CN105427043A - Improved nearest neighbor algorithm-based power grid alarm analysis method - Google Patents

Abstract

The invention discloses an improved nearest neighbor algorithm-based power grid alarm analysis method. The method comprises the following steps: carrying out classification and statistic on the data such as accident trip, alarm signal and measurement imbalance of a power grid scheduling automatic system to obtain alarms from the system to the regional monitors, wherein the alarms can be mainly divided into accident, abnormity, out of limit, deflection and notification; and combining the history alarm information to determine the weight of each alarm index, and adopting an improved nearest neighbor algorithm to judge the type of the current alarm so as to carry out corresponding processing. The method provided by the invention is capable of improving the analysis reliability and improving the working efficiency.

Description

A kind of electrical network alert analysis method of the nearest neighbor algorithm based on improving

Technical field

The present invention relates to power network schedule automation security fields, be specifically related to an electrical network alert analysis method based on the nearest neighbor algorithm improved.

Background technology

Along with the continuous expansion of electrical network scale, electric network composition is increasingly sophisticated, and system operation mode is quick and various, and operational management is faced with larger challenge.And power system monitor information category is various, data volume is huge, by the method monitoring power system monitor information of science, classification analysis finds that operation of power networks rule is to make a policy fast, seems very necessary.

Monitoring business is incorporated in regulation and control center service after " three collection five are large " adjustment, and each department are also in the exploratory stage about the treatment and analyses of pilot signal data before.Also mainly based on to the information collected, tabulate statistics is carried out to the Treatment Analysis of monitor data, lacks analysis means that is comprehensive, system and classified finishing is carried out to information, reduce work efficiency.

Summary of the invention

The present invention is directed to above problem, providing one can sort out data message and statistics, and then analyzes, and is convenient to differentiate alarm classification, improves and differentiate reliability, the electrical network alert analysis method based on the nearest neighbor algorithm improved of increasing work efficiency.

Technical scheme of the present invention is: comprise the following steps:

S1, by sorting out the data of automation system for the power network dispatching and adding up, the alarm data obtained to be classified, its classification be accident, abnormal, out-of-limit, conjugate and inform;

S2, in conjunction with history alarm information, determine the weights of each alarm index; Adopt the nearest neighbor algorithm improved to differentiate and the classification that current alarm belongs to process accordingly.

Step S2 comprises the following steps:

(1) Data Structure Design: set the warning information set in each area abstract as a tuple, be expressed as x _i=(x _i1, x _i2..., x _im), m is the dimension of sample object, and so history alarm information is expressed as X={x ₁, x ₂..., x _n, n is the number of sample object; x _ithe warning information of a certain regional known class that corresponding automation system for the power network dispatching collects;

(2) nearest neighbor algorithm improved is analyzed warning information, and step is as follows:

Step 1: build training sample set, by history alarm information X={x ₁, x ₂..., x _nas training sample set;

Step 2: establish defining K value, K concentrates at training sample the number choosing sample to be tested " neighbour ";

Step 3: concentrate at training sample and select the K nearest with a sample to be tested sample, " neighbour " between sample is measured by Euclidean distance; Suppose that sample to be tested is x _c=(x _c1, x _c2..., x _cm), now sample to be tested and training sample x _ibetween distance be:

$d^{\′}(x_i,x_c)=\sqrt[2]{w_1{(x_{i1}-x_{c1})}^2+w_2{(x_{i2}-x_{c2})}^2+\mathrm{...}+w_m{(x_{im}-x_{cm})}^2},$

Wherein w ₁, w ₂..., w _mbe respectively the weight that respective attributes index is corresponding, weights set is defined as W={w ₁, w ₂..., w _m; Wherein, the determination of weights is as follows:

Step 3.1: get w ₁, w ₂..., w _mvalue be all 1, the sample of several known classification is tested as test set, adds up difference number p between it and the actual classification value of sample;

Step 3.2: remove the jth attribute in sample successively, then obtain the estimation classification value of sample, add up difference number p between it and the actual classification value of sample _j(j=1,2 ..., m);

Step 3.3: calculate g _jrepresent that so the value of each element of weights set W is when lacking jth property value index to the influence degree of classifying j=1,2 ..., m, especially, if p _j=0 or p=0, put g _j=1;

Step 4: suppose, according to the process different to history alarm information, to be divided into the alarm of q class, to be expressed as S={s ₁, s ₂..., s _q; For sample to be tested x _c, x ₁, x ₂..., x _krepresent and x _ck nearest sample, if discrete objective function is f:x → s _i, wherein x represents certain alarm sample, s _irepresent i-th classification; represent f (x _c) estimation, be calculated as this function representation for making s when value is maximum, s ∈ S={s ₁, s ₂..., s _q, for if s=f is (x _i), otherwise so above formula just can export alarm classification s corresponding maximum in K neighbour of sample to be tested;

Step 5: namely be sample to be tested x _ccorresponding classification, and according to classification to x _ccarry out respective handling.

Alarm data in step S1 comprises emergency stop valve trip statistics, alarm signal statistics and measures uneven statistics.

The present invention is by accident, abnormal, out-of-limit, the collection that conjugates and inform five class warning information, and each alarm classification is composed with different weights, then in conjunction with history alarm information, determine the weights of each alarm index, and adopt the nearest neighbor algorithm improved to differentiate the classification that current alarm belongs to, carry out respective handling.

To the emergency stop valve trip of automation system for the power network dispatching, alarm signal and measure unbalanced data and sort out and add up, for the nearest neighbor algorithm of improvement provides effective Data support.

The present invention concludes, extracts crucial monitoring index data from the data such as numerous and diverse monitoring tripping operation, alarm and unbalancedness, and carry out analyzing, sort out, sum up, discovery historical law is summed up to help monitor staff, aid decision making, thus alleviate its heavy statistic analysis, the monitoring analysis ability of effective raising power system monitor department, and then increase work efficiency.

Accompanying drawing explanation

Fig. 1 is process flow diagram of the present invention;

Embodiment

The present invention as shown in Figure 1, comprises the following steps:

S1, by sorting out the data of automation system for the power network dispatching and adding up, the alarm data obtained to be classified, its classification be accident, abnormal, out-of-limit, conjugate and inform;

S2, in conjunction with history alarm information, determine the weights of each alarm index; Adopt the nearest neighbor algorithm improved to differentiate and the classification that current alarm belongs to process accordingly.

Step S2 comprises the following steps:

(1) Data Structure Design: set the warning information set in each area abstract as a tuple, be expressed as x _i=(x _i1, x _i2..., x _im), m is the dimension of sample object, and so history alarm information is expressed as X={x ₁, x ₂..., x _n, n is the number of sample object.M=5 in this example, represent respectively alarm classification be accident, abnormal, out-of-limit, conjugate and inform 5 ATTRIBUTE INDEX; x _ithe warning information of a certain regional known class that corresponding automation system for the power network dispatching collects, x _inamely the warning information that correspondence i-th is regional; N is then for collecting the sum of warning information;

(2) nearest neighbor algorithm improved is analyzed warning information, and step is as follows:

Step 1: build training sample set, by history alarm information X={x ₁, x ₂..., x _nas training sample set;

Step 2: establish defining K value, K concentrates at training sample the number choosing sample to be tested " neighbour ".In work, first determine an initial value, then constantly adjust according to the accuracy of classification, finally reach optimum; When the accuracy of classification results is lower than the threshold value pre-set, just increase the value of K, until make the accuracy of classification results reach requirement.

Step 3: concentrate at training sample and select the K nearest with a sample to be tested sample, " neighbour " between sample is measured by Euclidean distance, distance is less then represents nearer with the distance of sample to be tested.Suppose that sample to be tested is x _c=(x _c1, x _c2..., x _cm), now sample to be tested and training sample x _ibetween distance be:

$d^{\′}(x_i,x_c)=\sqrt[2]{w_1{(x_{i1}-x_{c1})}^2+w_2{(x_{i2}-x_{c2})}^2+\mathrm{...}+w_m{(x_{im}-x_{cm})}^2},$

Wherein w ₁, w ₂..., w _mbe respectively the weight that respective attributes index is corresponding, weights set is defined as W={w ₁, w ₂..., w _m; Wherein, the determination of weights is as follows:

Step 3.1: get w ₁, w ₂..., w _mvalue be all 1, the sample of several known classification is tested as test set, adds up difference number p between it and the actual classification value of sample;

Step 3.2: remove the jth attribute in sample successively, then obtain the estimation classification value of sample, add up difference number p between it and the actual classification value of sample _j(j=1,2 ..., m);

Step 3.3: calculate g _jrepresent that so the value of each element of weights set W is when lacking jth property value index to the influence degree of classifying j=1,2 ..., m, especially, if p _j=0 or p=0, put g _j=1;

Determine the weights of each alarm index in the present invention, improve the reliability analyzed and the reliability determining follow-up alarm classification, and then take corresponding process, save time, improve work efficiency.

Step 4: suppose, according to the process different to history alarm information, to be divided into the alarm of q class, to be expressed as S={s ₁, s ₂..., s _q; For sample to be tested x _c, x ₁, x ₂..., x _krepresent and x _ck nearest sample, if discrete objective function is f:x → s _i, wherein x represents certain alarm sample, s _irepresent i-th classification; represent f (x _c) estimation, be calculated as this function representation for making s when value is maximum, s ∈ S={s ₁, s ₂..., s _q, for if s=f is (x _i), otherwise so above formula just can export alarm classification s corresponding maximum in K neighbour of sample to be tested;

Step 5: namely be sample to be tested x _ccorresponding classification, and according to classification to x _ccarry out respective handling.

Alarm data in step S1 comprises emergency stop valve trip statistics, alarm signal statistics and measures uneven statistics.Wherein, emergency stop valve trip is added up, such as, by area, electric pressure statistics line tripping number of times; Alarm signal is added up, such as, by responsibility of maintenance district statistics warning information situation; Measurement unbalancedness is added up, and such as, by plant stand statistics, clicks the uneven situation can checking certain plant stand every day.

Specifically be implemented as follows:

Get K=5, the classification of alarm is S={s ₁, s ₂, s ₃, s ₄four classes, namely alarm is divided into four classes, carries out classification process.

(1) calculating accident, abnormal, out-of-limit, conjugate and inform weights corresponding to 5 ATTRIBUTE INDEX:

Step 1: the test sample book choosing 12 known alarm classifications, tests 12 test sample books respectively.S is belonged to for one ₂class testing sample is x, works as w ₁, w ₂..., w ₅when=1, utilize Euclidean distance to choose 5 nearest samples of training sample middle distance test sample book x namely " neighbour ", then utilize 5 " neighbour " obtain the estimation classification of test sample book x.The estimation classification of 12 samples can be obtained after the same method.The difference number of last statistical estimate classification and actual classification, supposes that difference number is 2.

Step 2: remove an ATTRIBUTE INDEX in sample respectively, then obtain the estimation classification value of sample, add up difference number between it and the actual classification value of sample.Suppose that difference number corresponding to 5 ATTRIBUTE INDEX is followed successively by 2,2,1,3,4, the weights that so five ATTRIBUTE INDEX are corresponding are $w_1=\frac{\frac{2}{2}}{\frac{2}{2}+\frac{2}{2}+\frac{1}{2}+\frac{3}{2}+\frac{4}{2}}=\frac{1}{6},$ Successively $w_2=\frac{1}{6},w_3=\frac{1}{12},w_4=\frac{1}{4},w_5=\frac{1}{3}.$

(2) utilize weights obtained above, obtain the belonging classification of sample to be tested:

For a sample to be tested x _c=(1,3,1,2,1), obtaining its 5 neighbours is x ₁=(2,1,0,2,1), x ₂=(1,2,1,1,1), x ₃=(1,3,1,0,1), x ₄=(0,4,1,2,1), x ₅=(1,2,0,3,1), wherein x ₁belong to s ₁, x ₂, x ₃, x ₄, x ₅belong to s ₂, then x can be obtained _c5 neighbours in belong to s ₂the sample of classification is maximum, then x _cbelong to s ₂.

Claims (3)

1., based on an electrical network alert analysis method for the nearest neighbor algorithm improved, it is characterized in that, comprise the following steps:

S1, by sorting out the data of automation system for the power network dispatching and adding up, the alarm data obtained to be classified, its classification be accident, abnormal, out-of-limit, conjugate and inform;

S2, in conjunction with history alarm information, determine the weights of each alarm index; Adopt the nearest neighbor algorithm improved to differentiate and the classification that current alarm belongs to process accordingly.

2. the electrical network alert analysis method of a kind of nearest neighbor algorithm based on improving according to claim 1, it is characterized in that, step S2 comprises the following steps:

(1) Data Structure Design: set the warning information set in each area abstract as a tuple, be expressed as x _i=(x _i1, x _i2..., x _im), m is the dimension of sample object, and so history alarm information is expressed as X={x ₁, x ₂..., x _n, n is the number of sample object; x _ithe warning information of a certain regional known class that corresponding automation system for the power network dispatching collects;

(2) nearest neighbor algorithm improved is analyzed warning information, and step is as follows:

Step 1: build training sample set, by history alarm information X={x ₁, x ₂..., x _nas training sample set;

Step 2: establish defining K value, K concentrates at training sample the number choosing sample to be tested " neighbour ";

Step 3: concentrate at training sample and select the K nearest with a sample to be tested sample, " neighbour " between sample is measured by Euclidean distance; Suppose that sample to be tested is x _c=(x _c1, x _c2..., x _cm), now sample to be tested and training sample x _ibetween distance be:

$d^{\′}(x_i,x_c)=\sqrt[2]{w_1{(x_{i1}-x_{c1})}^2+w_2{(x_{i2}-x_{c2})}^2+...+w_m{(x_{im}-x_{cm})}^2},$ Wherein w ₁, w ₂..., w _mbe respectively the weight that respective attributes index is corresponding, weights set is defined as W={w ₁, w ₂..., w _m; Wherein, the determination of weights is as follows:

Step 3.1: get w ₁, w ₂..., w _mvalue be all 1, the sample of several known classification is tested as test set, adds up difference number p between it and the actual classification value of sample;

Step 3.2: remove the jth attribute in sample successively, then obtain the estimation classification value of sample, add up difference number p between it and the actual classification value of sample _j(j=1,2 ..., m);

Step 3.3: calculate g _jrepresent that so the value of each element of weights set W is when lacking jth property value index to the influence degree of classifying especially, if p _j=0 or p=0, put g _j=1;

Step 4: suppose, according to the process different to history alarm information, to be divided into the alarm of q class, to be expressed as S={s ₁, s ₂..., s _q; For sample to be tested x _c, x ₁, x ₂..., x _krepresent and x _ck nearest sample, if discrete objective function is f:x → s _i, wherein x represents certain alarm sample, s _irepresent i-th classification; represent f (x _c) estimation, be calculated as this function representation for making s when value is maximum, s ∈ S={s ₁, s ₂..., s _q, for if s=f is (x _i), otherwise so above formula just can export alarm classification s corresponding maximum in K neighbour of sample to be tested;

Step 5: namely be sample to be tested x _ccorresponding classification, and according to classification to x _ccarry out respective handling.

3. the electrical network alert analysis method of a kind of nearest neighbor algorithm based on improving according to claim 1, is characterized in that, the alarm data in step S1 comprises emergency stop valve trip statistics, alarm signal statistics and measures uneven statistics.