CN108599152A - The key stato variable choosing method and device of power system transient stability assessment - Google Patents

Abstract

The present invention provides the key stato variable choosing method and device of power system transient stability assessment, the method includes：S1 is obtained multiple state variables in Power System Dynamic Simulation data, is pre-processed to the multiple state variable using fft algorithm；S2 obtains multiple key stato variables to carrying out Feature Selection by pretreated the multiple state variable；S3 carries out dimensionality reduction to each key stato variable.Selection and dimensionality reduction of the present invention by key stato variable can significantly shorten training time and the classification time of transient stability classification device, be more suitable for application on site in the case where not reducing transient stability arbiter nicety of grading.

Description

Key state variable selection method and device for power system transient stability assessment

technical field

The invention relates to the field of power system transient stability evaluation, and more specifically, to a method and device for selecting key state variables for power system transient stability evaluation.

Background technique

With the access of various new energy sources and the adoption of UHV transmission lines, the power system has become more complex. The power system is often subjected to various large disturbances during operation, especially the grounding and short-circuit faults of the operating line, which may cause the system to face transient instability problems. The destruction of transient stability is the main factor causing catastrophic accidents in the power grid. Therefore, fast and accurate transient stability evaluation methods are of great significance to ensure the safe and stable operation of the power grid.

The traditional numerical simulation method to evaluate the transient stability of the power system has the characteristics of complex model and slow speed. With the rapid rise of artificial intelligence, the accumulation of power grid operation data and the adoption of big data methods have brought new ideas to the prediction of power system transient stability based on machine learning. Machine learning transforms the transient stability evaluation of power systems into pattern recognition problems. There is a certain mapping relationship between the system transient stability and some feature quantities describing the operating state of the system, which is essentially an "offline simulation training, online The process of "matching application" is to obtain samples that can fully reflect the mapping relationship through offline simulation analysis, and select the functional relationship of the unknown mapping through the learning of the sample set. The transient stability of the system is classified and evaluated. The power system transient stability assessment method based on machine learning can speed up the power system simulation.

At present, the transient stability evaluation method of power system based on machine learning directly uses the dynamic simulation data of all operating state variables of the power grid for feature selection. Although it can make full use of the operating state information of the power system, it may cause "dimension disaster". The problem is not only difficult to analyze, but also takes a long time to train, and it is easy to ignore the information of key operating state variables, reducing the accuracy of transient stability evaluation.

Contents of the invention

The present invention provides a key state variable selection method and device for power system transient stability evaluation which overcomes the above problems or at least partly solves the above problems.

According to one aspect of the present invention, a key state variable selection method for power system transient stability assessment is provided, including:

S1. Obtain a plurality of state variables in the dynamic simulation data of the power system, and use an FFT algorithm to preprocess the plurality of state variables;

S2, performing feature selection on the plurality of preprocessed state variables to obtain a plurality of key state variables;

S3. Perform dimensionality reduction on each of the key state variables.

Wherein, the power system dynamic simulation data further includes: power system dynamic waveform data after fault removal.

Wherein, the step S1 further includes:

Use the FFT algorithm to extract the real and imaginary parts of the first n harmonics of the time series of each state variable in the power system dynamic simulation data, so that the time series attributes of each state variable can be converted into 2n real number attributes, where n is A natural number greater than 1.

Wherein, the step S2 further includes:

According to the 2n real number attributes of each state variable, the state variables with the highest weight are screened out by using the Relief algorithm, and multiple key state variables are obtained.

Wherein, according to the 2n real number attributes of each state variable, the step of obtaining a plurality of key state variables further includes:

S21, randomly select a dynamic simulation data R from a plurality of dynamic simulation data, and then search for the nearest neighbor dynamic simulation data H from a sample set of the same type as the dynamic simulation data R, and search for the nearest dynamic simulation data H from a sample set different from the dynamic simulation data R Neighboring dynamic simulation data M;

S22, update the weight of the dynamic simulation data R according to the following rules: if the distance between the dynamic simulation data R and the dynamic simulation data H on a certain state variable is smaller than the distance between the dynamic simulation data R and the dynamic simulation data M on the state variable , then increase the weight of the state variable; or, if the distance between the dynamic simulation data R and the dynamic simulation data H on a certain state variable is greater than the distance between the dynamic simulation data R and the dynamic simulation data M on the state variable, then reduce the weights of state variables;

S23, repeating the steps S1 and S2p times to obtain the weight values of multiple state variables, sort the state variables according to the weight values from large to small, and take the first q state variables with the largest weight values as key state variables;

Among them, the values of p and q are determined according to the requirements of power system transient stability assessment.

Wherein, the step S3 further includes:

The principal component analysis method is used to reduce the dimensionality of each of the key state variables.

Wherein, the step S3 further includes:

A popular learning method is used to reduce the dimensionality of each of the key state variables.

According to another aspect of the present invention, a key state variable selection device for power system transient stability assessment is provided, including:

A preprocessing module, configured to obtain a plurality of state variables in the power system dynamic simulation data, and use an FFT algorithm to preprocess the plurality of state variables;

A feature selection module, configured to perform feature selection on the preprocessed multiple state variables to obtain multiple key state variables;

A dimensionality reduction module is configured to perform dimensionality reduction on each of the key state variables.

Wherein, the preprocessing module is specifically used for:

Use the FFT algorithm to extract the real and imaginary parts of the first n harmonics of the time series of each state variable in the power system dynamic simulation data, so that the time series attributes of each state variable can be converted into 2n real number attributes, where n is A natural number greater than 1.

Wherein, the feature selection module is specifically used for:

According to the 2n real number attributes of each state variable, the state variables with the highest weight are screened out by using the Relief algorithm, and multiple key state variables are obtained.

The key state variable selection method and device for power system transient stability evaluation proposed by the present invention can significantly shorten the transient stability state without reducing the classification accuracy of the transient stability discriminator through the selection of key state variables and dimensionality reduction. The training time and classification time of the classifier are more suitable for online applications.

Description of drawings

Fig. 1 is a schematic flow chart of a key state variable selection method for power system transient stability assessment provided by an embodiment of the present invention;

Fig. 2 is a time-consuming analysis and comparison diagram of training using all dynamic data, training using an improved state variable extraction method, and training after dimensionality reduction provided by an embodiment of the present invention;

Fig. 3 is a schematic structural diagram of a key state variable selection device for power system transient stability assessment provided by another embodiment of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments It is only some embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.

As shown in Figure 1, it is a schematic flow diagram of a key state variable selection method for power system transient stability assessment provided by an embodiment of the present invention, including:

S1. Obtain a plurality of state variables in the dynamic simulation data of the power system, and use an FFT algorithm to preprocess the plurality of state variables;

S2, performing feature selection on the plurality of preprocessed state variables to obtain a plurality of key state variables;

S3. Perform dimensionality reduction on each of the key state variables.

Specifically, the power system transient stability assessment based on machine learning can be divided into four aspects. The first is the overall framework of transient stability assessment, the second is the selection of input data and the selection of features, the third is the selection of classification algorithms, and the last is the analysis of classification results and the corresponding improvements. The framework of transient stability assessment of power system based on machine learning is still the method of "offline training, online simulation". By mining the rules of offline data and matching with online data, the transient stability of the power system can be quickly judged. What the invention improves is the feature selection and input part of the classifier.

S1, when the power system transient stability assessment method based on machine learning is used to evaluate the transient stability of the power system, the dynamic dynamic simulation data is used. Dynamic simulation data contains more power system state variables, and at the same time simulates for a longer time length, so it is easy to cause the problem of "curse of dimensionality". For example, if there are 2000 selected state variables, the simulation time length is 10s, and the step size is 0.01s, then the result of each simulation is a 2000*1000 matrix. When training the transient stability evaluation classifier, if If 10,000 samples are used, it is easy to cause the samples to be too large to be analyzed, and the training takes a long time, which is not conducive to the multiple training parameter adjustment and online application of the transient stability evaluation classifier. Therefore, it is necessary to process the dynamic simulation data of the power system. The present invention first screens out key state variables from a plurality of state variables, compresses the size of the dynamic simulation data in terms of the number of state variables, and then combines the dimensionality reduction method to The dynamic simulation data is dimensionally reduced to further reduce its dimension, thereby speeding up the training speed of the classifier.

Selecting key state variables from multiple state variables requires the use of feature selection algorithms, and in traditional feature selection algorithms, the feature attributes of samples are real numbers. Since the dynamic simulation results of the power system can be expressed by the time series of m state variables, the time series of each state variable is the attribute of the sample. In order to use the traditional feature selection algorithm, we first use the FFT algorithm to preprocess the multiple state variables in the obtained power system dynamic simulation data, and convert the attributes of each state variable from time series to real numbers, so that The traditional feature selection algorithm can be used to select the features of the samples and screen out the key state variables.

S2, Feature Selection (Feature Selection), also known as Feature Subset Selection (FSS), or Attribute Selection (Attribute Selection), refers to selecting a feature subset from all features to make the constructed model better. In the practical application of machine learning, the number of features is often large, there may be irrelevant features, and there may be interdependence between features, which can easily lead to the following consequences: the more the number of features, the more features required for analyzing features and training models. the longer the time. The larger the number of features, the more likely it is to cause the "dimension disaster", the more complex the model will be, and its generalization ability will decrease. Feature selection can eliminate irrelevant or redundant features, so as to reduce the number of features, improve the accuracy of the model, and reduce the running time. On the other hand, picking out the truly relevant features makes it easy for researchers to understand the process of data generation. On the basis of step S1, the existing feature selection algorithm, such as the Relief algorithm, can be used to perform feature selection on the multiple state variables after preprocessing, so as to locate the power system transient stability evaluation. key state variables.

S3, after steps S1 and S2 are executed, key state variables for power system transient stability evaluation are selected. But for an actual power system, if there are 2000 state variables, simulate 10s with a step size of 0.01s. Even if 50 key state variables are selected from 2000 state variables, the result of each simulation will still be a 50*1000 matrix, the simulation result is still very large, and the problem of "curse of dimensionality" may still exist. Therefore, further dimensionality reduction is needed. There are two commonly used dimensionality reduction methods, one is a linear dimensionality reduction method, such as PCA. The other is a nonlinear dimensionality reduction method, such as manifold learning.

After the key state variables are screened out by the feature selection algorithm and the key state variables are reduced by the selected dimension reduction method, the dynamic simulation results of the power system can be reduced to a low-dimensional plane. Using the results of dimensionality reduction to train a transiently stable classifier can greatly shorten the time required for model training without reducing the accuracy of model training.

The key state variable selection method for power system transient stability evaluation proposed by the present invention can significantly shorten the transient stability classifier without reducing the classification accuracy of the transient stability discriminator through the selection of key state variables and dimensionality reduction The training time and classification time are more suitable for online applications.

Based on the above embodiment, the dynamic simulation data of the power system further includes: dynamic waveform data of the power system after fault removal.

Specifically, in the selection of input data, there are usually two types: one is the static data based on the power flow section before the fault, and the other is the dynamic data after the fault is removed. Using all the dynamic data after the fault is removed, because more information is taken into account, the accuracy in stability judgment is higher. Therefore, the present invention selects the dynamic waveform data of the power system after the fault is removed.

Based on the above-mentioned embodiment, the step S1 further includes:

Use the FFT algorithm to extract the real and imaginary parts of the first n harmonics of the time series of each state variable in the power system dynamic simulation data, so that the time series attributes of each state variable can be converted into 2n real number attributes, where n is A natural number greater than 1.

Specifically, FFT is a fast algorithm for discrete Fourier transform, which can transform a signal from the time domain to the frequency domain. The time series of each state variable in the obtained power system dynamic simulation data can be converted into real numbers through the FFT algorithm, that is, the real part and imaginary part of the first n harmonics of the time series of each state variable can be extracted through the FFT algorithm In this way, a time series can be converted into 2n real numbers, and the time series of each state variable can be converted into 2n real numbers. Since each sample of power system dynamic simulation data can be composed of time series of m state variables, the sample The attribute of corresponds to the time series of each state variable, that is, the attribute of the sample is converted to be composed of 2n real numbers of each state variable.

Based on the above-mentioned embodiment, the step S2 further includes:

According to the 2n real number attributes of each state variable, the state variables with the highest weight are screened out by using the Relief algorithm, and multiple key state variables are obtained.

Specifically, Relief (Relevant Features) is a well-known filtering feature selection method. Relief is a series of algorithms, including the earliest proposed Relief and the later expanded Relief-F and RRelief-F. The earliest proposed Relief is aimed at two For classification problems, the RRelief-F algorithm can solve multi-classification problems, and the RRelief-F algorithm is aimed at the regression problem where the target attribute is a continuous value. The original Relief algorithm was originally proposed by Kira. It is an algorithm based on feature weights, which can determine different weights of features according to the correlation between each feature of the sample and the category. If the weight of a feature is less than a certain threshold, then this feature will be removed. The weight of features in the Relief algorithm is determined by the feature's ability to distinguish close samples. The algorithm randomly selects a sample R from the training set D, and then finds the nearest neighbor sample H from the samples of the same type as R, which is called Near Hit, and finds the nearest neighbor sample M from the samples that are different from R, called NearMiss, and then Update the weight of each feature according to the following rules: If the distance between R and Near Hit on a certain feature is smaller than the distance between R and NearMiss, it means that this feature is beneficial to distinguish the nearest neighbors of the same class from different classes, then increase the The weight of the feature; conversely, if the distance between R and Near Hit on a certain feature is greater than the distance between R and Near Miss, it means that the feature has a negative effect on distinguishing the nearest neighbors of the same class from different classes, and the weight of the feature is reduced. The above process is repeated m times, and finally the average weight of each feature is obtained. The greater the weight of a feature, the stronger the classification ability of the feature, and vice versa, the weaker the classification ability of the feature. The running time of the Relief algorithm increases linearly with the number of sample sampling m and the number of original features, so the running efficiency is very high.

After step S1 is executed, the attributes of each sample in the power system dynamic simulation data are converted from the time series of m state variables to real numbers by the FFT algorithm, so the traditional Relief algorithm can be used for feature selection, according to each state variable The 2n real number attributes of , filter out the state variables with the highest weight ranking, so as to obtain multiple key state variables.

Based on the above-mentioned embodiments, the step of obtaining a plurality of key state variables specifically includes:

S21, randomly select a dynamic simulation data R from a plurality of dynamic simulation data, and then search for the nearest neighbor dynamic simulation data H from a sample set of the same type as the dynamic simulation data R, and search for the nearest dynamic simulation data H from a sample set different from the dynamic simulation data R Neighboring dynamic simulation data M;

S22, update the weight of the dynamic simulation data R according to the following rules: if the distance between the dynamic simulation data R and the dynamic simulation data H on a certain state variable is smaller than the distance between the dynamic simulation data R and the dynamic simulation data M on the state variable , then increase the weight of the state variable; or, if the distance between the dynamic simulation data R and the dynamic simulation data H on a certain state variable is greater than the distance between the dynamic simulation data R and the dynamic simulation data M on the state variable, then reduce the weights of state variables;

S23, repeating the steps S1 and S2p times to obtain the weight values of multiple state variables, sort the state variables according to the weight values from large to small, and take the first q state variables with the largest weight values as key state variables;

Among them, the values of p and q are determined according to the requirements of power system transient stability assessment.

The above has given the specific process of using the Relief algorithm to screen out key state variables from multiple state variables of the power system.

Based on the above-mentioned embodiment, the step S3 further includes:

The principal component analysis method is used to reduce the dimensionality of each of the key state variables.

Using principal component analysis to reduce the dimensionality of each of the key state variables specifically refers to using principal component analysis to reduce the dimensionality of the time series of each key state variable. The application of principal component analysis is very mature, so I won't go into details here.

Based on the above-mentioned embodiment, the step S3 further includes:

A popular learning method is used to reduce the dimensionality of each of the key state variables.

Specifically, the dimensionality reduction method of manifold learning assumes that the sample is located on a low-dimensional manifold, so a mapping can be found on the low-dimensional manifold to reduce the dimensionality of the sample. Thus, the distance of the samples on the low-dimensional manifold can be consistent with that in the high-dimensional space. For power systems, different post-fault dynamic waveforms can actually be regarded as manifolds under the combination of different operating conditions, different fault types and fault intensities, so the method of manifold learning can be used to control Perform dimensionality reduction.

After the improved feature selection algorithm and the dimension reduction method of manifold learning, the dynamic simulation results of the power system can be reduced to a low-dimensional plane. Using the results of dimensionality reduction to train a transiently stable classifier can greatly shorten the time required for model training without reducing the accuracy of model training.

The beneficial effect of the method provided by the present invention will be further described below in conjunction with a simulation example. The simulation test is carried out on the system of 10 machines and 39 nodes in the power system. 4000 samples are generated on the system of 10 machines and 39 nodes, among which 3200 samples are used as training set and 800 samples are used as test set. In the simulation samples, the characteristic attribute of each sample is the time series of 158 state variables such as voltage, current and power. Using the feature selection algorithm, the 20 state variables with the largest weight are screened from 158 state variables, as shown in Table 1.

Table 1 Selection results of key state variables

Use the selected time series of key state variables for dimensionality reduction. On this basis, the results of training using all dynamic data, training using the improved state variable extraction method, and training after dimensionality reduction are compared. Figure 2 shows the time-consuming analysis and comparison diagram of the three training methods.

For a classifier, there is generally a classification error. The evaluation index is an important part for measuring the performance of the classifier and guiding the parameter adjustment of the classification model. However, in the process of power system transient stability assessment, the losses to the system are different if the unstable samples are evaluated as stable and the stable samples are evaluated as unstable. Therefore, more indicators need to be used to measure the comprehensive performance of the classifier.

The following indicators are defined:

True case rate:

False positive rate:

False Negative Rate:

Correct rate:

In addition, the receiver characteristic curve (ROC) uses TPR as the vertical axis and FPR as the horizontal axis. It is used to evaluate the performance of the classifier, the larger the area of the plane spanned by the curve, the better the performance of the classifier. Therefore, the performance of a classifier can be measured by calculating the area of the envelope of the curve (AUC).

Table 2 shows the accuracy analysis results of the three training methods based on the above indicators.

Table 2 Accuracy analysis of three training methods

index FPR FNR Acc AUC Raw data 0.29% 0 99.71% 0.9981 feature selection data 1.01% 0.87% 98.11% 0.9921 Dimensionality reduction data 0.43% 1.01% 98.56% 0.9927

From the above data, it can be shown that although using all the dynamic data, the accuracy is slightly better than the result of key state variable selection and dimensionality reduction, but in terms of training time and prediction time, the selection of key state variables and the method after dimensionality reduction The time-consuming is much less than the method of using all dynamic data, even there is an order of magnitude difference.

As shown in FIG. 3 , it is a schematic structural diagram of a key state variable selection device for power system transient stability assessment provided by another embodiment of the present invention, including: a preprocessing module 31, a feature selection module 32 and a dimensionality reduction module 33, wherein ,

A preprocessing module 31, configured to obtain a plurality of state variables in the power system dynamic simulation data, and use an FFT algorithm to preprocess the plurality of state variables;

A feature selection module 32, configured to perform feature selection on the preprocessed plurality of state variables to obtain a plurality of key state variables;

A dimensionality reduction module 33, configured to perform dimensionality reduction on each of the key state variables.

Specifically, when the power system transient stability assessment method based on machine learning is used to evaluate the transient stability of the power system, the dynamic dynamic simulation data is used. Dynamic simulation data contains more power system state variables, and at the same time simulates for a longer time length, so it is easy to cause the problem of "curse of dimensionality". For example, if there are 2000 selected state variables, the simulation time length is 10s, and the step size is 0.01s, then the result of each simulation is a 2000*1000 matrix. When training the transient stability evaluation classifier, if If 10,000 samples are used, it is easy to cause the samples to be too large to be analyzed, and the training takes a long time, which is not conducive to the multiple training parameter adjustment and online application of the transient stability evaluation classifier. Therefore, it is necessary to process the dynamic simulation data of the power system. The key state variable selection device for power system transient stability evaluation provided by the embodiment of the present invention first screens out the key state variables from a plurality of state variables. In terms of numbers, the size of the dynamic simulation data is compressed, and combined with the dimensionality reduction method, the dimensionality reduction of the dynamic simulation data is further reduced, thereby speeding up the training speed of the classifier.

Selecting key state variables from multiple state variables requires the use of feature selection algorithms, and in traditional feature selection algorithms, the feature attributes of samples are real numbers. Since the dynamic simulation results of the power system can be expressed by the time series of m state variables, the time series of each state variable is the attribute of the sample. In order to use the traditional feature selection algorithm, the preprocessing module 31 first uses the FFT algorithm to preprocess the multiple state variables in the obtained power system dynamic simulation data, and converts the attributes of each state variable from time series to Real numbers, so that the traditional feature selection algorithm can be used to select the features of the samples, and the key state variables can be screened out.

Feature Selection (Feature Selection), also known as Feature Subset Selection (FSS), or Attribute Selection (Attribute Selection), refers to selecting a subset of features from all features to make the constructed model better. In the practical application of machine learning, the number of features is often large, there may be irrelevant features, and there may be interdependence between features, which can easily lead to the following consequences: the more the number of features, the more features required for analyzing features and training models. the longer the time. The larger the number of features, the more likely it is to cause the "dimension disaster", the more complex the model will be, and its generalization ability will decrease. Feature selection can eliminate irrelevant or redundant features, so as to reduce the number of features, improve the accuracy of the model, and reduce the running time. On the other hand, picking out the truly relevant features makes it easy for researchers to understand the process of data generation. After preprocessing by the preprocessing module 31, the feature selection module 32 can use the existing feature selection algorithm, such as the Relief algorithm, to perform feature selection on the multiple state variables after preprocessing, thereby locating the Key state variables for system transient stability assessment.

After being processed by the preprocessing module 31 and the feature selection module 32, the key state variables for power system transient stability evaluation are selected. But for an actual power system, if there are 2000 state variables, simulate 10s with a step size of 0.01s. Even if 50 key state variables are selected from 2000 state variables, the result of each simulation will still be a 50*1000 matrix, the simulation result is still very large, and the problem of "curse of dimensionality" may still exist. Therefore, the dimensionality reduction module 33 is required to further reduce the dimensionality of the key state variables that have been screened out. There are two commonly used dimensionality reduction methods, one is a linear dimensionality reduction method, such as PCA. The other is a nonlinear dimensionality reduction method, such as manifold learning.

After the key state variables are screened out by the feature selection algorithm and the key state variables are reduced by the selected dimension reduction method, the dynamic simulation results of the power system can be reduced to a low-dimensional plane. Using the results of dimensionality reduction to train a transiently stable classifier can greatly shorten the time required for model training without reducing the accuracy of model training.

Based on the above-mentioned embodiments, the preprocessing module 31 is specifically used for:

Use the FFT algorithm to extract the real and imaginary parts of the first n harmonics of the time series of each state variable in the power system dynamic simulation data, so that the time series attributes of each state variable can be converted into 2n real number attributes, where n is A natural number greater than 1.

Specifically, FFT is a fast algorithm for discrete Fourier transform, which can transform a signal from the time domain to the frequency domain. The time series of each state variable in the obtained power system dynamic simulation data can be converted into real numbers through the FFT algorithm, that is, the real part and imaginary part of the first n harmonics of the time series of each state variable can be extracted through the FFT algorithm In this way, a time series can be converted into 2n real numbers, and the time series of each state variable can be converted into 2n real numbers. Since each sample of power system dynamic simulation data can be composed of time series of m state variables, the sample The attribute of corresponds to the time series of each state variable, that is, the attribute of the sample is converted to be composed of 2n real numbers of each state variable.

Based on the foregoing embodiments, the feature selection module 32 is specifically used for:

According to the 2n real number attributes of each state variable, the state variables with the highest weight are screened out by using the Relief algorithm, and multiple key state variables are obtained.

Specifically, Relief (Relevant Features) is a well-known filtering feature selection method. Relief is a series of algorithms, including the earliest proposed Relief and the later expanded Relief-F and RRelief-F. The earliest proposed Relief is aimed at two For classification problems, the RRelief-F algorithm can solve multi-classification problems, and the RRelief-F algorithm is aimed at the regression problem where the target attribute is a continuous value. The original Relief algorithm was originally proposed by Kira. It is an algorithm based on feature weights, which can determine different weights of features according to the correlation between each feature of the sample and the category. If the weight of a feature is less than a certain threshold, then this feature will be removed. The weight of features in the Relief algorithm is determined by the feature's ability to distinguish close samples. The algorithm randomly selects a sample R from the training set D, and then finds the nearest neighbor sample H from the samples of the same type as R, which is called Near Hit, and finds the nearest neighbor sample M from the samples that are different from R, called NearMiss, and then Update the weight of each feature according to the following rules: If the distance between R and Near Hit on a certain feature is smaller than the distance between R and NearMiss, it means that this feature is beneficial to distinguish the nearest neighbors of the same class from different classes, then increase the The weight of the feature; conversely, if the distance between R and Near Hit on a certain feature is greater than the distance between R and Near Miss, it means that the feature has a negative effect on distinguishing the nearest neighbors of the same class from different classes, and the weight of the feature is reduced. The above process is repeated m times, and finally the average weight of each feature is obtained. The greater the weight of a feature, the stronger the classification ability of the feature, and vice versa, the weaker the classification ability of the feature. The running time of the Relief algorithm increases linearly with the number of sample sampling m and the number of original features, so the running efficiency is very high.

After being processed by the preprocessing module 31, the attributes of each sample in the power system dynamic simulation data are converted from the time series of m state variables to real numbers through the FFT algorithm, so the feature selection module 22 can use the traditional Relief algorithm to perform feature Select, according to the 2n real number attributes of each state variable, filter out the state variables with the highest weight ranking, so as to obtain multiple key state variables.

According to the 2n real number attributes of each state variable, use the Relief algorithm to screen out the state variables with the highest weight, and the steps to obtain multiple key state variables include:

S21, randomly select a dynamic simulation data R from a plurality of dynamic simulation data, and then search for the nearest neighbor dynamic simulation data H from a sample set of the same type as the dynamic simulation data R, and search for the nearest dynamic simulation data H from a sample set different from the dynamic simulation data R Neighboring dynamic simulation data M;

S22, update the weight of the dynamic simulation data R according to the following rules: if the distance between the dynamic simulation data R and the dynamic simulation data H on a certain state variable is smaller than the distance between the dynamic simulation data R and the dynamic simulation data M on the state variable , then increase the weight of the state variable; or, if the distance between the dynamic simulation data R and the dynamic simulation data H on a certain state variable is greater than the distance between the dynamic simulation data R and the dynamic simulation data M on the state variable, then reduce the weights of state variables;

S23, repeating the steps S1 and S2p times to obtain the weight values of multiple state variables, sort the state variables according to the weight values from large to small, and take the first q state variables with the largest weight values as key state variables;

Among them, the values of p and q are determined according to the requirements of power system transient stability assessment.

The key state variable selection device for power system transient stability evaluation proposed by the present invention can significantly shorten the transient stability classifier without reducing the classification accuracy of the transient stability discriminator through the selection of key state variables and dimensionality reduction The training time and classification time are more suitable for online applications.

Finally, the method of the present invention is only a preferred embodiment, and is not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims (10)

1. A method for selecting key state variables for power system transient stability assessment, characterized in that it includes: S1. Obtain a plurality of state variables in the dynamic simulation data of the power system, and use an FFT algorithm to preprocess the plurality of state variables; S2, performing feature selection on the plurality of preprocessed state variables to obtain a plurality of key state variables; S3. Perform dimensionality reduction on each of the key state variables. 2. The method according to claim 1, wherein the power system dynamic simulation data further comprises: power system dynamic waveform data after fault removal. 3. The method according to claim 1, wherein said step S1 further comprises: Use the FFT algorithm to extract the real and imaginary parts of the first n harmonics of the time series of each state variable in the power system dynamic simulation data, so that the time series attributes of each state variable can be converted into 2n real number attributes, where n is A natural number greater than 1. 4. The method according to claim 3, wherein said step S2 further comprises: According to the 2n real number attributes of each state variable, the state variables with the highest weight are screened out by using the Relief algorithm, and multiple key state variables are obtained. 5. method according to claim 4, is characterized in that, described according to 2n real number attributes of each state variable, utilizes the Relief algorithm to filter out the state variable with the highest weight, and the step of obtaining a plurality of key state variables further comprises : S21, randomly select a dynamic simulation data R from a plurality of dynamic simulation data, and then search for the nearest neighbor dynamic simulation data H from a sample set of the same type as the dynamic simulation data R, and search for the nearest dynamic simulation data H from a sample set different from the dynamic simulation data R Neighboring dynamic simulation data M; S22, update the weight of the dynamic simulation data R according to the following rules: if the distance between the dynamic simulation data R and the dynamic simulation data H on a certain state variable is smaller than the distance between the dynamic simulation data R and the dynamic simulation data M on the state variable , then increase the weight of the state variable; or, if the distance between the dynamic simulation data R and the dynamic simulation data H on a certain state variable is greater than the distance between the dynamic simulation data R and the dynamic simulation data M on the state variable, then reduce the the weight of the state variable; S23, repeating the steps S1 and S2p times to obtain the weight values of multiple state variables, sort the state variables according to the weight values from large to small, and take the first q state variables with the largest weight values as key state variables; Among them, the values of p and q are determined according to the requirements of power system transient stability assessment. 6. The method according to claim 1, wherein said step S3 further comprises: The principal component analysis method is used to reduce the dimensionality of each of the key state variables. 7. The method according to claim 1, wherein said step S3 further comprises: A popular learning method is used to reduce the dimensionality of each of the key state variables. 8. The key state variable selection device for power system transient stability assessment, characterized in that it includes: A preprocessing module, configured to obtain a plurality of state variables in the power system dynamic simulation data, and use an FFT algorithm to preprocess the plurality of state variables; A feature selection module, configured to perform feature selection on the preprocessed multiple state variables to obtain multiple key state variables; A dimensionality reduction module, configured to perform dimensionality reduction on the multiple key state variables. 9. The device according to claim 8, wherein the preprocessing module is specifically used for: Use the FFT algorithm to extract the real and imaginary parts of the first n harmonics of the time series of each state variable in the power system dynamic simulation data, so that the time series attributes of each state variable can be converted into 2n real number attributes, where n is A natural number greater than 1. 10. The device according to claim 9, wherein the feature selection module is specifically used for: According to the 2n real number attributes of each state variable, the state variables with the highest weight are screened out by using the Relief algorithm, and multiple key state variables are obtained.