CN107340456B - Power distribution network operating condition intelligent identification Method based on multiple features analysis - Google Patents

Abstract

The invention discloses an intelligent identification method of distribution network working conditions based on multi-feature analysis, which collects data on abnormal operating conditions of distribution network feeders through an online wave recording system of distribution network, and applies various feature extraction methods, including time Domain, frequency domain and wavelet transform signal feature extraction methods, extract a large number of signal features, such as current transient steady-state features, etc.; through training artificial neural network (ANN) model with adaptive learning characteristics, each type of abnormal working condition Identify, and establish a classification and identification process in the form of a decision tree to realize the effective identification of multiple working conditions in the distribution network.

Description

Intelligent identification method of distribution network working conditions based on multi-feature analysis

technical field

The invention relates to the field of abnormal state identification of distribution network operating conditions.

Background technique

The distribution network is an important part of the power system and a key link to ensure the quality of power supply and the efficient operation of the power grid. In order to ensure the highly intelligent operation of the distribution network, real-time monitoring of feeder operation data, timely warning of abnormal conditions, and rapid fault detection and processing are required. Among them, the identification of abnormal working conditions of the feeder is an important function of the intelligent distribution network. There are many reasons for the abnormal operation of distribution network feeders, such as phase-to-phase short circuit, single-phase grounding, excitation inrush current and so on.

The traditional recognition of abnormal feeder operation in distribution network is mostly based on setting thresholds and a certain logical relationship of a single factor. This identification mode is simple, it is difficult to consider the feeder voltage and current signals obtained under abnormal working conditions, and it is also affected by various random factors such as system operation mode, fault location, transition impedance and fault time, so there are certain defects. For the working conditions where the current changes slightly instantaneously and the characteristic quantity changes are not obvious, such as single-phase large resistance instantaneous grounding, etc., so far, the distribution system equipped with fault indicators and fault line selection technology is still unable to realize these working conditions. monitoring and identification.

Contents of the invention

The purpose of the present invention is to propose a method for intelligent identification of distribution network working conditions based on multi-feature analysis, through the distribution network online wave recording system to collect data on the abnormal operating conditions of distribution network feeders, and to apply a variety of feature extraction methods , including signal feature extraction methods in time domain, frequency domain and wavelet transform, extracting a large number of signal features, such as current transient steady state features, etc.; each type is performed by training an artificial neural network (ANN) model with adaptive learning characteristics Abnormal working conditions are identified, and a classification and identification process shaped like a decision tree is established to realize effective identification of multiple working conditions in the distribution network.

The technical solution adopted for realizing the object of the present invention is such, a kind of distribution network working condition intelligent identification method based on multi-feature analysis, it is characterized in that, comprises the following steps:

1) Through the online wave recording system of the distribution network, when the line fails, the three-phase synchronous wave recording is triggered to obtain the A, B, and C three-phase current wave recording signals and zero-sequence current signals.

2) Using multi-feature extraction method to extract signal features:

2-1) Extract time domain features

In the scope of time domain, each feature quantity shown in Table 1 is extracted from the current signal collected by the online wave recording system.

Table 1 Each feature component extracted in the time domain

Among them, I _p,i is the current signal collected by the detector, Indicates the three phases of A, B, and C, p represents the four phases of A, B, C, and Z, i represents the periodic sequence of the wave recording signal, j represents the number of sampling points in each cycle, and N is the number of sampling points.

2-2) Extract frequency domain features

In the frequency domain, for the DC and second harmonic components of the steady-state signal after the fault occurs, the discrete Fourier transform analysis is performed on i~(i+m) period data of the recorded wave signal. Taking a periodic current signal as an example, use Fourier series expansion to obtain the frequency domain transformation result of the second harmonic:

In the formula, _Ip,i is the current signal collected by the detector, p represents the four phases of A, B, C and Z, i is the periodic sequence of the current signal, N is the number of sampling points, n is the nth sampling point, j is the imaginary part indicator.

The characteristic quantities shown in Table 2 are sequentially extracted: DC component content I _dp , second harmonic component I _2x and second harmonic component content I _2xp .

Table 2 Each feature component extracted in the frequency domain

2-3) Analysis of transient signals by feature extraction method based on wavelet transform

Using wavelet transform to extract the transient signal at the time of abnormal operation of distribution network feeder. The original transient signal is decomposed into J different scales for analysis, and the low frequency and high frequency components of multiple frequency bands are extracted.

2-3-1) Decompose the collected abnormal signal, and extract the high-frequency component characteristics of the abnormal working condition signal.

Among them, A _J (k) is the low-frequency component coefficient obtained by J-order wavelet decomposition and reconstruction of the signal at time k, and D _i (k) is the ith-order high-frequency component coefficient. To unify the expression, replace A _J (k) with D _J+1 (k), and convert the expression to

The high-frequency component characteristic of the abnormal working condition signal is Df _p (i), and the statistical object is the sum of the absolute value of the high-frequency component coefficients of the half period before and after the abnormal signal appears, expressed as

2-3-2) Extract wavelet energy entropy and wavelet singular entropy

Wavelet energy entropy and wavelet singular entropy are used to represent the degree of chaos in the distribution of signal energy in different frequency bands during the period when abnormal conditions occur.

The wavelet energy entropy WEE is defined as follows:

Where p _i =E _i /E defines the signal energy spectrum at different scale i time k, E _i (k)=|D _i (k)| ² , is the energy sum of all moments on scale i, is approximately the total energy of the signal.

The coefficients D _i (k) reconstructed by the wavelet transform form a (J+1)×M matrix D _(J+1)×M , and the singular value decomposition of the matrix can obtain J+1 non-negative singular value σ _i , then the wavelet singular entropy WAE is defined as follows:

3) Perform tree clustering on the feature data set extracted according to step 2), and establish a multi-category recognition process in the form of a decision tree. Among them, short-circuit and grounding belong to fault conditions. In all working conditions, the two types of working conditions, short-circuit and grounding, need to follow the principle of "can be missed but not misjudged". First, identify these two working conditions. Missed judgment is to judge the original short-circuit (grounding) working condition as another working condition type, and misjudgment is to classify abnormal working conditions that do not belong to short-circuit (grounding) as fault working conditions.

4) The classifier in the multi-class recognition process is constructed using a three-layer ANN model.

4-1) Use feature data and abnormal operating condition categories to form a training data set, and train the ANN classifier. The ANN classifier uses a three-layer feed-forward neural network. The activation function of all neurons in the input layer and hidden layer is set to the tan-sig function, and the activation function of the output layer is set to the log-sig function. For binary classification recognition, classification will be required The actual marking types of the working condition set are 1 and 0, that is, an abnormal working condition and other kinds of working conditions that need to be identified.

4-2) Set the performance function of the model to mean square error:

Wherein, the cost function is equal to cost(h _w ( _xi ), y _i )=(h _w ( _xi )-y _i ) ² . X=(x ₁ ,x ₂ ,..., _xi ) is the input matrix, each column x _i is a set of input working condition parameters, h _w ( _xi ) is the output of the i-th input of the ANN model, y is the type of known labeled working conditions, and M is the number of training data sets.

4-3) In the coefficient training of ANN classifier 1 (grounding) and classifier 2 (short circuit), the cost function needs to follow the principle of "missing judgment but not misjudgement" for grounding and short circuit faults, adding a weight factor K to the function (K>1), the cost function becomes

cost(h _w ( _xi ),y _i )＝y _i ×(h _w ( _xi )-1) ² +K(1-y _i )×h _w ( _xi ) ²

5) After the multi-classification recognition model has been trained and tested, and the error is controlled within the allowable threshold range, it can be provided to the online wave recording system for multi-working condition recognition of the distribution network.

The technical effect of the present invention is unquestionable: firstly, according to the distribution network on-line wave recording system, the feeder abnormal operating condition data is collected, and the signal features are extracted by using various feature extraction methods such as time domain, frequency domain and wavelet transform, and sent to Into the ANN classifier for learning and training, and finally with the ANN classifier as the node to establish a distribution network multi-working condition intelligent identification model in the form of a decision tree, which can effectively identify abnormal working conditions of the distribution network.

Description of drawings

Figure 1 Recorded wave current under excitation inrush current condition;

Figure 2 Multi-type working condition identification process;

Fig. 3 Identification results of grounding conditions;

Figure 4. Identification results of short-circuit conditions.

Detailed ways

The present invention will be further described below in conjunction with the examples, but it should not be understood that the scope of the subject of the present invention is limited to the following examples. Without departing from the above-mentioned technical ideas of the present invention, various replacements and changes made according to common technical knowledge and conventional means in this field shall be included in the protection scope of the present invention.

A method for intelligent identification of distribution network operating conditions based on multi-feature analysis, characterized in that it includes the following steps:

1) Through the online recording system of the distribution network, when the line fails, the three-phase synchronous recording is triggered to obtain the current recording signals of each phase and the zero-sequence current signal. Taking the recorded wave current data of the excitation inrush current condition in Figure 1 as an example, it includes A, B, C three-phase current and zero-sequence (Z) current. Each phase current has 16 cycles in total, the sampling rate is 4kHz, and each cycle has a sampling point The number is 82;

2) Using multi-feature feature extraction method to extract signal features

2-1) Extract time domain features

In the time domain, the current signals collected by the online wave recording system are extracted according to the calculation method in Table 1.

Table 1 Each feature component extracted in the time domain

In the above table, I _p,i is the current signal collected by the detector, Indicates the three phases of A, B, and C, p represents the four phases of A, B, C, and Z, i represents the 1 to 16 cycles of the wave recording signal, j represents 1 to 82 sampling points in each cycle, and N is the sampling point The number is 82.

2-2) Extract frequency domain features

In the frequency domain, for the DC and second harmonic components of the steady-state signal after the fault occurs, the discrete Fourier transform analysis is performed on the 8th to 10th cycle data of the recorded wave signal.

Taking a periodic current signal as an example, use Fourier series expansion to obtain the frequency domain transformation result of the second harmonic:

In the formula, _Ip,i is the current signal collected by the detector, p represents the four phases of A, B, C and Z, i is the periodic sequence of the current signal, N is the number of sampling points, n is the nth sampling point, j is the imaginary part indicator.

According to the calculation method in Table 2, feature quantities are extracted sequentially: DC component content I _dp , second harmonic component I _2x and second harmonic component content I _2xp .

Table 2 Each feature component extracted in the frequency domain

2-3) The feature extraction method based on wavelet transform divides the transient signal

The original transient signal is decomposed into J different scales for analysis, and the low frequency and high frequency components of multiple frequency bands are extracted. Using wavelet transform to extract the transient signal at the time of abnormal operation of distribution network feeder.

2-3-1) Decompose the collected abnormal signal, and extract the high-frequency component characteristics of the abnormal working condition signal.

Among them, A _J (k) is the low-frequency component coefficient obtained by J-order wavelet decomposition and reconstruction of the signal at time k, and D _i (k) is the ith-order high-frequency component coefficient. To unify the expression, replace A _J (k) with D _J+1 (k), and convert the expression to

The high-frequency component characteristic of the abnormal working condition signal is Df _p (i), and the statistical object is the sum of the absolute value of the high-frequency component coefficients of the half period before and after the abnormal signal appears, expressed as

In order to effectively decompose the transient signal of the abnormal working condition of the distribution network feeder, the 2nd to 4th order wavelet transform is generally used, and the 3rd order db5 wavelet transform is used here.

2-3-2) Extract wavelet energy entropy and wavelet singular entropy

Wavelet energy entropy and wavelet singular entropy are used to represent the degree of chaos in the distribution of signal energy in different frequency bands during the period when abnormal conditions occur.

The wavelet energy entropy WEE is defined as follows,

Where p _i =E _i /E defines the signal energy spectrum at different scale i time k, E _i (k)=|D _i (k)| ² , is the energy sum of all moments on scale i, is approximately the total energy of the signal.

The coefficients D _i (k) reconstructed by the wavelet transform form a (J+1)×M matrix D _(J+1)×M , and the singular value decomposition of the matrix can obtain J+1 non-negative singular value σ _i , then the wavelet singular entropy WAE is defined as follows,

According to engineering experience, in order to be able to show the degree of chaos of the transient signal when the abnormal signal appears, in the calculation of wavelet energy entropy and wavelet singular entropy, the section of the signal used is the 10 sample points before and after the abnormal signal is detected.

3) Perform tree clustering on the feature data set extracted according to step 2), and establish a multi-category recognition process in the form of a decision tree. The categories of abnormal working conditions currently designed mainly include grounding, short circuit, inrush current, lightning strike, power failure and other categories. As shown in Figure 2, the classifiers in the recognition process are all binary classifiers. Among them, short-circuit and grounding belong to the fault working conditions. In all working conditions, the principle of "can be missed but not wrongly judged" should be followed for the two types of working conditions of short-circuit and grounding. For other types of working conditions, misjudgment is to classify abnormal working conditions that do not belong to short circuit (grounding) as fault working conditions, and firstly identify these two working conditions.

4) The classifier in the multi-class recognition process is constructed using a three-layer feed-forward artificial neural network (ANN) model.

4-1) Use feature data and abnormal operating condition categories to form a training data set, and train the ANN classifier. The ANN classifier adopts a three-layer network topology, the activation function of all neurons in the input layer and the hidden layer is set to the tan-sig function, and the activation function of the output layer is set to the log-sig function, and the working condition set to be classified is actually marked The types are 1 and 0, that is, an abnormal working condition and other kinds of working conditions that need to be identified.

4-2) The performance function of the model is the mean square error

Among them, the cost function is equal to cost(h _w ( _xi ),y _i )=(h _w ( _xi )-y _i ) ² , X=(x ₁ ,x ₂ ,..., _xi ) is the input matrix , each column _xi is a set of input working condition parameters, h _w ( _xi ) is the output of the i-th input of the ANN model, and y is the known label working condition type.

To train the ANN classifier, 2924 sets of working condition data (including all working conditions to be identified) were used, including 522 sets of grounding, 236 sets of short circuit, 560 sets of inrush current, 601 sets of lightning strike, 524 sets of power recovery, 293 sets of power failure, and others There are 188 groups of working conditions, and the training set and test set are allocated according to the ratio of 7:3. The training set is 2046 groups, and the test set is 878 groups. In order to prevent the performance function of the over-fitting model training process, a regularization item is added, and the regularization item coefficient Set to 0.00001.

4-3) In the training of ANN classifier 1 (ground) and 2 (short circuit) coefficients, the cost function needs to follow the principle that ground and short circuit faults can be missed and not misjudged, and the weight factor K (K>1) should be added to the function , the grounding and short-circuit conditions are marked as 1, and the cost function becomes,

cost(h _w ( _xi ),y _i )＝y _i ×(h _w ( _xi )-1) ² +K(1-y _i )×h _w ( _xi ) ²

During the training and testing of classifier 1 and classifier 2, the weighting factors were adjusted. As shown in Figures 3 and 4, from the training and test results, when the weight factor K increases from 1 to 4, the misjudgment error decreases to a certain extent. When K is equal to 4, it is already close to 0, and the total error is also Not much has changed.

5) After the multi-classification recognition model has been trained and tested, and the error is controlled within the allowable threshold range, it can be provided to the online wave recording system for multi-working condition recognition of the distribution network.

Since the working condition data is randomly mixed and redistributed, the average value of 10 experiments is taken as the final result. Table 3 shows the training and testing error results of the multi-working condition recognition model. It can be seen from the results that the multi-working condition classification process based on the decision tree model combined with the ANN classifier of each working condition can control the multi-working condition recognition error below 6%, and add weight factors to the model to meet the requirements as much as possible. Requirements not to misjudge failure conditions. So far, the trained multi-classification recognition model can be provided to the online wave recording system for multi-working condition recognition of the distribution network.

Table 3 Working condition recognition error results

Identify the type of working condition Training set error (%) Test set error (%) grounding 1.47 4.56 short circuit 1.07 2.59 Inrush current 1.04 1.79 lightning strike 0.94 1.23 call back 0.35 0.75 power failure 0.39 0.95 Multiple working conditions 4.99 5.12

Claims (1)

1. a distribution network operating condition intelligent identification method based on multi-feature analysis, is characterized in that, may further comprise the steps: 1) Through the online wave recording system of the distribution network, when the line fails, the three-phase synchronous wave recording is triggered, and the A, B, C three-phase current wave recording signals and zero-sequence current signals are obtained; 2) Using multi-feature extraction method to extract signal features: 2-1) Extract time domain features In the time domain, the current signal collected by the online wave recording system is extracted as follows: Periodic maximum I _max : I _max,p (i)=max(I _p,i (j)) Period minimum value I _min : I _min,p (i)=min(I _p,i (j)) Period mean value I _mean : Periodic variance I _var : Cycle mean square error I _rms : Maximum mean square error I _mar : I _mar,p ＝max(I _rms,p (i)) Minimum mean square error I _mir : I _mir,p ＝ min(I _rms,p (i)) The maximum value of the root mean square difference I _mard : I _mard,p ＝max(I _rms,p (2)-I _rms,p (1),...,I _rms,p (i+1)-I _{rms, p} (i)) Three-phase unbalance degree IDUB: Among them, I _p,i is the current signal collected by the detector, Indicates the three phases of A, B, and C, p represents the four phases of A, B, C, and Z, i represents the periodic sequence of the wave recording signal, j represents the number of sampling points in each cycle, and N is the number of sampling points; 2-2) Extract frequency domain features In the frequency domain, for the DC and second harmonic components of the steady-state signal after the fault occurs, the discrete Fourier transform analysis is performed on the i~(i+m) cycle data of the recorded wave signal; Taking the signal as an example, use Fourier series expansion to obtain the frequency domain transformation result of the second harmonic: In the formula, _Ip,i is the current signal collected by the detector, p represents the four phases of A, B, C and Z, i is the periodic sequence of the current signal, N is the number of sampling points, n is the nth sampling point, j is the imaginary part indicator; Extract the following features: DC component content I _dp : Second harmonic component I _2x : Second harmonic component content I _2xp : 2-3) Analysis of transient signals by feature extraction method based on wavelet transform Use wavelet transform to extract the transient signal at the time of abnormal operation of the distribution network feeder; decompose the original transient signal into J different scales for analysis, and extract the low-frequency and high-frequency components of multiple frequency bands; 2-3-1) Decompose the collected abnormal signal and extract the high-frequency component characteristics of the abnormal working condition signal: Among them, A _J (k) is the low-frequency component coefficient obtained by J-order wavelet decomposition and reconstruction of the signal at time k, and D _i (k) is the ith-order high-frequency component coefficient; it is a unified expression, using D _J+1 (k) Instead of A _J (k), the expression is converted to The high-frequency component characteristic of the abnormal working condition signal is Df _p (i), and the statistical object is the sum of the absolute value of the high-frequency component coefficients of the half period before and after the abnormal signal appears, expressed as 2-3-2) Extract wavelet energy entropy and wavelet singular entropy Wavelet energy entropy and wavelet singular entropy are used to indicate the degree of chaos in the distribution of signal energy in different frequency bands during the period when abnormal conditions occur; The wavelet energy entropy WEE is defined as follows: Where p _i =E _i /E defines the signal energy spectrum at different scale i time k, E _i (k)=|D _i (k)| ² , is the energy sum of all moments on scale i, is approximately the total energy of the signal; The coefficients D _i (k) reconstructed by the wavelet transform form a (J+1)×M matrix D _(J+1)×M , and the singular value decomposition of the matrix can obtain J+1 non-negative singular value σ _i , then the wavelet singular entropy WAE is defined as follows: 3) Perform tree clustering on the feature data sets extracted according to step 2), and establish a multi-classification identification process in the form of a decision tree; 4) The classifier in the multi-class recognition process is constructed using a three-layer ANN model; 5) After the multi-classification recognition model has been trained and tested, and the error is controlled within the allowable threshold range, it can be provided to the online wave recording system for multi-working condition recognition of the distribution network.