CN111401785A

CN111401785A - Power system equipment fault early warning method based on fuzzy association rule

Info

Publication number: CN111401785A
Application number: CN202010274132.9A
Authority: CN
Inventors: 马强; 王勇; 李磊; 管荑; 李慧聪; 田大伟; 耿玉杰; 刘勇; 林琳; 娄建楼; 李燕; 李建坡
Original assignee: State Grid Shandong Electric Power Co Ltd; Northeast Dianli University
Current assignee: State Grid Shandong Electric Power Co Ltd; Northeast Electric Power University
Priority date: 2020-04-09
Filing date: 2020-04-09
Publication date: 2020-07-10

Abstract

The invention discloses a power system equipment fault early warning method based on fuzzy association rules, which relates to a fault early warning method and comprises the following steps: determining the optimal partition number of the power equipment data through K mean value and information entropy mixed iteration so as to realize dynamic self-adaptive boundary partition; introducing a fuzzy set to divide a softening attribute boundary, and dividing a fuzzy interval by using a fuzzy C mean value; and (3) selecting a group of optimal minimum support degree and trust degree as main parameters of a mining algorithm by using an Apriori algorithm, and mining association rules according to the parameters to construct a rule base so as to analyze and predict the fault state of the power equipment. The method can quantitatively obtain the optimal partition number during attribute discretization, and realize dynamic self-adaptive attribute boundary division; compared with the traditional association rule method, the method can quickly and accurately detect the fault state of the equipment.

Description

Power system equipment fault early warning method based on fuzzy association rule

Technical Field

The invention relates to a fault early warning method, in particular to a power system equipment fault early warning method based on a fuzzy association rule.

Background

At present, with the continuous development of smart power grids, the scale of the power grids is continuously enlarged, and the requirements on the operation safety of power systems are higher and higher. In order to ensure the safe operation of the power system equipment, reduce the failure burst rate of the power system equipment and reduce the overhaul cost of the equipment, the method is of great importance to the state detection and the safe maintenance of the power system equipment.

In some domestic and foreign researches on the aspect of power system fault early warning, a density-based DBSCAN clustering algorithm is adopted to calculate the relative proximity of sampled data and historical fault data clusters so as to complete the classification of the data; some data samples are processed by three different normalization methods and are used as the input of a fuzzy C mean algorithm, and the fault type of the samples is determined by solving the membership degree; some transformers are clustered on different short circuit turns, axial displacement and radial deformation, and clustering results are used for explaining frequency response analysis to diagnose equipment faults; some methods adopt an exponential form of a membership function of a fuzzy c-means to obtain a judgment index of the distance, and the obtained membership matrix realizes the division of transformer fault data. Most of the existing fault early warning methods for the power equipment are realized through a simple clustering algorithm, the implicit correlation among data cannot be mined and analyzed, the fault trend cannot be detected as soon as possible, and more maintenance time cannot be strived for operation and maintenance personnel, so that serious loss is caused.

If timely and effective early warning is required to be carried out on the fault state of the power system equipment, the implicit association relation among the data must be deeply mined. Because the online monitoring data of the power system shows exponential growth trend every day, and the association rule algorithm has the advantage of being capable of mining out the rule which cannot be intuitively felt from a large amount of data in a centralized manner, and can often give out unexpected rule combinations, the method is widely applied to the fields of power system fault diagnosis, thermal power plant optimization, network safety and the like.

Disclosure of Invention

The invention mainly aims to provide a power system equipment fault early warning method based on a fuzzy association rule.

The technical scheme adopted by the invention is as follows: a power system equipment fault early warning method based on fuzzy association rules comprises the following steps:

determining the optimal partition number of the power equipment data through K mean value and information entropy mixed iteration so as to realize dynamic self-adaptive boundary partition;

introducing a fuzzy set to divide a softening attribute boundary, and dividing a fuzzy interval by using a fuzzy C mean value;

and (3) selecting a group of optimal minimum support degree and trust degree as main parameters of a mining algorithm by using an Apriori algorithm, and mining association rules according to the parameters to construct a rule base so as to analyze and predict the fault state of the power equipment.

Further, the method comprises the following steps:

the number and the center of the initial classes are solved by the improved K mean value and the information entropy mixed iteration, and the final clustering result is solved by FCM clustering;

set the sample set to be clustered as

The method for early warning the equipment fault of the power system based on the fuzzy association rule comprises the following specific steps:

s1: determining a range of initial cluster numbers

；

S2: in the process that the number of clusters is gradually increased, each pair corresponds to a cluster number j, the cluster center is calculated by using an improved K-means algorithm, and then the transition difference value of the information entropy is calculated on the basis of calculating the data deviation;

s3: at the determined

In sequence, obtained

The number k of clusters when the minimum value is reached and the cluster center at that time

；

S4: the optimal clustering number k and the class center

As an initialization parameter, initializing the FCM algorithm;

s5: updating membership matrix of fuzzy cluster

；

S6: if it is not

If so, the clustering algorithm stops and outputs the membership matrix

And class center

Otherwise, go to S5 to continue iteration;

s8: obtaining the final membership matrix

And class center

And dividing the data in the data set into corresponding classes.

The invention has the advantages that:

(1) the optimal partition number during attribute discretization can be obtained quantitatively, and dynamic self-adaptive attribute boundary division is realized.

(2) The KEFCM algorithm is used, the minimum support degree and the minimum confidence degree are selected, data edge information can be effectively reserved, rules with research values are prevented from being ignored in the process of mining association rules, and the KEFCM algorithm is high in classification accuracy.

(3) Compared with the traditional association rule method, the method can quickly and accurately detect the fault state of the equipment.

In addition to the objects, features and advantages described above, other objects, features and advantages of the present invention are also provided. The present invention will be described in further detail below with reference to the drawings.

Drawings

The accompanying drawings, which are incorporated in and constitute a part of this application, illustrate embodiments of the invention and, together with the description, serve to explain the invention and not to limit the invention.

FIG. 1 is a flow chart of a method of an embodiment of the present invention;

FIG. 2 is a graph of information entropy transition difference changes for an embodiment of the present invention;

FIG. 3 is a graph showing comparison results of three clustering algorithms according to the embodiment of the present invention;

FIG. 4 is a flow diagram of an embodiment of the present invention to obtain frequent 1-term streams;

FIG. 5 is a flow chart of obtaining frequent k-terms sets according to an embodiment of the present invention;

FIG. 6 is a graph showing the relationship between the mean matching rate and the confidence level in different support degrees according to the embodiment of the present invention;

FIG. 7 is a diagram of the relationship between the variance of the matching rate and the confidence level under different support degrees according to an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

The invention provides a power system equipment fault early warning method based on a fuzzy association rule.

Clustering algorithm

And (3) K-means clustering:

the K-means (K-means) is one of the most widely used methods in the clustering analysis, is a simple iterative clustering algorithm and aims to achieve the minimization of the intra-class data distance and the maximization of the inter-class data distance. For a given data set

Wherein n is the number of data and d is the dimension of data. Each data in x is divided into according to respective attributes

A plurality of different classes, each class having a cluster center, a class space for a jth class

Represented by the mean of all data in the class, computingThe formula is as follows:

(1)

given a cluster center of

The Euclidean distance is adopted among the data as a division index, and the calculation formula is as follows:

(2)

basic steps of the K-means algorithm:

s1: selecting k data in a data space as an initial clustering center;

s2: for each data in the data set x, according to the Euclidean distance dij between the data and the clustering centers, classifying the data into a class corresponding to the nearest clustering center according to the nearest criterion;

s3: taking the mean value corresponding to all the data in each category as the clustering center of the category, and updating the clustering center;

s4: judging whether the clustering center is changed or not, and if not, outputting a result; otherwise, return to S2 continues.

The K-means algorithm enables the obtained clusters to meet the condition that the similarity of data objects in the same class is high, and the similarity of data objects in different classes is low. However, the K-means algorithm determines an initial partition based on the initial cluster center and then optimizes the initial partition. The selection of the initial clustering center has a great influence on the clustering result, and once the initial value is not well selected, an effective clustering result cannot be obtained, which also becomes a main problem of the K-means algorithm. Therefore, the method for selecting the clustering centers by the K-means algorithm is improved to ensure that the mutual distance between the initial clustering centers is as far as possible. The cluster center initialization algorithm flow is as follows:

s1: randomly selecting a point from an input data point set x as a first clustering center;

s2: for each data in x:

(1) calculating their distance from the selected cluster center

；

(2) Calculating the probability p (xi) that each data is selected as the next cluster center;

(3)

(3) taking a random number r between [0,1], subtracting the probability of the first data by r, and selecting the random number r as the next clustering center if the subtraction result is less than or equal to 0; if the result is greater than 0, continuously subtracting the probability value of the next data, and repeating the process until the subtracted result is less than or equal to 0; selecting the data corresponding to the subtracted probability value as the next clustering center;

s3: s2 is repeated until all k cluster centers have been selected.

Information entropy:

shannon introduced the Entropy (Entropy) concept into the field of informatics in 1948, and used the Entropy to measure the amount of information contained in data. For a data set x divided into k class sets, Pij is used for representing the deviation degree of the ith sample point from the center of the jth class set, and the smaller the value of Pij is, the smaller the probability that the ith sample belongs to the jth class is, and the farther the ith sample is from the jth class is. Wherein, the calculation formula of the deviation Pij is as follows:

(4)

the overall information entropy calculation formula of the class is as follows:

(5)

the information entropy difference of jumping from the j-1 th state to the j state is called an information entropy jump value, and the calculation formula is as follows:

(6)

the difference value between the (j-1) th to (j) th state entropy jump values and the (j + 1) th state entropy jump values is called information entropy jump difference value, and the calculation formula is as follows:

(7)

with the increase of the number of clusters, the data amount in each class is reduced, the probability that each data belongs to one class is increased, and the information entropy of the class overall is increased. In the process that the number of classes is increased from small to large, class division is carried out according to the sequence from disorder to order to disorder, the initial disorder is that the clustering is too general and the overall characteristics of the data set cannot be known, and the final disorder is that the clustering is too fine and the overall knowledge of the data set is lacked. Thus, the data set information entropy transition difference value can be used

To determine the optimal number of clusters, i.e., from (k-1 classes → k classes) to (k classes → k +1 classes) in the number of clusters

At the lowest, this means that there has been no need to increase the dataset from k classes to k +1 classes, when k is the optimal number of clusters.

Fuzzy C-means clustering:

fuzzy C-means algorithm (FCM) is an unsupervised fuzzy clustering algorithm used to classify high-dimensional spatially distributed data into specific classes. The membership degree of each sample point to all class centers is obtained by optimizing an objective function, so that the sample points are determined to achieve the purpose of automatically classifying sample data. The basic idea of the FCM algorithm is to combine a given set of samples

Partitioning into k fuzzy clusters

The given objective function J is minimized. The objective function J is defined as follows:

(8)

in the formula:

in order to be an index of the blur,

and the ith data belongs to the membership matrix of the jth class.

In each iteration, a membership function is used to calculate a membership value and update the cluster center

And membership matrix

The basic steps of the algorithm are as follows:

s1: initializing cluster number k, fuzzy weighting index m and iteration termination threshold

Number of iterations

And membership matrix

；

S2: calculating a fuzzy clustering center:

(9)

s3: updating membership matrix of fuzzy cluster

：

(10)

S4: if the iteration end condition is satisfied

If so, the target function reaches the minimum value, the iteration is terminated, and a membership matrix is output

And class center, otherwise go to S2 to continue operation until the condition is satisfied.

Implementation of the KEFCM algorithm:

the whole algorithm is divided into two stages, the first stage is used for solving the number and the center of the initial class by the improved K mean value and the information entropy mixed iteration, and the second stage is used for solving the final clustering result by FCM clustering.

Referring to FIG. 1, as shown in FIG. 1, a sample set to be clustered is known as

The overall steps of the KEFCM (Fuzzy c-means algorithm based on K-means and Encopy) algorithm are:

s1: determining a range of initial cluster numbers

；

S2: in the process of gradually increasing the number of clusters, each pair corresponds to a cluster number j, the cluster center is solved by utilizing an improved K-means algorithm, and then the jump difference value of the information entropy is solved on the basis of calculating the data deviation

；

S3: at the determined

In sequence, obtained

；

S4: the optimal clustering number k and the class center

As an initialization parameter, initializing the FCM algorithm;

s5: updating membership matrix of fuzzy cluster

；

S6: if it is not

I.e. membership matrix

Membership matrix relative to last time

Is less than the iteration termination threshold, the clustering algorithm stops and outputs the membership matrix

And class center

Otherwise, go to S5 to continue iteration;

s8: obtaining the final membership matrix

And class center

And dividing the data in the data set into corresponding classes.

The invention can quantitatively obtain the optimal partition number during attribute discretization, and realize dynamic self-adaptive attribute boundary division;

the KEFCM algorithm is used, the minimum support degree and the minimum confidence degree are selected, so that data edge information can be effectively reserved, rules with research values are prevented from being ignored in the process of mining association rules, and the classification accuracy of the KEFCM algorithm is high;

compared with the traditional association rule method, the method can quickly and accurately detect the fault state of the equipment.

Example verification:

the validity of the algorithm provided by the text is verified by adopting a test data set, a UCI Wine data set is selected as test data, the data set comprises 178 samples and 13 characteristics (such as Alcohol, Malic acid and Ash), the data set is divided into 3 types in total, and partial sample data of the data set is shown in Table 1.

TABLE 1Wine partial sample data

The entropy transition difference obtained by combining the information entropy and the K-means iterative computation is shown in fig. 2, and it can be known from the figure that when the transition condition is 2, the transition difference of the system is the minimum, so the optimal classification number of the data set is 3, and the value is consistent with the actual classification number of the test data set.

In fact, the three types of samples of the Wine data set respectively contain 59 samples, 71 samples and 48 samples, and the Wine data set is classified by three algorithms of k-means, FCM and KEFCM respectively, and the result is shown in FIG. 3, and it can be seen that the classification accuracy of the KEFCM algorithm proposed herein is the highest.

Basic theory of Apriori algorithm:

the Apriori algorithm is a boolean-type management rule algorithm for finding a frequent item set, which is calculated using a layer-by-layer iteration method and generates the frequent item set based on a candidate item set, i.e., a (k-1) -item set L k-1 is used to generate a k-item set L k, a frequent 1-item set and a frequent k-item set, as shown in fig. 4 and 5, by scanning a database, counts of each item are accumulated to obtain an item satisfying the minimum support degree, a set of the frequent 1-item set is found and marked as L1, then a set L2 of the frequent 2-item set is found through a set L1 of the frequent 1-item set, and so on until an item set satisfying a condition cannot be obtained, at which time, the obtained item set is called as the maximum frequent item set.

The connecting step is to connect L k-1 with itself to generate a candidate k-item set, and is to mark Ck. the pruning step is to delete any (k-1) item subset of the candidate k-item set if the candidate k-item set does not exist in L k-1.

The validity of the association rule is determined by the support and trust. According to the definition of association rule, for database

Provided that A and B are

A subset transaction of, then

，

And is and

for the empty collection, then

The expression (A) and (B) are the front piece and the back piece of the association rule. Support degree is database

In

The percentage of (c) is shown in formula (12).

（11）

In the formula: a is a front piece of the association rule; b is a back piece of the association rule.

Confidence is a database

Wherein represents the probability of B when A appears, as shown in formula (13).

（12）

Experimental analysis:

preparing data:

the fault of the transformer is different, and the fault characteristic gas is different. In the existing GB/T7252-2001 'analysis and judgment guide rule for dissolved gas in transformer oil' in China, five attributes of characteristic gases influencing transformer fault generation, namely H2, CH4, C2H2, C2H4 and C2H6, are shown in Table 2, and 1000 groups of gas component historical normal data in 2017 and 600 groups of data before and after a fault recording point occurring in 5 months in 2018 are extracted for analysis.

TABLE 2 Transformer Properties

The 1000 groups of data of five continuous attributes are discretized by adopting a KEFCM clustering algorithm, and the discretization interval after the optimal classification number of each attribute is obtained is shown in table 3 (two effective fractions are reserved).

TABLE 3 discretization interval of Transformer Attribute

In view of the need of data mining, the sections belonging to different attributes need to be distinguished so as not to be repeated, and therefore, the data to be mined need to be numbered. For example, the value of H2 in a set of data is 14.88, i.e., the value falls in the fifth interval of the x0 attribute, so the data is labeled 05, and so on. The resulting form of the database to be mined is shown in table 4.

Table 4 database to be mined

4.2 Association rule base establishment

In order to enable the mined association rules to accurately express the relationship among the attributes of the transformer, the selection of the minimum support degree minSup and the minimum trust degree minConf is also the most critical step. The index of matching rate is used as an index for evaluating the accuracy of the association rule mined under a certain group of minSup and minConf by combining the mean value and the variance of the index, and a group of optimal minSup and minConf is found through a plurality of groups of experiments, and the rules mined under the parameters form an association rule base. The calculation formula of the matching rate is as follows:

(13)

in the formula:

the number of rules to which the current data conforms,

the number of rules for the rule antecedent and rule antecedent is only met,

is the degree of matching of the set of data with the rule base.

The larger the presentation rule, the more accurately reflects the intrinsic relationship of the set of data attributes. The average matching degree of the rule base and the training data is represented by the mean matching rate of all the matching data, and the stability degree of the rule base suitable for the training data is represented by the variance of the matching rate.

And establishing rule bases under different minSup and minConf, comparing the change of the mean value and the variance of the matching rate in different rule bases, determining a group of optimal minSup and minConf, and establishing the rule bases by using the optimal minSup and minConf as parameters of a mining algorithm. The results of the experimental analysis are shown in fig. 6 and 7 below.

Generally, the larger the support degree value and the higher the confidence degree, the smaller the mean value and the larger the variance, but the too large support degree and confidence degree can result in the suddenly reduced mean value and the suddenly increased variance. This is because too large support and trust leads to a drastic reduction in the number of rules, which reduces the coverage of the rule base, i.e., a large amount of data cannot find a rule matching therewith. As can be seen from FIGS. 6 and 7, when

，

The set of minimum support and confidence that is the best for the Apriori algorithm. On the basis, 2781 pieces of frequent attribute sets and 5546 pieces of association rules are mined, and the forms of partial frequent attribute sets and association rules are shown in tables 5 and 6.

TABLE 5 partial frequent Attribute set

TABLE 6 partial association rules

Taking the first association rule (11, 23 → 02) in table 8 as an example, the meaning is:

given a CH4 value in the first interval ([ 11.53, 12.61 ]), and a C2H2 value in the third interval ([ 2.29, 3.43 ]), the probability of the value of H2 falling in the second interval ([ 5.51, 8.35 ]) is 95.14%.

And (3) verifying the early warning effect:

in order to further verify the practicability of the fault early warning method, 600 groups of data before and after a fault recording point occurring in 5 months in 2018 are selected for validity verification (wherein the 300 th group of data is an initial fault point). The early warning results under the method and the early warning results under the conventional association rules are shown in table 7.

TABLE 7 Transformer Fault diagnosis results

The association rule is a representation of the relationship between the attributes of the device in a normal state, and in the early stage of a fault, the existing association relationship between the attributes is gradually broken and continuously worsened. The applicability of the original association rule to the current operating data is gradually reduced, so that an alarm occurs. From table 8, the diagnostic result of the fuzzy association rule detects that there is a failure trend in the 276 th group of data, whereas the conventional association rule method identifies the failure trend in the 293 th group of data, which indicates that the method can diagnose the failure state of the transformer more accurately. In summary, the experimental result verifies the effectiveness and the high efficiency of the fuzzy association rule in the fault early warning process.

The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.