CN111999591A

CN111999591A - Method for identifying abnormal state of primary equipment of power distribution network

Info

Publication number: CN111999591A
Application number: CN201910447860.2A
Authority: CN
Inventors: 戴义波; 姚蔷; 张建良
Original assignee: Beijing Inhand Network Technology Co ltd
Current assignee: Beijing Inhand Network Technology Co ltd
Priority date: 2019-05-27
Filing date: 2019-05-27
Publication date: 2020-11-27
Anticipated expiration: 2039-05-27
Also published as: CN111999591B

Abstract

The invention discloses a method for identifying an abnormal state of primary equipment of a power distribution network, which comprises the following steps: inputting the fault recording data of the power distribution network into a macro classifier to obtain working condition classification; inputting the fault recording data of the power distribution network into a micro classifier to obtain a mode label of the fault recording data; and inputting the obtained working condition classification and mode label into an abnormal state recognizer, selecting a corresponding abnormal state recognition sub-model according to the mode label, and recognizing the abnormal state of the primary equipment according to the working condition classification, the mode label, the line operation information and the line topology information to obtain the generation reason and the position of the abnormal state of the primary equipment.

Description

Method for identifying abnormal state of primary equipment of power distribution network

Technical Field

The invention relates to the technical field of power transmission and transformation, in particular to a method for identifying an abnormal state of primary equipment of a power distribution network.

Background

The abnormal primary equipment of the power distribution network causes the power grid not to operate in an expected mode and even causes serious consequences, such as: protection devices are abnormal to cause protection refusal or error protection, and critical devices are aged to cause faults to cause power failure. At present, the troubleshooting of abnormal equipment is often judged only according to the operation years and the appearance, the workload is large, the effect is not good, and the problem of identifying the abnormality of the equipment is the challenge facing the high-reliability power distribution network. At present, no effective technical method for identifying various equipment abnormalities exists in the industry, and the difficulty lies in that the equipment abnormalities are rich in types and various in abnormal modes, and various abnormal modes cannot be effectively detected by a single means. Only a single technical means can be used to monitor a certain type of abnormal pattern. For example, CN106597185 provides a method for monitoring substation primary equipment by using infrared imaging, but this method only can monitor abnormal primary equipment in which thermal change reaction occurs by using an infrared thermal imager, and cannot realize effective identification for other abnormal states. For another example, CN207379628U discloses a method for monitoring the temperature of a primary device by using an optical fiber sensor, so as to monitor the abnormal state of the primary device, which can only monitor the temperature of the primary device, but cannot effectively identify the abnormal state of the primary device.

The applicant believes that when certain equipment in the distribution network is abnormal, it will cause the line current and voltage to change in specific patterns, and if these specific patterns can be identified, it is possible to know what abnormality has occurred to the equipment. The transient recording type fault indicator is equipment capable of triggering recording according to current and voltage changes, and rich current and electric field information can be captured by adopting a high-precision and high-bandwidth current sensor and an electric field sensor for reflecting voltage changes. And storing the recording triggered by the transient recording type fault indicator in a main station database, and recording line events including equipment abnormity in the main station to form a case. After a large number of cases are formed, recording waves related to similar exceptions are analyzed, and modes only existing in specific exceptions in the recording waves are mined, so that the incidence relation between the specific modes in the recording waves and the equipment exceptions is established, and the purpose of identifying the equipment exceptions by using the recording waves is achieved. Therefore, there is a need in the art for a method for recognizing multiple abnormal modes of a primary device by using recording data, so as to monitor multiple abnormal modes of the primary device.

Disclosure of Invention

The invention aims to provide an identification method capable of accurately identifying the abnormal state of primary equipment of a power distribution network. The identification method accurately identifies the abnormal state of the primary equipment by utilizing the wave recording data of the power distribution network, and can complete synchronous monitoring and identification of the various abnormal states of the primary equipment.

In order to solve the technical problem, the invention provides a method for identifying an abnormal state of primary equipment of a power distribution network, which comprises the following steps:

inputting the fault recording data of the power distribution network into a macro classifier, extracting time domain and frequency domain characteristics from the fault recording data to form characteristic vectors, and inputting the characteristic vectors into a working condition classifier in the macro classifier to obtain working condition classification;

inputting fault recording data of the power distribution network into a micro classifier, extracting specific segments from the fault recording data, extracting wavelet energy entropy characteristics and wavelet singular entropy characteristics by using wavelet transformation, and inputting the specific segments, the wavelet energy entropy characteristics and the wavelet singular entropy characteristics into a mode classifier in the micro classifier to obtain a mode label of the fault recording data;

and inputting the obtained working condition classification and mode label into an abnormal state recognizer, selecting a corresponding abnormal state recognition sub-model according to the mode label, and recognizing the abnormal state of the primary equipment according to the working condition classification, the mode label, the line operation information and the line topology information to obtain the generation reason and the position of the abnormal state of the primary equipment.

In one embodiment, the macro classifier uses a classifier model that includes a plurality of three-layer feed-forward neural networks, and the condition classifications include power outage, power restoration, ground, short circuit, and magnetizing inrush current.

In one embodiment, the micro classifier uses a deep neural network model that includes a convolutional layer region and a full link layer region, and the mode labels include lightning arrester breakdown, insulator flashover, line discharge, switch vacuum bubble breakdown, line break, and switch short circuit fault protection.

In one embodiment, the abnormal state identifier includes a sub-model for a plurality of abnormal state identifications of a plurality of abnormal state types.

In one embodiment, the plurality of abnormal state identified submodels includes a decision tree model and a topology search model.

In one embodiment, the line operation information includes information such as temperature, wind, and rainfall; the line topology information includes line distribution and primary equipment location.

In one embodiment, the time domain features extracted from the fault recording data include: the three-phase unbalance degree of the three-phase current is determined according to the current cycle maximum value, the current cycle minimum value, the current cycle mean value, the current cycle variance, the current cycle root mean square, the current cycle minimum root mean square, the current cycle.

In one embodiment, the frequency domain features extracted from the fault recording data include: the dc component content, the second harmonic component and the second harmonic component content of the current.

According to another aspect of the present invention, there is also provided an apparatus for identifying an abnormal state of a primary device of a power distribution network, the apparatus including:

the macro classifier extracts time domain and frequency domain characteristics from the fault recording data to form a characteristic vector, and inputs the characteristic vector into the working condition classifier to obtain working condition classification;

the micro classifier extracts a specific segment from the fault recording data, extracts wavelet energy entropy characteristics and wavelet singular entropy characteristics by using wavelet transformation, and inputs the specific segment, the wavelet energy entropy characteristics and the wavelet singular entropy characteristics into the mode classifier to obtain a mode label of the fault recording data;

and the abnormal state identifier selects a corresponding abnormal state identification submodel according to the mode label, identifies the abnormal state of the primary equipment according to the working condition classification, the mode label, the line operation information and the line topology information, and obtains the generation reason and the position of the abnormal state of the primary equipment.

In one embodiment, the macro classifier uses a classifier model that includes a plurality of three-layer feed-forward neural networks.

In one embodiment, the micro classifier uses a deep neural network model that includes a convolutional layer region and a fully-connected layer region.

Abnormal State identifier >

Fig. 1 is a schematic flow chart of the abnormal state identification method of the present invention, in which the abnormal state identification model obtains abnormal segments and abnormal reasons through topology analysis by using all the working conditions of the circuit topology in which the macro classification model outputs recording waves and the micro classification model outputs mode labels, and combining with the circuit state information. The line status information varies for different anomaly types. The abnormal state identification model is not a single model, the equipment is rich in abnormal types, the abnormal modes are various, and the rules for identifying different abnormalities and the required data sources also have differences, so that the model and the data sources need to be selected according to the characteristics of specific abnormalities. Some abnormal states are only reflected in a macro scale, so that only the working condition of wave recording is needed, and some abnormal modes are only reflected in a certain wave recording segment, so that only mode tags are needed.

Macro classifier model >

The macro classification model is used for marking the upper working condition type for the recording, and the upper working condition type is finally marked for the recording by paying attention to the information of the large-scale visual angle in the recording, such as the change of the current amplitude value in the whole recording. The type of condition represents a certain line event in the present invention, and typically includes: power failure, power restoration, grounding, short circuit, magnetizing inrush current and the like. As shown in FIG. 2, the macro classifier model of the present invention comprises two steps when performing condition classification:

1. feature extraction and vectorization: extracting characteristics such as current electric field amplitude, transient state steady state harmonic component and the like from the recorded wave to form a characteristic vector;

2. and (3) working condition identification: and transmitting the characteristic vectors into a working condition classifier consisting of a plurality of neural network classifiers to obtain a working condition type.

In the present invention, a plurality of features are first extracted from the time domain and the frequency domain of the recording data, wherein the feature quantities extracted based on the time domain are shown in table 1 below.

TABLE 1

In the above tableI _p,i (j)Is a current signal collected by the detector, p＇the representation A, B, C is for three phases,prepresenting A, B, C, Z the four phases of the three-phase,irepresenting 1-16 cycles of the recording signal,jrepresenting 1-82 samples per cycle, and N is the number of samples 82.

A method for extracting characteristic quantity based on frequency domain mainly aims at direct current and second harmonic component of steady state signal after fault occurrence and adopts 8 th-10 th period data of wave recording signal to carry out DFT analysis. Taking a periodic current signal as an example, using Fourier series expansion to obtain a frequency domain transformation result of the second harmonic,

the extracted feature quantities are shown in table 2 below,

TABLE 2

The features obtained based on the time domain and the frequency domain as described above are grouped into a 4 × N matrix, F = [ a, B, C, D ]. Where A, B, C, Z correspond to a row of the matrix and each feature corresponds to a column of the matrix, e.g., A, B, C, Z has no corresponding feature set to 0. And N is the number of the features.

As shown in FIG. 3, for the identification of the working condition types, the invention firstly utilizes a physical model and engineering experience and carries out tree type on the working condition set through manual prejudgment so as to reasonably use a decision tree model with fewer classifiers and good classification effect and establish a more accurate classification model. The short circuit and the grounding belong to fault working conditions, and the two working conditions are firstly identified according to the principle that judgment can be missed and misjudgment can not be carried out on the two working conditions of the short circuit and the grounding in consideration of the importance of the short circuit and the grounding. The failure judgment judges the original short circuit (grounding) working condition as another working condition type, and the error judgment classifies the abnormal working condition which does not belong to the short circuit (grounding) as the fault working condition. And (3) realizing a classifier in the identification process, and using a three-layer feedforward neural network model with stronger nonlinear fitting capacity.

For each of the ANN classifiers is a function of,

the functions are tan-sig functions, the activation function of the output layer is a log-sig function, and the actual mark types of the working condition set needing to be classified are 0 and 1. The performance function of a generic model is the mean square error,

wherein the cost function is equal to,

Xfor an input matrix, each column

A set of input operating condition parameters is provided,

as an ANN modeliThe output of each of the inputs is,yis a known signature operating mode type. In the training of coefficients of the ANN classifier 1 and the ANN classifier 2, the cost function needs to follow the principle that the earth fault and the short-circuit fault can be missed and not misjudged, and a weight factor is added into the functionK(K>1) The ground and short circuit conditions are labeled 1, i.e.,

all neural network training algorithms adopt a self-adaptive learning speed algorithm, and the maximum training times are 2000 times.

In the training process, 2924 groups are shared by all working condition data, wherein a grounding 522 group, a short circuit 236 group, an excitation inrush current 560 group, a lightning stroke 601 group, a power restoration 524 group, a power failure 293 group and other working condition 188 groups, a training set and a test set are distributed according to the proportion of 7:3, the training set is 2046 groups, the test set is 878 groups, a regular term is added for preventing a performance function in the overfitting model training process, and the coefficient of the regular term is set to 0.00001. Since the condition data is randomly mixed and redistributed, fig. 4 shows the average result of 10 experiments, which shows the error of training and testing, i.e. the ratio of the sum of false and missed judgments to the sum of the total condition numbers, and the error of false judgments, wherein the weight factor K is adjusted to meet the functional requirements.

In fig. 4, it can be seen that the misjudgment error decreases to some extent when the weighting factor K increases from 1 to 4, and K is equal to 4, and is already close to 0, and the total error does not change too much. The error of the classifier used finally in the training and testing process, and the total error of the whole recognition process are shown in table 3.

TABLE 3

The result shows that the multi-working-condition classification flow based on the decision tree model is combined with the ANN classifiers of all the working conditions, the multi-working-condition identification error can be controlled to be less than 6%, and the weight factor is added into the model, so that the requirement of not misjudging the fault working condition as much as possible is met.

Micro classification model >

The micro classification model is mainly used for labeling a mode label on a specific segment in the recorded wave. Mode labels represent certain abnormal conditions in the present invention, for example, mode labels include lightning arrester breakdown, insulator flashover, line discharge, switch vacuum bulb breakdown, line breakage, and switch short circuit fault protection. The whole model comprises two steps:

preprocessing data. Extracting characteristic segments from the recorded waves, and extracting characteristics such as wavelet energy entropy and wavelet singular entropy by utilizing wavelet transformation;

mode classification. And transmitting the recording segments and the characteristics into a mode classifier to obtain a mode label. The mode classifier is a deep neural network mainly composed of 4 convolutional layers and 2 fully-connected layers.

In the present invention, wavelet transformation is an effective tool for analyzing transient signals, inThe fault diagnosis in various engineering fields has good application effect, and the principle is to decompose the original signal intoJAnd the low-frequency and high-frequency components of a plurality of frequency bands can be extracted by analyzing on different scales. Transient signals at the moment of abnormal operation of the feeder line of the power distribution network are extracted by using wavelet transformation, and the acquired abnormal signals are decomposed by using wavelet transformation.

Wherein

Is a low-frequency component coefficient obtained by reconstructing a signal through J-order wavelet decomposition,

are i-order high frequency component coefficients. For the sake of convenience in the subsequent description,

by using

Instead, then there are

The high frequency component of the anomaly is characterized by dfp (i),

the statistical object is the sum of the absolute values of the coefficients of the high-frequency components in half cycle before and after the abnormal signal appears. In order to effectively decompose transient signals of abnormal working conditions of the feeder line of the power distribution network, 3-order db5 wavelet transformation is used. In addition to extracting features using detail component coefficients in wavelet decomposition, Shannon information entropy representing the degree of information confusion is combined in a wavelet feature extraction algorithmThe wavelet energy entropy and the wavelet singular entropy are extracted and used for representing the degree of disorder of signal energy distributed in different frequency bands in the time period when the abnormal working condition occurs. Defined at different scalesiThe energy spectrum of the signal at time k is the sum of the energies at all time instants on the scale i. The energy entropy WEE can be defined as follows,

wherein

Approximately the total energy of the signal. Reconstructing the wavelet transform

Form a matrix of (J +1) xM

Performing singular value decomposition on the matrix to obtain J +1 nonnegative singular values

Then the wavelet singular entropy is defined as follows,

according to engineering experience, in order to show the chaos degree of a transient signal when an abnormal signal occurs, in the calculation of the wavelet energy entropy and the wavelet singular entropy, the sections of the signal are 10 sample points before and after the occurrence of the singular signal is detected.

The pattern classifier in the micro classifier of the present invention uses a deep neural network classifier including a convolutional layer region and a fully-connected layer region. The convolution layer area comprises an input convolution layer, a convolution block and an average pooling layer, the convolution operation related in the convolution layer adopts a convolution operation method known in the prior art, and a convolution kernel and related parameters used in the convolution operation are optimized hyper-parameters obtained through hyper-parameter machine training. The sampling points with small time interval of time sequence waveform have strong relativity, and the sampling points with larger time interval are weaker, so that the convolution layer is suitable for extracting features. And local-to-global feature extraction and abstract-to-concrete feature extraction are realized by arranging a plurality of convolutional layers in the convolutional layer region. And connecting a full connection area behind the convolution layer area, wherein the full connection area internally comprises two full connection layers and a softmax output layer. The first layer of the full connection layer can input the working condition identification and the line operation information, and the selection of the line operation information needs to be determined by the abnormal type. And finally obtaining an abnormal mode label after passing through a softmax output layer.

Fig. 5 shows a detailed structure of the model. Firstly, wavelet transformation is carried out on recorded wave data, and then the data size is reduced by utilizing a cubic spline interpolation method. In the deep neural network classifier, the width and length of convolution kernels in an input convolution layer are 6 × 5, and the number of the convolution kernels is 8. The convolution block i is a single-channel, double-layer convolution layer, where the width and length of the convolution kernel of the first convolution layer is 6 × 3, and the number is 8, and the width and length of the convolution kernel of the second convolution layer is 6 × 3, and the number is 16. The convolution block II is set to be a convolution layer with three channels, the channel a is a double-layer convolution layer, wherein the width and the length of convolution kernels of the first convolution layer are 6 multiplied by 2, the number of the convolution kernels is 16, the width and the length of convolution kernels of the second convolution layer are 6 multiplied by 3, and the number of the convolution kernels is 32. The channel b is a double-layer convolutional layer, wherein the width and length of the convolutional cores of the first convolutional layer are 6 multiplied by 3, the number of the convolutional cores is 32, and the width and length of the convolutional cores of the second convolutional layer are 6 multiplied by 3, and the number of the convolutional cores is 32. And the channel c is three convolutional layers, wherein the width and the length of a convolutional kernel of the first convolutional layer are 6 multiplied by 3, the number of the convolutional kernels is 16, the width and the length of a convolutional kernel of the second convolutional layer are 6 multiplied by 4, the number of the convolutional kernels is 16, the width and the length of a convolutional kernel of the third convolutional layer are 6 multiplied by 3, the number of the convolutional kernels is 32, and the sum of the results of the three channels of the convolutional block II is input into the convolutional block III. The convolution block III is provided as a convolution layer having 8 channels each of which is constituted by two convolution layers, wherein the convolution kernel of the first convolution layer has a width and length of 6 × 3 and the number of the convolution kernels is 32, and the convolution kernel of the second convolution layer has a width and length of 6 × 3 and the number of the convolution kernels is 64. The output of the 8 channels in volume block iii is then summed and input to the average pooling layer. And inputting the output result of the average pooling layer into a first full-connection layer, wherein the number of the neurons of the first full-connection layer is 24, the output result of the first full-connection layer is input into a second full-connection layer, and the number of the neurons of the second full-connection layer is set to be the same as the number of the abnormal types of the training set.

As shown in fig. 6, in the training method of the pattern classifier model of the present invention, case data is divided into a training data set and a test data set according to a ratio of 5:5, and an optimal hyper-parameter combination is obtained according to the training method shown in fig. 5. The parameters obtained through training comprise the number of convolution blocks, the length, the width and the number value of convolution kernels of convolution layers in each convolution block, the number of channels contained in each convolution block, the number of layers of convolution layers on each channel, and the number of neurons used in a full connection layer.

And taking the insulation weakness abnormity as an example, explaining the model training process. In the actual training process, 100 groups of insulation weak data and 200 groups of non-insulation weak data are adopted and then are evenly distributed into a test set and a verification set. The optimization method in the training process is batch Adam backward transmission, when the accuracy of the test data set is greater than 98% or the training exceeds 10000 rounds, the training is stopped, otherwise, the optimization is continued, and the combination with the highest accuracy of the verification data sets in the multiple hyper-parameter combination models is the optimal hyper-parameter combination model. The final training and testing errors are shown in table 4, and it can be seen that both errors of the weak insulation are lower than 2%, and the identification target of the weak insulation abnormity is correct, missing judgment and incorrect judgment, so that the actual application requirements are met.

TABLE 4 error results

Type (B)	Training set error (%)	Test set error (%)
			Weak insulation	1.5	1.8

Compared with the prior art, the invention has the following important invention points:

1. the invention uses the transient recording data of the wire network as the basis for judging the abnormal state of the primary equipment, improves the accuracy and completeness of the identification of the abnormal state of the primary equipment, and realizes the simultaneous identification of various abnormal states of various kinds of equipment.

2. According to the invention, the transient recording data is subjected to macro classification and periscopic classification, so that the working condition identification and the mode label of the abnormal state are obtained, and more valuable information in the recording data can be completely utilized, so that the accuracy of identifying the abnormal state is improved. This is one of the key points of the invention.

3. According to the invention, an abnormal state identification model is provided according to various abnormal states of different primary equipment, the model is a composite model, and aiming at different mode labels obtained by a micro classifier, the abnormal state identification model can complete the identification of different abnormal states, so that the synchronous identification of various abnormal states is realized at the same time. This is one of the key points of the invention.

Additional features and advantages of the invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. The objectives and other advantages of the invention will be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings.

Drawings

The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the principles of the invention and not to limit the invention. In the drawings:

fig. 1 is a flow chart of a method for identifying abnormal states of primary equipment of a power distribution network according to the present invention;

FIG. 2 is a schematic flow chart of the output duty type of the macro classifier according to the present invention;

FIG. 3 is a schematic diagram of a condition classification model according to the present invention;

FIG. 4 is a diagram illustrating the training results of a condition classification model according to the present invention;

FIG. 5 is a schematic diagram of a pattern classifier according to the present invention.

FIG. 6 is a schematic diagram of a pattern classifier training method according to the present invention.

Fig. 7 is a schematic flow chart of identifying an insulator flashover abnormality according to the first embodiment of the invention.

Fig. 8 is a schematic flow chart of identifying a line discharge abnormality according to the second embodiment of the present invention.

Fig. 9 is a schematic flow chart of identifying zero-sequence current abnormality according to the third embodiment of the present invention.

Fig. 10 is a schematic diagram of a line topology for identifying false triggering of switch protection according to a fourth embodiment of the present invention.

Fig. 11 is a schematic diagram of coloring a line topology for identifying false triggering of switch protection according to a fourth embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail below with reference to the accompanying drawings.

First embodiment

Fig. 1 is a schematic flow chart of a method for identifying an abnormal state of primary equipment of a power distribution network according to the embodiment. The method is described below with reference to fig. 1.

In this embodiment, the recording data is first input to the macro classifier and the micro classifier, and the working condition classification and the mode label are respectively obtained, where the working condition types output by the macro classifier in this embodiment include: power failure, power restoration, grounding, short circuit and excitation surge current; the pattern labels output by the onlooker classifier include: insulator flashover, line discharge, zero sequence current anomaly and switch short circuit protection false triggering.

And then, inputting the obtained working condition type and the mode label into the abnormal state identification model, and selecting the corresponding identification model according to the mode label result for judgment.

When the mode tag output is an insulator flashover, as shown in fig. 7, the present embodiment identifies an abnormal state by using a decision tree model, and the input values of the model are shown in table 5 below.

TABLE 5

Characteristic reference number	Characteristic name
		F1	Whether the type of operation is grounded
F2	Mode label is insulator flashing
		F3	Number of suspected insulator flashover within about 30 days
F4	Whether it rains or not

The results of the abnormal state recognition are shown in table 6.

TABLE 6

Output label	Output label
		T1	Non-insulator flashway
T2	Suspected insulator flash circuit
		T3	Insulator flashway

The identification process aiming at the abnormal flash circuit of the insulator is as follows:

1. judging whether the working condition type is grounding (F1), if not, directly outputting a non-insulator flash (T1); if yes, entering the next judgment;

2. judging whether the mode label is an insulator flashover (F2), if not, outputting a non-insulator flashover (T1), and if so, entering the next judgment;

3. judging the numerical relationship between the number of times of occurrence of the suspected insulator flash in nearly 30 days and a first threshold value N1, a second threshold value N2 and a third threshold value N3, and when the number of times of occurrence of the suspected insulator flash is smaller than the first threshold value N1, outputting an abnormal state judgment result as the suspected insulator flash (T2); when the number of times of the suspected insulator flash is larger than a third threshold value N3, outputting an abnormal state judgment result as an insulator flash (T3); and when the number of times of the suspected insulator flash is larger than a second threshold value N2, the next judgment is carried out.

4. Judging whether the recording occurs in a time period and raining, if not, outputting an abnormal state judgment result as a suspected insulator flash (T2); if yes, the judgment result of the abnormal state is output as an insulator flashover (T3).

In the above determination process, the time length of "the number of times of occurrence of the suspected insulator flashover within approximately 30 days" represented by the feature number F3 is determined according to the practical experience of the engineering, and may be adjusted to 10 days, 15 days, or 20 days according to the practical situation. The first threshold N1, the second threshold N2, and the third threshold N3 are obtained by model training of the decision tree model.

Second embodiment

When the mode tag output is line discharge, as shown in fig. 8, the present embodiment uses a decision tree model to identify an abnormal state. The input characteristics of the model are shown in table 7.

TABLE 7

Characteristic reference number	Characteristic name
		F1	Whether the operating condition is grounded
F2	Mode label is line discharge
		F3	Wind power class

Table 8 shows the output values of the model.

TABLE 8

Output label	Output label
		T1	Non-line discharge
T2	Line discharge

The identification process aiming at the line discharge abnormity comprises the following steps:

1. judging whether the working condition type is grounded (F1), if not, directly outputting non-line discharge (T1); if yes, entering the next judgment;

2. judging whether the mode label is line discharge (F2), if not, outputting a non-insulator flash (T1), and if so, entering the next judgment;

3. and judging whether the wind power level in the wave recording period exceeds a threshold value N, if not, outputting a non-insulator flash (T1), and if so, outputting a line discharge judgment result (T2).

Third embodiment

When the output of the mode tag is abnormal zero-sequence current, as shown in fig. 9, the present embodiment identifies the abnormal state by using a decision tree model. The input characteristics of the model are as in table 9.

TABLE 9

Characteristic reference number	Characteristic name
		F1	Whether the operating condition is grounded
F2	Zero sequence current amplitude

Table 10 shows the output values of the model.

Watch 10

Output label	Output label
		T1	Unknown zero sequence current state
T2	Zero sequence current anomaly

The identification process aiming at the zero sequence current abnormity comprises the following steps:

1. judging whether the working condition type is grounded (F1), if not, directly outputting the zero sequence current state (T1); if yes, entering the next judgment;

2. judging whether the mode label is abnormal zero sequence current (F2), if not, outputting the zero sequence current with unknown state (T1), if so, outputting the zero sequence current with abnormal state;

fourth embodiment

Such anomalies are usually due to improper switch protection settings, which result in a malfunction that cannot avoid the magnetizing inrush current. The abnormal conditions are represented by the fact that the recording wave before switching comprises the excitation inrush current working condition, and the recording wave after switching comprises the power restoration, the excitation inrush current and the power failure. Such anomalies need to be determined from topology analysis, and the required information includes: and outputting all the recording working conditions and the line topology under the topology, wherein the output comprises abnormal reasons and abnormal sections.

And when the mode label output is the switch short-circuit protection false triggering, the abnormal state identification model starts the switch short-circuit protection false triggering positioning search. As shown in fig. 10, in the logical structure, the line topology is a tree in which a substation (or a bus) is used as a root node and the wave recording device is used as a node, and the line topology description information also includes necessary information for obtaining a corresponding relationship between the wave recording device and the wave recording device. The abnormal recognition process comprises three steps of wave recording correspondence, topology coloring and topology searching:

1. and recording the waves correspondingly. The wave recording data corresponds to the wave recording equipment triggering the wave recording one by one;

2. and (5) coloring the macro topology. Marking the color of the wave recording equipment in the topology according to the working condition, wherein the working condition is only the excitation inrush current and is marked as R, the point of the working condition including the power restoration, the excitation inrush current and the power failure is marked as G, and the other types are marked as B, as shown in FIG. 11;

3. and searching the topology. The route containing the red node is found first, and then the whole route is searched in the order from the root node (substation) to the leaf node. Marking the last R node as an initial position S, marking the first G node after S as an end position E, and outputting 'switch short circuit protection misoperation exists between S and E'; if there is no green node after S, then output "there is a switch short circuit protection malfunction after S".

The above description is only an embodiment of the present invention, and the protection scope of the present invention is not limited thereto, and any person skilled in the art should modify or replace the present invention within the technical specification of the present invention.

Claims

1. A method for identifying abnormal states of primary equipment of a power distribution network is characterized by comprising the following steps:

2. The method for identifying the abnormal state of the primary equipment of the power distribution network according to claim 1, wherein the macro classifier uses a classifier model comprising a plurality of three-layer feedforward neural networks, and the working condition classification comprises power failure, power restoration, grounding, short circuit and excitation inrush current.

3. The method for identifying the abnormal state of the primary equipment of the power distribution network according to claim 1, wherein the micro classifier uses a deep neural network model comprising a convolution layer area and a full connection layer area, and the mode labels comprise lightning arrester breakdown, insulator flashover, line discharge, switch vacuum bubble breakdown, line breakage and switch short circuit error protection.

4. The method for identifying the abnormal state of the primary equipment of the power distribution network according to claim 1, wherein the abnormal state identifier comprises a plurality of abnormal state identification submodels aiming at a plurality of abnormal state types, and the plurality of abnormal state identification submodels comprise a decision tree model and a topology search model.

5. The method for identifying the abnormal state of the primary equipment of the power distribution network according to claim 1, wherein the line operation information comprises information such as temperature, wind power and rainfall; the line topology information includes line distribution and primary equipment location.

6. The method for identifying the abnormal state of the primary equipment of the power distribution network according to claim 1, wherein the time domain features extracted from the fault recording data comprise: the three-phase unbalance degree of the three-phase current is determined according to the current cycle maximum value, the current cycle minimum value, the current cycle mean value, the current cycle variance, the current cycle root mean square, the current cycle minimum root mean square, the current cycle.

7. The method for identifying the abnormal state of the primary equipment of the power distribution network according to claim 1, wherein the frequency domain features extracted from the fault recording data comprise: the dc component content, the second harmonic component and the second harmonic component content of the current.

8. An identification device for an abnormal state of primary equipment of a power distribution network is characterized by comprising:

9. The power distribution network primary equipment abnormal state recognition device as claimed in claim 8, wherein the macro classifier uses a classifier model comprising a plurality of three-layer feedforward neural networks.

10. The power distribution network primary equipment abnormal state recognition device according to claim 8, wherein the micro classifier uses a deep neural network model including a convolutional layer region and a full connection layer region.