CN112630564B

CN112630564B - Transformer DGA fault diagnosis method based on neighborhood rough set and AMPOS-ELM

Info

Publication number: CN112630564B
Application number: CN202011439354.8A
Authority: CN
Inventors: 周秀; 田天; 怡恺; 李刚; 杨鑫; 吴旭涛; 李秀广; 何宁辉; 罗艳; 刘威峰; 马云龙; 司文荣; 吴兴旺; 董盼; 丁学辉; 徐玉华; 牛勃; 傅晨钊; 杨海涛; 田禄
Original assignee: Electric Power Research Institute of State Grid Anhui Electric Power Co Ltd; Changsha University of Science and Technology; East China Power Test and Research Institute Co Ltd; Electric Power Research Institute of State Grid Ningxia Electric Power Co Ltd
Current assignee: Electric Power Research Institute of State Grid Anhui Electric Power Co Ltd; Changsha University of Science and Technology; East China Power Test and Research Institute Co Ltd; Electric Power Research Institute of State Grid Ningxia Electric Power Co Ltd
Priority date: 2020-12-07
Filing date: 2020-12-07
Publication date: 2023-02-28
Anticipated expiration: 2040-12-07
Also published as: CN112630564A

Abstract

The transformer DGA fault diagnosis method based on the neighborhood rough set and the AMPOS-ELM comprises the steps of establishing a transformer initial fault characteristic quantity set by combining each sample initial characteristic quantity of a transformer with each DGA fault diagnosis standard, wherein each sample initial characteristic quantity is the content of one most main volatile gas generated due to transformer faults and the volume ratio of two main gases generated due to the same transformer faults; obtaining key characteristic quantity with high attribute importance after analyzing by adopting a neighborhood rough set; and constructing an AMPOS-ELM model, and taking the key characteristic quantity screened out from the neighborhood rough set as the input of the AMPOS-ELM network model to carry out fault diagnosis on the transformer. Aiming at the problems that the accuracy of the existing DGA-based transformer fault diagnosis method is easily influenced by input characteristics and the selection of parameters of an extreme learning machine is difficult, the accuracy of the diagnosis result of the existing intelligent algorithm is improved.

Description

Transformer DGA fault diagnosis method based on neighborhood rough set and AMPOS-ELM

Technical Field

The invention relates to the technical field of monitoring of fault states of transformers of power systems, in particular to a transformer DGA fault diagnosis method based on a neighborhood rough set and AMPOS-ELM.

Background

The power transformer is one of important and expensive electrical equipment in a power grid system, plays an indispensable role in the power system, and plays a core role in the links of conversion and transmission of electric energy. Therefore, it is necessary to ensure safe operation of the power transformer.

After long-term transformer operation and maintenance practices and extensive fault investigation and analysis, it is found that various gases dissolved in transformer oil reflect early signs if a potential fault exists in the transformer or in the initial stage of fault formation. The Analysis of Dissolved Gas in oil (DGA) is just to detect the components and contents of these fault-characteristic gases, so as to analyze and judge the operation condition and hidden fault trouble of the transformer.

The transformer insulating oil is one of mineral oils, and mainly contains unsaturated hydrocarbons containing carbon-carbon double bonds or triple bonds and other hydrocarbons. In the internal discharge fault or heating fault of the transformer, some carbon-carbon bonds or carbon-carbon bonds in some oil molecules are broken, so that a trace amount of active hydrogen atoms and hydrocarbon free radicals are generated, and the free hydrogen atoms and the free radicals are combined again through chemical reaction to finally form hydrocarbon gas compounds such as H2, CH4, C2H6, C2H4, C2H2 and the like.

The power transformer has a complex structure and numerous devices, and once a fault occurs, the fault type cannot be found out in time, so that great inconvenience is brought to subsequent maintenance. At present, boundary conditions set by the IEC three-ratio method are harsh, so that misjudgment, missed judgment and the like are easily caused, the transformer can run for a long time with faults, and potential risks are buried in normal running of a power system.

At present, many achievements of intelligent algorithms in the field of transformer fault diagnosis have been obtained, which mainly include methods such as a manual neural network method (BPNN), a Support Vector Machine (SVM) (Extreme Learning Machine, ELM), and an Extreme Learning Machine, but the methods still have shortcomings. Firstly, in the aspect of characteristic quantity selection, the selection of input characteristics has great influence on classification results, too high degree of dimension of the selected input characteristics leads to too many unnecessary variables, so that not only is the prediction accuracy reduced, but also the model participating in training is too complex, and when the selected input characteristic quantity is less, enough information representation output characteristics are difficult to obtain, so that the selection of proper input characteristics is of great importance to the reliability of subsequent diagnosis. Secondly, in the aspect of algorithm, the BPNN has strong fault-tolerant and nonlinear mapping capability, but has the problems of low convergence speed, easy overfitting and the like. The SVM can better process local minimum values and has strong generalization capability, but the selection of the kernel parameters and the penalty factors limits the classification performance of the SVM, and if the selected parameters are not appropriate, the problems of larger error of a diagnosis result and the like occur. The ELM has the characteristics of high learning speed, excellent generalization capability, high classification accuracy and the like, and is widely applied to the field of transformer fault diagnosis, but because the ELM classification result is influenced by the initial randomly generated parameters, if the parameters are not optimized by a proper optimization algorithm, the problems of larger loss function and poorer robustness are easily caused. Although the particle swarm algorithm has a good diagnosis effect, the particle swarm algorithm has the defect of easy falling into local optimum and needs further optimization.

Disclosure of Invention

The invention provides a transformer DGA fault diagnosis method based on a neighborhood rough set and AMPOS-ELM, aiming at the problem that the existing intelligent algorithm of intelligent transformer faults based on DGA data is low in transformer fault diagnosis accuracy, so that the transformer fault diagnosis accuracy is improved.

A transformer DGA fault diagnosis method based on a neighborhood rough set and AMPOS-ELM comprises the following steps:

step S001, establishing a transformer initial fault characteristic quantity set by combining each sample initial characteristic quantity of a transformer with each DGA fault diagnosis standard, wherein each sample initial characteristic quantity is the content of one most main volatile gas generated due to transformer faults and the volume ratio of two main gases generated due to the same transformer faults;

s002, obtaining key characteristic quantity with high attribute importance after analyzing by adopting a neighborhood rough set;

and S003, constructing an AMPOS-ELM model, and performing transformer fault diagnosis by using the key characteristic quantity screened out from the neighborhood rough set as the input of the AMPOS-ELM network model.

Has the advantages that: the transformer DGA fault diagnosis method based on the neighborhood rough set and the AMPOS-ELM, which is disclosed by the invention, aims at solving the problems that the accuracy of the existing transformer fault diagnosis method based on DGA data is easily influenced by input characteristics and the selection of parameters of an extreme learning machine is difficult, and provides a transformer fault diagnosis method based on the neighborhood rough set and a self-adaptive variation particle swarm extreme learning machine algorithm, so that the accuracy of the diagnosis result of the existing intelligent algorithm is improved.

Detailed Description

The following description of the embodiments of the present application is provided by way of specific examples, and other advantages and effects of the present application will be readily apparent to those skilled in the art from the disclosure herein. The present application is capable of other and different embodiments and its several details are capable of modifications and/or changes in various respects, all without departing from the spirit of the present application. It should be noted that the features in the following embodiments and examples may be combined with each other without conflict.

and S003, constructing an AMPOS-ELM model, and performing transformer fault diagnosis by taking the key characteristic quantity screened out from the neighborhood rough set as the input of the AMPOS-ELM network model.

Further, in step S001, "establishing a transformer initial fault feature set by combining each sample initial feature of the transformer with each DGA fault diagnosis standard" includes the following steps:

step S101, establishing 7 fault types for each sample initial characteristic quantity of the transformer according to DGA fault diagnosis standards, wherein the fault types are respectively a normal state group, a partial discharge group, a low-energy discharge state group, a high-temperature overheat state group, a discharge and overheat group and a medium-low temperature overheat state group, namely all sample initial characteristic quantities can be classified into the 7 fault types;

step S102, according to a rule of constructing fault characteristics by a ratio method, taking the content of one most main volatile gas generated due to transformer faults and the volume ratio of two main gases generated due to the same transformer faults as an initial fault characteristic quantity set.

Further, in step S002, "obtaining the key feature quantity with high attribute importance after analyzing the neighborhood rough set" includes the following steps:

step S201, establishing a decision information table by taking the transformer fault type as a decision attribute and the initial fault characteristic quantity as a condition attribute, and determining an attribute importance lower limit and a neighborhood radius set;

step S202, obtaining an optimized fault characteristic quantity set through a neighborhood rough set algorithm and obtaining the attribute importance of each optimized fault characteristic;

step S203, the attribute importance of each optimized fault feature quantity and the attribute importance of the feature quantity selected by the traditional transformer fault diagnosis are sequenced from large to small, the fault feature quantity corresponding to the optimized fault feature quantity with the attribute importance larger than the attribute importance of the feature quantity selected by the traditional transformer fault diagnosis is used as the DGA feature quantity, and the DGA feature quantity is the key feature quantity with the high attribute importance.

The specific steps of the neighborhood rough set algorithm are as follows:

step S301, inputting a decision system NDT = { U, a = V, N, f }, where U is a set of fault types and initial fault feature amounts, a is a set of initial fault feature amounts, V is a set of fault types, N is the number of fault types, and f is an information function that specifies an attribute value of each object in the set U, and sets an attribute importance lower limit value to 0.2 for obtaining a more optimized fault feature amount;

step S302, a subset B of the set A is established, namely A and B are two equivalent relation families of U,

and subset B is independent of a, subset B is denoted red;

step S303, order

For any a _n E (A-red) using formula

Calculating Positive Domain pos _B (V) and selecting a _n Make positive domain pos _B (V) is maximum, wherein

δ _B (X _i )＝{X _i |X _j ∈U,Δ(X _i ,X _j )≤δ}，

X _g ∈U，

δ _B (X _i ) Is X _g Is the neighborhood of (1), i.e. set delta _B (X _i ) All fault signatures in (1) are associated with X _g Similarly, Y _g Is an equivalent subset of U;

step S304, defining the dependency of the set V on B:

the initial fault characteristic quantity a belongs to B, and the attribute importance formula of the initial fault characteristic quantity a to the decision attribute V is as follows: sig (a, B, V) = γ _B (V)-γ _B-{a} And (V) if the calculated result is greater than the attribute importance lower limit value of 0.2, outputting red, otherwise, outputting the initial fault characteristic quantity a until all the initial fault characteristic quantities are classified, thereby obtaining an optimized fault characteristic quantity set and obtaining the attribute importance of each optimized fault characteristic.

The method for diagnosing the transformer faults by constructing the AMPOS-ELM model and taking the key characteristic quantity screened out from the neighborhood rough set as the input of the AMPOS-ELM network model comprises the following steps:

step S401: determining an ELM topological structure and a training sample, initializing a particle swarm, and selecting a proper c _max And c _min 、k _max 、ω、M；

Step S402: setting the current fitness F of the POS algorithm _i And = σ. Comparing the fitness value of each particle with the optimal position pbest that the particle has experienced if F _i If pbest is less, replacing F with pbest _i Otherwise, maintaining the current situation;

step S403: comparison F _i If the size of the gbest is better than the former, the current gbest is used as the latest global optimum position;

step S404: updating the particle weight α according to equation (1) _i And simultaneously judging whether the particle swarm is premature according to the formula (2), if so, giving escape operation to the particles according to the formula (5), and otherwise, updating the particle speed and the position according to the formulas (3) and (4). The steps are circulated, and the operation times reach k _max If the number of maximum iterations of =160 or gbest reaches stability, the program is exited, and the current optimal individual and fitness thereof are returned;

in the formula: alpha is alpha _i Is the weight of the particle iCoefficient, i =1,2,3 _i As the best position for the particle, gbest is the global best position.

In the formula: delta is the deviation of the mean adaptation value of the particle swarm, f _i Is the fitness of the particle i, f _avg Is the average fitness of the current population.

In the formula, k is iteration times; omega is the inertial weight; c. C ₁ 、c ₂ Is a particle learning factor; r is ₁ 、r ₂ ∈rand[0,1]；

Respectively the speed, the position, the individual optimal position and the global optimal position of the jth dimension variable of the hyperparameter i in the kth iteration.

In the formula: r is an element of [0,1 ]]A uniform random number above; z is a radical of ^k Is an escape control factor.

Step S405: outputting an input weight corresponding to the optimal fitness and a hidden layer threshold, and calculating an optimal output weight matrix;

step S406: according to the output weight matrix, establishing a transformer fault diagnosis model based on ELM;

step S407: and inputting the test sample set into the model established in the step S406 for transformer fault diagnosis.

The invention adopts 417 groups of confirmed transformer fault type samples provided by a certain electric academy, and can be divided into 7 states according to related regulation diagnosis results, which are as follows: (1) normal state (N); (2) Partial Discharge (PD) (3) low energy discharge state (D1); (4) a high energy discharge state (D2); (5) a high temperature superheated state (T3); (6) discharge and superheat (TD); (7) a medium-low superheat condition (T12). The distribution of samples is shown in table 1, and there are 417 groups of samples, 317 groups are randomly selected from them as training set, and the remaining 100 groups are test set.

The sample distribution is shown in table 1.

TABLE 1 Transformer Fault samples

The gas generated by the transformer fault has the following existing forms:

2. screening key attribute features

317 training data are used for establishing a decision table, attribute reduction is carried out on the decision table based on a neighborhood rough set algorithm, and a final minimum fault feature set and related importance are obtained and shown in a table 3. And comparing the characteristic quantity selected by the traditional transformer fault diagnosis, and selecting the fault characteristics of CH4/H2, C2H4/C2H6 and C2H2/C2H4 with the attribute importance degrees ranked at the top in the table 3 as the DGA characteristic quantity optimization result.

TABLE 3 Key Attribute feature set and Attribute importance

3. ELM parameter optimization and fault diagnosis

And (4) fault diagnosis and identification are carried out on the test set, the key characteristic indexes are used as input characteristic quantities of the AMPOS-ELM, and the AMPOS-ELM diagnosis results are shown in the following table 4.

TABLE 4 AMPOS-ELM diagnostic results

As can be seen from the table, the highest test accuracy of the AMPOS-ELM model is 89%, so that the AMPOS-ELM diagnostic model provided by the method has high reliability.

While the invention has been described with reference to a preferred embodiment, it will be understood by those skilled in the art that various changes in form and detail may be made therein without departing from the spirit and scope of the invention.

Claims

1. A transformer DGA fault diagnosis method based on a neighborhood rough set and AMPOS-ELM is characterized in that: the method comprises the following steps:

step S001, combining initial characteristic quantities of each sample of the transformer with each DGA fault diagnosis standard to establish a transformer initial fault characteristic quantity set, wherein the initial fault characteristic quantity set consists of CH ₄ 、C ₂ H ₂ 、C ₂ H ₄ 、C ₂ H ₆ 、H ₂ 、CH、C ₂ H ₂ /H ₂ 、C ₂ H ₂ /C ₂ H ₄ 、C ₂ H6/C ₂ H ₂ 、C ₂ H ₂ /CH ₄ 、C ₂ H ₄ /C ₂ H ₆ 、C ₂ H ₄ /CH ₄ 、C ₂ H ₄ /H ₂ 、C ₂ H ₆ /CH ₄ 、C ₂ H ₆ /H ₂ 、CH ₄ /H ₂ 、C ₂ H ₂ /CH、H _2/ CH+H ₂ 、C ₂ H ₄ /CH、CH ₄ /CH、C ₂ H ₆ /CH、CH ₄ +C ₂ H ₄ a/CH composition;

s002, analyzing by using a neighborhood rough set to obtain a key characteristic quantity, wherein the key characteristic quantity is represented by C ₂ H ₄ /C ₂ H ₆ 、C ₂ H ₂ /C ₂ H ₄ 、(CH ₄ +C ₂ H ₄ )/(CH)、CH ₄ /H ₂ 、H ₂ /(H ₂ +CH)、C ₂ H ₆ /C ₂ H ₂ 、C ₂ H ₄ 、C ₂ H ₆ /CH ₄ 、H ₂ 、C ₂ H ₂ Forming;

s003, constructing an AMPOS-ELM model, and performing transformer fault diagnosis by using the key characteristic quantity screened out by the neighborhood rough set as the input of the AMPOS-ELM network model;

the transformer fault diagnosis method comprises the following steps:

step S401: determining an ELM topological structure and a training sample, initializing a particle swarm, and selecting a proper c ₁ And c ₂ 、k _max 、ω、M；

Step S402: setting the current fitness F of the POS algorithm _i = σ, compare fitness value of each particle to the optimal position pbest that the particle has experienced if F _i If < pbest, replacing F with pbest _i Otherwise, maintaining the current situation;

step S404: updating the particle weight α according to equation (1) _i And simultaneously judging whether the particle swarm is premature according to the formula (2), if so, giving escape operation to the particles according to the formula (5), otherwise, updating the particle speed and position according to the formulas (3) and (4), and circulating the steps until the operation frequency reaches k _max If the number of maximum iterations of =160 or gbest reaches stability, the program is exited, and the current optimal individual and fitness thereof are returned;

in the formula: alpha (alpha) ("alpha") _i M, M is the population size, pbest, which is the weight coefficient of particle i, i =1,2,3 _i The best position of the particle is, and the gbest is the global best position;

in the formula: delta is the deviation of the mean adaptation value of the particle swarm, f _i As the fitness of the particle i, f _avg The average fitness of the current population;

Respectively the speed, the position, the individual optimal position and the global optimal position of a jth dimension variable of the hyper-parameter i during the kth iteration;

in the formula: r is an element of [0,1 ]]A uniform random number above; z is a radical of ^k Is an escape control factor;

step S405: outputting an input weight corresponding to the optimal fitness and a hidden layer threshold value, and calculating an optimal output weight matrix;

2. The method of claim 1 for diagnosing DGA faults of a transformer based on a coarse neighborhood set and an AMPOS-ELM, wherein: in step S001, 7 fault types are established for each sample initial characteristic quantity of the transformer according to the DGA fault diagnosis standard, and are respectively a normal state group, a partial discharge group, a low-energy discharge state group, a high-temperature overheat state group, a discharge and overheat group, and a medium-low temperature overheat state group, that is, all sample initial characteristic quantities can be classified into the 7 fault types.

3. The method of claim 1 for diagnosing DGA faults of a transformer based on a coarse neighborhood set and an AMPOS-ELM, wherein: the specific steps of the neighborhood rough set algorithm are as follows:

step S301, inputting a decision system NDT = { U, AUC V, N, f }, wherein U is a set of fault types and initial fault characteristic quantities, A is a set of initial fault characteristic quantities, V is a set of fault types, N is the number of fault types, f is an information function, the information function specifies an attribute value of each object in the set U, and in order to obtain more optimized fault characteristic quantities, a lower limit value of attribute importance is set to be 0.2;

and subset B is independent of A, subset B is marked red;

step S303, order

For any a _n E (A-red) utilizing formula

Calculating the pos of the positive region _B (V) and selecting a _n Make positive domain pos _B (V) is maximum, wherein

δ _B (X _i )＝{X _i |X _j ∈U,Δ(X _i ,X _j )≤δ}，

X _g ∈U，

δ _B (X _i ) Is X _g Of (c), i.e. set delta _B (X _i ) All fault signatures in (1) are associated with X _g Similarly, Y _g Is an equivalent subset of U;

step S304, defining the dependency degree of the set V on B: