CN102063626A - Power quality disturbance mode discrimination method - Google Patents

Abstract

Power quality disturbance mode discrimination means monitoring on the current or voltage of certain nodes in a power grid. When abnormal phenomena (such as surge, dip, harmonic waves, flickers, etc. of power or voltage) occur, the types of the abnormal phenomena can be quickly discriminated, and furthermore, aid decision is provided for power grid operation management. The invention discloses a power quality disturbance mode discrimination method, and a feature selection method based on evolutionary computation is provided in the power quality disturbance mode discrimination method, aiming at the characteristics of high dimensionality and great information quantity of a feature vector extracted from an electrical signal. By applying the feature selection method, a large amount of redundant features in the feature vector can be eliminated. A support vector machine based mode discrimination method is provided to discriminate the mode of the extracted features of the electrical signal. In order to increase the robustness of the system, a data fusion method is provided to carry out data fusion on mode discrimination results of a plurality of support vector machines, and the final mode discrimination result is obtained. In the invention, evolutionary computation and the support vector machine technology in the field of computational intelligence are combined, the speed of power quality disturbance mode discrimination can be increased, and the accuracy of mode discrimination is increased.

Description

A kind of electrical energy power quality disturbance pattern discrimination method

Technical field

The present invention relates to a kind of electrical energy power quality disturbance pattern discrimination method, belong to computer utility and automation field.

Background technology

(Power Quality Disturbances PQD) is meant in voltage in the network system, the electric current and deviation occurs electrical energy power quality disturbance, thereby causes subscriber equipment to break down or the electrical problems of operation exception.It will cause direct or indirect adverse effect to electric company, user and consumer.Under the ideal state, submit to user's voltage or electric current and should have sinusoidal waveform, stable frequency, stable amplitude and stable phasing degree.Yet the nonlinear load of some in the electrical network as computing machine, governor motor etc., all can bring influence to electrical network, produces problems such as the rising sharply of voltage or electric current, rapid drawdown, harmonic wave, concussion.The consequently decline of the quality of power supply, waste of electric energy or electricity charge over-expense, even can cause damage to other electric equipment in the electrical network.

In view of the harmfulness of electrical energy power quality disturbance, in network system, should be real-time the quality of power supply is monitored.When the electrical energy power quality disturbance problem took place, system can discern it and classify apace, and then took measures to reduce the negative effect that quality disturbance brings.Carry out power quality analysis, at first will carry out feature extraction, then the proper vector of extracting is classified, finally obtain classification results the detected electric signal of watch-dog.Existing feature extracting method comprises small wave converting method, Fourier transformation method, t-transform method etc.Yet characteristics such as the proper vector of extracting has the dimension height, contain much information, and often contain a large amount of redundancies and noise characteristic value in the proper vector from electric signal.Therefore utilize the feature of all extracting that electric signal is classified, shortcoming such as it is long often to have an elapsed time, and classification accuracy is low.At the problems referred to above, the present invention proposes a kind of feature selecting model based on evolutionary computation.In order to classify, designed a kind of many support vector machine classifiers, after electric signal is carried out feature extraction and feature selecting, the feature after simplifying is submitted to a plurality of support vector machine classifiers.In order to improve the robustness of system, a kind of data fusion method has been proposed, the classification results of a plurality of support vector machine is carried out data fusion, obtain final classification results.Because the evolutionary computation method that is proposed can be eliminated bulk redundancy and noise information in the proper vector, so this method can reduce the training time and the recognition time of support vector machine.Because designed data fusion method can be assessed the classification results of many support vector machine, so this method can be guaranteed the accuracy of identification of system.

Summary of the invention

Fundamental purpose of the present invention provides a kind of electrical energy power quality disturbance recognition methods based on evolutionary computation and support vector machine.At the proper vector of extracting from electric signal have the dimension height, characteristics such as contain much information, designed a kind of feature selection approach based on evolutionary computation.In the evolutionary computation method, proposed a kind of based on binary-coded feature selecting scheme method for expressing, and proposed a kind of based on Ba Ta just in the candidate scheme evaluation method of inferior distance (Bhattacharyya distance).Optimized Algorithm can be chosen optimized Algorithm commonly used in the evolutionary computation, as genetic algorithm, Artificial Immune Algorithm, particle cluster algorithm etc.Feature extracting method based on evolutionary computation can be eliminated redundancies a large amount of in the proper vector and noise information, can reduce the training and the recognition time of sorter; For finishing classification task, adopt a plurality of support vector machine to classify; Because the randomness of support vector machine classification, the sorting technique of a plurality of support vector machine might not be in full accord, therefore, designed a kind of data anastomosing algorithm.This method can improve the robustness of sorter, improves the accuracy rate of classification.

The present invention is achieved by the following technical solutions:

1. feature extraction.Electric signal from monitoring equipment extracts extracts feature with two kinds of technology, and a kind of is wavelet transformation technique (Wavelet Transform).Classical wavelet transformation is defined as:

$\mathrm{Wf}(a,b)=<f\left(t\right),{\left|a\right|}^{-\frac{1}{2}}\mathrm{\&psi;}\left(\frac{t-b}{a}\right)>=\&Integral;{\left|a\right|}^{-\frac{1}{2}}\mathrm{\&psi;}\left(\frac{t-b}{a}\right)f\left(t\right)\mathrm{dt}---\left(1\right)$

Wherein, ψ (t) is basic small echo and satisfies enabled condition

With non-zero condition ω ≠ 0, parameter b plays phorogenesis in the basic small echo, and parameter alpha changes the size of window function.By signal is carried out wavelet transformation, can obtain the detailed information of signal

And the equal value information of afterbody decomposition

When the present invention adopts wavelet transformation technique that electric signal is carried out feature extraction, choose the db6 small echo, at first continuous signal is sampled, then sample sequence is carried out three grades of wavelet decomposition of one dimension, obtain detailed information as basic small echo With equal value information

Then with

As proper vector, that is:

$V=[{\stackrel{\&OverBar;}{a}}_3,{\stackrel{\&OverBar;}{d}}_3,{\stackrel{\&OverBar;}{d}}_2,{\stackrel{\&OverBar;}{d}}_1]---\left(2\right)$

Another kind of Feature Extraction Technology is double frequency-band demodulation techniques (Double Side-Band Demodulation).The double frequency-band demodulation techniques can be expressed as with formula

DSB (s (kT))=filter (s (kT) sin (ω kT)) (3) wherein ω is the angular frequency of electric signal under the perfect condition, and filter () is a wave filter.In the present invention, adopt three rank Butterworth filters (3-order Butterworth filter), cut frequency is set to 30Hz.

2. based on the feature selecting of evolutionary computation.When the method for utilizing evolutionary computation was carried out feature selecting, each candidate scheme all was encoded into one " individuality ", and several individualities have constituted " colony ".When methods such as utilizing genetic algorithm, Artificial Immune Algorithm is optimized operation, always produce some individualities (being initial solution) at random.According to designed objective function each individuality is assessed, provided fitness value.Based on this fitness value, select individuality to duplicate the next generation.The individuality that chooses then is combined into a new generation through Optimizing operation.So iteration produces until optimum solution.Utilize the evolutionary computation method to carry out the feature screening, relate to two problems, one of them is that candidate scheme is encoded, and another problem is the design of objective function.The present invention designed a kind of based on binary-coded method and a kind of based on Ba Ta just in the fitness function computing method of inferior distance.

(1) based on binary coding method.A proper vector to be screened S feature to be screened arranged.Adopt the binary coding mode,

Its structure is as follows:

c＝[r ₁，r ₂，…，r _s]

R wherein _i∈ 0,1}, (1≤f≤s), r _i=1 expression character pair v _iScreened, r _i=0 expression character pair v _iNot screened.Article one, coding represent a candidate scheme, and the synoptic diagram of constructing new proper vector by a candidate scheme as shown in Figure 1.

(2) based on Ba Ta just in the fitness function of inferior distance.Two factors are considered in the design of fitness function:

The first minimize vector v ' dimension, i.e. Minimize dim; It two is the difference of maximization between all kinds of electrical energy power quality disturbance sample of signal collection, promptly Maximize ∑ dis (i, j), (i j) is the difference measurement function of sample set i and sample set j to dis; For first factor, but the number of gene 1 in the statistical coding.For second factor, we adopt Ba Ta just in inferior distance (Bhattacharyya Distance) distance estimate difference between the two class sample sets.Be provided with two class sample sets:

${\mathrm{\&Omega;}}_i=\left|\begin{array}{cccc}{\stackrel{\&RightArrow;}{\mathrm{\&omega;}}}_1^i& {\stackrel{\&RightArrow;}{\mathrm{\&omega;}}}_2^i& \&CenterDot;\&CenterDot;& {\stackrel{\&RightArrow;}{\mathrm{\&omega;}}}_m^i\end{array}\right|=\left|\begin{array}{cccc}{\mathrm{\&omega;}}_{11}^i& {\mathrm{\&omega;}}_{21}^i& \&CenterDot;\&CenterDot;& {\mathrm{\&omega;}}_{m1}^i\\ {\mathrm{\&omega;}}_{12}^i& {\mathrm{\&omega;}}_{22}^i& \&CenterDot;\&CenterDot;& {\mathrm{\&omega;}}_{m2}^i\\ \&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\\ {\mathrm{\&omega;}}_{1n}^i& {\mathrm{\&omega;}}_{2n}^i& \&CenterDot;\&CenterDot;& {\mathrm{\&omega;}}_{\mathrm{mn}}^i\end{array}\right|$

${\mathrm{\&Omega;}}_j=\left|\begin{array}{cccc}{\stackrel{\&RightArrow;}{\mathrm{\&omega;}}}_1^j& {\stackrel{\&RightArrow;}{\mathrm{\&omega;}}}_2^j& \&CenterDot;\&CenterDot;& {\stackrel{\&RightArrow;}{\mathrm{\&omega;}}}_m^j\end{array}\right|=\left|\begin{array}{cccc}{\mathrm{\&omega;}}_{11}^j& {\mathrm{\&omega;}}_{21}^j& \&CenterDot;\&CenterDot;& {\mathrm{\&omega;}}_{m1}^j\\ {\mathrm{\&omega;}}_{12}^j& {\mathrm{\&omega;}}_{22}^j& \&CenterDot;\&CenterDot;& {\mathrm{\&omega;}}_{m2}^j\\ \&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\\ {\mathrm{\&omega;}}_{1n}^j& {\mathrm{\&omega;}}_{2n}^j& \&CenterDot;\&CenterDot;& {\mathrm{\&omega;}}_{\mathrm{mn}}^j\end{array}\right|$

Wherein m is the capacity of sample set, and n is the dimension of each sample in the sample set.Note:

$M_i=\left|\begin{array}{c}{M}_1^i\\ M_2^i\\ \&CenterDot;\&CenterDot;\\ M_n^i\end{array}\right|=\left|\begin{array}{c}\frac{1}{m}\underset{k=1}{\overset{m}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_{k1}^i\\ \frac{1}{m}\underset{k=1}{\overset{m}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_{k2}^i\\ \&CenterDot;\&CenterDot;\\ \frac{1}{m}\underset{k=1}{\overset{m}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_{\mathrm{kn}}^i\end{array}\right|$ With ${\mathrm{\&Sigma;}}_i=\left|\begin{array}{cccc}{({\mathrm{\&omega;}}_{11}^i-M_1^i)}^2& 0& \&CenterDot;\&CenterDot;& 0\\ 0& {({\mathrm{\&omega;}}_{22}^i-M_2^I)}^2& \&CenterDot;\&CenterDot;& 0\\ \&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\\ 0& 0& \&CenterDot;\&CenterDot;& {({\mathrm{\&omega;}}_{\mathrm{nn}}^i-M_n^i)}^2\end{array}\right|$

Then the Ba Ta between two class sample set i and the j just in inferior distance available below formula calculate:

$B_{\mathrm{ij}}=\frac{1}{8}{(M_i-M_j)}^t{\left\{\frac{{\mathrm{\&Sigma;}}_i-{\mathrm{\&Sigma;}}_j}{2}\right\}}^{-1}(M_i-M_j)+\frac{1}{2}\ln \left\{\frac{\left|\frac{{\mathrm{\&Sigma;}}_i+{\mathrm{\&Sigma;}}_j}{2}\right|}{{\left|{\mathrm{\&Sigma;}}_i\right|}^{1/2}{\left|{\mathrm{\&Sigma;}}_j\right|}^{1/2}}\right\}---\left(3\right)$

Two, one candidate code [r of Consideration one and factor ₁, r ₂..., r _s] fitness function can calculate with following formula:

$f\left(\right[{\mathrm{<figure-callout\; id="1"\; label="r"\; filenames="HSA00000401791600012.png"\; state="\{\{state\}\}">r</figure-callout>}}_1,r_2,\&CenterDot;\&CenterDot;\&CenterDot;,r_s\left]\right)=\underset{i\&NotEqual;j}{\mathrm{\&Sigma;}}{B}_{\mathrm{ij}}/n---\left(4\right)$

Wherein n is coding [r ₁, r ₂..., r _s] in 1 number.

3. classify based on the signal of support vector machine.The basic thought that support vector machine is found the solution classification problem is: at first select a Nonlinear Mapping, the input space is transformed to a high-dimensional feature space, in this higher dimensional space, utilize structural risk minimization, the structure optimal decision function is sought the nonlinear relationship between input variable and the output variable.In order to improve the robustness of algorithm, the present invention adopts a plurality of support vector machine that electric signal is classified.Set up m multi-category support vector machines, the proper vector after simplifying is submitted to each support vector machine respectively, obtain m classification results by this m multi-category support vector machines afterwards.

4. data anastomosing algorithm.At electrical energy power quality disturbance classification, to an input signal, classify respectively by this m sorter, obtain m classification results after, adopt a kind of temporal voting strategy, that is categorized as the classification results of this sorter the who gets the most votes.

5. based on the electrical energy power quality disturbance sorting algorithm flow process of evolutionary computation and support vector machine.Electrical energy power quality disturbance sorting technique structural representation proposed by the invention as shown in Figure 2.

Description of drawings

The synoptic diagram that Fig. 1 screens proper vector for the candidate scheme coding

Fig. 2 is an electrical energy power quality disturbance pattern discrimination method synoptic diagram

Embodiment

Below concrete steps of the present invention are summarized:

Step (1): the electric signal that the monitoring point is provided carries out feature extraction;

Step (2): utilize the method for evolutionary computation to carry out feature selecting;

Step (3):, utilize a plurality of support vector machine classifiers that electric signal is classified based on the selected feature that arrives in the step (2);

Step (4): the classification results to a plurality of support vector machine carries out data fusion;

Step (5): output category result;

Step (6):, then change step (8) if the classification results shows signal is normal; Otherwise change step (7);

Step (7):, start corresponding quality disturbance handling procedure according to classification results;

Step (8): judge whether classification finishes.If finish, then output monitoring report; Otherwise change step (1), next group signal is monitored again.

Claims (8)

1. an electrical energy power quality disturbance pattern discrimination method comprises the steps: at least

Step (1): the electric signal that the monitoring point is provided carries out feature extraction;

Step (2): utilize the method for evolutionary computation to carry out feature selecting;

Step (3):, utilize a plurality of support vector machine classifiers that electric signal is classified based on the selected feature that arrives in the step (2);

Step (4): the classification results to a plurality of support vector machine carries out data fusion;

Step (5): output category result;

Step (6):, then change step (8) if the classification results shows signal is normal; Otherwise change step (7);

Step (7):, start corresponding quality disturbance handling procedure according to classification results;

Step (8): judge whether classification finishes, if finish, then output monitoring report; Otherwise change step (1), next group signal is monitored again.

2. according to claim 1 described a kind of electrical energy power quality disturbance pattern discrimination method, it is characterized in that: the electric signal that the monitoring point is provided carries out feature extracting method small echo method of changing, double frequency-band demodulation method.

3. according to claim 1 described a kind of electrical energy power quality disturbance pattern discrimination method, it is characterized in that: utilize the method for evolutionary computation that the feature of being extracted is optimized selection.

4. according to claim 1 described a kind of electrical energy power quality disturbance pattern discrimination method, it is characterized in that: in the method for evolutionary computation, the candidate feature selection scheme is encoded with binary vector;

5. according to claim 1 described a kind of electrical energy power quality disturbance pattern discrimination method, it is characterized in that: in the evolutionary computation method, inferior distance was come the quality of evaluate candidate scheme in employing Ba Ta was proper.

6. according to claim 1 described a kind of electrical energy power quality disturbance pattern discrimination method, it is characterized in that: the method for evolutionary computation can adopt genetic algorithm, particle swarm optimization algorithm, Artificial Immune Algorithm.

7. according to claim 1 described a kind of electrical energy power quality disturbance pattern discrimination method, it is characterized in that: utilize a plurality of support vector machine to carry out repeatedly Classification and Identification, obtain a plurality of classification results.

8. according to claim 1 described a kind of electrical energy power quality disturbance pattern discrimination method, it is characterized in that: utilize data anastomosing algorithm that the classification results of a plurality of support vector machine is carried out data fusion, come the state of electric signal is judged.

Images