability and comprehensive performance, and is conducive to improving the stability and reliability of power grid equipment.
[0107] In a possible implementation, the transformer state monitoring model is trained with the goal of maximizing the function value of the objective function, specifically including:
[0108] Initialize the population, which includes multiple individuals X , each individual X Represents a set of possible parameters i .
[0109] Calculate the function value of each individual's objective function, and determine the current optimal solution based on the function value of each individual's objective function X max .
[0110] Update the position of each individual according to the following formula:
[0111]
[0112] in, X t+1 Indicates t +1 individual position at iteration, X t Indicates t The individual position at the iteration,L Represents the search path, represents point multiplication, l Represents the position update parameters, X max represents the current optimal solution, r t Indicates t The step size adjustment factor for the iteration.
[0113] In the present invention, extensive exploration can be performed in the parameter space, so that the best possible parameter set can be found, and the interaction between different parameters is taken into account, so as to facilitate finding the best combination.
[0114] Among them, the step length adjustment coefficient r The method of determining is:
[0115]
[0116] in, r t Indicates t The distance control parameter at the iteration, r 0 represents the initial distance, t Indicates the current iteration number, T Indicates the maximum number of iterations.
[0117] In the present invention, the step length adjustment coefficient r The introduction of allows the search step size to be adaptively adjusted according to the number of iterations. In the early stages of the iteration, the step size adjustment coefficient r Larger, which helps to conduct a wide search, and the step size adjustment coefficient in the later and early stages r Smaller, which helps with fine adjustments.
[0118] The search path L The method of determining is:
[0119]
[0120] in, L Represents the search path, r u and r v Both represent random numbers that satisfy the normal distribution.
[0121] In the present invention, a normally distributed random number is introduced r u and r v To determine the search path L , which increases the randomness of the algorithm, helps to escape from the local optimal solution and further improve the effect of global search.
[0122] Continue iterating until the maximum number of iterations is reached T , the individual that outputs the function value of the objective function at this time is the optimal model parameter.
[0123] In the present invention, the training method with the goal of maximizing the objective function fully utilizes the advantages of the heuristic algorithm, can better find the optimal model parameters, improve the performance and generalization ability of the model, and help to improve the accuracy and reliability of the transformer condition monitoring model.
[0124] Compared with the prior art, the present invention has at least the following beneficial technical effects:
[0125] (1) In the present invention, based on multiple types of transformer status data such as dissolved gas in oil, partial discharge, core grounding current, oil temperature and oil pressure, the transformer status is automatically and comprehensively monitored through the transformer status monitoring model, which can more comprehensively evaluate the status of the transformer without human intervention, saving time and effort, avoiding the influence of subjective factors, and not being disturbed by environmental noise. The status monitoring has strong interpretability, thereby improving the accuracy of transformer monitoring.
[0126] (2) In the present invention, a preliminary monitoring is first performed through a recurrent neural network. When it is determined that the preliminary monitoring results of the transformer are faulty, feature fusion is further performed with a convolutional neural network to more accurately monitor the transformer status. If the preliminary monitoring results indicate that the transformer status is normal, time and computing resources can be saved, and there is no need for more complex feature extraction and analysis, which can reduce unnecessary calculations and data processing.
[0127] (3) In the present invention, the recurrent neural network can capture the long-term dependencies in the time series, while the convolutional neural network can capture the local features. Through feature fusion, the data can be analyzed more comprehensively, thereby improving the accuracy of transformer monitoring.
[0128] Example 2
[0129] In one embodiment, the reference specification Figure 2 , showing a schematic structural diagram of a transformer state monitoring system provided by the present invention.
[0130] The present invention provides a transformer state monitoring system, comprising a processor 201 and a memory 202 for storing executable instructions of the processor 201. The processor 201 is configured to call the instructions stored in the memory 202 to execute the transformer state monitoring method in embodiment 1.
[0131] A transformer state monitoring system provided by the present invention can implement the steps and effects of the transformer state monitoring method in the above-mentioned embodiment 1. To avoid repetition, the present invention will not go into details.
Claims
1. A transformer status monitoring method, characterized in that: include: S1: continuously acquiring transformer status data, wherein the transformer status data includes multiple types of status data: dissolved gas in oil, partial discharge, core grounding current, oil temperature and oil pressure; S2: performing dimension reduction processing on the transformer state data through principal component analysis; S3: constructing a transformer state monitoring model, and monitoring the state of the transformer through the transformer state monitoring model; S3 specifically includes: S301: extracting global features of the transformer state data through a recurrent neural network; S302: Preliminarily monitor the status of the transformer according to the global feature, and determine whether the preliminary monitoring result of the transformer indicates a fault. If so, proceed to the next step; otherwise, return to S1; S303: extracting local features of the transformer state data through a convolutional neural network; S304: Fusing the local feature with the global feature to obtain a fused feature; S305: Based on the fusion feature, the transformer status is carefully monitored to determine whether the preliminary monitoring result of the transformer has a fault. If so, an alarm is issued; otherwise, the process returns to S1; Wherein, the S2 specifically includes: S201: Combining the transformer state data into a transformer state matrix A ; S202: Standardize the transformer state matrix to obtain a standardized matrix C : in, c ik Represents the first i Class k Status data, a ik Represents the first i Class k Status data, Indicates i The average value of class state data, σ i Indicates i Standard deviation of class status data; S203: According to the standardized matrix, solve the correlation coefficients between various types of state data to obtain a correlation coefficient matrix R : in, r ij Indicates i Class status data and j The correlation coefficient of class status data, c ik Represents the first i Class k Status data, Represents the first i The average value of class state data, c jk Represents the first j Class k Status data, Represents the first j The average value of class state data, , K Indicates the total number of data; S204: Solving the eigenvalues and eigenvectors of the correlation coefficient matrix: in, R represents the correlation coefficient matrix, λ Represents the discriminant matrix A The characteristic value of p Represents the discriminant matrix A The characteristic vector of I Represents the unit matrix, and arranges the eigenvalues in descending order as , n represents the total number of eigenvalues; S205: Calculate the variance contribution of each principal component: in, α i Indicates i The variance contribution of the principal components is λ i It means that the eigenvalues are arranged in descending order. i eigenvalues; S206: Calculate the cumulative contribution rate of the principal component: in, β j Before j The cumulative contribution rate of the principal components, λ j It means that the eigenvalues are arranged in descending order. j eigenvalues; S207: When the cumulative contribution rate reaches a preset contribution rate, retain the main component and discard other components; The recurrent neural network includes: an input layer, a state layer and an attention layer; S301 specifically includes: S3011: At the input layer, input the transformer status data sequence ,in, C i Indicates i Transformer status data, m Indicates the total number of transformer status data, ; S3012: In the state layer, calculate i The hidden state of the transformer state data: in, h i Indicates i The hidden state of the transformer state data, Indicates i The previous state of the forward loop of the transformer state data, Indicates i The previous state of the backward loop of transformer state data, Indicates i -1 The previous state of the forward loop of the transformer state data, Indicates i +1 The previous state of the forward loop of the transformer state data, GRU() represents the nonlinear calculation after the recurrent neural network, u i express The weight coefficient of v i express The weight coefficient of b i Indicates i A bias term for transformer status data; S3013: In the attention layer, weights are assigned to each transformer state data, and the global features of the transformer state data are obtained by accumulation: in, H 1 represents the global characteristics of transformer status data, γ i Indicates i The weight of each transformer status data, h i Indicates i The hidden state of the transformer state data; Wherein, the S302 specifically includes: S3021: Outputting a preliminary monitoring value of the transformer state according to the global characteristics of the transformer state data: in, S 1 represents the initial monitoring value of the transformer state, Softmax() represents the Softmax activation function, W 1 represents the weight coefficient of preliminary monitoring, H 1 represents the global characteristics of transformer status data, B 1 represents the bias term for preliminary monitoring; S3022: When the initial monitoring value of the transformer state is greater than the first preset value, it is determined that the transformer is faulty and the next step is entered; otherwise, the process returns to S1; The convolutional neural network includes: an input layer, a convolutional layer, a pooling layer and a fully connected layer; S303 specifically includes: S3031: In the input layer, input the transformer status data; S3032: In the convolution layer, extract data features of the transformer state data: in, Indicates the current convolutional layer j The output of each channel, Represents the previous convolutional layer j In the channel i The output of the convolution kernel, Represents the convolution kernel weight of the current convolution layer, represents the bias term of the current convolutional layer, represents the selected input feature map, f c () represents the activation function of the convolutional layer; S3033: In the pooling layer, the features extracted by the convolutional layer are compressed by dimensionality reduction: in, Indicates the current pooling layer j The output of each channel, f p () represents the pooling layer activation function, Represents the multiplication bias of the current pooling layer, f down () represents the downsampling function, Indicates the previous pooling layer j The output of each channel, Represents the additive bias of the current pooling layer; S3034: In the fully connected layer, the output of the pooling layer is summarized to obtain local features of the transformer state data; Wherein, the S304 is specifically: According to the following formula, the local feature and the global feature are fused: in, H Represents the fusion features of transformer status data, H 1 represents the global characteristics of transformer status data, μ represents the weight of the global feature, H 2 represents the local characteristics of transformer status data; Wherein, the S305 specifically includes: S3051: Outputting a detailed monitoring value of the transformer state according to the fusion characteristics of the transformer state data: in, S 2 represents the fine monitoring value of the transformer state, Softmax() represents the Softmax activation function, W 2 represents the weight coefficient of fine monitoring, H Represents the fusion features of transformer status data, B 2 represents the bias term for fine monitoring; S3052: When the fine monitoring value of the transformer state is greater than the second preset value, it is determined that the transformer is faulty and an alarm is issued; otherwise, the process returns to S1; The training method of the transformer state monitoring model includes: With the highest monitoring accuracy as the goal, the objective function of the transformer condition monitoring model is constructed: in, F represents the objective function, max() represents the maximum value function, Accuracy It represents the detection accuracy, 1 Indicates the number of samples classified correctly, Total represents the total number of samples, θ represents the set of model parameters, , a total of 11 model parameters, θ 1 represents the number of layers of the convolutional neural network. θ 2 represents the convolution kernel size of the convolutional neural network. θ 3 represents the number of convolution kernels of the convolutional neural network. θ 4 represents the number of layers of the recurrent neural network. θ 5 represents the hidden layer dimension of the recurrent neural network, θ 6 represents the discard ratio of the discard layer of the recurrent neural network. μ represents the weight of the global feature, W 1 represents the weight coefficient of preliminary monitoring, W 1 represents the weight coefficient of preliminary monitoring, B 1 represents the bias term for preliminary monitoring, W 2 represents the weight coefficient of fine monitoring, B 2 represents the bias term for fine monitoring; Taking the maximum function value of the objective function as the goal, training the transformer state monitoring model; The transformer state monitoring model is trained with the goal of maximizing the function value of the objective function, specifically including: Initialize a population, the population includes multiple individuals X , each individual X Represents a set of possible parameters θ ; Calculate the function value of each individual's objective function, and determine the current optimal solution based on the function value of each individual's objective function X max ; Update the position of each individual according to the following formula: in, X t+1 Indicates t + Individual positions at iteration 1, X t Indicates t The individual position at the iteration, L Represents the search path, represents point multiplication, l Represents the position update parameters, X max represents the current optimal solution, ρ t Indicates t The step size adjustment coefficient for the iteration; Among them, the step length adjustment coefficient ρ The method of determining is: in, ρ t Indicates t The distance control parameter at the iteration, ρ 0 represents the initial distance, e represents the base of natural logarithms, t Indicates the current iteration number, T Indicates the maximum number of iterations; The search path L The method of determining is: in, L Represents the search path, r u and r v All represent random numbers that satisfy the normal distribution; Continue iterating until the maximum number of iterations is reached T , the individual that outputs the function value of the objective function at this time is the optimal model parameter.
2. A transformer status monitoring system, characterized in that: It comprises a processor and a memory for storing processor executable instructions; the processor is configured to call the instructions stored in the memory to execute the transformer state monitoring method according to claim 1.
Citation Information
Patent Citations
Transformer state evaluation and fault detection method based on multi-source data fusion
CN109856488A
Converter transformer fault diagnosis method based on deep convolutional neural network
CN110929847A
Method for identifying partial discharge fault type through convolutional neural network
CN113283579A
Transformer health state monitoring method and system
CN116011243A
Power transformer fault diagnosis method and system of multi-modal information fusion network
CN116310551A
Cited By
Online monitoring system and method for grounding current of transformer iron core
CN120703633A
