CN104967097A - Excitation surge current identification method based on support vector classifier - Google Patents

Abstract

Provided is an excitation surge current identification method based on a support vector classifier. The method selects seven characteristics of a secondary second harmonic, a third harmonic, a current dead angle, a wave width, a waveform distortion amount, a waveform correlation coefficient and an excitation side measured impedance as inputs of a support vector machine, then various running states of a transformer are trained, and a decision function identifying an excitation surge current and a fault current is constructed. When the transformer is out of order, data collected by a protection device collection system is calculated to obtain seven characteristics, the seven characteristics are put into the decision function and determination of an excitation surge current and a fault current is carried out. The identification method integrates advantages of principles of harmonic wave braking, a dead angle, waveform similarity and the like and avoids respective limitation, and the surge current identification credibility is raised. An algorithm can be converted into a convex optimization problem finally, and the local minimum problem which cannot be solved by a neural network is avoided. The method is free from a transformer wiring mode and is free from model parameters, the applicability is strong and the flexibility is good.

Description

Based on the excitation flow recognition method of support vector classification

Technical field

The present invention relates to a kind of excitation flow recognition method, specifically a kind of excitation flow recognition method based on support vector classification.

Background technology

In recent years, along with the construction of superhigh pressure, bulk power grid, increasing high-power transformer comes into operation, and this effect also making transformer play in power grid operation all the more important, makes tranformer protection subject increasing test.But research shows, the performance factor of 220kv and above power trans-former protection is hovered 70% ~ 80% always, far below the performance factor of generator and 220kV and Above Transmission Lines protection.Trace it to its cause and be, tranformer protection can not distinguish magnetizing inrush current and fault current very accurately, and this becomes gear at the mountain improving tranformer protection action rate.

Numerous scholar both domestic and external is devoted to the research distinguishing magnetizing inrush current and fault current for many years always, has made brilliant achievements.The advantage that each method has it original, but also have some limitations simultaneously.As:

Secondary harmonic brake realizes convenient and simple, most widely used in tranformer protection, but the impact of the shunt capacitance of reactive power compensation and ultra-high-tension power transmission line distributed capacitance makes also there is higher second harmonic during transformer fault; When the reduction of high-power transformer saturation flux makes magnetizing inrush current, secondary harmonic component is reduced to less than 10% sometimes.

Interval angle recognition principle is also conventional excitation flow recognition method, but it is saturated by CT, aperiodic component, the impact such as sample frequency is larger, artificial neural network method be subject to local minizing point impact and to training sample interdependency higher, fuzzy multi criteria principle is difficult to determine to the weight that each criterion is got.

Equivalent circuit method avoids the impact that differential current is brought, but is subject to transformer transient Model, the isoparametric restriction of transformer leakage inductance.

So, find one and can inherit absorption forefathers achievement, its method with shortcoming can be overcome again necessary.

Summary of the invention

In order to solve the limitation of said method when identifying magnetizing inrush current, the object of the present invention is to provide a kind of advantage that can either utilize independent characteristic quantity criterion, can avoid again the limitation of single criterion, criterion is accurately based on the excitation flow recognition method of support vector classification.

The present invention solves the problems of the technologies described above taked technical scheme: a kind of excitation flow recognition method based on support vector classification, and it is characterized in that, the method comprises the following steps:

The input feature vector amount of a, selection SVMs,

The content of second harmonic, triple-frequency harmonics in a1, collection three-phase differential current, with discriminant equation $\left\{\begin{array}{c}{I}_{2nd}>K_{2xb}*I_{1st}\\ I_{3rd}>K_{3xb}*I_{1st}\end{array}\right.$ Based on, order $\left\{\begin{array}{c}{K}_2=I_{2nd}/I_{1st}\\ K_3=I_{3nd}/I_{1st}\end{array}\right.,$ By K ₂as first characteristic quantity input of SVMs, by K ₃as second characteristic quantity input of SVMs; Wherein, I _2ndfor the second harmonic in every phase differential current; I _3rdfor the triple-frequency harmonics in every phase differential current; I _1stfor the difference stream first-harmonic of corresponding phase, k _2xbfor secondary harmonic brake coefficient setting value; k _3xbfor third harmonic restrained coefficient setting value;

Flow into the differential current of differential relay when a2, collection power transformer interior fault, take absolute value again after differential, | i ' ₂|, with | i ' ₂| < KI ' _m2angle θ corresponding to duration as the 3rd characteristic quantity input of SVMs, wherein I ' _m2for | i ' ₂| amplitude, K=0.25;

A3, gather four the characteristic quantity input of the wide Width=of ripple (t/T) * 360 as SVMs of magnetizing inrush current;

A4, gather five the characteristic quantity input of integrated value S as SVMs of differential current, wherein wherein i (k) represents the value of a kth sampled point, and N represents the sum of a cycle sampled point, and Ts represents the time difference of adjacent two sampled points;

A5, half cycles before differential current is set to x (t), later half cycle negate is also set to y (t), and the two does correlation analysis, adopts waveform correlation coefficient J as the 6th feature input variable of SVMs, wherein k wherein in x (k) and y (k) represents the number of sampled point, and N represents the sum of a cycle sampled point;

A6, using seven the characteristic quantity input of the change degree σ Z of measurement impedance Z (t) as SVMs, wherein k represents the number of sampled point, and N represents the sum of a cycle sampled point, and Ts represents the time difference of adjacent two sampled points, and T represents all wave period;

B, train the various running statuses of transformer, the decision function of conformation identification magnetizing inrush current and fault current, described training step is:

B1, obtain the true dynamic model data under transformer various running status by dynamic simulation test, and draw seven input feature vector values in step a accordingly,

If x _i∈ χ=R ⁷, x _i=(K _2i, K _3i, θ _i, W _i, S _i, J _i, σ Z _i) ^t, y _i∈ γ=1 ,-1},

Y during magnetizing inrush current _i=1; Y during fault _i=-1;

Then negative class sample is expressed as:

T ₁＝{((K _2,1,K _3,1,θ ₁,W ₁,S ₁,J ₁,σZ ₁) ^T,-1),…

(x _a,-1)}∈(χ×γ) ^a

Wherein, a is negative class sample number;

Positive class sample is expressed as:

T ₂＝{((K _2,a+1,K _3,a+1,θ _a+1,W _a+1,S _a+1,J _a+1,σZ _a+1) ^T,+1),

…(x _a+b,+1)}∈(χ×γ) ^b

Wherein, a is negative class sample number, and b is positive class sample number;

B2, introducing kernel function if σ=0.07, C=1, structure optimum programming equation, brings optimum ruleization equation into by parameter

$\underset{\mathrm{\&alpha;}}{min}\frac{1}{2}\underset{i=1}{\overset{a+b}{\&Sigma;}}\underset{j=1}{\overset{a+b}{\&Sigma;}}{y}_i{y}_j{\mathrm{\&alpha;}}_i{\mathrm{\&alpha;}}_j\exp \{-{\frac{|x_i-x_j|}{{\mathrm{\&sigma;}}^2}}^2\}-\underset{j=1}{\overset{a+b}{\&Sigma;}}{\mathrm{\&alpha;}}_j;$

$\underset{i=1}{\overset{a+b}{\&Sigma;}}{y}_i{\mathrm{\&alpha;}}_j=0;$

0≤α _i≤C,i＝1,…,a+b；

Obtain optimal solution ${\mathrm{\&alpha;}}^{*}={({\mathrm{\&alpha;}}_1^{*},...{\mathrm{\&alpha;}}_l^{*})}^T;$

B3, choose α ^*a positive component calculated threshold

B4, structure decision function $f\left(x\right)=sgn(\underset{i=1}{\overset{l}{\&Sigma;}}{y}_i{\mathrm{\&alpha;}}_i^{*}K(x\&CenterDot;x_i)+b^{*});$

C, after transformer has an accident, the data collection of protective device acquisition system come calculate in the decision function brought into after seven kinds of input feature vector values in b4 through step a, if

Further, in step b1, if there is newly-increased training sample, sample will be increased newly for people's factor of influence (ES) estimation formulas:

$\begin{array}{c}ES\left(x\right)=0.5+\frac{1}{1+e^{-\left|g\left(x\right)\right|}}\\ =0.5+\frac{1}{1+e^{-\left|\underset{i=1}{\overset{76}{\&Sigma;}}{y}_i{\mathrm{\&alpha;}}_i^{*}\exp \{-\frac{|x_i-x{|}^2}{{0.007}^2}\}+0.8821\right|}}\end{array};$

The sample of ES value between 0.8-1.0 is retained.

The invention has the beneficial effects as follows:

1, the present invention using wide to second harmonic, triple-frequency harmonics, electric current interval angle, ripple, wave distortion amount, waveform correlation coefficient and these seven characteristic quantities of excitation side measurement impedance as the criterion identifying magnetizing inrush current, introduce kernel function, seven characteristic quantities are mapped to the septuple space simultaneously, the difference of magnetizing inrush current and fault current is found in the septuple space, so both comprehensive harmonic brakings, interval angle, the advantage of the principles such as wave-form similarity, avoid again the limitation of its principle separately, add the confidence level of magnetizing inrush current identification.

2, by gathering the various running status sample of transformer, training being carried out to sample and obtains strong adaptability, the reliable criterion of magnetizing inrush current identification that generalization ability is good.

3, the method is not by the impact of transformer connection mode, do not restrict by model parameter, and applicability is strong, and flexibility is good.

4, adopt optimal sample labelling method constantly to revise newly-increased sample, make this magnetizing inrush current identical criterion will be more accurate.

5, this method is carried out to the test of accuracy, choose the sample of various running state of transformer, test result is as shown in the table:

Can obtain after tested, the method well can identify no-load switchon surge, general internal fault and drop in alternate and earth fault, and also can well identify for the turn-to-turn short circuit of dropping in more than 10%, correct recognition rata reaches 96.2%.

Accompanying drawing explanation

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

Fig. 2 is experimental transformer system model schematic diagram of the present invention.

Embodiment

Below in conjunction with accompanying drawing, specific embodiments of the invention are described in detail:

As shown in Figure 1, a, first selection SVMs input feature vector amount

The method of current identification magnetizing inrush current is roughly divided into three classes: based on waveform (harmonic braking principle, waveform degree of correlation principle, principles of mathematical morphology, amplitude change principle, mathematical morphology, improves mathematical morphology etc.), based on model (equivalent circuit equation principle, power differential principle etc.), and based on artificial intelligence (artificial neural net and based on fuzzy nearness principle etc.).In line with realizing principle easily and efficiently, existing excitation flow recognition method is selected, waveform is inputted as SVMs characteristic quantity.

A1, harmonic content identification magnetizing inrush current

Adopt the content of second harmonic, triple-frequency harmonics in three-phase differential current to identify magnetizing inrush current, discriminant equation is as follows:

$\left\{\begin{array}{c}{I}_{2nd}>K_{2xb}*I_{1st}\\ I_{3rd}>K_{3xb}*I_{1st}\end{array}\right.$

Wherein I _2nd, I _3rdbe respectively the second harmonic in every phase differential current and triple-frequency harmonics, I _1stfor the difference stream first-harmonic of corresponding phase, k _2xb, k _3xbbe respectively second harmonic and third harmonic restrained coefficient setting value.Order

$\left\{\begin{array}{c}{K}_2=I_{2nd}/I_{1st}\\ K_3=I_{3nd}/I_{1st}\end{array}\right.$

By K ₂, K ₃respectively as first input feature vector amount and second input feature vector amount of SVMs.

The wide identification magnetizing inrush current of a2, interval angle and ripple

There will be interval angle in the waveform of magnetizing inrush current, and the stable state differential current flowing into differential relay during power transformer interior fault is sinusoidal wave, there will not be interval angle.By differential current after differential, then take absolute value, | i ' ₂|, I ' _m2for | i ' ₂| amplitude; K=0.25.With | i ' ₂| < KI ' _m2angle θ corresponding to duration as the 3rd characteristic quantity input of SVMs.

Increase the assistant criteria that a reflection ripple is wide, will the correctness identified be improved.The wide Width=of ripple (t/T) * 360, can be used as the 4th characteristic quantity input of SVMs.

A3, wave distortion identification magnetizing inrush current

During fault, difference stream is power frequency sine wave substantially.And during magnetizing inrush current, having a large amount of harmonic components to exist, waveform distorts, and is interrupted, asymmetric.Utilize algorithm identified to go out this distortion, can magnetizing inrush current be identified.

According to this criterion, using five input of the integrated value S of differential current as SVMs.Algorithm is as follows:

$S=\&lsqb;\frac{1}{2}i\left(0\right)+\underset{k=1}{\overset{N-1}{\&Sigma;}}i\left(k\right)+\frac{1}{2}i\left(N\right)\&rsqb;T_S$

A4, waveform correlation coefficient

Half cycles before differential current is set to x (t), and later half cycle negate is also set to y (t), and the two does correlation analysis.Now form factor is:

Wherein: ρ (X, Y) is the covariance coefficient between x (t) and y (t), σ ²x variance that () is x (t), the coefficient correlation J of fault current is close to 1, and the coefficient correlation of magnetizing inrush current is close to zero.So waveform correlation coefficient can be adopted as the 7th feature input variable of SVMs, the algorithm of mapping:

$J=\frac{\underset{K=0}{\overset{N/2-1}{\&Sigma;}}x\left(k\right)y\left(k\right)-\frac{2}{N}\underset{K=0}{\overset{N/2-1}{\&Sigma;}}x\left(k\right)\underset{K=0}{\overset{N/2-1}{\&Sigma;}}y\left(k\right)}{\underset{K=0}{\overset{N/2-1}{\&Sigma;}}{x}^2\left(k\right)-\frac{2}{N}{\left(\underset{K=0}{\overset{N/2-1}{\&Sigma;}}x\left(k\right)\right)}^2}$

A5, excitation measurement impedance method

Measurement impedance substantially constant when normal operation and fault, be then jumpy when magnetizing inrush current.Therefore, the change of measurement impedance can be utilized differentiate and shove and fault.Wherein U, I are the voltage, the electric current calculating amplitude at a time that are obtained by filtering algorithm.Its mathematic expectaion is mean square deviation is the change degree that in formula, σ Z (t) is measurement impedance Z (t).

If σ is Z _setfor threshold value, then

Accordingly, can using seven the characteristic quantity input of the change degree σ Z of measurement impedance Z (t) as SVMs.Corresponding algorithm is as follows:

$\mathrm{\&sigma;}Z=\frac{T_S}{T}\{\frac{1}{2}{\&lsqb;Z\left(0\right)-E\left(Z\right)\&rsqb;}^2+\underset{k=1}{\overset{N-1}{\&Sigma;}}{\&lsqb;Z\left(k\right)-E\left(Z\right)\&rsqb;}^2+\frac{1}{2}{\&lsqb;Z\left(N\right)-E\left(Z\right)\&rsqb;}^2\}.$

B, training

B1, obtain the true dynamic model data under transformer various running status by dynamic simulation test, and draw seven input feature vector values in step a accordingly.

Below in conjunction with experimental transformer, this step is described in detail:

The true dynamic model data under the various running status of a large amount of transformers are obtained, for SVMs two deformation algorithm identification magnetizing inrush current of classifying provides training desired data by dynamic simulation test.

As shown in Figure 2, the three-phase transformer that experimental transformer T2 is made up of three single-phase DMB-10 type transformers, wiring group is Yn, d11.Its relevant parameter is: (1) capacity: 10kVA; (2) low-pressure side rated voltage: 380V; (3) low-pressure side rated current: 25.3A; (4) high-pressure side rated voltage: 1000V; (5) high-pressure side rated current: 10A; (6) no-load current: 1.45%; (7) no-load loss: 100W; (8) short circuit loss: 0.35%; (9) short-circuit voltage: 9%-15%.(10), when carrying out data record ripple, each cycle gathers 24 points.In Fig. 2, T1 is voltage regulating transformer, CT is current transformer, PT is voltage transformer.

Table one lists simulation test system to the various running status training sample of transformer.

The each running status sample of table one transformer

If x _i∈ χ=R ⁷, x _i=(K _2i, K _3i, θ _i, W _i, S _i, J _i, σ Z _i) ^t, y _i∈ γ={ 1 ,-1}, y during magnetizing inrush current _i=1; Y during fault _i=-1.

Then negative class sample is expressed as:

T ₁＝{((K _2,1,K _3,1,θ ₁,W ₁,S ₁,J ₁,σZ ₁) ^T,-1),…

(x ₇₀₀₀,-1)}∈(χ×γ) ⁷⁰⁰⁰

Positive class sample is expressed as:

T ₂＝{((K _2,7001,K _3,7001,θ ₇₀₀₁,W ₇₀₀₁,S ₇₀₀₁,J ₇₀₀₁,σZ ₇₀₀₁) ^T,+1),

…(x ₁₃₀₀₀,+1)}∈(χ×γ) ⁶⁰⁰⁰

B2, introducing kernel function.

Because being input as 7 dimension variablees, the space do not corresponded in reality, also cannot find physical plane H, so introduce a kernel function K (xx'), observation data being mapped to higher dimensional space, asking optimal solution by finding hyperplane (sphere).

The also the most frequently used kernel function of current most study has: 1, the kernel function of polynomial form; 2, RBF type (RBF) kernel function; 3, sigmoid function kernel function etc.

This trains employing as kernel function, the support vector obtained is a kind of radial basis function classifiers.The basic distinction of it and conventional radial basic function (RBF) method is, the center of each basic function corresponds to a support vector here, they and export weights and all automatically determined by algorithm.

If σ=0.07, C=1, and by above kernel function, σ=0.07, C=1, and seven kinds of input feature vector value parameters substitute into the optimum programming equation constructed

$\underset{\mathrm{\&alpha;}}{\min }\frac{1}{2}\underset{i=1}{\overset{a+b}{\&Sigma;}}\underset{j=1}{\overset{a+b}{\&Sigma;}}{y}_i{y}_j{\mathrm{\&alpha;}}_i{\mathrm{\&alpha;}}_j\exp \{-{\frac{|x_i-x_j|}{{\mathrm{\&sigma;}}^2}}^2\}-\underset{j=1}{\overset{a+b}{\&Sigma;}}{\mathrm{\&alpha;}}_j;$

$\underset{i=1}{\overset{a+b}{\&Sigma;}}{y}_i{\mathrm{\&alpha;}}_j=0;$

0≤α _i≤C,i＝1,…,a+b；

: $\underset{\mathrm{\&alpha;}}{min}\frac{1}{2}\underset{i=1}{\overset{13000}{\&Sigma;}}\underset{j=1}{\overset{13000}{\&Sigma;}}{y}_i{y}_j{\mathrm{\&alpha;}}_i{\mathrm{\&alpha;}}_j\exp \{-\frac{|x_i-x_j{|}^2}{{0.007}^2}\}-\underset{j=1}{\overset{13000}{\&Sigma;}}{\mathrm{\&alpha;}}_j;s.t.\underset{i=1}{\overset{13000}{\&Sigma;}}{y}_i{\mathrm{\&alpha;}}_j=0;0\&le;{\mathrm{\&alpha;}}_i\&le;1,i=1,...,l.;$

Solve through training and obtain 76 support vectors.3 support vectors are wherein had to bear class sample and 73 positive class samples of support vector represent respectively:

T ₁＝{((K _2,1,K _3,1,θ ₁,W ₁,S ₁,J ₁,σZ ₁) ^T,-1),…(x ₃,-1)}∈(χ×γ) ³T ₂＝{(K _2,j,K _3,j,θ _j,W _j,S _j,J _j,σZ _j) ^T,+1}∈(χ×γ) ^1…73

Its corresponding optimal solution value as shown in Table 2:

Table two α optimal solution

Numbering	1	2	3	4	5	6	7	8
									α value	0.7120	0.5643	0.8454	-0.0045	-0.0776	-0.0112	-0.0236	-0.0341
Numbering	9	10	11	12	13	14	15	16
									α value	-0.0222	-0.0020	-0.0824	-0.0019	-0.0712	-0.0162	-0.0076	-0.0053
Numbering	17	18	19	20	21	22	23	24
									α value	-0.0243	-0.0081	-0.0010	-0.0378	-0.0190	-0.0081	-0.0411	-0.0027
Numbering	25	26	27	28	29	30	31	32
									α value	-0.0355	-0.0012	-0.0354	-0.0107	-0.0262	-0.0149	-0.0843	-0.0274
Numbering	33	34	35	36	37	38	39	40
									α value	-0.0176	-0.0291	-0.0037	-0.0298	-0.0170	-0.0298	-0.0556	-0.0181
Numbering	41	42	43	44	45	46	47	48
									α value	-0.0877	-0.0483	-0.0931	-0.0155	-0.0058	-0.0016	-0.0614	-0.0424
Numbering	49	50	51	52	53	54	55	56
									α value	-0.0502	-0.0119	-0.0169	-0.0148	-0.0175	-0.0135	-0.0154	-0.0252
Numbering	57	58	59	60	61	62	63	64
									α value	-0.0104	-0.0023	-0.0165	-0.0043	-0.0295	-0.0050	-0.0079	-0.0054
Numbering	65	66	67	68	69	70	71	72
									α value	-0.0227	-0.0521	-0.0180	-0.0120	-0.1016	-0.0785	-0.0190	-0.0750
Numbering	73	74	75	76	77	78	79	80
									α value	-0.0537	-0.0089	-0.0386	-0.0998

B3, choose α ^*a positive component, and calculated threshold accordingly:

$b^{*}=y_i-\underset{i=1}{\overset{78}{\&Sigma;}}{y}_i{\mathrm{\&alpha;}}_i^{*}\exp \{-\frac{|x_i-x_j{|}^2}{{0.007}^2}\},$ Obtain b ^*=0.8821.

The decision function of b4, conformation identification magnetizing inrush current and fault current

$\begin{array}{c}f\left(x\right)=sgn\left(g\right(x)),\\ =\mathrm{sgn}\left(\underset{i=1}{\overset{76}{\&Sigma;}}{y}_i{\mathrm{\&alpha;}}_i^{*}\exp \right\{-\frac{|x_i-x{|}^2}{{0.007}^2}\}+0.8821)\end{array},$

Only have one 7 in formula and tie up variable x the unknown, when input feature vector amount once confirm, namely this formula available carries out the identification of magnetizing inrush current.

C, after transformer has an accident, the data collection of protective device acquisition system come calculate in the decision function brought into after seven kinds of input feature vector values in b4 through step a, if

For making this magnetizing inrush current identical criterion will be more accurate, need constantly to revise newly-increased sample.Above-mentioned magnetizing inrush current decision function f (x) is only determined by 76 support vectors trained, if there is newly-increased training sample, no matter derive from dynamic simulation test or actual motion, marked by optimum sample labeling method, retain the sample larger to classification results, remainingly to be rejected, to reach minimizing amount of calculation, do not affected the object of recognition effect simultaneously.Optimum labelling method performing step is as follows:

Sample will be increased newly for people's factor of influence (ES) estimation formulas:

$\begin{array}{c}ES\left(x\right)=0.5+\frac{1}{1+e^{-\left|g\left(x\right)\right|}}\\ =0.5+\frac{1}{1+e^{-\left|\underset{i=1}{\overset{76}{\&Sigma;}}{y}_i{\mathrm{\&alpha;}}_i^{*}\exp \{-\frac{|x_i-x{|}^2}{{0.007}^2}\}0.8821\right|}}\end{array}$

Its ES, more close to 1, illustrates that the amount of information that it contains is larger, and its possibility belonging to optimal sample is larger, and such sample is retained.Preferably, the sample that ES value is positioned at 0.8-1.0 is retained.

Test below, to illustrate whether this decision function f (x) has higher discrimination.

The sample choosing various running state of transformer in addition carries out the test of accuracy.Test result is as shown in Table 3:

Table three test result

Can obtain after tested, the method well can identify no-load switchon surge, general internal fault and drop in alternate and earth fault, and also can well identify for the turn-to-turn short circuit of dropping in more than 10%, correct recognition rata reaches 96.2%.

The above has made detailed description to the preferred embodiment of the present invention and embodiment by reference to the accompanying drawings; but the present invention is not limited to the above-described embodiment and examples; for those skilled in the art; without departing from the inventive concept of the premise; can also make some improvement and modification, these improve and modification also should be considered as protection scope of the present invention.

Claims (2)

1. based on an excitation flow recognition method for support vector classification, it is characterized in that, the method comprises the following steps:

The input feature vector amount of a, selection SVMs;

The content of second harmonic, triple-frequency harmonics in a1, collection three-phase differential current, with discriminant equation $\left\{\begin{array}{c}{I}_{2nd}>K_{2xb}*I_{1st}\\ I_{3rd}>K_{3xb}*I_{1st}\end{array}\right.$ Based on, order $\left\{\begin{array}{c}{K}_2=I_{2nd}/I_{1st}\\ K_3=I_{3nd}/I_{1st}\end{array}\right.,$ By K ₂as first characteristic quantity input of SVMs, by K ₃as second characteristic quantity input of SVMs; Wherein, I _2ndfor the second harmonic in every phase differential current; I _3rdfor the triple-frequency harmonics in every phase differential current; I _1stfor the difference stream first-harmonic of corresponding phase, k _2xbfor secondary harmonic brake coefficient setting value; k _3xbfor third harmonic restrained coefficient setting value;

Flow into the differential current of differential relay when a2, collection power transformer interior fault, take absolute value again after differential, | i ' ₂|, with | i ' ₂| < KI ' _m2angle θ corresponding to duration as the 3rd characteristic quantity input of SVMs, wherein I ' _m2for | i ' ₂| amplitude, K=0.25;

A3, gather four the characteristic quantity input of the wide Width=of ripple (t/T) * 360 as SVMs of magnetizing inrush current;

A4, gather five the characteristic quantity input of integrated value S as SVMs of differential current, wherein wherein i (k) represents the value of a kth sampled point, and N represents the sum of a cycle sampled point, and Ts represents the time difference of adjacent two sampled points;

A5, half cycles before differential current is set to x (t), later half cycle negate is also set to y (t), and the two does correlation analysis, adopts waveform correlation coefficient J as the 6th feature input variable of SVMs, wherein k wherein in x (k) and y (k) represents the number of sampled point, and N represents the sum of a cycle sampled point;

A6, using seven the characteristic quantity input of the change degree σ Z of measurement impedance Z (t) as SVMs, wherein k represents the number of sampled point, and N represents the sum of a cycle sampled point, and Ts represents the time difference of adjacent two sampled points, and T represents all wave period;

B, train the various running statuses of transformer, the decision function of conformation identification magnetizing inrush current and fault current, described training step is:

B1, obtain the true dynamic model data under transformer various running status by dynamic simulation test, and draw seven input feature vector values in step a accordingly,

If x _i∈ χ=R ⁷, x _i=(K _2i, K _3i, θ _i, W _i, S _i, J _i, σ Z _i) ^t, y _i∈ γ=1 ,-1},

Y during magnetizing inrush current _i=1; Y during fault _i=-1;

Then negative class sample is expressed as:

T ₁＝{((K _2,1,K _3,1,θ ₁,W ₁,S ₁,J ₁,σZ ₁) ^T,-1),···

(x _a,-1)}∈(χ×γ) ^a

Wherein, a is negative class sample number;

Positive class sample is expressed as:

T ₂＝{((K _2,a+1,K _3,a+1,θ _a+1,W _a+1,S _a+1,J _a+1,σZ _a+1) ^T,+1),

···(x _a+b,+1)}∈(χ×γ) ^b

Wherein, a is negative class sample number, and b is positive class sample number;

B2, introducing kernel function if σ=0.07, C=1, structure optimum programming equation, brings optimum ruleization equation into by parameter

$\underset{{}_{\mathrm{\&alpha;}}}{\min }\frac{1}{2}\underset{i=1}{\overset{a+b}{\&Sigma;}}\underset{j=1}{\overset{a+b}{\&Sigma;}}{y}_i{y}_j{\mathrm{\&alpha;}}_i{\mathrm{\&alpha;}}_j\exp \left\{-\frac{\text{|}{x}_i-x_j{\text{|}}^{\text{2}}}{{\mathrm{\&sigma;}}^2}\right\}-\underset{j=1}{\overset{a+b}{\&Sigma;}}{\mathrm{\&alpha;}}_j;$

$\underset{i=1}{\overset{a+b}{\&Sigma;}}{y}_i{\mathrm{\&alpha;}}_j=0;$

0≤α _i≤C,i＝1,···,a+b；

Obtain optimal solution ${\mathrm{\&alpha;}}^{\text{*}}{\left({\mathrm{\&alpha;}}_{\text{2}}^{\text{*}},...{\mathrm{\&alpha;}}_l^{*}\right)}^T;$

B3, choose α ^*a positive component calculated threshold $b^{*}=y_i-\underset{i=1}{\overset{l}{\&Sigma;}}{y}_i{\mathrm{\&alpha;}}_i^{*}K(x_i\&CenterDot;x_j);$

B4, structure decision function $f\left(x\right)=sgn\left(\underset{i=1}{\overset{1}{\&Sigma;}}{y}_i{\mathrm{\&alpha;}}_i^{*}K\right(x\&CenterDot;x_i)+b^{*});$

C, after transformer has an accident, the data collection of protective device acquisition system come calculate in the decision function brought into after seven kinds of input feature vector values in b4 through step a, if

2. the excitation flow recognition method based on support vector classification according to claim 1, is characterized in that, in step b1, if there is newly-increased training sample, then will increase sample newly for people's factor of influence (ES) estimation formulas:

$\begin{array}{c}ES\left(x\right)=0.5+\frac{1}{1+e^{-\left|g\left(x\right)\right|}}\\ =0.5+\frac{1}{1+e^{-|\underset{i=1}{\overset{76}{\&Sigma;}}{y}_i{\mathrm{\&alpha;}}_i^{*}\exp \left\{-\frac{|x_i-x{|}^2}{{0.007}^2}\right\}+0.8821|}}\end{array};$

The sample of ES value between 0.8-1.0 is retained.