CN116451170A - A distribution network detection method based on multi-source heterogeneous data - Google Patents

Abstract

The invention discloses a power distribution network detection method based on multi-source heterogeneous data, which comprises the following steps: acquiring multi-source heterogeneous data of a power distribution network history, and classifying according to corresponding equipment types and voltage levels; for any type of multi-source heterogeneous data, carrying out feature extraction on the basis of data layer fusion, and carrying out fusion processing on a feature layer to obtain a fused data feature set; training a power distribution network detection model by using a data characteristic set; and inputting data acquired in real time to detect the power distribution network by using the power distribution network detection model after training. According to the method, for the multi-source heterogeneous data of the power distribution network equipment acquired from different paths, the feature set in the data is extracted after the multi-source heterogeneous data are fused, the feature set is used for training a power distribution network detection model to detect the power distribution network, the association fusion of the data is not needed, the processing speed of the acquired data is higher, and the power distribution network detection efficiency is higher.

Description

A distribution network detection method based on multi-source heterogeneous data

technical field

The invention relates to the technical field of distribution network, in particular to a detection method of distribution network based on multi-source heterogeneous data.

Background technique

The distribution network is at the end of the power system, and has the distinctive characteristics of wide geographical distribution, large-scale power grid, various types of equipment, various network connections, and variable operation modes. With the popularization and application of application systems such as distribution automation and power consumption information collection, for large-scale distribution networks, massive heterogeneous and polymorphic data will be generated exponentially, and traditional single data analysis technology can no longer meet the demand. When testing the distribution network, due to the different types and methods of data collection for the equipment in it, the collected data are different, the source of the original data is not uniform, and the data storage structure is diversified, resulting in the same power grid equipment in the same network. There are large differences in data structures such as codes, types, names, and associations in different systems, and it is impossible to directly conduct comprehensive query and statistics of ledger information, operating information, and spatial information on the same device. Multi-source heterogeneous Subsequent operations can only be performed after data fusion processing. The traditional data fusion method mainly relies on management methods, coordinates various data management departments to realize the associated fusion of stock data through manual correspondence, and integrates the business processes of each data management department to realize the associated fusion of incremental data. Although this method can achieve data fusion, for a large number of equipment and various data sources in the distribution network, it takes a lot of manpower and material resources to integrate. At the same time, when new equipment is connected or new data needs to be collected , and the corresponding data association fusion method must be updated again, which seriously affects the efficiency of distribution network detection.

"A classification method based on multi-source data fusion" disclosed in Chinese patent literature, its publication number is CN107247787A, and the publication date is 2017-10-13, through data combing, data personalized classification, and multi-source data fusion classification step-by-step realization; data sorting: respectively sorting out government data, social data, Internet data producers and data; data personalized classification: according to the respective attributes of government data, social data, and Internet data, classify them according to different dimensions; Multi-source data fusion classification: According to the respective classifications of government data, social data, and Internet data, find common classification dimensions, carry out fusion classification according to themes and industries, establish public theme and industry classification systems and their respective personalized classification dimensions, and realize government data , social data, and Internet data are deeply integrated. However, this technology is only a classification of multi-source data fusion in the general framework and general direction. When it comes to the multi-source heterogeneous data fusion in the specific distribution network detection process, it still needs to realize the correlation fusion of multi-source heterogeneous data through manual correspondence. , time-consuming and labor-intensive and because the data processing process takes a long time, it seriously affects the efficiency of distribution network detection.

Contents of the invention

The present invention aims to overcome the problem in the prior art that a large number of equipment collection data in the distribution network still needs manual assistance to realize the association fusion of multi-source heterogeneous data, thereby affecting the detection efficiency of the distribution network, and provides a multi-source heterogeneous Distribution network detection method of structured data, for the multi-source heterogeneous data of distribution network equipment collected from different channels, the feature set in the data is extracted after the fusion of multi-source heterogeneous data, and the feature set is used to train The detection model of the distribution network is used to detect the distribution network, which does not require artificial correlation and fusion of data, the processing speed of the collected data is faster, and the efficiency of distribution network detection is higher.

In order to achieve the above object, the present invention adopts the following technical solutions:

A distribution network detection method based on multi-source heterogeneous data, including:

Obtain multi-source heterogeneous data of distribution network history, and classify according to the corresponding equipment type and voltage level;

For any type of multi-source heterogeneous data, feature extraction is performed on the basis of data layer fusion, and the feature layer is fused to obtain the fused data feature set;

Use the data feature set to train the distribution network detection model;

Use the distribution network detection model that has been trained to input real-time collected data for distribution network detection.

In the present invention, the multi-source heterogeneous data is first classified according to the corresponding equipment type and voltage level, that is, all the multi-source heterogeneous data collected from the equipment of the same voltage level and equipment type are grouped into one category, and each type of data Carry out data fusion and establish corresponding distribution network detection models for training and detection, that is, each distribution network detection model only corresponds to detecting the fault or status of the same type of equipment, and then the detection results of all distribution network detection models Comprehensively obtain the faults and status of all types of equipment in the distribution network, thus reflecting the detection results of the entire distribution network; for the real-time collection data input into the distribution network detection model, advanced classification and multi-source heterogeneous data are also required The fusion and feature extraction are input into the model to obtain the corresponding detection results; the whole process does not require manual fusion of data, and the detection results can be obtained directly by using the distribution network detection model. The data processing speed is fast and the efficiency is high.

Preferably, for the multi-source heterogeneous data of distribution network history, preprocessing is performed on each type of multi-source heterogeneous data after classification;

Perform time registration on multi-source heterogeneous data, and use Lagrangian difference method to fill the data;

Data mutual verification is performed for the same type of multi-source heterogeneous data collected from different sources.

The preprocessing for multi-source heterogeneous data in the present invention includes evaluating, cleaning, filtering and reorganizing the data, etc., which belong to the existing common technologies and therefore will not be described in detail; in addition, these data are collected and can be obtained It is considered to be time-series data collected in chronological order; due to the different sampling periods of different collection channels, multi-source heterogeneous data will have time mismatch problems, so it is necessary to time-register the asynchronous data; data verification: yes In view of the characteristics of wide data sources and multiple channels, data from different sources are mutually checked to ensure data consistency, completeness, and accuracy, including mutual checking of electricity and quantity measurement, and mutual checking between different data systems , Mutual verification of data with different structures, etc.

Preferably, the process of performing fusion processing on the feature layer to obtain the fused data feature set includes:

The data after data layer fusion is divided into several subspaces according to different data types;

For the data set in the subspace s Extract its feature data set Y _s = X _s A, /> A is the projection matrix, n _s and k _s are the dimensions of X _s and Y _s respectively, and n is the number of data in the subspace;

Fuse the feature data sets of each subspace to get the final data feature set

In the present invention, the dimension of the data set after data layer fusion is relatively high, and the types and data characteristics of different variables may be different, and redundant information often exists. In order to remove redundant information and noise and extract features more accurately, it is necessary to divide the data set after data layer fusion into subspaces according to different data types; divide data of the same data type and similar data characteristics into the same Subspace, and then use the feature extraction method to extract features in each subspace, and fuse the features, and use the projection matrix to obtain the feature data set, which can not only maintain the global structure of the data set, but also maintain the local structure of the data set.

Preferably, the process of extracting feature data sets in the subspace includes:

The global structure of the data set in the subspace s is represented by the mutual information matrix MI(X _s ); and the objective function is constructed:

J＝ _Jg / _Jl

J _g ＝maxA ^T (MI(X _s ))A

Where y _i and y _j represent any data in the feature data set and its neighbors, W _ij is a weight matrix describing the relationship between any data and its neighbors;

The projection matrix A is solved based on the Laplace multiplier, so that the feature data set Y _s is obtained from the data set X _s .

In the present invention, the Gaussian kernel distance is kept unchanged during the extraction process from the original data set to the feature data set, so that the adjacent relationship in the original data set is maintained, and in order to characterize the global structure of the data set, mutual information is used to describe the overall data set If the relationship between the two variables is close, the mutual information value is larger, otherwise the mutual information value is smaller; the dimension of the feature data set calculated by the projection matrix is smaller than the original data set, which can reduce the dimension and pass the dimensionality reduction Eliminate noise.

Preferably, the data layer is fused to convert multi-source heterogeneous data into data in a unified data format;

Correspond each heterogeneous data source in the multi-source heterogeneous data to a JSON data schema document;

Use JSON Schema to map the data source, and project the data of the data source to a JSON string.

In the present invention, JSON is a light-weight data exchange format with language independence and platform independence, which not only makes it easy for people to read and write, but also facilitates generation and analysis; the various characteristics of JSON make it an ideal data exchange Language, as a general data format, overcomes the interaction between heterogeneous data sources, greatly improving the speed and efficiency of data interaction processing; JSON, as a data transmission format, is independent and transparent between heterogeneous data sources, without knowing other data source information, It can flexibly represent objects and express various types of data. As a general structured data representation, it can effectively realize format conversion and fusion of heterogeneous data sources.

As a preference, in the fusion process of multi-source heterogeneous data, a heterogeneous data evaluation formula based on authenticity and error omission is established:

q(x)=αrp _i +(1-α)ra _i

x is the data source; α is the proportional factor; rp _i is the degree of authenticity, indicating the ratio of the number of true values provided by the data source to the number of all true values; ra _i is the degree of error omission, indicating the number of true values provided by the data source The ratio of the number of data that is not true to all the numbers that are not true; according to the evaluation formula, the data is selected when the fusion of multiple heterogeneous data conflicts.

In the present invention, when performing multi-source heterogeneous data fusion, there may be a problem of data source conflict, so it is necessary to evaluate multi-source heterogeneous data. When data fusion conflicts, data can be calculated based on the evaluation results of conflicting data sources The choice of source, so as to solve the problem of data conflict; the degree of truth is to evaluate the data source from the perspective of the correct truth value contained in the data source, and the degree of error and omission is to evaluate the data source from the perspective of the wrong truth value.

Preferably, the process of training the distribution network detection model with the data feature set includes:

Divide the data feature set into a training set and a test set, and build a learning model based on a deep neural network;

Use the training set to train the learning model of the deep neural network, and use the test set to test the trained learning model;

When the detection accuracy of the model tested by the test set meets the detection requirements, save the model as the distribution network detection model.

In the present invention, train by the learning model of deep neural network, divide training set and test set according to certain ratio, obtain the corresponding parameter of learning model with training set training, then test the detection precision and accuracy rate of learning model with test set; If required, the training is completed to obtain the distribution network detection model; in the subsequent detection of the distribution network, the data collected in real time is preprocessed and then the data layer fusion and feature layer fusion are performed to obtain the data feature set of real-time data, and input into the power distribution network The detection results are obtained in the network detection model.

As a preference, after performing time registration and data filling on multi-source heterogeneous data, it is also necessary to denoise the data:

Among them, pmu _i is the real value of the time series data at time i, y _i is the smoothing value of the first step of the data, trend _i is the second smoothing value of time i, h is the forecast step size, β ₁ and β ₂ are the trend smoothing parameters respectively .

In the present invention, due to the different sampling frequencies of the acquisition devices corresponding to different multi-source heterogeneous data, in order to ensure the same density of time-series data, it is necessary to fill in the corresponding time-series data; the noise reduction of the time-series data is performed on the basis of completing the time-series filling, for subsequent Data fusion provides the basis.

The present invention has the following beneficial effects: for the multi-source heterogeneous data of distribution network equipment collected from different ways, the feature set in the data is extracted after the fusion of multi-source heterogeneous data, and the distribution network is trained using the feature set. The power grid detection model detects the distribution network without the need for artificial data correlation and fusion, the processing speed of the collected data is faster, and the efficiency of distribution network detection is higher; through time registration, all data are synchronized in time, At the same time, the time series data filling provides the basis for the subsequent data fusion; first through the fusion of the data layer and then the fusion of the feature layer, and the use of the projection matrix can not only maintain the global structure of the data set in the process of feature extraction, but also maintain the data set local structure.

Description of drawings

Fig. 1 is a flowchart of a distribution network detection method based on multi-source heterogeneous data in the present invention.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

Embodiment 1, as shown in Figure 1, a distribution network detection method based on multi-source heterogeneous data, including:

Obtain multi-source heterogeneous data of distribution network history, and classify according to the corresponding equipment type and voltage level;

For the multi-source heterogeneous data of distribution network history, after classification, preprocess each type of multi-source heterogeneous data; perform time registration on multi-source heterogeneous data, and use Lagrange difference method to fill in the data;

Data mutual verification is performed for the same type of multi-source heterogeneous data collected from different sources.

For any type of multi-source heterogeneous data, feature extraction is performed on the basis of data layer fusion, and the feature layer is fused to obtain the fused data feature set;

The data layer is fused into data that converts multi-source heterogeneous data into a unified data format;

Correspond each heterogeneous data source in the multi-source heterogeneous data to a JSON data schema document; when the data source changes, you only need to modify the corresponding schema file, which can realize the plug-and-play of the data source;

Use JSON Schema to map the data source, and project the data of the data source to a JSON string.

The process of fusing the feature layer to obtain the fused data feature set includes:

The data after data layer fusion is divided into several subspaces according to different data types;

For the data set in the subspace s Extract its feature data set Y _s = X _s A, /> A is the projection matrix, m _s and k _s are the dimensions of X _s and Y _s respectively (also the number of observed variables), n is the number of data in the subspace; before extracting features, the data samples are first normalized Processing, that is, each variable is subtracted from its respective mean and then divided by its respective standard deviation; the feature data sets of each subspace are fused to obtain the final data feature set/>

Use the data feature set to train the distribution network detection model;

Use the distribution network detection model that has been trained to input real-time collected data for distribution network detection.

The process of training the distribution network detection model with the data feature set includes:

Divide the data feature set into a training set and a test set, and build a learning model based on a deep neural network;

Use the training set to train the learning model of the deep neural network, and use the test set to test the trained learning model;

When the detection accuracy of the model tested by the test set meets the detection requirements, save the model as the distribution network detection model.

The process of extracting feature datasets within the subspace includes:

The global structure of the data set in the subspace s is represented by the mutual information matrix MI(X _s ); and the objective function is constructed:

J＝ _Jg / _Jl

J _g ＝maxAT(MI(X _s ))A

Among them, y _i and y _j represent any data in the feature data set and its neighbors, W _ij is the weight matrix describing the relationship between any data and its neighbors; the Gaussian kernel is used to describe the distance between the data and its neighbors, and the original space to the feature space strives to The Gaussian kernel distance remains constant:

both for /> samples in .

The objective function J has constraints:

A ^T X ^T (DW)XA＝I

Based on Laplacian multipliers can be obtained by

(MI(X _s ))A＝λX ^T (DW)XA

Solve the projection matrix A, so as to obtain the feature data set Y _s =X _s A according to the data set X _s .

After time registration and data filling of multi-source heterogeneous data, the data needs to be denoised:

Among them, pmu _i is the real value of the time series data at time i, y _i is the smoothing value of the first step of the data, trend _i is the second smoothing value of time i, h is the forecast step size, β ₁ and β ₂ are the trend smoothing parameters respectively .

In the present invention, the multi-source heterogeneous data is first classified according to the corresponding equipment type and voltage level, that is, all the multi-source heterogeneous data collected from the equipment of the same voltage level and equipment type are grouped into one category, and each type of data Carry out data fusion and establish corresponding distribution network detection models for training and detection, that is, each distribution network detection model only corresponds to detecting the fault or status of the same type of equipment, and then the detection results of all distribution network detection models Comprehensively obtain the faults and status of all types of equipment in the distribution network, thus reflecting the detection results of the entire distribution network; for the real-time collection data input into the distribution network detection model, advanced classification and multi-source heterogeneous data are also required The fusion and feature extraction are input into the model to obtain the corresponding detection results; the whole process does not require manual fusion of data, and the detection results can be obtained directly by using the distribution network detection model. The data processing speed is fast and the efficiency is high.

The preprocessing for multi-source heterogeneous data in the present invention includes evaluating, cleaning, filtering and reorganizing the data, etc., which belong to the existing common technologies and therefore will not be described in detail; in addition, these data are collected and can be obtained It is considered to be time-series data collected in chronological order; due to the different sampling periods of different collection channels, multi-source heterogeneous data will have time mismatch problems, so it is necessary to time-register the asynchronous data; data verification: yes In view of the characteristics of wide data sources and multiple channels, data from different sources are mutually checked to ensure data consistency, completeness, and accuracy, including mutual checking of electricity and quantity measurement, and mutual checking between different data systems , Mutual verification of data with different structures, etc.

In the present invention, the dimension of the data set after data layer fusion is relatively high, and the types and data characteristics of different variables may be different, and redundant information often exists. In order to remove redundant information and noise and extract features more accurately, it is necessary to divide the data set after data layer fusion into subspaces according to different data types; divide data of the same data type and similar data characteristics into the same Subspace, and then use the feature extraction method to extract features in each subspace, and fuse the features, and use the projection matrix to obtain the feature data set, which can not only maintain the global structure of the data set, but also maintain the local structure of the data set.

In the present invention, the Gaussian kernel distance is kept unchanged during the extraction process from the original data set to the feature data set, so that the adjacent relationship in the original data set is maintained, and in order to characterize the global structure of the data set, mutual information is used to describe the overall data set If the relationship between the two variables is close, the mutual information value is larger, otherwise the mutual information value is smaller; the dimension of the feature data set calculated by the projection matrix is smaller than the original data set, which can reduce the dimension and pass the dimensionality reduction Eliminate noise.

In the present invention, JSON is a light-weight data exchange format with language independence and platform independence, which not only makes it easy for people to read and write, but also facilitates generation and analysis; the various characteristics of JSON make it an ideal data exchange Language, as a general data format, overcomes the interaction between heterogeneous data sources, greatly improving the speed and efficiency of data interaction processing; JSON, as a data transmission format, is independent and transparent between heterogeneous data sources, without knowing other data source information, It can flexibly represent objects and express various types of data. As a general structured data representation, it can effectively realize format conversion and fusion of heterogeneous data sources.

In the present invention, train by the learning model of deep neural network, divide training set and test set according to certain ratio, obtain the corresponding parameter of learning model with training set training, then test the detection precision and accuracy rate of learning model with test set; If required, the training is completed to obtain the distribution network detection model; in the subsequent detection of the distribution network, the data collected in real time is preprocessed and then the data layer fusion and feature layer fusion are performed to obtain the data feature set of real-time data, and input into the power distribution network The detection results are obtained in the network detection model.

In the present invention, due to the different sampling frequencies of the acquisition devices corresponding to different multi-source heterogeneous data, in order to ensure the same density of time-series data, it is necessary to fill in the corresponding time-series data; the noise reduction of the time-series data is performed on the basis of completing the time-series filling, for subsequent Data fusion provides the basis.

In the embodiment of the present invention, before performing time registration, it is necessary to detect the inferior value of the abnormal measurement. It is mainly based on the comparison between the correction value and the threshold value to determine whether the measured value is an inferior value of the abnormal measurement. The correction form is:

Among them, Z(k) is the correction value of the k-th index, r _new (k) is the new information, and X is the judgment parameter of abnormal measurement; the measured value can be obtained through the correction value, and it is used as the weight of the real estimated value, Express the weighting function form as:

λ(k)=e ^-c·μ(k)

Where c is a constant, and -c·μ(k) is the weight of the estimated value μ.

On the above basis, data registration is carried out, and the least square method is used to perform time registration on the two types of sensor data of A and B. The sampling period of A is τ, the sampling period of B is T, and the proportional coefficient of the sampling period is set to an integer. n; A's latest sampling time is recorded as (k-1)τ, and the current time is recorded as kτ=(k-1)τ+nT. In terms of time registration, the measured value obtained by sensor B is mainly recorded as a virtual measurement value and fuse that measurement with the measurement from Sensor A data.

Record the measurement sequence of B as Z _n = (z ₁ ,z ₂ ,…,z _n ) ^T , where z _n represents the measurement value of sensor A, (z ₁ ,z ₂ ,…,z _n ) ^T is expressed as n A measurement set composed of fusion values of estimated values and their reciprocals, and the new expression of the measurement is written as:

z _i ＝Z _n +(in)TZ _n +v _i

v _i is the noise value that appears in the data measurement process; and the measurement vector of sensor B is expressed as:

Where T' is the fusion time, after the above calculation, the collected data will be time-registered.

The time series data is filled with the Lagrangian interpolation method, and the corresponding Lagrangian polynomial function is expressed as:

Among them, Function(t) _scada is the Lagrangian difference function formula corresponding to the time series data, l _j (t) is the interpolation basis function, t _s is the time corresponding to the time series data s, and t _j is the time series data. Perform time series data denoising after filling the time series data.

Embodiment 2. On the basis of Embodiment 1, in the fusion process of multi-source heterogeneous data, a heterogeneous data evaluation formula based on authenticity and error omission is established:

q(x)=αrp _i +(1-α)ra _i

x is the data source; α is the scaling factor;

rp _i is the degree of authenticity, which means the ratio of the number of true values of the data provided by the data source to the number of all true values:

Ra _i is the degree of error omission, which means the ratio of the number of data not provided by the data source that is not true to all the numbers that are not true:

Among them, tp _i represents the number of data provided by data source i that is true, fp _i represents the number of data provided by data source i that is not true, and tn _i represents the number of data not provided by data source i that is not true The number, fn _i represents the number of true values in the data not provided by the data source i; according to the evaluation formula, the data is selected when the multivariate heterogeneous data fusion conflicts.

In the present invention, when performing multi-source heterogeneous data fusion, there may be a problem of data source conflict, so it is necessary to evaluate multi-source heterogeneous data. When data fusion conflicts, data can be calculated based on the evaluation results of conflicting data sources The choice of source, so as to solve the problem of data conflict; the degree of truth is to evaluate the data source from the perspective of the correct truth value contained in the data source, and the degree of error and omission is to evaluate the data source from the perspective of the wrong truth value.

The above-mentioned embodiment is a further elaboration and description of the present invention, so as to facilitate understanding, and is not any limitation to the present invention. Any modification, equivalent replacement and improvement made within the spirit and principle of the present invention shall be included in this document. within the scope of protection of the invention.

Claims (8)

1. The utility model provides a distribution network detection method based on heterogeneous data of multisource, which is characterized by comprising the following steps:

acquiring multi-source heterogeneous data of a power distribution network history, and classifying according to corresponding equipment types and voltage levels;

for any type of multi-source heterogeneous data, carrying out feature extraction on the basis of data layer fusion, and carrying out fusion processing on a feature layer to obtain a fused data feature set;

training a power distribution network detection model by using a data characteristic set;

and inputting data acquired in real time to detect the power distribution network by using the power distribution network detection model after training.

2. The method for detecting the power distribution network based on the multi-source heterogeneous data according to claim 1, wherein the multi-source heterogeneous data of the power distribution network history is subjected to preprocessing after being classified;

performing time registration on the multi-source heterogeneous data, and filling the data by adopting a Lagrangian difference method;

and carrying out data mutual check on the same type of multi-source heterogeneous data of different acquisition sources.

3. The method for detecting the power distribution network based on the multi-source heterogeneous data according to claim 1 or 2, wherein the process of fusing the feature layers to obtain the fused data feature set comprises the following steps:

dividing a plurality of subspaces for the data after the fusion of the data layers according to different data types;

data sets within subspace sExtracting its characteristic dataset Y _s ＝X _s A，/>A is projection matrix, m _s And k _s Respectively X _s And Y _s Dimension of (2)The degree, n, is the number of data in the subspace;

fusing the feature data sets of each subspace to obtain a final data feature set

4. A method of detecting a power distribution network based on heterogeneous multi-source data according to claim 3, wherein the process of extracting the feature data set in the subspace comprises:

in a mutual information matrix MI (X _s ) Representing the global structure of the dataset within subspace s; and constructing an objective function:

J＝J _g /J _l

J _g ＝maxA ^T (MI(X _s ))A

wherein y is _i And y _j Representing any one of the feature data set and its neighbors, W _ij A weight matrix for describing any one data and its neighbor relation;

solving the projection matrix A based on Laplace multiplier, thereby from the data set X _s Obtaining a characteristic data set Y _s 。

5. The method for detecting the power distribution network based on the multi-source heterogeneous data according to claim 1, 2 or 4, wherein the data layer is fused into data which converts the multi-source heterogeneous data into a unified data format;

each heterogeneous data source in the multi-source heterogeneous data corresponds to one JSON data mode document;

and mapping the data source by using the JSON schema, and projecting the data of the data source to the JSON string.

6. The method for detecting the power distribution network based on the multi-source heterogeneous data according to claim 5, wherein in the process of fusion of the multi-source heterogeneous data, a heterogeneous data evaluation formula based on the fidelity and the error omission is established:

q(x＝αrp _i +(1-α)ra _i

x is a data source; alpha is a scale factor; rp (rp) _i For the fidelity, the data provided by the data source is represented as the ratio of the number of true values to the number of all true values; ra _i The error omission degree is the ratio of the number of data provided by the data source and not being true value to the number of all data not being true value; and (5) accepting and rejecting the data when the fusion of the multi-element heterogeneous data is in conflict according to an evaluation formula.

7. A method for detecting a power distribution network based on heterogeneous multi-source data according to claim 1, 2, 4 or 6, wherein training the power distribution network detection model with the data feature set comprises:

dividing the data characteristic set into a training set and a testing set, and building a learning model based on a deep neural network;

training a learning model of the deep neural network by using the training set, and testing the trained learning model by using the testing set;

and when the detection precision of the model tested by the test set meets the detection requirement, storing the model as a power distribution network detection model.

8. The method for detecting a power distribution network based on multi-source heterogeneous data according to claim 2, wherein after time registration and data filling are performed on the multi-source heterogeneous data, noise reduction is further performed on the data:

therein pmu _i Is the real value of time sequence data of time i, y _i A smoothed value for a data one step size, trend _i A secondary smooth value of time i, h is a prediction step length, beta ₁ And beta ₂ The trend smoothing parameters are respectively.