CN107590604B

CN107590604B - Coherent unit grouping method and system combining S transformation and 2DPCA

Info

Publication number: CN107590604B
Application number: CN201710821583.8A
Authority: CN
Inventors: 徐振华; 苏毅; 黄道姗; 程鑫; 苏清梅; 吴丹岳; 余秀月; 黄霆; 宋少群; 江伟; 仲悟之; 宋新立; 叶小晖; 王涛; 杨越; 赵存璞; 顾雪平
Original assignee: State Grid Corp of China SGCC; China Electric Power Research Institute Co Ltd CEPRI; North China Electric Power University; Electric Power Research Institute of State Grid Fujian Electric Power Co Ltd; State Grid Fujian Electric Power Co Ltd
Current assignee: State Grid Corp of China SGCC; China Electric Power Research Institute Co Ltd CEPRI; North China Electric Power University; Electric Power Research Institute of State Grid Fujian Electric Power Co Ltd; State Grid Fujian Electric Power Co Ltd
Priority date: 2017-09-13
Filing date: 2017-09-13
Publication date: 2020-09-04
Anticipated expiration: 2037-09-13
Also published as: CN107590604A

Abstract

The invention relates to a coherent unit grouping method and system combining S transformation and 2 DPCA. The method comprises the steps of obtaining a real-time power angle of a system unit by using a WAMS phasor measuring device, converting a power angle signal of each generator into a time-frequency characteristic module value matrix by adopting rapid S transformation, reducing the dimension of the time-frequency characteristic module value matrix by using 2DPCA, converting the time-frequency characteristic module value matrix into a low-dimension characteristic index matrix, and inputting the characteristic index matrix into a self-organizing neural network for clustering identification. The method can divide the units of the system on line in real time along with different faults and different fault places under different operation modes, comprehensively considers the time-frequency domain information of the power angle, has high identification accuracy and stable clustering result, provides necessary premise for simplifying the power grid, determining the system oscillation center and controlling the splitting, and ensures the safe, efficient and stable operation of the power grid.

Description

Coherent unit grouping method and system combining S transformation and 2DPCA

Technical Field

The invention relates to the technical field of coherent unit identification, in particular to a coherent unit grouping method and system combining S transformation and 2 DPCA.

Background

With the implementation of the engineering project of 'west-east power transmission, south-north interconnection', national power grids are developing towards large-area interconnection power grids. However, while economic benefits are improved, the problem of grid safety is more and more prominent, and small disturbance may cause cascading failure, resulting in large-area power failure. The active splitting control is the last ring for avoiding the power grid paralysis, and the searching of the splitting section needs to be based on accurate unit coherent grouping. In addition, along with the expansion of the scale of the power grid, the system has various fault modes, and the system equivalence is a single-machine infinite system or a two-machine system during fault analysis, so that the requirement of transient stability analysis cannot be completely met. The dynamic equivalence method is a good equivalence simplification method of a system model, is beneficial to quickly analyzing the dynamic stability of the system and provides reference for active splitting, and the unit coherent grouping is the core technology of the dynamic equivalence method. Therefore, the quick and accurate unit coherent grouping has important significance.

When a power system fails, some generators in the system exhibit similar dynamic characteristics, referred to as unit coherence. The coherent grouping method of the units mainly comprises three types: identifying a coherent machine group according to a maneuvering parameter of a generator rotor; analyzing the unit coherence characteristics according to the disturbance system state matrix trajectory characteristic root; and thirdly, performing unit homodyne identification according to the time-frequency characteristics of the power angle swing curve of the disturbance generator. The first method needs to define the parameters of the system generator in advance, the accuracy of the unit coherent grouping depends on the accuracy of the generator parameters and the like to a great extent, but the accurate generator parameters are not easy to obtain sometimes. In addition, the method cannot consider the influence of the system operation condition and the accident type on the grouping result of the unit. In the second method, before the unit coherence identification, the power grid system needs to be simplified in a linearization manner, which may result in low accuracy of the unit clustering result and is not suitable for a large-scale interconnection system. The third method is originally restricted by obtaining the power angle information of the generator in real time, but with the development of a wide area measurement system and an optical fiber communication technology, the real-time power angle information of the generator can be easily obtained, and then the coherent cluster is identified according to the effective time-frequency characteristics of the power angle information. The method not only can realize real-time grouping by considering the system operation condition and the fault type, but also is suitable for a large-scale interconnection system.

And identifying the coherent cluster according to the power angle information of the generator, wherein effective characteristics of the power angle need to be extracted. At present, the power angle information feature extraction method comprises the following steps: fourier transform, wavelet transform, empirical mode decomposition, hilbert-yellow transform, and the like. However, these methods have disadvantages, such as poor effect of fourier transform on unstable signal processing, difficulty in selecting wavelet basis functions in wavelet transform, modal aliasing phenomenon in inherent modal components in empirical mode decomposition, and complex recursive operation required for hilbert-yellow transform.

Disclosure of Invention

The invention aims to provide a coherent unit grouping method and system combining S transformation and 2DPCA, so as to overcome the defects of complex operation, low efficiency and accuracy and the like in the prior art.

In order to achieve the purpose, the technical scheme of the invention is as follows: a coherent unit grouping method combining S transformation and 2DPCA,

the method comprises the following steps:

step S1: acquiring power angle information increment after the M generators have faults through a synchronous phasor measuring device;

step S2: obtaining corresponding power angle rocking curve information according to the power angle information increment of each generator obtained in the step S1, and performing rapid S transformation to obtain an mxn-dimensional time-frequency domain information matrix;

step S3: reducing the dimension of the m × n-dimensional time-frequency domain information matrix of the step S2 by adopting a 2DPCA algorithm to obtain a q × p-dimensional characteristic index matrix Ts; wherein s is a generator label;

step S4: and clustering M samples in the characteristic index matrix obtained in the step S3 to obtain a coherent unit.

In an embodiment of the present invention, in the step S1, the increment dimension of the power angle information is N, and an nxm-dimensional power angle information database is obtained.

In an embodiment of the present invention, in the step S4, the M samples in the characteristic index matrix obtained in the step S3 are clustered by using a self-organizing neural network.

In an embodiment of the present invention, the clustering by using the self-organizing neural network is implemented according to the following steps:

step S41: the number of input neurons is the number of elements of the characteristic index matrix Ts, and the input neurons are used as a sample characteristic index matrix; the output layer neuron is initially one; will connect the weight W_ijInitially setting an arbitrary sample characteristic index matrix, wherein i is an input layer neuron label, and j is an output layer neuron label;

step S42: recording a division precision threshold used for determining the accuracy degree of clustering as lambda;

step S43: inputting a new sample X, calculating Euclidean distance between the new sample and a corresponding clustering center of each output neuron, and recording the minimum distance as D_jJ is the output layer neuron label corresponding to the minimum distance, and the formula is as follows:

in the formula: x_iIs an input vector; connection weight W_ijA cluster center of a jth class mode;

step S44: if D is_j<λ, then the jth output neuron wins, the sample is taken into the jth neuron, and the neuron connection weight is updated as follows:

W_ij(n+1)＝W_ij(n)+η(n)(X_i-W_ij(n))

wherein η (n) is the reciprocal of the total number of samples contained in the neuron;

if D is_j>Lambda, the sample does not belong to any existing output neuron, an output neuron is newly added, and the initial connection weight of the output neuron and the input layer is a sample characteristic index matrix;

step S45: and sequentially inputting all samples according to the steps, wherein the number of final output layer neurons is the number of clustering clusters, each output layer neuron comprises a sample which is a same cluster sample, and the connection weight of the sample and the input layer is the clustering center of the cluster.

In an embodiment of the present invention, the method further includes a step S5: and introducing an internal effective index-CH index to determine the optimal cluster number.

In one embodiment of the present invention, in the step S5, the internal valid indicator — CH indicator is based on a measure of intra-class dispersion matrix and inter-class dispersion matrix of all samples, as follows:

wherein k is the number of clusters, n is the number of all samples, tr (B (k)) is the trace of the inter-class dispersion matrix, and tr (W (k)) is the trace of the intra-class dispersion matrix; and k corresponding to the maximum value obtained by the CH index is the optimal cluster number.

In an embodiment of the present invention, the method further includes a step S6: the synchrophasor measurement device acquires the power angle information increment of the M generators after the fault once again every certain time, refreshes the power angle information database, and repeats the steps S2, S3, and S4.

Further, a coherent unit grouping system combining S transformation and 2DPCA is also provided, including:

the synchronous phasor measurement unit is used for acquiring power angle information increment after the M generators are in fault through the synchronous phasor measurement device;

the S transformation unit is used for carrying out rapid S transformation on the power angle swing curve information of each generator acquired by the synchronous phasor measurement unit and acquiring an m multiplied by n dimensional time-frequency domain information matrix;

a dimension reduction unit based on a 2DPCA algorithm, which is used for reducing the dimension of the m multiplied by n dimension time-frequency domain information matrix through the 2DPCA algorithm to obtain a q multiplied by p dimension characteristic index matrix Ts, wherein s is the generator label;

and the clustering unit is used for clustering the characteristic index matrix Ts to obtain the coherent unit.

In an embodiment of the present invention, the method further includes: and an optimal clustering number calculating unit for determining the optimal clustering number according to an internal effective index-CH index.

Compared with the prior art, the invention has the following beneficial effects: the invention provides a coherent unit grouping method and a coherent unit grouping system combining S transformation and 2DPCA, which aim at the problems of feature extraction one-sided and complex and mutual restriction of calculation processes faced by coherent group identification. The multi-resolution time-frequency information of the power angle swing curve of the generator is obtained by utilizing the rapid S transformation and the 2DPCA, the characteristics are comprehensively extracted, and effective characteristics can be effectively mined. And the self-organizing neural network is used as a classifier to perform cluster analysis on the generator characteristic index matrix, so that a cluster result is stable. The method can acquire the grouping state of the system unit in real time, provide necessary premises for simplifying the power grid, determining the system oscillation center and controlling the splitting, and ensure the safe, efficient and stable operation of the power grid; the invention is easy to operate and suitable for practical engineering application.

Drawings

Fig. 1 is a flow chart of a coherent unit identification method using fast S-transform and 2DPCA according to the present invention.

FIG. 2 is a graph illustrating power angle incremental swing of an IEEE-39 generator in accordance with an embodiment of the present invention.

Fig. 3 is a schematic diagram illustrating the information content of the index matrix of different dimensional characteristics according to an embodiment of the present invention.

Fig. 4 is a graph comparing the trend of the power angle swing curve of the generator according to an embodiment of the present invention.

Detailed Description

The technical scheme of the invention is specifically explained below with reference to the accompanying drawings.

The invention provides a coherent unit grouping method combining S transformation and 2DPCA, and the invention is explained in detail with the accompanying drawings and the specific embodiments.

As shown in fig. 1, a flow chart of a coherent unit identification method combining fast S-transform and 2DPCA includes the following steps:

step S1: and obtaining the power angle information increment after the M generators have faults by the synchronous phasor measuring device, wherein the dimensionality is N, and obtaining an NxM dimensional power angle information database.

Step S2: and obtaining corresponding power angle swing curve information according to the power angle information increment of each generator obtained in the step S1, and performing S transformation to obtain an m × n-dimensional time-frequency domain information matrix.

S3, adopting 2DPCA algorithm to reduce the dimension of the time-frequency domain information matrix of the step S2, and finally obtaining a q × p-dimension characteristic index matrix T_sAnd s is the generator number.

Step S4: and clustering the M samples obtained in the step S3, namely the characteristic index matrix, by using a self-organizing neural network, wherein the samples clustered into the same cluster have high similarity, and the change rules of the corresponding power angle swing curves are similar, so that the unit coherence is obtained.

Step S5: and introducing an internal effective index-CH index to determine the optimal cluster number.

Step S6: and the PMU (phasor measurement Unit) acquires the power angle data of the system unit once again every specific time, refreshes the database, and repeats the steps S2, S3 and S4. Preferably, the specific time period is 10 ms.

Furthermore, in this embodiment, the adopted S transform can perform multi-resolution analysis on the unstable signal, and the width of the gaussian window is adjustable, so that the method has good time-frequency transformation capability. The power angle data information is converted into a high-dimensional matrix through S conversion, and the matrix contains a large amount of redundant information. Yang et al propose two-dimensional principal component analysis (2 DPCA for short) can extract the main information characteristic of two-dimensional matrix, have high, advantage such as being simple and easy of accuracy with high efficiency.

Further, in this embodiment, the internal valid indicator — CH indicator is based on the measure of the intra-class dispersion matrix and the inter-class dispersion matrix of all samples, and is defined by the following formula:

wherein k is the number of clusters, n is the total number of samples, tr (B (k)) is the trace of the inter-class dispersion matrix, and tr (W (k)) is the trace of the intra-class dispersion matrix. And k corresponding to the maximum value obtained by the CH index is the optimal cluster number.

Further, in this embodiment, the specific steps of using the self-organizing neural network algorithm are as follows:

step S41: and (5) initializing the network. The number of input neurons is the number of elements of the characteristic index matrix. The output layer neurons are initially one. Connection weight W_ij(i is the input layer neuron label, j is the output layer neuron label) is initially an arbitrary sample characteristic index matrix.

Step S42: a division accuracy threshold λ is set. The accuracy degree of clustering is determined by a threshold lambda, and the smaller the threshold lambda is, the more accurate the clustering result is and the more the classification is.

Step S43: and (5) identifying clusters. Inputting a new sample X, calculating Euclidean distance between the new sample and a corresponding clustering center of each output neuron, and recording the minimum distance as D_jAnd j is the output layer neuron label corresponding to the minimum distance. The formula is as follows:

in the formula: x_iIs an input vector; connection weight W_ijThe cluster center of the jth class pattern.

Step S44: and updating the network. If D is_j<λ, then the jth output neuron wins, the sample is taken into the jth neuron, and the neuron connection weight is updated as follows:

W_ij(n+1)＝W_ij(n)+η(n)(X_i-W_ij(n)) (3)

η (n) is the reciprocal of the total number of samples contained in the neuron.

If D is_j>And lambda, if the sample does not belong to any existing output neuron, adding a new output neuron, and taking the initial connection weight of the new output neuron and the input layer as a sample characteristic index matrix.

Step S45: and (5) performing cyclic training. And sequentially inputting all samples according to the steps, wherein the number of final output layer neurons is the number of clustering clusters, each output layer neuron comprises a sample which is a same cluster sample, and the connection weight of the sample and the input layer is the clustering center of the cluster.

Further, in this embodiment, a coherent group grouping system combining S transform and 2DPCA is further provided, including:

In this embodiment, the synchrophasor measurement unit obtains, through the synchrophasor measurement device, power angle information increments after failures of M generators, where the dimensionality is N, and obtains an nxm-dimensional power angle information database. After the grouping of the coherent units is completed, the synchrophasor measurement unit acquires the power angle information increment of the M generators after the fault once again every 10ms, refreshes the power angle information database, and repeats the step S2, the step S3 and the step S4.

And the clustering unit clusters and groups the M samples in the characteristic index matrix by adopting a self-organizing neural network to obtain a coherent unit. The self-organizing neural network adopted for clustering is realized according to the following steps:

step A: the number of input neurons is the number of elements of the characteristic index matrix Ts, and the input neurons are used as a sample characteristic index matrix; the output layer neuron is initially one; will connect the weight W_ijInitially setting an arbitrary sample characteristic index matrix, wherein i is an input layer neuron label, and j is an output layer neuron label;

and B: recording a division precision threshold used for determining the accuracy degree of clustering as lambda;

and C: inputting a new sample X, calculating Euclidean distance between the new sample and a corresponding clustering center of each output neuron, and recording the minimum distance as D_jJ is the output layer neuron label corresponding to the minimum distance, and the formula is as follows:

step D: if D is_j<λ, then the jth output neuron wins, the sample is taken into the jth neuron, and the neuron connection weight is updated as follows:

W_ij(n+1)＝W_ij(n)+η(n)(X_i-W_ij(n))

step E: and sequentially inputting all samples according to the steps, wherein the number of final output layer neurons is the number of clustering clusters, each output layer neuron comprises a sample which is a same cluster sample, and the connection weight of the sample and the input layer is the clustering center of the cluster.

Further, in this embodiment, the method further includes: and an optimal clustering number calculating unit for determining the optimal clustering number according to an internal effective index-CH index. The internal effective index, the CH index, is based on the measure of the intra-class dispersion matrix and the inter-class dispersion matrix of all samples, as follows:

Furthermore, in order to make those skilled in the art further understand the method proposed in the present invention, the IEEE-39 classic system is used as an example, and G is used as an example₃₀The generator is used as a reference, a three-phase open circuit fault occurs on a line bus3-bus18, the fault is removed after the fault lasts for 0.14s, the front and the back of a power grid structure are kept consistent, a power angle change curve of a system generator at a time period of 4.0s from the fault occurrence time is monitored, simulated power angle data are derived from PSDEdit, power angle increment data are obtained through Matlab processing, a power angle swing curve is drawn, and the power angle increment swing curve of an IEEE-39 generator is shown in an attached figure 2.

Furthermore, rapid S conversion is carried out on the power angle increment time sequence number sequence of each generator, the sampling frequency is 100Hz, the total number of samples is 400, the corresponding frequency range of S conversion time frequency information matrix rows is 0 Hz-50 Hz, the interval is 0.25Hz, the corresponding time range of the columns is 0S-4S, and the interval is 0.01S. And 2DPCA dimension reduction processing is carried out on the time-frequency information matrix after S transformation, when the main component numbers p and q of the rows and the columns take different values respectively, the information content of the characteristic index matrix is different, and the information content of the characteristic index matrix with different dimensions is shown in an attached figure 3.

According to analysis, when p is 2 and q is 1 in 2DPCA, the principal component already contains 81% of the information content of the time-frequency information matrix, so that the self-organizing neural network input sample is selected as a 2 × 1 dimensional characteristic index matrix, the samples are clustered, the number of clusters changes along with the change of the threshold value λ, and the clustering result is shown in table 1.

TABLE 1 ST-2DPCA coherent unit grouping results

Number of classification	ST-2DPCA clustering results
		9	G₃₁/G₃₂/G₃₃/G₃₄/G₃₅/G₃₆/G₃₇/G₃₈/G₃₉
8	G₃₁/G₃₂/G₃₃/G₃₄/G₃₅G₃₆/G₃₇/G₃₈/G₃₉
		7	G₃₁/G₃₂/G₃₃G₃₄/G₃₅G₃₆/G₃₇/G₃₈/G₃₉
6	G₃₁/G₃₂/G₃₃G₃₄G₃₅G₃₆/G₃₇/G₃₈/G₃₉
		5	G₃₁/G₃₂/G₃₃G₃₄G₃₅G₃₆/G₃₇/G₃₈G₃₉
4	G₃₁G₃₇/G₃₂/G₃₃G₃₄G₃₅G₃₆/G₃₈G₃₉
		3	G₃₁G₃₇/G₃₂G₃₈G₃₉/G₃₃G₃₄G₃₅G₃₆
2	G₃₁G₃₂G₃₇G₃₈G₃₉/G₃₃G₃₄G₃₅G₃₆
		1	G₃₁G₃₂G₃₃G₃₄G₃₅G₃₆G₃₇G₃₈G₃₉

The CH index varies with the number of clusters, as shown in table 2. When k is 3, the CH index takes the maximum value, and therefore the optimal cluster number is 3.

TABLE 2 index values for different cluster numbers CH

Number of clusters clustered	2	3	4	5	6	7	8
								CH index	89.7	119.6	110.2	91.2	62.6	32.9	7.9

After the failure, the database is refreshed every 10ms, the grouping result is continuously changed along with the updating of the database,

when the units are divided into 3 groups, the change process of the grouping result at different times is shown in table 3:

TABLE 3 comparison of clustering results at different times

Time t(s)	Clustering results
		0.2	G₃₂/G₃₁G₃₇G₃₈G₃₉/G₃₃G₃₄G₃₅G₃₆
0.5	G₃₁G₃₂G₃₈G₃₉/G₃₃G₃₄G₃₅G₃₆/G₃₇
		1.0	G₃₁G₃₇/G₃₂G₃₈G₃₉/G₃₃G₃₄G₃₅G₃₆
1.5	G₃₁G₃₇/G₃₂G₃₈G₃₉/G₃₃G₃₄G₃₅G₃₆
		2.0	G₃₁G₃₇/G₃₂G₃₈G₃₉/G₃₃G₃₄G₃₅G₃₆
2.5	G₃₁G₃₇/G₃₂G₃₈G₃₉/G₃₃G₃₄G₃₅G₃₆

It can be seen from the table that after the fault occurs for 1s, the clustering result tends to be stable, because the power angle increment information in the database contains more and more power angle increment information along with the increase of time, and the actual characteristics of the unit can be reflected more and more.

The characteristic indexes extracted by the ST-2DPCA method and the characteristic indexes extracted by the PCA, ICA and NMF methods provided by the text are respectively input into the self-organizing neural network for clustering, and the optimal clustering result obtained according to the CH index is shown in Table 4.

TABLE 4 comparison of coherent unit grouping results

Feature extraction method	Clustering results
		ST-2DPCA	G₃₁G₃₇/G₃₂G₃₈G₃₉/G₃₃G₃₄G₃₅G₃₆
PCA	G₃₁G₃₂G₃₇/G₃₃G₃₄G₃₅G₃₆/G₃₈G₃₉
		ICA	G₃₁G₃₂G₃₇/G₃₃G₃₄G₃₅G₃₆/G₃₈G₃₉
NMF	G₃₁G₃₂/G₃₃G₃₄G₃₅G₃₆/G₃₇G₃₈G₃₉

Defining a unit coherence criterion formula as follows:

k may take any constant such that S_ijMinimum, let:

the following can be found:

and (5) verifying the effect of the 4 clustering results by utilizing a coherence criterion formula. The maximum criterion among all clusters of samples in the ST-2DPCA feature extraction method grouping result is 0.6721, and the maximum criterion among all clusters of samples in the PCA, ICA and NMF feature extraction method grouping result is 0.9652, 0.9652 and 1.2745 respectively.

The four unit grouping methods all group the generators {33,34,35,36}, and it can be seen from fig. 4(2) that the power angle swing curves are similar and have strong coherence. As can be seen from FIG. 4(1) and FIG. 4(3), the oscillation frequencies of the generators {32,38,39} are similar, and the homologies of the generators {32} and {38,39} are stronger than those of the generators {32} and {31,37 }. Looking at fig. 4(4), it is clear that the generators {37} are not in tune with the generators {38,39 }.

Furthermore, the criterion and the trend of the power angle curve all show that the feature extraction method in the embodiment can effectively extract the time-frequency features of the power angle signals compared with the PCA, ICA and NMF methods. The method provided by the invention is verified to have higher accuracy.

Further, in this embodiment, a 30 # generator is used as a reference, different fault types are set at different positions, the fault is removed after the duration lasts for 0.12s, the power angle of the system generator after the fault starts is monitored, the coherent unit clustering is performed according to the method, and the clustering result is shown in table 5.

TABLE 5 comparison of clustering results for different accident types

Type of accident	Clustering results
		Three-phase open circuit of line 4-14	G₃₁G₃₇/G₃₃G₃₄G₃₅G₃₆/G₃₂G₃₈G₃₉
Line 25-26 three-phase short circuit	G₃₁G₃₂/G₃₃G₃₄G₃₅G₃₆/G₃₇/G₃₈G₃₉
		Three-phase short circuit of lines 16-24	G₃₁G₃₂/G₃₃G₃₄G₃₅G₃₆/G₃₇G₃₈G₃₉
Three-phase break of line 9-39	G₃₁G₃₂/G₃₃G₃₅G₃₆G₃₇/G₃₄G₃₈/G₃₉

The grouping results in table 5 show that the method can perform unit grouping on the system in real time along with the change of different faults and different fault locations in different operation modes, and the model-based grouping method (performing unit grouping according to system parameters) cannot realize the function.

The above are preferred embodiments of the present invention, and all changes made according to the technical scheme of the present invention that produce functional effects do not exceed the scope of the technical scheme of the present invention belong to the protection scope of the present invention.

Claims

1. A coherent unit grouping method combining S transformation and 2DPCA is characterized by comprising the following steps:

step S2: obtaining corresponding power angle swing curve information according to the power angle information increment of each generator obtained in the step S1, and performing rapid S transformation to obtain an m × n-dimensional time-frequency domain information matrix;

step S3: for said step S2 by using 2DPCA algorithmm×nReducing the dimension of the dimension time-frequency domain information matrix to obtainq ×pFeature index matrix of dimensionTs(ii) a Wherein the content of the first and second substances,sthe generator is labeled;

step S4: clustering M samples in the characteristic index matrix obtained in the step S3 by using a self-organizing neural network to obtain a coherent unit;

the clustering by adopting the self-organizing neural network is realized according to the following steps:

step S41: the number of input neurons is the number of elements of the characteristic index matrix Ts, and the input neurons are used as a sample characteristic index matrix; the output layer neuron is initially one; will connect the weightW _ijInitially, an arbitrary sample characteristic index matrix is obtained,ifor the input layer neuron labels,jis the output layer neuron label;

step S42: the partition accuracy threshold, which is used to determine the accuracy of the clustering, isλ；

Step S43: inputting a new sample X, calculating Euclidean distance between the new sample and a corresponding clustering center of each output neuron, and recording the minimum distance asD _j，jFor the output layer neuron label corresponding to the minimum distance, the formula is as follows:

in the formula:X _iis an input vector; connection weightW _ijIs as followsjA cluster center of class group patterns;

step S44: if it isD _j<λ, thenjAn output neuron wins, the sample is includedjIn each neuron, updating the neuron connection weight value, wherein the formula is as follows:

wherein the content of the first and second substances,η(n)is the reciprocal of the total number of samples contained in the neuron;

if it isD _j >λIf the sample does not belong to any existing output neuron, adding an output neuron, and taking the initial connection weight of the output neuron and the input layer as a sample characteristic index matrix;

2. The coherent unit clustering method combining S-transform and 2DPCA according to claim 1, wherein in step S1, the power angle information increment dimension is N, and a power angle information database with N × M dimensions is obtained.

3. The coherent group clustering method combining the S-transform and the 2DPCA according to claim 1, further comprising a step S5: and introducing an internal effective index-CH index to determine the optimal cluster number.

4. The coherent group clustering method combining S-transform and 2DPCA according to claim 3, wherein in the step S5, the internal valid index-CH index is based on the measure of the intra-class dispersion matrix and the inter-class dispersion matrix of all samples, as follows:

wherein the content of the first and second substances,kto be the number of clusters to be clustered,nfor the purpose of the total number of samples,tr(B(k))is the trace of the inter-class separation difference matrix,tr(W(k))is the trace of the intra-class dispersion matrix; the CH index corresponding to the maximum valuekNamely the optimal clustering number.

5. The coherent group clustering method combining the S-transform and the 2DPCA according to claim 1, further comprising a step S6: the synchrophasor measurement device acquires the power angle information increment of the M generators after the fault once again every certain time, refreshes the power angle information database, and repeats the steps S2, S3, and S4.

6. A coherent unit clustering system combining S-transform and 2DPCA, comprising:

a clustering unit, which clusters the characteristic index matrix Ts by using a self-organizing neural network to obtain a coherent unit;

7. The coherent team clustering system combining S-transform and 2DPCA according to claim 6, further comprising: and an optimal clustering number calculating unit for determining the optimal clustering number according to an internal effective index-CH index.