CN-122020131-A - Transformer service life prediction method and system

CN122020131ACN 122020131 ACN122020131 ACN 122020131ACN-122020131-A

Abstract

The invention discloses a method and a system for predicting the service life of a transformer, which relate to the technical field of transformer maintenance, collect multi-source data of the transformer, extract macroscopic features and microscopic features of the multi-source data, evaluate the association degree of the microscopic features to the macroscopic features through mutual information and random forest double indexes, construct degradation indexes according to association degree division scenes, verify and guarantee reliability through monotonicity, robustness and correlation, finally predict the residual service life by using a gating circulation unit or a support vector regression model field Jing Bianya machine, realize the accuracy and early stage of residual service life prediction through multi-source data fusion and intelligent algorithm innovation, can identify aging symptoms in advance, adapt different transformers, and provide basis for preventive maintenance, and reduce fault risks and operation and maintenance costs.

Inventors

GUO SHOUPENG
ZHANG JINYU
WANG HAINING
Wen Haijie

Assignees

江苏威科变压器有限公司

Dates

Publication Date: 20260512
Application Date: 20260410

Claims (7)

1. The method for predicting the service life of the transformer is characterized by comprising the following steps of: Step S1, multi-source data of a transformer are collected, wherein the multi-source data comprise state monitoring data, static data and microscopic data; The step S1 comprises the following sub-steps: Step S101, multi-source data in the whole life cycle of a transformer are collected; the state monitoring data comprise winding leakage current, insulation resistance absorption ratio, micro water content in oil, insulation oil dielectric loss, furfural content in oil, total hydrocarbon relative gas production rate, total hydrocarbon volume fraction, hydrogen volume fraction and acetylene volume fraction; extracting the static data from the equipment file, wherein the static data comprises rated capacity, voltage level, ambient temperature and ambient humidity; the microscopic data comprise high-frequency electromagnetic noise and dynamic behavior of microbubbles in oil; the dynamic behavior of the microbubbles in the oil comprises microbubble images, bubble concentration, size distribution, rising rate and coalescence frequency; s2, extracting degradation features from the state monitoring data, extracting static features from the static data, extracting micro features from the micro data, taking the degradation features and the static features as macro features, performing relevance evaluation on the micro features and the macro features, and obtaining relevance between the micro features and the macro features, wherein the relevance comprises a first relevance and a second relevance; the step S2 comprises the following sub-steps: Step S201, extracting degradation characteristics of the state monitoring data by improving a convolutional neural network; Step S202, constructing a similarity matrix for the preprocessed static data, standardizing the Laplace matrix, clustering K-means into K' clusters, and extracting a cluster center as a static feature; -taking the degradation features and static features as macroscopic features; Step S203, extracting the micro features of the micro data through the graph attention network, specifically including: Constructing a graph structure, wherein nodes of the graph structure are transformer components, edges of the graph structure are physical interaction relations among the components, and microscopic data are used as node attributes of the nodes in the graph structure; calculating node attention weights by adopting a graph attention network, and outputting micro-characteristics; Step S204, carrying out relevance evaluation on the micro-feature and the macro-feature, wherein the specific logic is as follows: calculating the mutual information value of each micro feature and each macro feature by adopting a K neighbor estimation method, and counting the micro feature quantity duty ratio of which the mutual information value is larger than a second threshold value; Taking a transformer aging stage as a label, taking micro features and macro features as inputs, training a random forest model, and counting the average importance ratio of the micro features, wherein the transformer aging stage comprises a health stage, an early aging stage and a critical aging stage; step S205, obtaining the association degree of the micro-feature and the macro-feature, wherein the association degree comprises a first association degree and a second association degree; If the micro-feature quantity ratio of the mutual information value larger than the second threshold value is smaller than or equal to a first percentage, and the average importance ratio of the micro-features is smaller than or equal to a second percentage, the association degree of the micro-features and the macro-features is the first association degree; If the micro-feature quantity ratio of the mutual information value larger than the second threshold value is larger than the first percentage, or the average importance ratio of the micro-features is larger than the second percentage, the association degree of the micro-features and the macro-features is the second association degree; Step S3, when the association degree is a first association degree, a first degradation index is constructed according to the degradation characteristic and the static characteristic, and when the association degree is a second association degree, a second degradation index is constructed according to the degradation characteristic, the static characteristic and the micro characteristic; S4, when the first degradation index is used, predicting the first residual service life through a pre-trained gating cycle unit time sequence model, and when the second degradation index is used, inputting the second degradation index and micro-characteristics into a pre-trained support vector regression model, and outputting the second residual service life; The step S4 includes the following sub-steps: Step S401, when the first degradation index is used, predicting a first degradation index sequence in a future unit time through a pre-trained gating cycle unit time sequence model, wherein a moment when a first degradation index which is met for the first time is found to be greater than or equal to a preset failure threshold value in the first degradation index sequence is defined as a first failure time, and the time difference between the first failure time and the current time is the first residual service life; and step S402, inputting the second degradation index and the micro-feature into a pre-trained support vector regression model when the second degradation index is used, outputting a second residual service life, and inputting the updated micro-feature again if the micro-feature fluctuates, and outputting a new second residual service life by the support vector regression model in real time.
2. The method for predicting the life of a transformer according to claim 1, wherein said step S1 further comprises the sub-steps of: Preprocessing the state monitoring data and static data comprises the following steps: Adopting a3 sigma criterion to respectively eliminate abnormal values of the state monitoring data and the static data, respectively complementing the missing values of the state monitoring data and the static data through linear interpolation, respectively mapping the state monitoring data and the static data to a [0,1] interval based on normalization, and deleting redundant variables with the pearson correlation coefficient larger than a first threshold; Preprocessing the microscopic data includes: denoising the high-frequency electromagnetic noise by adopting a wavelet threshold value, and converting a time domain signal into a time-frequency matrix through short-time Fourier transform; And removing image noise from the microbubble image by adopting Gaussian filtering, extracting a bubble region by threshold segmentation, and removing tiny impurities by using a morphological operation method.
3. The method for predicting the life of a transformer according to claim 2, wherein said step S201 comprises the sub-steps of: Step S2011, constructing an improved convolutional neural network, wherein the improved convolutional neural network comprises a two-channel processing structure, and the two-channel processing structure comprises a first channel and a second channel; the first channel is a one-dimensional depth separable convolutional neural network, and the second channel is a two-dimensional transposed convolutional neural network; Step S2012, performing a channel convolution operation and a point convolution operation through the first channel, and processing an original time sequence signal in the multi-source data, where the original time sequence signal includes a sequence of a winding leakage current changing with time and a time sequence of a micro water content in oil, outputting a time sequence feature vector, performing a transpose convolution on a reconstruction matrix signal through the second channel, and outputting a spatial feature vector; Step S2013, merging the time sequence feature vector and the space feature vector to obtain a space-time feature vector, inputting the space-time feature vector into a depth residual error shrinkage network, and executing identity mapping through a residual error block of the depth residual error shrinkage network; Filtering noise in the space-time feature vector after identity mapping through a soft threshold module, carrying out weight distribution on the space-time feature vector after denoising through an attention mechanism, and outputting a preliminary degradation feature; step S2014, optimizing and improving super parameters of the convolutional neural network by adopting a whale optimizing algorithm, and outputting degradation characteristics, wherein the method specifically comprises the following steps of: Mapping the super-parameter space into whale individual positions with the aim of minimizing the prediction error, and iteratively optimizing the time step, the number of hidden layer neurons and the learning rate by surrounding the adjustment parameters and calculating the spiral update positions, wherein the iteration termination condition is that the preset maximum iteration times or error convergence is achieved.
4. A method of predicting the life of a transformer as claimed in claim 3, wherein said step S3 comprises the sub-steps of: Step S301, constructing a degradation index according to the relevance degree division scene, and constructing a first degradation index according to the degradation characteristic and the static characteristic when the relevance degree is the first relevance degree; the calculation formula of the first degradation index is as follows: ; Wherein, the As an index of the first degradation, a first degradation indicator, As a characteristic of the degradation, For the preset weight of the degraded features, As a feature of the static state, Is a preset static feature weight.
5. The method for predicting the life of a transformer according to claim 4, wherein said step S3 further comprises the sub-steps of: step S302, when the association degree is a second association degree, a second degradation index is constructed according to the degradation characteristic, the static characteristic and the micro characteristic; the calculation formula of the second degradation index is as follows: ; Wherein, the As an index of the second degradation, a second degradation indicator, In order to correct the weight of the degraded feature, As a characteristic of the degradation, As a feature of the static state, For a preset static feature weight, Is the weight of the micro-feature which is preset, Is a microscopic feature; Step S303, performance verification is carried out on the monotonicity, the robustness and the correlation of the degradation indexes.
6. The method of claim 5, wherein the specific logic for performance verification comprises: And when the monotonicity of the degradation indicator is smaller than the third threshold, the robustness is larger than the fourth threshold and the correlation is smaller than the fifth threshold, the degradation indicator is unqualified, and the step S204 is returned to readjust the first percentage and the second percentage of the correlation judgment.
7. A transformer life prediction system applied to the method for predicting the life of the transformer according to any one of claims 1 to 6, wherein the system comprises an acquisition module, an extraction module, a construction module and a prediction module; the acquisition module is used for acquiring multi-source data of the transformer; the extraction module is used for extracting macro-features and micro-features of the multi-source data, carrying out relevance evaluation on the micro-features and the macro-features, and obtaining the relevance of the micro-features and the macro-features; the construction module is used for constructing a degradation index according to the relevance degree scene; and the prediction module is used for predicting the residual service life of the transformer based on the degradation index.

Description

Transformer service life prediction method and system Technical Field The invention relates to the technical field of transformer maintenance, in particular to a method and a system for predicting service life of a transformer. Background The power transformer is used as an important energy conversion device in a power system and is mainly used for adjusting the voltage and converting the transmission form of electric energy. The interior of the power transformer contains complex electrical and mechanical components including insulation, coils and cooling systems. These components age gradually over time due to voltage, thermal and mechanical stresses, resulting in reduced performance and increased risk of failure of the power transformer. In order to ensure stable operation of the power transformer, timely maintenance and updating are important. However, if the point in time for maintenance and replacement is improperly selected, unplanned outages and reduced efficiency of equipment use may result. In more serious cases, the safe operation of the whole power grid is threatened, so that the stability of power supply and the economic benefit of the power company are further influenced, and the research on the maintenance technology of the power transformer is always the focus of the power company. At present, china invention with the application number of CN118246246B discloses a life prediction method and a life prediction system of a transformer based on multi-source data fusion, wherein a running state evaluation system is constructed by acquiring test data and running condition data of a power transformer, comprehensive health indexes are calculated, life prediction under maintenance-free and maintenance states is realized based on a Weibull distribution model and a failure rate correction coefficient, the limitation of the prior art is mainly that the data dimension is insufficient, the utilization of microscopic data such as high-frequency electromagnetic noise, micro-bubble dynamic behaviors in oil and the like is mainly depended on traditional macroscopic parameters, the capturing capacity of early degradation characteristics is limited and real early warning is difficult to realize, the characteristic extraction capacity is limited, the weight calculation combines a subjective and objective method, the characteristic extraction is still depended on a static index system, a deep degradation information cannot be mined from complex data in a self-adapting mode, the anti-noise interference capacity is weak, the failure rate correction is based on a Boolean distribution and empirical formula, the correction process depends on historical data, the maintenance data is insufficient for the transformer under variable working conditions or the new running modes, the real-time degradation environment is not balanced with the real-time degradation requirements is difficult to be satisfied, and the real-time degradation requirements are not satisfied, and the real-time degradation conditions are not well-time is difficult to be satisfied. Disclosure of Invention The invention solves the technical problems that the prior art is difficult to solve the pain points of insensitivity of early aging detection, incomplete feature extraction, insufficient prediction precision and poor adaptability in the traditional prediction method. In order to solve the technical problems, the invention provides the following technical scheme that the method for predicting the service life of the transformer comprises the following steps: Step S1, multi-source data of a transformer are collected, wherein the multi-source data comprise state monitoring data, static data and microscopic data; s2, extracting degradation features from the state monitoring data, extracting static features from the static data, extracting micro features from the micro data, taking the degradation features and the static features as macro features, performing relevance evaluation on the micro features and the macro features, and obtaining relevance between the micro features and the macro features, wherein the relevance comprises a first relevance and a second relevance; Step S3, when the association degree is a first association degree, a first degradation index is constructed according to the degradation characteristic and the static characteristic, and when the association degree is a second association degree, a second degradation index is constructed according to the degradation characteristic, the static characteristic and the micro characteristic; And S4, when the first degradation index is used, predicting the first residual service life through a pre-trained gating cycle unit time sequence model, and when the second degradation index is used, inputting the second degradation index and micro-characteristics into a pre-trained support vector regression model, and outputting the second residual service life. Preferably, the step S1 comprises the follo