CN111008584B - Power quality measurement missing repair method for fuzzy self-organizing neural network - Google Patents

Abstract

The application discloses a method for repairing the loss of electric energy quality measurement data based on a fuzzy self-organizing neural network, which is implemented by a computer program: the method comprises the following steps: 1) Inputting a one-dimensional measurement data set containing the missing electric energy quality; 2) Sectionally intercepting the sampled one-dimensional waveform data according to N power quality sampling periods; 3) Converting the matrix x into an image by adopting a graying method; 4) Extracting the characteristic value of the two-dimensional harmonic gray scale map, and carrying out normalization processing; 5) Determining the optimal clustering number of the harmonic data of the power grid; 6) Constructing an objective function by taking the weighted square sum of the distances from each sample to all cluster centers as a target; 7) Repairing the two-dimensional gray scale map is completed; 8) The data of each layer of repaired data are fused, and experimental data analysis is carried out, so that the method provided by the application has lower repair error and higher signal to noise ratio under the conditions of low data loss rate and high data loss rate compared with the existing algorithm under the conditions of random loss or continuous loss.

Description

A method for repairing power quality measurement defects using fuzzy self-organizing neural network

Technical field

The present invention relates to a method for repairing missing measurement data, and further relates to a method for repairing missing data for power quality measurement, and in particular to a method for repairing missing data for power quality measurement using a fuzzy self-organizing neural network.

Background technique

The Ubiquitous Electric Internet of Things (UEIoT) realizes comprehensive perception and intelligent measurement of the power system, providing strong information support for the safety, stability, and economic operation of the power grid. The ubiquitous sensing big data underlying UEIoT is the basis for situational awareness and status identification of the entire system. Among them, power grid harmonic monitoring data is the key to mastering harmonic laws, achieving harmonic control, and improving power quality. However, whether the classic Nyquist sampling or compressed sensing sampling method is used, there is often a problem of losing part of the collected harmonic signals due to faults in sensors, transmission equipment, conversion equipment, etc.; or in communication channels, such as Power line carrier, the phenomenon of data loss due to channel interference during propagation. Due to the non-repeatability of power grid data collection, when there is insufficient redundancy, the conclusions drawn using missing harmonic data for analysis will undoubtedly deviate greatly from the correct rules; while when compressed sampling is used to reconstruct the signal , Since each sampling point contains a large amount of information, the loss of each sampling value will have a huge impact on signal reconstruction. Therefore, how to accurately and effectively repair missing data and restore the original appearance of the collected data is the focus of harmonic waveform data management.

The repair strategy provided by the present invention is not only based on the significant measurement timing characteristics of the power grid power quality data, but also based on the data autocorrelation and harmonic change regularity, and through the similar relationship between different data to repair the missing data, which greatly improves the repair strategy. Reduced repair errors.

Contents of the invention

The technical problem to be solved by the present invention is to provide a fuzzy self-organizing neural network power quality measurement data missing repair method, whether in the case of random missing or continuous missing, under low data loss rate and high data loss rate. Lower repair errors and higher signal-to-noise ratio.

The technical solution adopted by the present invention is: a repair method suitable for missing power quality measurement data, which includes the following stages:

1) Input the missing one-dimensional power quality measurement data set;

2) Intercept the sampled one-dimensional waveform data in segments according to N power quality sampling periods (N is 20 according to experimental analysis), and use the mapping rule of the two-dimensional matrix x x(i,j)=x( n)|n=(i-1)n _l +jj≤n _l , map it into rows or columns in the two-dimensional matrix x, and perform two-dimensional truncation and reorganization of the one-dimensional signal;

3) Calculate the spatial gray value P(i,j)=[(255-(254*(m-x(i,j)/(m-n))]], and convert the matrix x into an image using grayscale method;

4) Extract the eigenvalues X _j =[X _1,j ,X _2,j ,…,X _m+l,j ] of the two-dimensional harmonic grayscale image and perform normalization processing;

5) Determine the optimal number of clusters for power grid harmonic data, including:

(1)Coefficient for calculating the degree of aggregation and the coefficient that represents the degree of dispersion between classes/>

(2) Calculate the index of overall clustering effect λ = α _max + (1-β _min );

(3) When ||λ(k)-λ(k-1)||<ε, the total number of clustering iterations before convergence is determined to be the optimal number of classifications k for clustering; otherwise, return to (1);

6) Construct an objective function based on the weighted sum of squares of distances from each sample to all cluster centers: And construct the membership matrix U=[u _ij ], to/> As constraints, train the optimal clustering method. The parameter update within the objective function includes the following steps:

(1) Update membership degree and blur index/>

(2) Update learning efficiency

(3) Update neuron node weights

(4) Update weight vector If ||Δw|| ₂ =||w(t+1)-w(t)|| ₂ >ε, return (1); otherwise, end the loop.

7) Complete the repair of the two-dimensional grayscale image, including:

(1) Traverse all data, search and record the position sequence and hierarchical sequence of each missing point;

(2) Search for the same missing value in all layers, find the maximum number of available information around the missing value, and extract the location information and layer information of the missing point;

(3) The missing points with the largest amount of available information around them are first repaired in this layer;

(4) Delete the location information and layer information of the repaired point. If there is still missing data, go to step (2) and search again; otherwise, go to 8).

8) Fusion the repaired data of each layer and restore the image information into waveform signals.

beneficial effects

A method of repairing missing power quality measurement data of the present invention has the following characteristics:

The electrical measurement data carried by the ubiquitous power Internet of Things is disturbed in various aspects such as collection, transmission, and conversion, resulting in missing data, which affects the accuracy of state estimation and the stable operation of the system. In view of the shortcomings of the traditional repair strategy that only considers the horizontal distribution pattern of one-dimensional measurement data, resulting in low data repair accuracy, the present invention fully considers the neighborhood data of the missing points of the power system measurement data and the periodic change pattern of the measurement data. A method for repairing missing power quality measurement data based on fuzzy self-organizing neural network is proposed. In the early stage of this invention, the one-dimensional measurement data of power quality is mapped into a two-dimensional grayscale image to improve the analysis of spatio-temporal correlation between data. Later, the artificial intelligence FSOM neural network algorithm was used to cluster the original data, deconstruct the multi-layer feature values of the data, and perform hierarchical repair of the clustered data. Through experimental data analysis, whether in the case of random deletion or continuous deletion, the FSOM repair algorithm proposed by the present invention has lower repair error and higher confidence than the existing algorithm under low data loss rate and high data loss rate. noise ratio.

Description of the drawings

The drawings described here are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation to the embodiments of the present invention. In the attached picture:

Figure 1 is a two-dimensional truncation and reorganization of one-dimensional data;

Figure 2 is the FSOM neural network mapping;

Figure 3 is the algorithm implementation process;

Figure 4 is a comparison of the repair effects of voltage swell and sag missing measurements;

Figure 5 is a comparison of the repair effects of missing voltage harmonic measurement data.

Detailed ways

Specific implementations of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate the invention but are not intended to limit the scope of the invention.

The present invention is a repair method suitable for missing power quality measurement data. In the early stage, the one-dimensional power quality measurement data is mapped into a two-dimensional grayscale image to improve the spatio-temporal correlation analysis between data. Then the artificial intelligence FSOM neural network algorithm is used to cluster the original data, deconstruct the multi-layer feature values of the data, and finally perform hierarchical repair on the clustered data.

Step 1: Input the missing one-dimensional power quality measurement data set

Step 2: Two-dimensional truncation and reorganization of waveform data

The voltage or current waveform at the monitoring point is sampled at high frequency to obtain a one-dimensional signal. Using the truncation and reorganization method as shown in Figure 1, the sampled one-dimensional waveform data is segmented according to N power quality sampling periods (through experimental analysis, N is 20 is more appropriate), and mapped into a two-dimensional matrix. rows or columns to achieve two-dimensional truncation and recombination of signals. The mapping rules are as follows:

x(i,j)＝x(n)|n＝(i-1)n _l +jj≤n _l (1)

Step 3: Grayscale conversion of two-dimensional matrix space

The matrix is converted into an image using the grayscale method, and the original two-dimensional matrix value of the power quality signal is mapped into a spatial grayscale value. The mapping rules are as follows:

Step 4: Extract the eigenvalues of the two-dimensional harmonic grayscale image

The time domain characteristic values of the waveform are extracted from each sampling period, and a characteristic matrix is established based on the statistical characteristics. The eigenvalue matrix can be expressed as:

The selected time domain eigenvalue indicators are:

Table 1 Time domain eigenvalue indicators

The time domain characteristic values T ₁ to T ₆ are calculated respectively in each harmonic signal period and normalized.

Step 5: Use FSOM neural network to cluster the waveform data, and cluster all samples into k layers. The data set P can be expressed as: P＝P ₁ ∪P ₂ ∪...∪P _k , i≠j.

1) Initialize the clustering parameters of the harmonic missing data set, including: the number of output layer nodes, the fuzzy index d _t , the number of initialization iterations t=1, the maximum number of iterations T _max , and the initial neuron node weight v _t .

2) Perform harmonic missing data set clustering, and construct an objective function based on the weighted sum of squares of distances from each sample to all cluster centers as the goal:

3) Construct a membership matrix U = [u _ij ] to represent the relationship between each power quality cycle data and clustered waveform data, where the membership matrix needs to satisfy

4) Determine the optimal number of clusters for power grid harmonic data, including:

(1) Coefficients for calculating the degree of aggregation of power grid harmonic data sets and the coefficient that represents the degree of dispersion between classes/>

(2) Calculate the index λ = α _max + (1-β _min ) for the overall clustering effect of the power grid harmonic data set.

(3) When ||λ(k)-λ(k-1)||<ε, the total number of clustering iterations before convergence is determined to be the optimal number of clustering k; otherwise, return to (1).

5) Use the Lagrange multiplier method to optimize the objective function to As constraints, the parameter update within the objective function includes the following steps:

(1) Update the membership matrix and blur index/>

(2) Update learning efficiency

(3) Update neuron node weights

(4) Update weight vector If ||Δw|| ² =||w(t+1)-w(t)|| ² >ε, return (1); otherwise, end the loop.

Step 6: Complete the repair of the two-dimensional grayscale image, including:

(1) Traverse all data and search for the position sequence x(i,j) of the missing point and the hierarchical sequence q _r (r is the number of clustering hierarchies).

(2) Search for the same missing value in all layers, find the maximum value S _max of the number of available information around the missing value, and extract the position information x'(i,j) of the missing point and the layer information q' _r .

(3) For the missing point with the largest amount of available information around it, first press formula (3) in this layer Make repairs.

(4) Delete the position information x'(i,j) and the layer information q' _r of the repaired point. If there is still missing data, go to step (2) and search again; otherwise, go to step (5).

(5) Re-press the repaired data of each layer according to formula (4) Perform fusion.

In order to verify the effectiveness of the method of repairing missing power quality measurement data based on fuzzy self-organizing neural network of the present invention, the method of the present invention is applied to the original harmonic missing measurement data to analyze the data repair effect.

The FSOM neural network mapping method shown in Figure 2 is used to cluster the power grid harmonic data set according to the characteristic values of the waveform. The algorithm flow chart is shown in Figure 4.

Use the power quality standard signal and disturbance signal models in Table 2 below to establish the original data set. Experiment 1 repaired the missing abnormal power quality data including voltage surges and dips. The original signal-to-noise ratio after mixing in Gaussian white noise was 17db. The data in Experiment 2 are missing harmonic voltages. The maximum sampling frequency is 20kHZ. The missing methods of each group of experimental data are divided into two types, namely continuous missing and random missing.

Table 2 Power quality standard signal and disturbance signal model

The present invention selects multiple evaluation indicators to evaluate the quality of data repair effects, including: mean absolute deviation (MAD), which can avoid the problem of mutual cancellation of errors and better reflect the actual situation of repair errors; signal noise The signal-to-noise ratio (SNR) reflects the repair accuracy of the noisy signal waveform; the root mean square error (RMSE) reflects the discreteness of the repair results. Calculation formula:

Using a method of repairing missing power quality measurement data based on fuzzy self-organizing neural network of the present invention, the missing measurement data are repaired respectively. The comparison of the repair effects of voltage swell and sag missing measurement is shown in Figure 4. The comparison of the repair effects of missing voltage harmonic measurement data is shown in Figure 5.

The abscissa in Figures 4 and 5 is the proportion of missing data. In the case of a continuous missing pattern with a data loss rate of 30%, the average absolute error of the algorithm proposed by the present invention is reduced by 60.71% compared to the MARS algorithm; the signal-to-noise ratio is increased by 47.3%; and the root mean square error is reduced by 56.76%. No matter in the case of random missing or continuous missing, the FSOM repair algorithm proposed by the present invention has lower repair error and higher signal-to-noise ratio than the time dynamic matrix decomposition method, the multivariate adaptive regression spline method and the KNN repair method. .

Claims (4)

1. A method for repairing missing power quality measurement data of a fuzzy self-organizing neural network, the method being executed by a computer program: the method comprises the following steps:

step 1), inputting a one-dimensional measurement data set containing the missing electric energy quality;

step 2) carrying out segmentation interception on the sampled one-dimensional waveform data according to N power quality sampling periods and carrying out mapping rules of x (i, j) =x (N) |n= (i-1) N through a two-dimensional matrix x _l +j j≤n _l Mapping the signals into rows or columns in a two-dimensional matrix x, and carrying out two-dimensional truncation recombination on one-dimensional signals;

step 3) calculating a space gray value P (i, j) = [ (255- (254 (m-x (i, j)/(m-n)) ], and converting the matrix x into an image by using a gray method;

step 4) extracting characteristic value X of two-dimensional harmonic gray scale _j ＝[X _1,j ,X _2,j ,…,X _m+l,j ]And normalized；

Step 5) determining the optimal clustering number of the harmonic data of the power grid; the step 5) is used for determining the optimal clustering number of the harmonic data of the power grid, and comprises the following steps:

(1) Calculating coefficients of degree of aggregationAnd a coefficient characterizing the degree of dispersion from class to class ∈>

(2) Calculating the index λ=α of the overall clustering effect _max +(1-β _min )；

(3) When I lambda (k) -lambda (k-1) I < epsilon, judging that the total number of clustering iterations before convergence is the optimal classification number k of the clusters; otherwise, returning to the step 5) (1);

step 6) taking the weighted square sum of the distances from each sample to all cluster centers as a target to construct an objective function:and constructing a membership matrix U= [ U ] _ij ]To->Training an optimal clustering mode for constraint conditions;

step 7) repairing the two-dimensional gray scale map; and 7) finishing the repair of the two-dimensional gray scale map, which comprises the following steps:

(1) Traversing all data, searching and recording the position sequence x (i, j) and the layering sequence q of each missing point _r ；

(2) Searching the same missing point in all layers, finding out the maximum value of the number of available information around the missing point, and extracting the position information and the layer information of the missing point;

(3) Repairing the defect point with the highest number of surrounding available information in the layer firstly;

(4) Deleting the position information and the layer information of the repaired points, and if the position information and the layer information of the repaired points have missing data, entering the step 7), and searching again in the step (2); otherwise enter 8);

and 8) fusing the repaired data of each layer, and restoring the image information into waveform signals.

2. The method for repairing the loss of power quality measurement data of the fuzzy self-organizing neural network according to claim 1, wherein the method comprises the following steps: the updating of the parameters in the objective function in the step 6) comprises the following steps:

(1) Updating membershipAnd ambiguity index->

(2) Updating learning efficiency

(3) Updating neuron node weights

(4) Updating weight vectorsIf Deltaw is | ₂ ＝||w(t+1)-w(t)|| ₂ >Epsilon, returning to (1); otherwise, the cycle is ended.

3. The method for repairing missing power quality measurement data of a fuzzy self-organizing neural network according to claim 1, wherein when repairing the two-dimensional gray scale map in step 7), if the corresponding data is not missing, the data is not processed.

4. The method for repairing missing power quality measurement data of a fuzzy self-organizing neural network according to claim 1, wherein the final clustering number k determined in the step 5) is substituted into the step 6) to perform similarity clustering.