CN105654229A - Power grid automation system and equipment running state risk assessment algorithm - Google Patents

Abstract

The invention relates to a power grid automation system and equipment running state risk assessment algorithm. The algorithm comprises the following steps of 1.1 feature extraction based on a Relief algorithm: equipment is abnormal because of interaction of one factor or a plurality of factors of the equipment during an equipment operation process; and the feature extraction means that from a lot of information collected from the equipment, the factor influencing a type of abnormity is extracted and irrelevant factors are rejected so that latter data analysis is performed; 1.2 similar sample gathering based on kmeans algorithm: through the Relief algorithm in the step1.1, a feature index sequence influencing an equipment operation risk is acquired; and through the k-means algorithm, a threshold scope of each feature index under each risk grade is calculated. By using the algorithm, there are the following advantages that aiming at each equipment, a specific state analysis threshold knowledge database is set so that a threshold error caused by an extraneous environment and an equipment self condition is avoided; and accuracy of alarm and early warning is ensured.

Description

A kind of grid automation system and equipment running status risk assessment algorithm

Technical field

The invention belongs to dispatching automation of electric power systems technical field, utilize data mining algorithm, relation between analysis automated device service data and defect, generate equipment running status secure threshold experts database, and then equipment is carried out monitoring analysis with interim running status in real time, potentially dangerous in being run by equipment carries out early warning in advance, reduces the system fault and occurs.

Background technology

Along with the expansion day by day of electrical network scale, the safe and stable operation of electrical network more and more be unable to do without the support of automation system, expanding gradually also for the daily fortune dimension of automatization professional brings challenges of automation system scale and covering scope, how effectively and rationally management and control works the key that the system having under its command is automatization maintenance work.

Automatic equipment is the basis of system, and the healthy operation of equipment is the prerequisite of system stable operation, and automatic equipment One's name is legion, relation is complicated, and management difficulty is big. Although present stage the existing supervisory system for automatic equipment, but mostly being alarm afterwards, and alarm decision rule is single does not possess adaptability, so there is the situation that alarm is inaccurate, still needing a large amount of manpower manually to judge.

It is thus desirable to a kind of more intelligent monitor mode, for each equipment, the metrics-thresholds scope that it is independent is set, thus avoid the threshold value difference that equipment causes because installing the externalities such as environment, the operation time limit, and in conjunction with data mining algorithm, the history run situation of Analytical equipment, generates the metrics-thresholds being applicable to this equipment automatically, and the continuous service real-time update along with equipment, reduce the artificial cost that manpower safeguards large number quipments threshold value while the accuracy of support equipment metrics-thresholds.

Summary of the invention

It is an object of the invention to application data mining algorithm, by the analysis to device history operating index data, the running state analysis threshold model of structure equipment, and the real time execution situation of equipment compared judgement with this model, current running status and following state to equipment carry out interpretation and application, push alarm and early warning, it is achieved the intelligent monitoring of automatic equipment.

(1) determination of equipment state influence index item.

Dissimilar equipment has its operating index separately, and the index that the equipment that affects runs is different, it is necessary to a large amount of historical data is analyzed, filters out the bigger index of equipment disturbance degree as index of correlation.

(2) equipment runs automatic generation and the renewal of index of correlation threshold value.

Automatic equipment substantial amounts, people building site arranges its index of correlation threshold value one by one needs a large amount of manpowers, and can not accomplish to upgrade in real time, it is thus desirable to realize, by algorithm, the automatic generation that equipment runs index of correlation threshold value, and the running condition real-time update according to equipment, ensure reasonableness and the real-time of threshold value.

The present invention is directed to the deficiency of artificial setting threshold value, this scheme adopts bonding apparatus service data and device exception information, utilize data mining algorithm, extraction equipment operation characteristic value, and the method that equipment state carries out cluster analysis dynamically determines early warning alarm threshold. Along with the continuous service of equipment, constantly update equipment early warning alarm threshold, enable the working order threshold value of threshold value accurate response equipment. Concrete technical scheme is as follows:

Compared with prior art, tool of the present invention has the following advantages in the present invention:

1, analysis process is undertaken by system, when without setting up perfect rational equipment running status analysis knowledge storehouse when human input.

2, its distinctive state analysis threshold value knowledge base can be set for each equipment, avoid the threshold error that external environment and equipment self-condition cause, it is ensured that the accuracy of alarm, early warning.

Accompanying drawing explanation

Fig. 1 is that the specific embodiment of the invention calculates each index finally obtained to equipment state weighing factor figure through Relief algorithm.

Fig. 2 is the design sketch of specific embodiment of the invention cluster.

Embodiment

This scheme combines Relief and kmeans two data mining algorithms and system information is analyzed, and extracts for last early warning alarm threshold.

1.Relief algorithm

Effect: equipment is abnormal, equipment interacts by equipment factor or multiple factor in operational process and to cause often. From numerous information that equipment gathers, extracting the factor affecting such exception, reject irrelevant factor, for data analysis below, this is exactly that feature is extracted. Relief is exactly such a algorithm. Relief algorithm is proposed by Kira the earliest, is confined to the classification problem of two class data at first. Relief algorithm is a kind of feature weight algorithm (Featureweightingalgorithms), gives feature different weights according to the dependency of each characteristic sum classification, and the feature that weight is less than certain threshold value will be removed. In Relief algorithm, the dependency of characteristic sum classification is that feature based is to the separating capacity of closely sample. Algorithm is random from training set D selects a sample R, then from the sample similar with R, nearest samples H is found, it is called NearHit, nearest samples M is found from the sample with R inhomogeneity, it is called NearMiss, then the weight of each feature is upgraded according to following rule: if the distance that the distance of R and NearHit in certain feature is less than on R and NearMiss, then illustrate that differentiation nearest neighbour that is similar and inhomogeneity is useful by this feature, then increase the weight of this feature; Otherwise, if the distance that R and NearHit is greater than on R and NearMiss in the distance of certain feature, illustrate that differentiation nearest neighbour that is similar and inhomogeneity is played negative effect by this feature, then reduce the weight of this feature. Above process Repeated m time, finally obtains the average weight of each feature.The weight of feature is more big, represents that the classification capacity of this feature is more strong, otherwise, represent that this tagsort ability is more weak. The working time of Relief algorithm linearly increases along with the increase of the sampling number of times m and primitive character number N of sample, and thus operational efficiency is very high.

In electrical network system, algorithm, for the service data of computer equipment, is described by certain below.

If the attribute that operation information comprises is as shown in the table:

Information attribute	Explanation	Feature number
			hostname	Host name	Nothing
cpu_1_load	The average load of 1min cpu	1
			cpu_5_load	The average load of 5min cpu	2
cpu_15_load	The average load of 15min cpu	3
			mem_total	Internal memory total amount	4
mem_free	The idle space of internal memory	5
			mem_used_rate	Internal memory utilization ratio	6
disk_total	Disk total amount	7
			disk_free	Disk slack space	8
disk_used_rate	Disk rate of utilization	9
			status	Equipment real-time status

In table, the information gathered being carried out feature number, each information gathered is attribute. The object of relief algorithm is exactly extract, from so many acquisition attributes, the attribute affecting Host Status, and these attributes affecting Host Status are exactly feature.

Using 400 main frame information collected as training set, each information all comprises information attribute listed in table. If the sampling number of times of sample is m, weight selects threshold value to be x. The following describes algorithm steps

1. according to real-time device state status, data are divided into normally and abnormal two classes;

2. the weight of each attribute of initialize is 0, and namely W (A)=0, A is feature number 1-9;

3. a random selection sample R from 400 data;

4. from similar sample, find nearest samples H, from inhomogeneity sample, find nearest samples M. Distance between sample and sample is Euclid distance, and nearest neighbour and Euclid distance differ minimum sample.

5. set A as feature number, use formula

W (A)=W (A)-diff (A, R, H)/m+diff (A, R, M)/m

Calculating the weight of feature number each attribute from 1 to 9 successively, wherein diff (A, R, H) is the nearest neighbor distance of this sample and homogeneous data, and what diff (A, R, M) was this sample and inhomogeneity data faces distance recently;

6. repeating step 3 arrives step 5m time, finally obtains the weight of each attribute;

7. selecting threshold value x to compare the weight of the weight of each attribute and setting successively, the great attribute in x of preference considers index as what affect this equipment operation risk.

2.kmeans algorithm

Effect: after the feature of previous step is extracted, have found the factor really affecting equipment state, but due to sample quantity very little (only an information), the early warning alarm threshold of equipment cannot be carried out scientific and effective extraction. Therefore, sample quantity now should be increased, the random error caused when avoiding threshold value to extract because sample number is very few. And the not random increase of sample, only meaningful to having the sample of relation to carry out analyzing ability when occurring with equipment is abnormal. At this time need kmeans algorithm to be brought together by the sample similar to this sample, extract for early warning alarm threshold below.

Algorithm: be the similar stroke of method given on data nature due to cluster algorithm, it is desired to the cluster obtained be the inner data of each cluster similar as much as possible and will big-difference as much as possible between cluster. So a kind of yardstick of definition weighs similarity just seems extremely important. In general, there is the method for two kinds of definition similarities. The first method is the distance between definition of data, description be the difference of data. 2nd kind of method is the similarity between direct definition of data. Here is the method for several frequently seen definition distance:

1.Euclidean distance, this is a kind of traditional distance concept, is suitable for 2,3 dimension spaces.

2.Minkowski distance is the expansion of Euclidean distance, it is possible to understand that be the distance of N dimension space.

Cluster algorithm has a variety of, can require to select suitable cluster algorithm according to the application of the object of involved data type, cluster and tool when needed. Introduce K-means cluster algorithm below:

K-means algorithm is a kind of conventional cluster algorithm based on division. K-means algorithm take k as parameter, and n object is divided into k bunch, has higher similarity in making bunch, and bunch between similarity lower. Then the data instance above, feature through the first step is extracted, assume to have found cpu_1_load, cpu_5_load, cpu_15_load, mem_free, mem_used_rate as characteristic value sequence, by these attributes as operational state of mainframe considered index, so in ensuing cluster, just only these attributes can be carried out cluster, get rid of the interference of other attributes. Therefore, new data form is:

Attribute	Feature number
		cpu_1_load	1
cpu_5_load	2
		cpu_15_load	3
mem_free	5
		mem_used_rate	6

Below for such data 400, arranging k according to risk class is 3 (3 kinds of risk class: devoid of risk, low risk, excessive risk), carries out the algorithmic descriptions of k-means:

1. random in 400 data select 3 samples as the barycenter of initial 3 bunches;

2. calculate the distance of other samples to these three barycenter successively, by sample dispensing to that bunch nearest from it;

3., after all data are all assigned to one bunch, recalculate the barycenter of each bunch. And replace original barycenter with new barycenter.

4. repeating step 1-3, until barycenter is not in change.

Bunch centroid calculation formula be:

$Z_j=\frac{1}{N_j}\underset{x\∈w_i}{\Σ}x$

Wherein Nj represents the number of data in bunch, in bunch arithmetic average a little be barycenter. Object generally adopts Euclidean distance to the distance of barycenter.

So by after k-means Algorithm Analysis, 400 data are finally divided into 3 bunches according to k value, namely devoid of risk bunch, low risk bunch, excessive risk bunch. Extract the metrics-thresholds of eigenwert bound as such state of each bunch.

The use of Relief algorithm and result are described by one group of test data below. As shown in Figure 1, it is assumed that whole index item of certain equipment are 1,2,3 ... 10, calculate each index of finally obtaining to equipment state weighing factor as shown in Figure 1 through Relief algorithm, as we know from the figure, index 6 is main influence factor. If the threshold value of feature weight has been defined as 0.2, so index 1,6,8 can be classified as influence index, and other low disturbance degree indexs remaining can be rejected.

When influence index and index weight are determined, then according to eigenwert, data can be carried out cluster analysis, below provide the design sketch of a cluster, as shown in Figure 2. By cluster, the data similar to sampled data can form one bunch, defines 4 bunches in fig. 2, and the data difference between bunch is big, and bunch in difference little. By analyzing with the coexist data of cluster of sampled data, it is determined that early warning alarm threshold.

Claims (3)

1. a grid automation system and equipment running status risk assessment algorithm, it is characterised in that, described algorithm comprises the steps:

1.1 extract based on the feature of Relief algorithm;

Equipment is abnormal, and equipment interacts by equipment factor or multiple factor in operational process and to cause often, from numerous information that equipment gathers, extract the factor affecting such exception, rejecting irrelevant factor, for data analysis below, this is exactly that feature is extracted;

1.2 assemble based on the similar sample of kmeans algorithm;

Get the characteristic index sequence affecting this equipment operation risk through step 1.1Relief algorithm, calculate the threshold range of each characteristic index under each risk class by k-means algorithm.

2. a kind of grid automation system according to claim 1 and equipment running status risk assessment algorithm, it is characterised in that, described step 1.1 is as follows based on the feature extraction concrete grammar of Relief algorithm:

Extraction equipment runs relevant index item, as affect equipment operation characteristic index can option, each can be numbered by option, option can extract the important indicator affecting equipment state by relief algorithm from all, as characteristic index;

Using the equipment operating data set in for some time as training set, every bar service data all comprises all index item operation values of this equipment, if the sampling number of times of sample is m, weight selects threshold value to be x, namely the final weight that calculates is greater than the index of x as characteristic index, and algorithm steps is as follows:

A data are classified by () according to real-time device state status;

B the weight of each attribute of () initialize is 0, namely W (A)=0, A is index number;

C () is random from training set selects a sample R;

D () finds nearest samples H from similar sample, find nearest samples M from inhomogeneity sample; Distance between sample and sample is Euclid distance, and nearest neighbour and Euclid distance differ minimum sample;

E () sets A as index number, use formula

W (A)=W (A)-diff (A, R, H)/m+diff (A, R, M)/m

Calculating the weight of each index successively, wherein diff (A, R, H) is the nearest neighbor distance of this sample and homogeneous data, and what diff (A, R, M) was this sample and inhomogeneity data faces distance recently;

F () repeating step (c), to step (e) m time, finally obtains the weight of each attribute;

G the weight of the weight of each attribute and setting is selected threshold value x to compare by () successively, the great attribute in x of preference considers index as what affect this equipment operation risk.

3. a kind of grid automation system according to claim 1 and equipment running status risk assessment algorithm, it is characterised in that, described step 1.2 concrete grammar is as follows:

Using the equipment operating data in for some time as training set, every bar service data comprises the operation value of characteristic index, and it is { k1, k2 according to risk class by Data Placement ... kn}n class, algorithm steps is as follows:

(1) select n sample as the barycenter of individual bunch of initial n from training set at random;

(2) distance of other samples to this n barycenter is calculated successively, by sample dispensing to that bunch nearest from it;

(3) after all data are all assigned to one bunch, the barycenter of each bunch is recalculated; And replace original barycenter with new barycenter;

(4) repeating step (1)-(3), until barycenter no longer changes;

Bunch centroid calculation formula be:

Wherein Nj represents the number of data in bunch, in bunch arithmetic average a little be barycenter; Object generally adopts Euclidean distance to the distance of barycenter;

So by, after k-means Algorithm Analysis, training set data is finally divided into n bunch according to risk class, comprises the service data item under such risk class in each bunch; Extract the metrics-thresholds of characteristic index extreme value as such risk class of service data in each bunch.