CN116031879A - A hybrid intelligent feature selection method for power system transient voltage stability assessment - Google Patents

Abstract

The invention discloses a hybrid intelligent feature selection method suitable for transient voltage stability evaluation of a power system, which mainly comprises the following steps: s1, sample generation: the method comprises the steps of obtaining a high-dimensional time sequence sample data set through stability related factor analysis, original feature set construction and multi-scene transient time domain simulation; s2, feature effectiveness measurement and preliminary screening based on a T-Relief algorithm; s3, feature selection and stability evaluation based on an improved population intelligent algorithm: the search performance of the group intelligent optimization algorithm is enhanced through the feature validity metric value obtained in the step S2, and an improved group intelligent optimization algorithm is obtained; based on the algorithm, a packaged feature selection scheme is built, a ConvGRU evaluation model is embedded as a subset evaluator, feature subset optimization is further achieved on the basis of primary screening of features in step S2, and feature redundancy is reduced. The method can directly process the time sequence characteristics, and can improve the searching efficiency as much as possible while guaranteeing the screening quality of the characteristic subsets.

Description

A hybrid intelligent feature selection method for transient voltage stability assessment of power systems

Technical Field

The invention relates to the technical field of power systems, and in particular to a hybrid intelligent feature selection method adapted to transient voltage stability assessment of power systems.

Background Art

The large-scale grid connection of new energy and the commissioning of multiple UHV DC transmission projects have made the new power system "double high and one low" and the "hollowing out" of load centers obvious, and the uncertainty of source-load has intensified. The high degree of power electronics has caused profound changes in the dynamic characteristics of the power grid, and at the same time, the fault tolerance has been reduced, which has increased the risk of transient voltage instability in the system. With the development and maturity of big data theory and artificial intelligence technology, as well as the rapid popularization of power grid measurement devices, the response data-driven artificial intelligence method has provided a new idea for the rapid realization of transient voltage stability assessment. However, there are a large number of various components in the actual power system, the scale of the power grid is huge, the electrical characteristics are high-dimensional data, and the redundancy between features will greatly affect the classification performance and efficiency of the assessment model. In recent years, in order to fully explore the mechanism evolution information of key features after the system is disturbed and thus improve the accuracy of stability assessment, some literatures have proposed using dynamic time series data for transient stability analysis, which undoubtedly increases the computational burden and overfitting risk of the model. Therefore, exploring appropriate feature selection schemes to reduce the original feature dimension is a key issue in the application of artificial intelligence methods in transient voltage stability assessment of power systems.

In previous studies, manual methods were often used to select key features, that is, relying on expert experience to select variable features that conform to prior knowledge for stability analysis; currently, many scholars have conducted research on feature selection based on data mining, mainly in three types: filtering, embedded, and encapsulated. The filtering method often selects the features most relevant to the target attribute by calculating the correlation between the feature variable and the label. Typical correlation measurement criteria include relief statistics, information measurement, Fisher statistics, etc. The embedded method is to realize feature selection synchronously during the training of the learning model. Typical embedded feature selection models include decision trees. The encapsulated method is to find the best feature subset that matches it through the mutual cooperation of the optimization algorithm and the learning model in an iterative optimization manner.

Since the fault characteristics of many new dynamic response devices in the current power system are unclear, and the mechanism analysis of transient voltage instability is also unclear, it is difficult to adapt to the characteristics of complex large power grids and easily cause information omissions by relying solely on manual experience to select variable features. In the feature selection research based on data mining, although the filtering method has fast calculation speed and strong operability, its feature screening principle is relatively simple, and it cannot effectively reduce feature redundancy, and cannot be effectively applied to time series feature selection; the embedded method is too dependent on the learning model and is difficult to adapt to the requirements of rapid evaluation of transient voltage stability; the encapsulation method has a high accuracy rate in feature evaluation, but the algorithm calculation cost is high, and the operation efficiency needs to be improved. Therefore, the existing feature selection methods are still insufficient in terms of balancing screening efficiency and screening accuracy.

Summary of the invention

In view of the problem that it is difficult to achieve both feature subset screening efficiency and classification performance in current power system analysis feature selection methods, the present invention provides a hybrid intelligent feature selection method suitable for power system transient voltage stability assessment.

The hybrid intelligent feature selection method adapted to transient voltage stability assessment of power systems provided by the present invention comprises the following steps:

S1. Sample generation:

By analyzing stability-related factors, constructing original feature sets, and simulating transient time domains in multiple scenarios, a high-dimensional time series sample data set is obtained. The time domain simulation obtains transient voltage sample data after the power system is disturbed, and completes sample stability labeling based on practical engineering criteria, thereby constructing training and test data sets for feature selection and evaluation models, namely, high-dimensional time series sample data sets.

S2. Feature validity measurement and preliminary screening based on T-Relief algorithm.

The Relief algorithm has the advantage of high computational efficiency, but it is not directly applicable to time series input feature selection. Therefore, it is improved by time series hierarchical processing to obtain the T-Relief algorithm. This algorithm is used for the preliminary screening of original features. On the one hand, the feature classification effectiveness metric is calculated to enhance the search performance of the subsequent step S3. On the other hand, it realizes preliminary dimensionality reduction and reduces the iterative computation cost of step S3. This step specifically includes the following sub-steps:

S21. Assume that the original feature data set has a total of M samples, each sample has d feature attributes, and the number of time points of data records is N; perform time series stratification on M time series high-dimensional feature data with N time steps to form N stratified high-dimensional feature matrices;

S22, respectively calculating the Euclidean distances between M samples and other samples in N hierarchical high-dimensional feature matrices, and constructing a Euclidean distance matrix;

S23, according to the Euclidean distance matrix, find the neighboring samples of the high-dimensional feature samples of each layer; calculate the relevant statistics of each feature at the corresponding time;

S24, averaging the relevant statistics obtained at different time layers to obtain the effectiveness measurement value δ of each feature of the comprehensive time series information;

S25, setting a threshold τ, screening out valid features that meet the conditions and proceeding to the next step. The set threshold τ may be the number of screened features or a specific value of a related statistic, etc.

S3. Feature selection and stability evaluation based on improved swarm intelligence algorithm.

This step uses the feature validity metric obtained in step S2 to enhance the search performance of the swarm intelligence optimization algorithm (BGOA) to obtain the improved swarm intelligence optimization algorithm (IBGOA). Based on this algorithm, a packaged feature selection scheme is built, and the ConvGRU evaluation model is embedded as a subset evaluator, which can fully consider the temporal evolution information and internal combination relationship of the features, and further realize feature subset optimization based on the initial screening features in step S2 to reduce feature redundancy.

This step includes the following sub-steps:

S31, using improved binary locust optimization algorithm for feature selection;

The effectiveness metric value δ obtained in step S24 is used to enhance the search performance of the binary locust optimization algorithm to obtain an improved binary locust optimization algorithm; in the improved binary locust optimization algorithm, the improved locust position initialization formula and iterative update formula are as follows:

In the formula, a∈(0,1) is the weight coefficient, r∈[0,1] is a uniformly distributed random number; Round(.) represents the rounding function;

In the formula, β, η, γ are weight coefficients, r∈[0,1] is a uniformly distributed random number;

S32. The multi-dimensional time series binary classification model embedded with ConvGRU evaluates the classification performance of the feature subset. The fitness function is used as a comprehensive evaluation index to determine whether the comprehensive evaluation index meets the iteration number requirement. If the iteration number requirement is met, the optimal feature subset is output. If the iteration number requirement is not met, the improved binary locust optimization algorithm is repeatedly used for feature selection.

In this step, the calculation formula of ConvGRU is as follows:

R _t =σ(W _xr *X _t +W _hr *H _t-1 )

Z _t =σ(W _xz *X _t +W _hz *H _t-1 )

表示当前候选值；H_t-1表示上一时刻隐藏状态；W_xr、W_hr、W_xz、W_hz、W_xh、W_hh分别表示对应连接的权重系数。Wherein, σ and tanh represent activation functions; ⊙ represents element-wise multiplication; * represents convolution operation; R _t represents the reset gate; Z _t represents the update gate; X _t represents the input data at the current moment;

represents the current candidate value; H _t-1 represents the hidden state at the previous moment; W _xr , W _hr , W _xz , W _hz , W _xh , and W _hh represent the weight coefficients of the corresponding connections respectively.

The fitness function used is as follows:

In the formula, β∈(0,1) is the weight coefficient, |R| is the dimension of the corresponding feature subset, |N| is the original feature dimension; P _c is the comprehensive misjudgment rate indicator;

The calculation formula of the comprehensive misjudgment rate index _Pc is as follows:

P _c = αP _fs + (1-α)P _fus

Among them, the missed rate

式中，α为权重系数，F_s表示错误判定为稳定的样本数(漏判)；F_us表示错误判定为失稳的样本数(误判)；T_s表示正确判定为稳定的样本数；T_us表示正确判定为失稳的样本数。False positive rate

Where α is the weight coefficient, _Fs represents the number of samples incorrectly judged as stable (missed judgment); _Fus represents the number of samples incorrectly judged as unstable (misjudgment); _Ts represents the number of samples correctly judged as stable; _Tus represents the number of samples correctly judged as unstable.

Compared with the prior art, the present invention is beneficial in that:

(1) By improving the time series of the Relief algorithm, we obtain the T-Relief algorithm, which is suitable for measuring the effectiveness of high-dimensional time series features and still has the advantage of high computational efficiency.

(2) The feature validity metric is integrated to improve the initialization and iterative update formula of the swarm intelligent optimization algorithm, effectively improving the efficiency and performance of encapsulated feature selection based on this intelligent algorithm;

(3) The hybrid intelligent feature selection method based on double screening can effectively achieve feature dimensionality reduction. The initial screening stage provides feature validity measurement and preliminary dimensionality reduction, which improves the efficiency of subsequent subset optimization. The encapsulated stage embedded ConvGRU timing evaluation model based on the improved group intelligence optimization algorithm can fully consider the time-varying characteristics of the features and has higher classification accuracy. Compared with other commonly used feature selection methods, the feature subset classification performance selected by the method of the present invention is better.

Other advantages, objectives and features of the present invention will be embodied in part through the following description, and in part will be understood by those skilled in the art through study and practice of the present invention.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1. Block diagram of hybrid intelligent feature selection algorithm.

FIG2 is a flow chart of the improved T-Relief algorithm of the present invention.

Figure 3. Schematic diagram of the ConvGRU unit structure.

Figure 4. Comparison of iterative convergence curves of IBGOA and BGOA.

DETAILED DESCRIPTION

The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

As shown in Figure 1, the hybrid intelligent feature selection method adapted to the transient voltage stability assessment of the power system provided by the present invention mainly includes three parts: sample generation, feature validity measurement and preliminary screening based on the T-Relief algorithm, and feature subset optimization and stability assessment based on the swarm intelligence optimization algorithm.

The principle of the T-Relief algorithm is as follows:

The Relief algorithm is a filtering-based feature selection method proposed for binary classification problems. This method calculates relevant statistics based on the ability of features to distinguish close-range samples, thereby evaluating the effectiveness of different features. The specific steps are as follows:

(1) Based on the Euclidean distance, the sample closest to the specified sample is found among the same samples as the "guessed neighbor", and the sample closest to the specified sample is found among the different samples as the "wrongly guessed neighbor". Then, the relevant statistics of each feature are calculated based on the selected "guessed neighbors" and "wrongly guessed neighbors". Finally, the relevant statistics of all samples in the training set are summed to obtain the overall feature importance measurement result. The calculation formula is as follows:

的取值为：In the formula, the superscript j represents the j-th dimension feature of the corresponding sample, δ is the list of relevant statistics of all features, _xi is the current calculation sample, xi _,nh is the "guessed neighbor" of the current sample, and xi _,nm is the "wrongly guessed neighbor" of the current sample. Since the power system response data is a continuous variable,

The value of is:

(2) Sort the relevant statistics of each feature, select a suitable threshold, and choose features with higher importance for classification tasks.

In the method of the present invention, the Relief algorithm is improved in time series by time series layering processing, and the improved T-Relief algorithm is obtained, so that it can directly screen high-dimensional time series features. Assuming that the original feature data set has a total of M samples, each sample has d feature attributes, and the number of time points of data records is N, the calculation process of the T-Relief algorithm is shown in Figure 2, which is as follows:

① Perform time-series stratification on M time-series high-dimensional feature data with N time steps to form N stratified high-dimensional feature matrices;

② Calculate the Euclidean distances between M samples and other samples in N hierarchical high-dimensional feature matrices respectively, and construct a Euclidean distance matrix;

③ Based on the Euclidean distance matrix, find the neighboring samples of each layered high-dimensional feature sample, normalize the original feature data, and calculate the relevant statistics of each feature at the corresponding time; the calculation formula is as follows:

④Average the relevant statistics obtained at different time levels to obtain the effectiveness measurement values of each feature of the comprehensive time series information;

⑤ Set a threshold value τ to screen out valid features that meet the conditions for the next step of screening research. The threshold value τ in this method can be the number of screened features or a specific value of related statistics.

The principle of the improved swarm intelligence optimization algorithm is as follows:

The encapsulated feature selection method based on swarm intelligence is an important means to find the approximate solution of the optimal feature subset. The present invention selects the relatively robust and reliable Binary Grasshopper Optimization Algorithm (BGOA) as the search strategy for encapsulated feature selection.

BGOA is a swarm intelligence optimization algorithm proposed to simulate locust foraging behavior and mathematically model the social forces between individuals. The algorithm has a simple structure and stable performance. The mathematical model is:

_Xi ＝ _Si + _Gi + _Ai

In the formula, the subscript i represents the i-th locust in the population; _Xi represents its position; _Si is the social force exerted on it by other agents; _Gi is its own gravity; _Ai is the wind force it is subjected to, and gravity and wind are often ignored in practical applications. The calculation formula of _Si is as follows:

In the formula, d _ij = | _{xi -xj} | represents the distance between two locusts, s(.) is the social force strength function, positive numbers represent attraction, and negative numbers represent repulsion. The calculation formula is as follows:

In the formula, f is the gravitational intensity parameter, l is the gravitational range parameter, f=0.5, l=1.5, r is an independent variable with no special meaning, which is equivalent to the code of _dij in the above formula.

In the feature selection problem, the position vector _Xi of BGOA is a binary sequence of length D, indicating the selection of D features, that is, _Xid = 1 means that the d-th dimension feature is selected, and vice versa. Therefore, combined with the social force, the step vector of locust position update is defined as _Δxi , and the S-type transfer function is used to convert it into the probability of selection. The specific calculation method is as follows:

In the formula, ^ud and ^ld are the upper and lower limits of the feature value respectively; t represents the iteration round; T(.) is a function; c is the adaptive adjustment coefficient, and the adjustment method is as follows:

c＝c _max −t(c _max −c _min )/L

Wherein, c _max and c _min are the maximum adjustment coefficient and the minimum adjustment coefficient respectively, and L represents the maximum number of iterations.

In the present invention, in order to integrate the guiding role of the feature validity metric value of the T-Relief stage on the iterative update of locusts, the locust position initialization formula and iterative update formula are improved as follows:

In the formula, a∈(0,1) is the weight coefficient, r∈[0,1] is a uniformly distributed random number. Round(.) represents the rounding function.

In the formula, β, η, and γ are weight coefficients, and r∈[0,1] is a uniformly distributed random number.

The structure of the ConvGRU unit is shown in Figure 3. The evaluation of the quality of the feature subset to be selected mainly considers two aspects, namely, the classification effect of using the feature subset for stable evaluation and the dimension of the feature subset.

In order to evaluate the performance of the classifier, the accuracy rate P _acc , missed detection rate P _fs , and false positive rate P _fus indicators are defined to evaluate the performance of the temporary stable classifier (ConvGRU). At the same time, the weight coefficient α>1 is introduced to balance the importance of false positives and missed detections, and the comprehensive false positive rate indicator P _c is defined to evaluate the classification performance of the feature subset.

P _c = αP _fs + (1-α)P _fus

In order to maximize the classification accuracy and minimize the dimension of the feature subset, the objective function of the stable evaluation feature selection (that is, the optimization fitness function of the improved BGOA feature selection algorithm) is defined as:

Where β∈(0,1) is the weight coefficient, |R| is the dimension of the corresponding feature subset, and |N| is the original feature dimension.

Performance comparison analysis:

(1) Performance comparison between the improved BGOA algorithm (IBGOA) and the original BGOA algorithm

In order to verify that the improved BGOA algorithm (IBGOA) of the present invention has better performance in the subset iterative optimization process, a comparative experiment is conducted on the two. Experimental setup: BGOA and IBGOA are used to perform feature screening on the original feature set, and the number of iterations is 50 times. The average fitness value of each iteration is obtained by multiple experiments. The final fitness value iterative convergence curve is shown in Figure 4. It can be clearly seen from the figure that under the IBGOA method, the optimization search efficiency of the subset is higher, and the fitness value of the optimal subset is lower, which fully proves the effectiveness of the present invention in improving BGOA.

(2) Comparison of transient voltage stability assessment results before and after feature selection

The effects of feature selection were compared and demonstrated from the perspectives of stable evaluation accuracy, missed detection rate, misjudgment rate, feature dimension and model training time.

Experimental details: After the original features (370 dimensions) were calculated by T-Relief, 131 dimensions of effective features with effectiveness metrics greater than 0.6 were retained; after IBGOA iterative optimization, 18 dimensions of key features were finally retained (average values were taken from multiple experiments, and each experiment was iterated 50 times). The experimental results are shown in Table 1. As can be seen from Table 1, after the feature selection method of the present invention was used, only 18 key features were finally retained, and the model training time was even reduced by 76.41%. This will greatly reduce the difficulty of model training when the model complexity increases and the sample size increases sharply. In addition, although the dimension of the features was greatly reduced, the accuracy of the model evaluation increased by 2.73%, and the false positive rate and missed positive rate were both reduced by 1.365%. This shows that the feature selection method proposed in the present invention can effectively remove invalid features in high-dimensional features, thereby improving the accuracy of the transient voltage stability assessment model.

Table 1. Comparison of stable evaluation effects before and after feature selection

The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technician familiar with this profession can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still fall within the scope of the technical solution of the present invention.

Claims (4)

1. A hybrid intelligent feature selection method suitable for transient voltage stability evaluation of a power system is characterized by comprising the following steps:

s1, sample generation:

the method comprises the steps of obtaining a high-dimensional time sequence sample data set through stability related factor analysis, original feature set construction and multi-scene transient time domain simulation;

s2, feature effectiveness measurement and preliminary screening based on a T-Relief algorithm, comprising the following substeps:

s21, assuming that an original characteristic data set has M samples, wherein each sample has d characteristic attributes, and the number of time points of data record is N; performing time sequence layering on M time sequence high-dimensional characteristic data with N time steps to form N layered high-dimensional characteristic matrixes;

s22, respectively calculating Euclidean distances between M samples and other samples in the N layered high-dimensional feature matrices, and constructing an Euclidean distance matrix;

s23, searching for neighbor samples of each layered high-dimensional characteristic sample according to the Euclidean distance matrix; calculating the relevant statistics of each feature at the corresponding moment;

s24, averaging the related statistics obtained under different time layering to obtain each characteristic effectiveness metric delta of the comprehensive time sequence information;

s25, setting a threshold tau, screening out effective features meeting the conditions, and carrying out the next step;

s3, feature selection and stability evaluation based on an improved population intelligent algorithm, comprising the following substeps:

s31, adopting an improved binary locust optimization algorithm to perform feature selection;

the binary locust optimization algorithm is subjected to search performance enhancement through the effectiveness metric delta obtained in the step S24, so that an improved binary locust optimization algorithm is obtained; in the improved binary locust optimization algorithm, an improved locust position initialization formula and an iterative update formula are as follows:

wherein a epsilon (0, 1) is a weight coefficient, and r epsilon [0,1] is a uniformly distributed random number; round () represents a rounding function;

wherein, beta, eta and gamma are weight coefficients, and r E [0,1] is a uniformly distributed random number;

s32, classifying performance evaluation is carried out on the feature subsets by a multi-dimensional time sequence classification model embedded with ConvGRU, an fitness function is used as a comprehensive evaluation index, whether the comprehensive evaluation index meets the iteration number requirement is judged, if the comprehensive evaluation index meets the iteration number requirement, the optimal feature subset is output, and if the comprehensive evaluation index does not meet the iteration number requirement, an improved binary locust optimization algorithm is repeatedly adopted again to carry out feature selection.

2. The hybrid intelligent feature selection method for adapting to power system transient voltage stability assessment according to claim 1, wherein in step S1, transient voltage sample data after power system disturbance is obtained through time domain simulation, and sample stability labeling is completed according to engineering practical criteria, so as to construct training and testing data sets of feature selection and assessment models, namely high-dimensional time sequence sample data sets.

3. The hybrid intelligent feature selection method for adapting to power system transient voltage stability assessment according to claim 1, wherein in step S32, the calculation formula of convglu is as follows:

R _t ＝σ(W _xr *X _t +W _hr *H _t-1 )

Z _t ＝σ(W _xz *X _t +W _hz *H _t-1 )

wherein σ and tanh represent excitation functions; the addition of the elements is carried out; * Is convolution operation; r is R _t Schematic weightSetting a door; z is Z _t Schematic update gate; x is X _t Representing the current time input data;

representing the current candidate value; h _t-1 Representing the hidden state at the previous moment; w (W) _xr 、W _hr 、W _xz 、W _hz 、W _xh 、W _hh Respectively representing the weight coefficients of the corresponding connections.

4. The hybrid intelligent feature selection method for adapting to transient voltage stability assessment of a power system according to claim 3, wherein in step S32, the fitness function employed is as follows:

wherein, beta epsilon (0, 1) is a weight coefficient, r| is the dimension of the corresponding feature subset, and N| is the original feature dimension; p (P) _c Is a comprehensive erroneous judgment rate index;

P _c the calculation formula is as follows:

P _c ＝αP _fs +(1-α)P _fus

wherein, alpha is a weight coefficient,

miss rate P _fs The calculation formula of (2) is as follows:

false positive rate P _fus The calculation formula of (2) is as follows:

wherein F is _s Indicating the number of samples that were erroneously determined to be stable; f (F) _us A sample number indicating that the error determination is unstable; t (T) _s Representation ofCorrectly judging the stable sample number; t (T) _us The number of samples correctly determined as unstable is represented.

Images