CN110609200A - A Ground Fault Protection Method for Distribution Network Based on Fuzzy Metric Fusion Criterion - Google Patents

Abstract

The invention discloses a distribution network grounding fault protection method based on fuzzy metric fusion criterion, which includes: obtaining historical samples composed of various fault characteristic quantities of the protected feeder, and obtaining samples to be tested by the same method; adopting a clustering algorithm Divide n historical sample clusters into faulty and non-faulty classes; take membership degree, distance proportion measure and angle proportion measure as the similarity measure criterion respectively, calculate the fault measure value and non-fault measure value of the sample to be tested, and Construct the fuzzy metric fusion criterion matrix of the sample to be tested; take the fuzzy metric fusion criterion matrix as the evaluation set of the evaluation index system, construct the factor set of the evaluation index system with all the similarity measurement criteria, and preset the factor set of each element in the factor set The weight coefficient is used to judge the fuzzy metric fusion criterion matrix to obtain whether the distribution network to be tested is faulty or not. The invention widens the fault judgment interval, effectively avoids misjudgment caused by power system vibration and other factors, and improves robustness.

Description

A Ground Fault Protection Method for Distribution Network Based on Fuzzy Metric Fusion Criterion

technical field

The invention relates to the technical field of power systems, in particular to a distribution network grounding fault protection method based on fuzzy metric fusion criteria.

Background technique

my country's 6-66kV medium-voltage distribution network generally adopts the neutral point ungrounded and arc suppression coil grounded methods. When a single-phase ground fault occurs, the line voltage of the system remains symmetrical, so the continuous power supply to the load is not affected, and the system can still continue to run for 1 to 2 hours. However, especially for small current grounding systems, due to low voltage levels, weak fault currents, and are easily affected by factors such as arc instability and load harmonic interference, the problem of distribution network feeder grounding protection has not been effectively resolved, making small current grounding systems The ground fault detection protection in the field has become a recognized problem in the industry.

With the increase of the number of cable lines in the distribution network, the capacitive current gradually increases, and the long-term operation with faults will easily cause the fault to develop into a phase-to-phase fault or a multi-point fault; nonlinear loads and extensive access to power electronic equipment make interference factors more sensitive to fault detection. The impact of the arc light grounding fault is easy to cause the overvoltage of the whole system, causing multiple sets of transformers and switch cabinets to burn out, endangering equipment and personal safety. Therefore, it is necessary to study a distribution network ground fault protection method with high precision and strong robustness to ensure the safe and reliable operation of the power system.

After long-term efforts of experts and scholars at home and abroad, a variety of distribution network grounding protection methods have been proposed, which can be roughly divided into: protection methods based on steady-state characteristic criteria, protection methods based on transient characteristic criteria, and signal injection methods. The protection criterion of the above method is only based on the analysis of a single characteristic quantity, but the operation mode of the distribution network is complex and changeable, the fault conditions cannot be predicted, and the single protection criterion cannot cover all grounding conditions, so the accuracy of the protection action is not good. High (only 20% to 30%).

In recent years, the rise of SDG (Smart Distribution Grid) has greatly promoted the development of Advanced Distribution Automation (ADA). The ground fault protection method based on intelligent algorithm has become a research hotspot in the field of relay protection. These methods mainly include: neural network method, Bayesian method, genetic algorithm and intelligent fault protection method based on rough set theory. These methods rely on good data processing capabilities to improve the accuracy and adaptability of fault protection schemes to a certain extent, but the formation process of protection criteria generally lacks a clear physical mechanism, and fault judgment is only completed by training a large number of samples. In the case of changing operation mode, it is impossible to achieve a comprehensive description of the system operation state, and the fault judgment results are one-sided, which cannot meet the adaptive needs of relay protection in the dynamic environment of distribution network.

Contents of the invention

Based on the above-mentioned technical problems in the prior art, the present invention provides a distribution network grounding fault protection method based on fuzzy metric fusion criteria, which can broaden the fault judgment interval, effectively avoid misjudgment caused by factors such as power system oscillations, and improve robustness.

In order to realize the above-mentioned technical purpose, the present invention adopts following technical scheme:

A distribution network ground fault protection method based on fuzzy metric fusion criterion, comprising the following steps:

Step 1, construct a historical sample set, divide the historical sample set into faulty and non-faulty classes and calculate the cluster center;

Step A1, obtaining various fault feature quantities of the protected feeder under the known operating state of the distribution network to form a historical sample; repeating this step until n historical samples are obtained, and forming the n historical samples into a historical sample set;

Step A2, using a clustering algorithm to divide the sample clusters in the historical sample set into faulty and non-faulty classes, and calculate the faulty cluster centers and non-faulty cluster centers in the historical sample set;

Step 2, construct incremental sample set, divide incremental sample set into fault class and non-fault class and calculate cluster center;

Step B1, according to step A1, obtain various fault characteristic quantities of the protected feeder when it is in the running state to be tested, and form a sample to be tested; and form an incremental sample set with the sample to be tested and n historical samples;

Step B2, using a clustering algorithm to divide the sample clusters in the incremental sample set into faulty and non-faulty classes, and calculate the faulty cluster centers and non-faulty cluster centers in the incremental sample set;

Step 3, constructing the fuzzy metric fusion criterion matrix of the sample to be tested;

Step C1, use the fault cluster center and non-fault cluster center of the incremental sample set to calculate the membership degree of the sample to be tested relative to the fault class and non-fault class respectively; use the fault cluster center and the fault class cluster center of the historical sample set The non-fault class clustering center calculates the distance proportion measure of the sample to be tested relative to the fault class and the non-fault class, and calculates the angle proportion measure of the test sample relative to the fault class and the non-fault class respectively;

Step C2, the membership degree, distance proportion measure and angle proportion measure of the sample to be tested relative to the fault class, and the membership degree, distance proportion measure and angle proportion measure of the test sample relative to the non-fault class are used to construct the fuzzy Metric fusion criterion matrix;

Step 4, judging the running state of the protected feeder;

The fuzzy metric fusion criterion matrix is used as the evaluation set of the evaluation index system, and the factor set of the evaluation index system is constructed by the degree of membership, the distance proportion measure and the angle proportion measure, and the weight coefficient of each element in the factor set is preset, and the evaluation index system is used to treat The fuzzy metric of the test sample is fused with the criterion matrix to judge whether the protected feeder is faulty or not.

This program extracts a variety of fault feature quantities of the protected feeder to form samples, and through clustering and division of a large number of samples, and using membership degree, distance proportion measure and angle proportion measure as similarity measure criteria respectively, the sample to be tested is calculated3 One fault measure relative to fault class (namely membership degree relative to fault class, distance proportion measure and angle proportion measure) and three non-fault measures relative to non-fault class (namely membership degree relative to non-fault class, distance Proportion measure and angle proportion measure), so as to construct fuzzy metric fusion criterion matrix, and further use the evaluation index system to judge the fuzzy metric fusion criterion matrix to know whether the distribution network is faulty or not.

Since the fuzzy metric fusion criterion matrix of the sample to be tested is the fusion of three similarity metric criteria, i.e. membership degree, distance proportion measure and angle proportion measure, the robustness of the distribution network fault judgment method of the present invention is improved.

The horizontal comparison of the fault degree of each feeder in the existing distribution network fault judgment method is transformed into a vertical comparison of a certain feeder belonging to the fault category or non-fault category, and thus the fault fuzzy metric fusion criterion is obtained, The fault judgment interval is widened, which can effectively avoid misjudgment caused by power system vibration and other factors, and improve robustness.

Further, the evaluation index system is a multi-level evaluation index system, and the factor set composed of degree of membership, distance proportion measure and angle proportion measure is the final factor set of the evaluation index system;

The judging rules are as follows: take the fuzzy metric fusion criterion matrix as the last-level criterion matrix, and calculate the judgment matrix of each level factor set to the judgment set in order from the last level to the first level, and finally obtain the first-level judgment matrix As a comprehensive judgment matrix for fault metrics B ₁ =(b ₁ ,b ₂ ), if b ₁ >b ₂ , the distribution network to be tested is in a fault state, otherwise the distribution network to be tested is in a non-fault state;

Among them, the calculation method of the evaluation matrix B _q output at the qth level is: B _q = A _q R _q , and R _q-1 = B _q , A _q represents the input of the qth level and corresponds to the qth level factor set The preset weight coefficient set, R _q represents the criterion matrix of the input qth level.

The multi-level evaluation index system of this scheme can weaken the non-fault metric of the fault sample to be tested, and at the same time make the fault metric stand out, so that the fault sample can still be accurately identified when the fault signal is weak and affected by strong interference.

Further, the multiple fault feature quantities include 4 steady-state feature quantities and 3 transient-state feature quantities, and the 4 steady-state feature quantities are: zero-sequence impedance angle, negative-sequence current, zero-sequence current, grounding Fault resistance, the three kinds of transient characteristic quantities are: zero-sequence current db4 modulus maxima after wavelet transformation, modulus maxima after cubic B-spline wavelet transform, zero within half power frequency cycle value of the sequence energy function.

By extracting multiple fault feature quantities of the protected feeder, it is possible to avoid the influence of partial feature quantity extraction distortion (caused by interference factors such as nonlinear loads and fault signal distortion), thereby improving the accuracy of fault judgment.

Further, the evaluation index system is a two-level evaluation index system, and the first-level factor set includes two factors: steady-state feature quantity and transient state feature quantity.

Further, the specific process of step C1 is:

First, the membership degree μ _k ′ ₁ of the cluster sample x _k in the incremental sample set relative to the fault class and the membership degree μ′ _k2 relative to the non-fault class are calculated according to the following formulas:

In the formula, i represents the category of the cluster, c represents the number of categories and the value is c=2, i=1 represents the fault class, i=2 represents the non-fault class; μ′ _ki represents the cluster sample x _k belongs to the first The degree of membership of i clustering categories, p′ ₁ represents the fault cluster center of the incremental sample set, p′ ₂ represents the non-fault cluster center of the incremental sample set, and d is the weighted index;

Then, the membership degree μ′ _g1 of the sample x _g to be tested relative to the fault class is taken as the first fault measure, and the membership degree μ′ _g1 of the sample x _g to be tested relative to the non-fault class is taken as the first non-fault measure.

Further, the specific process of step C2 is:

First, select the Euclidean distance in the distance discriminant method to measure the distance between the sample x _g to be tested and the i-th cluster center p _i :

In the formula: d _g1 represents the distance between the test sample x _g and the fault cluster center p ₁ of the historical sample set; d _g2 represents the distance between the test sample x _g and the non-fault cluster center p ₂ of the historical sample set x _gj represents the j-th sample data of the sample x _g to be tested, p _ij represents the j-th sample data of the i-th cluster center p _i ;

Then, calculate the second failure metric of the sample under test relative to the failure class: and a second non-fault metric relative to the non-fault class:

Further, the specific process of step C3 is:

First, calculate the cosine of the angle between the sample x _g to be tested and the i-th cluster center p _i :

In the formula: cosθ _g1 represents the cosine of the angle between the sample x _g to be tested and the fault cluster center p ₁ of the historical sample set; cosθ _g2 represents the angle between the sample x _g to be tested and the non-fault cluster center p ₂ of the historical sample set Cosine, x _gj represents the jth sample data of the sample x _g to be tested, p _ij represents the jth sample data of the i-th cluster center p _i ;

Then, calculate the third failure metric of the sample under test relative to the failure class: and a third non-fault metric relative to the non-fault class:

Further, the kth historical sample x′ _k obtained in step 1 is expressed as: x′ _k =(x′ _k1 ,x′ _k2 ,…,x′ _ks ); where, x′ _k1 ,…, x′ _ks is the s fault feature quantity of the kth historical sample x′ _k , and the jth fault feature quantity is expressed as x′ _kj ; the historical sample set composed of n historical samples is: X′={x′ ₁ ,x′ ₂ ,...,x′ _n } ^T ;

Before step 2, normalization preprocessing of the historical sample set X' is also included:

In the formula, x _kj is the sample data after normalization processing; is the sample mean value of the jth fault feature quantity before normalization processing; S(x′) _j is the sample standard deviation of the jth fault feature quantity before normalization processing;

After normalization preprocessing, the kth historical sample is expressed as: x _k ＝(x _k1 ,…,x _kj ,…,x _ks ), and the historical sample set is expressed as: X={x ₁ ,x ₂ ,…, x _n } ^T .

Further, in step 2, the fuzzy c-means clustering algorithm is used to cluster and divide the n historical samples. The specific method is: through the following optimization objective function (4), by balancing the iterative equations (5) and (6) for all historical The samples are dynamically clustered, and the cluster centers of faulty and non-faulty classes are obtained:

In the formula: c is the number of cluster categories, take c=2; μ _ki ∈ [0,1] represents the membership degree of the cluster sample x _k belonging to the i-th cluster type, satisfying p _i is the cluster center, expressed as: p _i = (p _i1 , p _i2 ,..., p _is ); · is the matrix norm representing the spatial distance between the cluster samples and the cluster center; d is the weighted index, Take d=2; U represents the membership degree matrix composed of the membership degree μ _ki of all cluster samples x _k : U=[μ _ki ] _n×c ; P represents the cluster center matrix composed of all cluster centers p _i P=[p _i ]; J _d represents the clustering loss function; M _fc represents the fuzzy c partition space representing cluster samples x _k The termination condition of the iterative process of optimizing the objective function (4) is: w is the current iteration number, ε is the iteration stop threshold, and ε=1.0e-6.

Beneficial effect

This program extracts a variety of fault feature quantities of the protected feeder to form samples, and through clustering and division of a large number of samples, and using membership degree, distance proportion measure and angle proportion measure as similarity measure criteria respectively, the sample to be tested is calculated3 One kind of fault metrics relative to the fault category and three kinds of non-fault metrics relative to the non-fault category, so as to construct the fuzzy metric fusion criterion matrix, and further use the evaluation index system to judge the fuzzy metric fusion criterion matrix. Is it malfunctioning.

Because the fuzzy measurement fusion criterion matrix of the sample to be tested is the fusion of three similarity measurement criteria, the advantages of different similarity measurement criteria under different fault conditions can be fully utilized, and more fault situations are covered, which improves the efficiency of the invention. Robustness of power grid fault judgment method.

The horizontal comparison of the fault degree of each feeder in the existing distribution network fault judgment method is transformed into a vertical comparison of a certain feeder belonging to the fault category or non-fault category, and thus the fault fuzzy metric fusion criterion is obtained, It can effectively integrate the advantages of multiple fault criteria, thereby broadening the fault judgment range, effectively avoiding misjudgment caused by factors such as power system oscillations, and improving robustness.

Description of drawings

Fig. 1 is a schematic diagram of a method for constructing a historical sample set of the present invention;

Fig. 2 is a schematic diagram of a two-level evaluation index system model of the present invention;

Fig. 3 is a schematic diagram of a distribution network grounding protection model of the present invention;

Fig. 4 is a schematic diagram of a ground fault simulation model of the power distribution system of the present invention.

Detailed ways

The following is a detailed description of the embodiments of the present invention. This embodiment is carried out based on the technical solution of the present invention, and provides detailed implementation methods and specific operation processes to further explain the technical solution of the present invention.

Step 1, extract various feature quantities of different operating states of the distribution network, and construct a historical sample set

Under different operating states of the distribution network (including fault state and non-fault state), using the distribution automation terminal unit (FTU) installed at the outlet of the protected feeder, as shown in Fig. Fault characteristic quantity. Among them, the multiple fault feature quantities include 4 steady-state fault feature quantities (respectively: zero-sequence impedance angle, negative-sequence current, zero-sequence current, ground fault resistance) and 3 transient fault feature quantities (respectively: zero-sequence Current db4 modulus maxima after wavelet transformation, modulus maxima after cubic B-spline wavelet transform, zero-sequence energy function value within half power frequency cycle after fault initiation). Define the s fault feature quantities measured in the kth operating state to form the kth historical sample x′ _k :

x' _k = (x' _k1 ,x' _k2 ,...,x' _ks );

In the formula: x′ _k1 , ..., x′ _ks are the specific values of s fault feature quantities extracted by FTU respectively.

As shown in Figure 1, n historical samples are extracted under n operating states to form a historical sample set X′={x′ ₁ ,x′ ₂ ,…,x′ _n } ^T .

Perform normalized preprocessing on the historical sample set to obtain:

In the formula, x _kj is the jth historical feature quantity of the kth historical sample obtained by normalization preprocessing; is the sample mean of the jth historical feature quantity; S(x′ _j ) is the sample standard deviation of the jth historical feature quantity.

After normalization preprocessing, the kth historical sample is expressed as:

x _k ＝(x _k1 ,...,x _kj ,...,x _ks );

The feature index matrix containing n historical samples is expressed as:

X={x ₁ ,x ₂ ,...,x _n } ^T .

Step 2, perform fuzzy clustering on historical samples

The fuzzy c-means clustering algorithm is used to train the historical samples in the feature index matrix. The process is mathematically described as, for the normalized preprocessed feature index matrix X=(x _k ∈ R ^s :k=1,...,n), by optimizing the objective function (4), each historical sample is used as the aggregation By balancing iteration equations (5) and (6), the dynamic clustering of each historical sample is realized, and the cluster centers of faulty and non-faulty classes are obtained.

In the formula: c is the number of clustering categories, c=2 is taken in this paper; μ _gi ∈ [0,1] represents the membership degree of the clustering sample x _g belonging to the i-th clustering type, satisfying p _i is the cluster center, expressed as: p _i =(p _i1 ,p _i2 ,…,p _is ); ||·|| is the matrix norm that characterizes the spatial distance between the cluster sample and the cluster center; d As the weighting index, take d=2. U represents the membership degree matrix composed of the membership degree μ _ki of all cluster samples x _k : U=[μ _ki ] _n×c ; P represents the cluster center matrix P ₌ [p _i ]; J _d represents the clustering loss function; M _fc represents the fuzzy c partition space representing cluster samples x _k The termination condition of the iterative process of optimizing the objective function (4) is: w is the current iteration number, ε is the iteration stop threshold, and ε=1.0e-6.

The two clustering categories obtained by clustering in the present invention are respectively faulty and non-faulty, and the characteristic quantity between the cluster center representing the faulty class and the clustering center representing the non-faulty class is very obvious. Those skilled in the art can directly determine which is the fault cluster center and which is the non-fault cluster center, and then know which category is the fault category and which category is the non-fault category.

Step 3, extract the sample to be tested

When a fault occurs in the distribution network, various methods are used to extract various fault feature quantities (four kinds of steady-state fault feature quantities: zero-sequence impedance angle, negative-sequence current, zero-sequence current, ground fault resistance, three kinds of transient fault features Quantity: zero-sequence current db4 modulus maximum value after wavelet transformation, three times B-spline wavelet transformation modulus maximum value, zero-sequence energy function value in the half power frequency cycle after fault initiation, a total of 7 fault characteristic quantities) Sample x _g to be tested.

Step 4, according to the sample similarity measurement criterion, construct the fuzzy measurement fusion criterion matrix of the sample to be tested

To judge whether the sample to be tested is a faulty sample or a non-faulty sample, it is necessary to quantify the similarity between the sample to be tested and the historical sample, and quantitatively describe the similarity relationship between the sample to be tested and the historical sample. Reasonable fault measurement constitutes fuzzy measurement fusion criterion. The present invention adopts three criteria of membership degree, distance proportion measure and angle proportion measure as fuzzy test fusion criterion.

Criterion 1: Criterion of membership degree;

Treat the sample x _g to be tested as the n+1th historical sample, that is, take k=g=n+1, build an incremental sample set together with the previous n historical samples, and use the same clustering method as the historical sample set , divide the incremental sample set into faulty and non-faulty, calculate the faulty cluster center p′ ₁ and the non-faulty clustering center p′ ₂ of the incremental sample set, and calculate the relative incremental sample set of the sample to be tested The membership degree of fault class and non-fault class.

Define the membership degree μ′ _g1 of the test sample x _g to the fault class as the first fault measure of the test sample x _g , and the membership degree μ’ _g2 of the non-fault class as the first non-fault measure of the test sample x _g .

Criterion 2. Criterion of distance specific gravity;

Considering a sample as a definite point in a high-dimensional space, the degree of similarity between samples can be measured by the distance between two points in a high-dimensional space.

Calculate the fault cluster center p ₁ and the non-fault cluster center p ₂ of the historical sample set X, and select the Euclidean distance in the distance discriminant method to measure the distance between the test sample x _g and the i-th cluster center p _i distance:

In the formula: d _g1 represents the distance between the test sample x _g and the fault cluster center; d _g2 represents the distance between the test sample x _g and the non-fault cluster center.

Define the second failure metric of the sample x _g to be tested as: The second non-failure metric is:

Criterion 3. Criterion of Angular Specific Gravity;

Considering any two samples x _i and x _j as two vectors starting from the coordinate origin in a high-dimensional space, the cosine cosθ _ij of the angle between the vectors is the angular similarity measure between samples.

Calculate the cosine of the angle between the vector represented by the sample x _g to be tested and the vector represented by the i-th cluster center p _i :

In the formula: cosθ _g1 represents the cosine of the angle between the sample x _g to be tested and the center of the fault class; cosθ _g2 represents the cosine of the angle between the sample x _g to be tested and the center of the non-fault class.

The third failure metric defining the sample x _g to be tested is: The third non-failure metric is:

Then, the three similarity measure criteria of membership degree, distance proportion measure and angle proportion measure of the sample to be tested are combined into a fuzzy measure fusion criterion matrix with their respective fault measure values and non-fault measure values.

Step 5, set up a multi-level evaluation index system, evaluate the fuzzy measurement fusion criterion matrix, and know whether the distribution network is faulty;

A multi-level evaluation index system is set to evaluate the fuzzy metric fusion criterion matrix and output a comprehensive criterion matrix to realize the vertical fusion of various fault test criteria of the samples to be tested and the horizontal fusion of different types of fault feature quantities. In the present invention, the model of the two-level evaluation index system is set as shown in FIG. 2 .

Among them, the multi-level evaluation index system includes: evaluation set, fault judgment factor set and weight coefficient set.

The setting of the evaluation set:

Let V={V ₁ , V ₂ } be the evaluation set, where V ₁ represents the fault measurement, and V ₂ represents the non-fault measurement; the evaluation set is applicable to the fault judgment factor sets of all levels.

The fault judgment factor set is set according to the classification level:

Let U be a factor set including all fault judgment factors, where the factors are divided into l groups:

U＝{U ₁ ,...,U _t ,...,U _l }

In the formula, U _t ={U _t1 , U _t2 ,...,U _tm }, where m represents the number of single factors contained in the t-th group of fault judgment factor sets.

It can be seen that each group of fault judgment factor sets is divided into multiple levels: U is the highest level factor set, U _t (t=1,2,...,l) is the second-highest factor set, and so on. The actual number of grading layers depends on the specific situation.

The weight coefficient set is set correspondingly according to the classification level of the fault judgment factor set:

Let A _t ＝{a _t1 ,a _t2 ,...,a _tm } be the weight coefficient set of each single factor in the sub-advanced factor set U _t to the evaluation set V, a _tk (t=1,2,... ,m) According to the distribution of the importance degree of the corresponding single factor in the sub-advanced factor set U _t , satisfy A＝ _{ a ₁ ,a ₂ ,...,am } is the weight coefficient set of each single factor in U to the evaluation set V, satisfying

First, the fuzzy metric fusion criterion matrix is used as the judgment set, and the composite algorithm is used to obtain the comprehensive judgment result of the second-level factors:

B _q =A _q R _q =(b _q1 ,b _q2 );

in 1≤q≤2.

Then judge the first-level factor set, and use the second-level factor evaluation output matrix B _q to form the single-factor evaluation matrix R of the first-level factor set U:

R=[B ₁ ,B ₂ ,...,B _L ] ^T ;

Therefore, the final comprehensive criterion matrix of the first-level factor set U is:

B=A·R=(b ₁ ,b ₂ );

In order to transform the fuzzy metric fusion criterion in the fault metric comprehensive criterion matrix into actual protection judgment, this paper selects the maximum membership degree criterion as the protection action criterion: b ₁ >b ₂ .

Example:

Using PSCAD/EMTDC simulation software to simulate and analyze a 35kV power distribution system, the simulation model is shown in Figure 4. The system adopts the grounding mode of the neutral point through the arc suppression coil, and the bus is powered by a 110kV system through a △/Y transformer.

Extract three kinds of steady-state fault feature quantities: zero-sequence impedance angle x _k1 , negative-sequence current x _k2 , zero-sequence current x _k3 , and three kinds of transient fault feature quantities: zero-sequence current db4 modulus maximum after wavelet transform x _k4 , modulus maxima x _k5 after cubic B-spline wavelet transform, and zero-sequence energy function value x _k6 in half power frequency cycle after the start of the fault, a total of 6 fault feature quantities are used as feature quantities for fuzzy cluster analysis. When the fault occurs on feeder 3 and feeder 4, the characteristic samples collected at the exit of feeder 4 constitute the historical sample set. Due to limited space, the most representative 16 groups of historical samples are preferably listed in Table 1. In the case of a single-phase ground fault on feeder 4, a set of fault feature quantities are extracted as samples to be tested, see Table 2.

Table 1 Historical sample set

Table 2 Samples of fault characteristics to be tested

The fault metric and non-fault metric of various sample similarity metric criteria are calculated for the steady-state fault eigenvalue and transient fault eigenvalue of the sample to be tested respectively, and the fuzzy metric fusion criterion matrix is established. From the data in Table 3, it can be seen that the membership degree criterion, the angle proportion measure criterion and the distance proportion criterion of the transient feature quantity all show that the fault measure of the sample to be tested is greater than the non-fault measure, and the test sample is accurately judged as a fault sample; The distance-proportion criterion of the steady-state feature quantity shows that the sample to be tested is a non-faulty sample. The reason is that the extraction process of fault samples to be tested is affected by interference factors such as nonlinear loads, and the fault signal is distorted, resulting in distortion of feature extraction and affecting the accuracy of fault judgment. If the traditional protection scheme is used to protect and identify the tested samples, the protection judgment will fail.

Table 3 Two-level index system for ground fault judgment of distribution network

Establish a two-level evaluation index system to analyze the fuzzy metric fusion criterion matrix of the fault samples to be tested, see Table 3. The weight coefficient sets A and A ₁ are formulated according to the protection judgment accuracy of all levels of criteria obtained from a large number of simulation calculations and field experiments, where A=(0.5,0.5), A ₁ =A ₂ =(0.5,0.3,0.2).

From the data in Table 3, the fuzzy metric fusion criterion matrix can be obtained, which are:

Calculate the comprehensive evaluation matrix of U ₁ :

B ₁ =A ₁ ·R ₁ =(0.5145,0.4855)

Similarly, the comprehensive evaluation matrix of U ₂ can be obtained:

B ₂ =A ₂ ·R ₂ =(0.6755,0.3245)

Utilize B ₁ and B ₂ to construct the first-level single-factor evaluation matrix: R=[B ₁ ,B ₂ ] ^T , and calculate the comprehensive evaluation matrix B of the first-level factor set U, that is, the comprehensive evaluation matrix of fault measurement:

B＝A·R＝(0.5950,0.4050)

From b ₁ >b ₂ , it can be seen that the fuzzy metric fusion criterion matrix is corrected through the two-level evaluation index system, and the samples to be tested are accurately judged as faulty samples.

The calculation results show that the two-level evaluation index system weakens the non-fault metrics of the fault samples to be tested, and at the same time makes the fault metrics stand out. When the fault signal is weak and affected by strong interference, this method can accurately identify fault samples.

The above embodiments are preferred embodiments of the present application, and those skilled in the art can also perform various transformations or improvements on this basis, and without departing from the general concept of the application, these transformations or improvements should all belong to the present application. within the scope of the application.

Claims (9)

1. A power distribution network ground fault protection method based on fuzzy metric fusion criterion is characterized by comprising the following steps:

step 1, constructing a historical sample set, dividing the historical sample set into a fault class and a non-fault class and calculating a clustering center;

a1, acquiring various fault characteristic quantities of a protected feeder line in a known power distribution network operation state to form a history sample; repeating the steps until n historical samples are obtained, and forming a historical sample set by the n historical samples;

step A2, clustering the samples in the historical sample set into fault classes and non-fault classes by adopting a clustering algorithm, and calculating fault class centers and non-fault class centers in the historical sample set;

step 2, constructing an increment sample set, dividing the increment sample set into a fault class and a non-fault class and calculating a clustering center;

b1, acquiring various fault characteristic quantities of the protected feeder line in the running state to be tested according to the A1 to form a sample to be tested; forming an incremental sample set by the sample to be detected and the n historical samples;

step B2, clustering the samples in the incremental sample set into fault classes and non-fault classes by adopting a clustering algorithm, and calculating fault class centers and non-fault class centers in the incremental sample set;

step 3, constructing a fuzzy measurement fusion criterion matrix of the sample to be tested;

step C1, calculating the membership degrees of the samples to be detected relative to the fault class and the non-fault class respectively by using the fault class center and the non-fault class center of the incremental sample set; calculating distance specific gravity measurement of the sample to be measured relative to the fault class and the non-fault class and angle specific gravity measurement of the sample to be measured relative to the fault class and the non-fault class respectively by utilizing the fault class center and the non-fault class center of the historical sample set;

step C2, constructing a fuzzy metric fusion criterion matrix of the sample to be tested according to the membership degree, the distance specific gravity metric and the angle specific gravity metric of the sample to be tested relative to the fault class and the membership degree, the distance specific gravity metric and the angle specific gravity metric of the sample to be tested relative to the non-fault class;

step 4, judging the running state of the protected feeder line;

the fuzzy metric fusion criterion matrix is used as a judgment set of a judgment index system, a factor set of the judgment index system is constructed by using membership, distance proportion measurement and angle proportion measurement, weight coefficients of all elements in the factor set are preset, and the fuzzy metric fusion criterion matrix of a sample to be tested is judged by using the judgment index system to obtain whether the protected feeder line is in fault or not.

2. The method according to claim 1, wherein the evaluation index system is a multi-level evaluation index system, and a factor set consisting of a membership degree, a distance weight metric and an angle weight metric is a final factor set of the evaluation index system;

the evaluation rule is as follows: taking the fuzzy measurement fusion criterion matrix as a final-stage criterion matrix, sequentially obtaining the judgment matrix of each stage factor set to the judgment set according to the sequence from the final stage to the first stage, and taking the finally obtained first-stage judgment matrix as a fault measurement comprehensive judgment matrix B₁＝(b₁,b₂) If b is₁>b₂The power distribution network to be distributed is in a fault state, otherwise the power distribution network to be distributed is in a non-fault state;

wherein, the evaluation matrix B output by the q level_qThe calculation method comprises the following steps: b is_q＝A_q·R_qAnd R is_q-1＝B_q，A_qA set of predetermined weighting factors, R, representing the input q-th level and corresponding to the q-th level factor set_qA matrix of criteria representing the input q-th level.

3. The method according to claim 2, wherein the plurality of fault characteristic quantities include 4 kinds of steady-state characteristic quantities and 3 kinds of transient characteristic quantities, and the 4 kinds of steady-state characteristic quantities are respectively: zero sequence impedance angle, negative sequence current, zero sequence current, earth fault resistance, 3 kinds of transient characteristic quantity are respectively: zero-sequence current db4 wavelet transform post-modulus maximum, cubic B-spline wavelet transform post-modulus maximum, and zero-sequence energy function value in half of power frequency cycle after fault initiation.

4. The method of claim 3, wherein the evaluation index system is a two-level evaluation index system, and the first-level factor set comprises 2 factors: a steady state characteristic quantity and a transient characteristic quantity.

5. The method according to claim 1, wherein the specific process of step C1 is as follows:

firstly, calculating clustering samples x in the incremental sample set according to the following formula_kMembership mu 'to failure class'_k1And membership μ 'to non-failure classes'_k2：

In the formula, i represents the category of the cluster, c represents the number of categories and has a value of c-2, i-1 represents a fault class, and i-2 represents a non-fault class; mu's'_kiRepresenting a cluster sample x_kDegree of membership, p 'to the ith cluster category'₁Fault class center, p ', representing incremental sample set'₂Representing a non-fault class center of the incremental sample set, wherein d is a weighted index;

then, the sample x to be measured_gMembership mu 'to failure class'_g1As the first fault measure, the sample x to be measured_gMembership mu 'to non-failed classes'_g1As a first non-failure metric.

6. The method according to claim 1, wherein the specific process of step C2 is as follows:

firstly, the Euclidean distance in the distance discrimination method is selected to measure the sample x to be measured_gWith class i centres p_iThe distance between:

in the formula: d_g1Representing the sample x to be measured_gFault clustering center p with historical sample set₁The distance between them; d_g2Representing the sample x to be measured_gNon-fault cluster center p with historical sample set₂Distance between, x_gjRepresenting the sample x to be measured_gJ sample data, p_ijRepresenting class i centres p_iThe jth sample data of (1);

then, calculating a second fault metric of the sample to be tested relative to the fault class:and a second non-fault metric with respect to the non-fault class:

7. the method according to claim 1, wherein the specific process of step C3 is as follows:

first, a sample x to be measured is calculated_gWith class i centres p_iCosine of the included angle therebetween:

in the formula: cos θ_g1Representing the sample x to be measured_gFault clustering center p with historical sample set₁Cosine of the included angle of (c); cos θ_g2Representing the sample x to be measured_gNon-fault class center p with historical sample set₂Cosine of (x)_gjRepresenting the sample x to be measured_gJ sample data, p_ijRepresenting class i centres p_iThe jth sample data of (1);

then, calculating a third fault metric of the sample to be tested relative to the fault class:and a third non-fault metric with respect to the non-fault class:

8. the method of claim 1, wherein the kth historical sample x 'obtained in step 1'_kExpressed as: x'_k＝(x′_k1,x′_k2,…,x′_ks) (ii) a Wherein, x'_k1、…、x′_ksIs the kth historical sample x'_kS fault characteristic quantities, and the jth fault characteristic quantity is expressed as x'_kj(ii) a The history sample set composed of n history samples is as follows: x '═ X'₁,x′₂,…,x′_n}^T；

Before step 2, the method further comprises the following steps of normalizing the historical sample set X':

in the formula, x_kjThe sample data after normalization processing;the sample mean value of the jth fault characteristic quantity before normalization processing is obtained; s (x')_jThe standard deviation is the sample standard deviation of the jth fault characteristic quantity before normalization processing;

after normalization preprocessing, the kth historical sample is obtained and is represented as: x is the number of_k＝(x_k1,…,x_kj,…,x_ks) The historical sample set is represented as: x ═ X₁,x₂,…,x_n}^T。

9. The method according to claim 8, wherein the step 2 adopts a fuzzy c-means clustering algorithm to cluster and divide the n historical samples, and the specific method is as follows: dynamically clustering all historical samples by means of balanced iterative equations (5) and (6) through an optimized objective function (4) as follows, and obtaining the clustering centers of fault classes and non-fault classes:

in the formula: c is the number of clustering categories, and c is 2; mu.s_ki∈[0,1]Representing a cluster sample x_kMembership degree belonging to the ith cluster typep_iIs the cluster center, expressed as: p is a radical of_i＝(p_i1,p_i2,…,p_is) (ii) a I | · | is a matrix norm representing a spatial distance between the clustering samples and the clustering center; d is a weighting index, and d is taken to be 2; u denotes the sample x of all clusters_kDegree of membership mu of_kiForming a membership matrix: u ═ mu_ki]_n×c(ii) a P denotes the center P of all clusters_iForming a cluster center matrix P ═ P_i]；J_dRepresenting a clustering loss function; m_fcRepresenting a cluster sample x_kFuzzy c ofThe end conditions of the iterative process of optimizing the objective function (4) are as follows:w is the current iteration number, epsilon is an iteration stop threshold, and epsilon is 1.0 e-6.