Electric power system transient power angle and voltage integrated stability evaluation method based on two-stage feature selection

Through the combination of two-stage feature selection and convolutional gated cycle unit (ConvGRU), the problem of interleaving power angle and voltage in the power system transient stability evaluation is solved, and the integration of transient power angle and voltage stability of the power system is realized. Evaluation improves the accuracy and efficiency of the evaluation model.

CN119004005APending Publication Date: 2024-11-22CHINA THREE GORGES UNIV
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202410967518.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-18
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

In the current technology, in the transient stability evaluation of power systems, traditional methods are difficult to effectively analyze the interleaving problem of work angle and voltage, resulting in low evaluation efficiency and poor accuracy, and high-dimensional characteristics lead to overfitting of the model, which is unable to adapt to the rapidity of the new power system. Assess the needs.

Method used

Using a two-stage feature selection method, features are screened through class separability and Pearson correlation coefficients, and combined with a convolutional gating recurrent unit (ConvGRU) to perform multi-task transient stability integrated evaluation, reducing feature dimensions and improving the evaluation model Accuracy and efficiency.

Benefits of technology

The integrated evaluation of the transient power angle and transient voltage stability is achieved, which reduces the characteristic dimensions, improves the accuracy and applicability of the evaluation model, adapts to the rapid evaluation requirements of the new power system, and reduces misjudgment and leakage. judgment rate.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119004005A_ABST
    Figure CN119004005A_ABST
Patent Text Reader

Abstract

The invention discloses a transient power angle and voltage integrated stability evaluation method of a power system based on two-stage feature selection. The method comprises the following steps: constructing a comprehensive feature set representing a transient power angle stability problem and a transient voltage stability problem; a two-stage transient stability feature selection method is provided. In the first stage, an original comprehensive feature set is preliminarily screened based on class separability scores and Pearson's correlation coefficients. In the second stage, binarizing a competitive search algorithm (CSA), and embedding a multi-task transient stability integrated evaluation model based on a convolutional gating cycle unit ConvGRU in a selection process to further optimize a feature subset; and on the basis of an optimal feature subset obtained by a two-stage transient stability feature selection method, integrated evaluation of transient power angle stability and transient voltage stability is realized by using the constructed multi-task transient stability integrated evaluation model. According to the method, the feature dimension can be effectively reduced, and the effect of the transient power angle and transient voltage stability integrated evaluation model is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of transient stability assessment of electric power systems, and in particular to a method for evaluating the transient power angle and voltage integration stability of electric power systems based on two-stage feature selection. Background Art

[0002] With the formation of my country's super-large AC / DC hybrid power grid, my country's power grid presents a new network form. The characteristics and operation characteristics of the power grid will undergo major changes. The low fault tolerance capability makes the stable operation of the system face severe challenges. Among the traditional transient stability assessment methods, the time domain simulation method has high assessment accuracy but takes a long time to calculate; the direct method can quickly provide transient stability assessment results, but can only give approximate results and is difficult to use in complex power systems. Therefore, the applicability of the traditional time domain simulation method and the direct method in the current situation is challenged.

[0003] In recent years, wide area measurement systems (WAMS) and phase measurement units (PMU) have been widely used, and data-driven artificial intelligence methods have provided new solutions and research directions for transient stability assessment of power systems.

[0004] Transient power angle stability and transient voltage stability are two major problems faced by power systems during operation. When the system is unstable, power angle instability may cause voltage collapse, and voltage collapse may also cause power angle instability. These two problems are often intertwined, but most of them are currently analyzed by separating the power angle problem from the voltage problem. In addition, due to the rapid development of data measurement devices and transmission and storage technologies, power system data has shown an "explosive" growth in feature dimensions and sample quantity. Directly inputting the original time series data into the model will significantly increase the computational burden, making the evaluation process slow and inefficient. High-dimensional features will also cause the model to overfit, making it impossible to accurately evaluate the transient stability of the power system. Therefore, analyzing the important features that affect the power angle problem and the voltage problem under the same framework has important theoretical significance for exploring the key variables that affect the power angle voltage integration, and has important practical significance for improving the effect of the power angle voltage integration evaluation model and building a security defense system that adapts to the new power system.

[0005] In terms of feature sampling, current research mostly uses a multi-point non-continuous method. Although it has certain advantages in theoretical analysis and model construction, it faces challenges in actual engineering applications. Since data in actual scenarios presents large-scale and continuous influx characteristics, it is difficult to extract irregular data and it is difficult to meet the speed and accuracy of smart grid evaluation.

[0006] Previous studies on stability assessment usually use single-task learning models. However, since transient power angle stability and voltage stability are inseparable in mechanism, it is difficult to share information between tasks in a single-task learning model, resulting in low efficiency in assessing multiple stability issues. Multiple single-task learning models not only increase the consumption of storage and computing resources, but also limit the accuracy and efficiency of the assessment. Summary of the invention

[0007] In order to analyze the important electrical characteristics that jointly affect the power angle problem and the voltage problem in the same framework, the integrated evaluation of the transient power angle and transient voltage stability of the power system is realized. The present invention proposes a method for evaluating the integrated stability of the transient power angle and voltage of the power system based on two-stage feature selection. This method can effectively reduce the feature dimension and improve the effect of the integrated evaluation model of transient power angle and transient voltage stability.

[0008] The technical solution adopted by the present invention is:

[0009] The method for evaluating the transient power angle and voltage integrated stability of a power system based on two-stage feature selection includes the following steps:

[0010] Step 1: Construct a comprehensive feature set to characterize transient power angle stability problems and transient voltage stability problems;

[0011] Step 2: A two-stage transient stability feature selection method is proposed: in the first stage, the original comprehensive feature set is preliminarily screened based on the class separability score and the Pearson correlation coefficient; in the second stage, the competitive search algorithm CSA is binarized, and a multi-task transient stability integrated evaluation model based on the convolutional gated recurrent unit ConvGRU is embedded in the selection process to further optimize the feature subset;

[0012] Step 3: Based on the optimal feature subset obtained by the two-stage transient stability feature selection method, the multi-task transient stability integrated evaluation model constructed in step 2 is used to realize the integrated evaluation of transient power angle stability and transient voltage stability, and perform transient integrated evaluation based on the phasor measurement unit PMU data.

[0013] In the step 1, the transient power angle stability problem and the transient voltage stability problem are collectively studied, and a set of comprehensive and integrated feature sets are constructed by considering the time series sampling characteristics of data in actual engineering;

[0014] In the comprehensive feature set, the features related to the generator include G ω , G δ , G △f , G Pm , G P , G Q, respectively represent the angular velocity, power angle, frequency deviation, mechanical power, active power, and reactive power of different generators;

[0015] The characteristics related to the busbar include B u , B θ , respectively represent the voltage amplitude and phase angle of all buses;

[0016] Line-related features include C P , C Q , respectively represent the active power and reactive power of all lines.

[0017] In step 1, in the feature selection process, the dynamic time series of the characteristic quantity in the power system is the key, so its characteristic quantity can form a multidimensional time series (Multivariate Time Series, MTS) composed of a univariate time series (Univariate Time Series, UTS), and the multidimensional time series can be expressed as a D×L matrix, as shown in formula (1):

[0018]

[0019] In formula (1), UTS d represents the time series data corresponding to the dth feature, d∈{1,2,…,D}, D represents the feature dimension in the MTS matrix, that is, the total number of features; L is the length of the time series data segment, It represents the measurement value of feature d at time l; C represents the category label corresponding to the MTS matrix.

[0020] For the selection of multi-dimensional time series feature sampling intervals, since the data changes significantly near the time of fault occurrence and fault removal, it reflects the impact of the fault on the system and plays a key role in the evaluation process. Therefore, in order to more realistically reflect the actual situation in the project, the present invention adopts a multi-point continuous sampling method to construct the time series feature.

[0021] In step 1, the power angle stability is the synchronous stability. An important sign of the normal operation of the power system is that all synchronous generators in the system are synchronously operated. Therefore, the present invention uses the transient angle stability index (TASI) η TASI The power angle stability is judged as shown in formula (2):

[0022]

[0023] In formula (2), △δ max represents the maximum power angle difference between any two generators during the simulation time; when ηTASI >0 is considered stable, and the corresponding category label is 1; otherwise, it is considered unstable, and the corresponding category label is 0.

[0024] In step 1, the transient voltage stability is judged using engineering practical criteria: the bus node voltage is less than 0.75pu for no more than 1 second after a fault. The present invention does not consider partitioning for the time being. It is considered that if one of the monitored buses meets the voltage instability criterion, the system is judged to be unstable, and the corresponding category label is 0; otherwise, it is judged to be stable, and the corresponding category label is 1.

[0025] In step 2, the degree of discreteness between samples is reflected by the mean and standard deviation of the distance. Since the physical quantity of the power system is time series data, the dynamic time warping (DTW) algorithm is used to calculate the distance between time series;

[0026] For a time series training set sample (v,c)∈(UTS D ,C), v represents the time series data in the training set, c represents the label corresponding to the sample, UTS D Represents the time series data corresponding to a single feature, UTS D There is a D-dimensional feature space, C = {1, 2, ..., c} is the class label set, c means there are c different categories;

[0027] For the single feature time series set corresponding to feature k D represents the total number of features; represents the measured value of feature k at the starting time; Then it represents the measured value of feature k at time l; represents the measured value of feature k at the end time; L is the length of the time series data segment;

[0028] Calculating UTS k The mean distance between two sequences in the same category i and standard deviation As shown in formula (2) and formula (3);

[0029]

[0030] In the above formula, N i Represents the number of time series of feature k in class i; and They represent the pth time series and the qth time series of feature k in class i respectively; is the DTW distance between the p-th and q-th time series of feature k in class i;

[0031] Recalculate UTS kThe mean distance between two sequences in different classes (i, j) and standard deviation As shown in formula (4) and formula (5).

[0032]

[0033] In the above formula, N i and N j Respectively represent the number of time series of feature k in class i and class j; represents the pth time series of feature k in class i; represents the qth time series of feature k in class j; is the DTW distance between the p-th and q-th time series of feature k in different classes (i, j);

[0034] Each category introduces a weight coefficient ω i and ω j , the weight of each category is inversely proportional to the number of samples in each category, as shown in equations (7) and (8);

[0035]

[0036] In the above formula, N is the total number of samples; N i Represents the number of samples in the i-th class; N j represents the number of samples in the jth class.

[0037] The score evaluation formula (9) is used as an indicator to evaluate the classification performance of feature k. The larger the score value, the better the classification performance. By introducing the weight coefficient, the influence of the minority class in calculating the mean and standard deviation can be increased, so that more attention is paid to the minority class when calculating the score, reducing the impact of sample imbalance.

[0038]

[0039] In formula (9), score(k) represents the score corresponding to feature k; ω i and ω j Represent the weights of class i and class j respectively; Represents the mean distance between sequences of feature k in the same category i; Represents the mean distance between sequences of feature k in the same class j; Represents the mean distance between sequences of feature k in different classes (i, j); Represents the standard deviation of the distance between sequences in the same category i for feature k; It represents the mean standard deviation of the distance between sequences of feature k in the same class j; Represents the standard deviation of the distance between feature k sequences in different classes (i, j).

[0040] Based on multidimensional time series MTS, UTS d represents the time series data corresponding to the dth feature, D represents the total number of features; L is the length of the time series data segment; Then it represents the measured value of feature d at time l;

[0041] Assume that there are d-th and q-th features of the multidimensional time series MTS, which are represents the measured value of feature d at time l, and L is the length of the time series data segment; represents the measured value of feature q at time l, and L is the length of the time series data segment;

[0042] UTS d and UTS q The correlation between them can be expressed as:

[0043]

[0044] In formula (10): R(UTS d ,UTS q ) indicates feature UTS d and characteristic UTS q The correlation between represents the measured value of feature d at time l; represents the measured value of feature q at time l; L is the length of the time series data segment; and Represents the characteristics UTS d and characteristic UTS q The average value of:

[0045]

[0046] In formula (11) and formula (12): represents the measured value of feature d at time l; represents the measured value of feature q at time l; L is the length of the time series data segment.

[0047] Combining the previously calculated score of the feature k classification performance and the Pearson correlation coefficient between the features, we can obtain the evaluation criteria for selecting features as shown in formula (13):

[0048]

[0049] In formula (13), MTS represents the set of unselected features; S represents the set of selected features; score (UTS d ) indicates UTS d The corresponding feature score; |R(UTS d ,UTS q )|Indicates feature UTS d and characteristic UTS q The absolute value of the Pearson correlation coefficient between |S| represents the number of variables in the selected feature subset; UTS d Indicates the dth feature; UTS q represents the qth feature.

[0050] In step 2, according to the evaluation criteria shown in formula (13), the forward method is used to perform the first stage feature selection, and the first stage feature selection process is as follows:

[0051] ①. First, set the feature subset to an empty set, and calculate the mean and standard deviation of the distance between different samples in the same class for each feature quantity according to formula (3) and formula (4), and calculate the mean and standard deviation of the distance between different samples in different classes for each feature quantity according to formula (5) and formula (6);

[0052] ②. Calculate the score of each feature according to formula (9) and arrange them in descending order, and add the feature with the highest score to the feature subset;

[0053] ③. Calculate the Pearson correlation coefficient matrix between the feature quantities according to formula (10);

[0054] ④. The remaining feature quantities are calculated according to formula (13) to calculate the ratio of their scores to the average value of the Pearson correlation coefficient between the selected feature quantities, and the feature quantity with the largest result is added to the feature subset;

[0055] ⑤. Set the threshold K and repeat the selection process in step ④. When the number of feature subsets reaches the threshold K, the feature subset data is output to conduct the second stage feature selection study.

[0056] In step 2, in the present invention, the continuous CSA is converted into a binary version using an S-shaped transfer function. The S-shaped transfer function and the contestant conversion are shown in formula (14) and formula (15) respectively. The converted competitive search algorithm is called a binary competitive search algorithm (Binary Competitive Search Algorithm, BCSA).

[0057]

[0058] In formula (14), e is the base of natural logarithm; represents the value of the jth evaluation index of the i-th player in the t+1th iteration, that is, the position information in the j-th dimension; The continuous variable Convert to probability representation of binary variables;

[0059]

[0060] In formula (15), is a continuous variable Convert to probability representation of binary variables; represents the value of the jth evaluation index of the i-th player in the t+1th iteration after binary conversion, that is, the position information in the j-th dimension; rand(0,1) represents a random number that follows a uniform distribution in the interval [0,1];

[0061] In step 2, the Convolution Gated Recurrent Unit (ConvGRU) is an improved version of the gated recurrent unit, and its gated recurrent mechanism remains unchanged, still including the update gate and the reset gate. However, in order to fully capture the spatial correlation between local pixels of the data, ConvGRU uses a convolutional network structure to replace the fully connected network structure of the ordinary gated recurrent unit to realize feature calculation, while considering the spatial characteristics (across features) and temporal characteristics in time series data. The calculation formula of ConvGRU is shown in equations (16) to (19):

[0062] R t =σ(W xr *X t +W hr *H t-1 ) (16);

[0063] Z t =σ(W xz *X t +W hz *H t-1 ) (17);

[0064]

[0065] The above formula ignores the bias vector. In the above formula, σ and tanh are activation functions; W is the weight parameter; * is the convolution operation; ⊙ is the Hadamard product; R t With Z t is the reset gate and update gate at time t; X t is the input data at the current moment; Indicates the current candidate value; H t With H t-1 Represents the hidden state at time t and the previous time.

[0066] In step 2, a multi-task transient stability integrated evaluation model based on a convolutional gated recurrent unit ConvGRU is constructed, transient power angle stability evaluation is defined as task 1, transient voltage stability evaluation is defined as task 2, and the two tasks are linked through multi-task learning. In order to fully extract the local spatial features of electrical measurement data and construct a temporal relationship, a convolutional gated recurrent unit ConvGRU network is used to process the electrical measurement data; the features extracted by the convolutional gated recurrent unit ConvGRU network are input into the classifiers corresponding to task 1 and task 2 to obtain the evaluation results, and the information sharing between the two tasks 1 and 2 and the transient stability joint evaluation are completed.

[0067] In step 2, the setting of the subset evaluation criteria for the second stage encapsulated feature selection, i.e., setting the iterative optimization objective function, is mainly considered from two aspects: the feature subset dimension and the classification performance of the classification model on different feature subsets. Considering the sample imbalance problem existing in the power system, i.e., the imbalance between the number of samples and the importance, the evaluation criteria are more biased towards minority samples (i.e., unstable samples) to reduce the number of missed judgments of the model. Therefore, the recall rate R is defined according to the evaluation confusion matrix shown in Table 1. ec , precision P rec and f β The objective function is formed as shown in equations (20) to (25):

[0068] Table 1 Confusion matrix

[0069]

[0070]

[0071] In Table 1, TP and TN are the number of correctly classified stable samples and unstable samples, respectively; FN is the number of stable samples misclassified as unstable samples; FP is the number of unstable samples missed as stable samples;

[0072] F fitness =α(1-f1)+(1-α)f2 (20);

[0073] f1=γf β功角 +(1-γ)f β电压 (twenty one);

[0074]

[0075] Where: F fitnessrepresents the total fitness value; f1 represents the fitness value corresponding to the transient stability integrated assessment result; f2 represents the fitness value corresponding to the feature subset dimension; |S| represents the dimension of the selected feature subset; |X| represents the total dimension of the feature set; TP and TN are the number of correctly classified stable samples and unstable samples, respectively; FN is the number of stable samples misclassified as unstable samples; FP is the number of unstable samples missed as stable samples; R ec represents the “recall rate” of unstable samples; P rec represents the “precision rate” of unstable samples; α, γ, β are weight coefficients; f β功角 、f β电压 They are the f corresponding to transient power angle stability assessment and transient voltage stability assessment. β Value; f β The value is a comprehensive consideration of the recall rate and precision rate. The larger β is, the greater the weight given to the recall rate is, and the number of missed judgments of the multi-task transient stability integrated evaluation model of the convolutional gated recurrent unit ConvGRU is minimized.

[0076] In step 2, the process of the second stage encapsulated feature selection is as follows:

[0077] ①. The results of the first-stage preliminary feature screening are sent to the BCSA feature selection algorithm part of the second-stage feature selection; ②. The position vectors of the contestants in each iteration round are updated according to the BCSA algorithm, and the feature subset data corresponding to the position vector is input into the multi-task transient stability integrated assessment model;

[0078] ③. According to the classification results of the multi-task transient stability integrated assessment model, the fitness function value corresponding to the feature subset is calculated using equations (20) to (25), and iterative optimization is performed with the goal of minimizing the fitness value shown in equation (20);

[0079] ④. Repeat the above steps ② and ③ until the required number of iterations is met, and finally output the optimal feature subset.

[0080] In step 3, the present invention adopts a multi-task learning method to simultaneously evaluate multiple stability issues. On the basis of two-stage feature selection, a multi-task transient stability integrated evaluation model based on convolutional gated recurrent unit ConvGRU is used to realize the integrated evaluation of transient power angle and transient voltage stability of the power system, and the learning efficiency and effect of model parameters are jointly improved by optimizing the loss functions of different tasks.

[0081] In the step 3, the transient power angle stability assessment is defined as task 1, and the transient voltage stability assessment is defined as task 2. The two tasks are linked through multi-task learning, and a multi-task transient stability integrated assessment model based on the convolutional gated recurrent unit ConvGRU is established to achieve the integrated assessment of transient power angle and transient voltage stability. In order to fully extract the local spatial features of the electrical measurement data and construct the temporal relationship, the convolutional gated recurrent unit ConvGRU network is first used to process the electrical measurement data. Then, the features extracted by the convolutional gated recurrent unit ConvGRU network are input into the classifiers corresponding to task 1 and task 2 respectively to obtain the assessment results, thereby completing the information sharing and joint assessment between the two tasks 1 and 2.

[0082] The present invention provides a method for evaluating the transient power angle and voltage integrated stability of a power system based on two-stage feature selection, and the technical effects are as follows:

[0083] 1) The method of the present invention can screen out the important features of the system and obtain the key variables that affect the transient power angle and transient voltage evaluation. The performance of the evaluation model is not proportional to the acquired feature quantity. Only a small number of key features play an important role in model training. Selecting features can reduce feature dimensions and enhance the applicability of the evaluation model in large-scale systems.

[0084] 2) The class separability criterion proposed in the first stage of the present invention measures the classification importance of features in the case of sample imbalance, and pays more attention to unstable samples with greater impact but fewer numbers; when constructing the feature subset evaluation criteria in the second stage, a greater weight is given to the recall rate of unstable samples, and BCSA improves the problems of local stagnation and low optimization efficiency. In the process of feature selection, the sample imbalance in the power system is fully considered, and a targeted solution is proposed for the problem of different costs of misjudgment / omission.

[0085] 3) The multi-task transient stability integrated evaluation model based on two-stage feature selection and convolution gated recurrent unit ConvGRU proposed in the present invention comprehensively considers the temporal and spatial characteristics of the features and their joint impact on transient power angle stability and voltage stability, and realizes the integrated evaluation of transient power angle stability and transient voltage stability. By comparing before and after feature selection, the method of the present invention improves the accuracy and reliability of the evaluation while achieving effective feature dimensionality reduction. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 This is a schematic diagram of time series data sampling.

[0087] Figure 2 It is the internal unit structure of ConvGRU.

[0088] Figure 3Schematic diagram of the multi-task transient stability integrated evaluation model based on the convolutional gated recurrent unit ConvGRU.

[0089] Figure 4 Flowchart for the second stage packaged feature selection.

[0090] Figure 5 The framework is a two-stage feature selection method.

[0091] Figure 6 is the feature score distribution diagram.

[0092] Figure 7 Pearson correlation coefficient heat map between features.

[0093] Figure 8 This is a graph showing the relationship between the number of features and the evaluation accuracy.

[0094] Fig. 9 It is the convergence curve of fitness value of different optimization algorithms.

[0095] Fig.10 These are the transient power angle stability assessment results for different assessment models.

[0096] Fig.11 Transient voltage stability assessment results of different assessment models. DETAILED DESCRIPTION

[0097] The integrated stability assessment method of transient power angle and voltage of power system based on two-stage feature selection first integrates a set of comprehensive features characterizing power angle stability and voltage stability based on existing research. Secondly, in the first stage of feature selection, the filtering method and the sample imbalance problem are combined to propose a class separability criterion to measure the importance of features to classification, and the Pearson correlation coefficient is used to remove redundancy between features; in the second stage, the competitive search algorithm (CSA) is binarized as a search strategy, and the convolution gated recurrent unit (ConvolutionGated Recurrent Unit, ConvGRU) is used as a classifier to evaluate the quality of subsets, and an evaluation criterion is constructed for sample imbalance. Finally, an analysis is conducted on the New England 10-machine 39-node system, which verifies that the method proposed in this invention can effectively reduce the feature dimension and improve the effect of integrated assessment of transient power angle and transient voltage stability of power system.

[0098] Below in conjunction with the accompanying drawings, the preferred embodiment is described in detail:

[0099] (I) Construction of comprehensive characteristics of transient power angle and transient voltage stability:

[0100] Research on transient stability assessment of power systems usually considers power angle and voltage separately. Since the power angle and voltage are closely coupled during the transient process of the power system, the characteristics related to the power angle indirectly affect the voltage stability, and at the same time, the characteristics closely related to the voltage stability also indirectly affect the power angle.

[0101] Therefore, the present invention considers both the transient power angle problem and the transient voltage problem and selects a relatively complete set of comprehensive features. In the feature set, the features related to the generator are G ω , G δ , G △f , G Pm , G P , G Q , respectively representing the angular velocity, power angle, frequency deviation, mechanical power, active power, and reactive power of different generators; the characteristics related to the bus are B u , B θ , respectively represent the voltage amplitude and phase angle of all buses; the line-related characteristics are C P , C Q , respectively represent the active power and reactive power of all lines.

[0102] In terms of feature sampling, current research mostly adopts a multi-point non-continuous method. Although it has certain advantages in theoretical analysis and model construction, it faces challenges in actual engineering applications. Since data in actual scenarios presents the characteristics of large-scale and continuous influx, it is difficult to extract irregular data and it is difficult to meet the rapidity and accuracy of smart grid assessment. In order to make full use of the time series data of the power system under disturbance, this paper will carry out transient stability feature selection research based on time series data, aiming to improve the effect of transient stability assessment model driven by data.

[0103] In the feature selection process, the dynamic time series of the characteristic quantity in the power system is the key, so its characteristic quantity can form a multidimensional time series (MTS) composed of a univariate time series (UTS). MTS can be expressed as a D×L matrix, as shown in formula (1):

[0104]

[0105] In formula (1), UTS d represents the time series data corresponding to the dth feature, D represents the feature dimension in the MTS matrix, that is, the total number of features; L is the length of the time series data segment, It represents the measurement value of feature d at time l; C represents the category label corresponding to the MTS matrix.

[0106] The transient stability assessment problem of the power system studied in this paper is a supervised classification problem. The time series data of each electrical characteristic quantity collected during the fault process are arranged in the format of the MTS matrix. The matrix is ​​called a sample. Each sample contains D electrical characteristic quantities. The system stability result corresponding to the sample is the category label C, which is used to describe the category of the sample.

[0107] For the selection of multi-dimensional time series feature sampling intervals, such as Figure 1 As shown. Figure 1 In the time series data, the fault occurs at time t f and the fault clearing time t c The data nearby changes significantly, reflecting the impact of the fault on the system, and plays a key role in the evaluation process. Therefore, in order to more truly reflect the actual situation in the project, the present invention adopts a multi-point continuous sampling method to construct the time series feature.

[0108] Figure 1 Where U and θ represent bus voltage and phase angle respectively. The data before the fault occurs is used to provide steady-state information of the system, while the data after the fault is removed reflects the process of system recovery to steady state and fully reflects the dynamic response process of the system. However, different fault removal times t c Will cause the failure period T f To ensure that the data input into the model has a unified time dimension, simplify the model training process, and make full use of the effective information in the time series data, the method of adaptive filling of steady-state data is adopted. The sampling period T set by the PMU is 0.01 seconds, the total length of the fixed time series data is 0.4 seconds, and the data length after the fault removal is 0.08 seconds. For each different fault removal time t c , according to the fault period T f The actual length of the sample is filled by selecting the steady-state data of the corresponding length. Specifically, for each resection time t c , and its corresponding sampling interval can be expressed as [t c -32T,t c +8T], in practical applications, the model detection time can be adjusted according to specific needs to further reduce the risk of missed detection.

[0109] Power angle stability is synchronous stability. An important sign of normal operation of the power system is that all synchronous generators in the system are synchronously operated. Therefore, the present invention uses the transient angle stability index (TASI) η TASI The power angle stability is judged as shown in formula (2):

[0110]

[0111] In formula (2), △δ max represents the maximum power angle difference between any two generators during the simulation time. TASI >0 is considered stable, and the corresponding category label is 1; otherwise, it is considered unstable, and the corresponding category label is 0.

[0112] The transient voltage stability is judged using engineering practical criteria: the bus node voltage is less than 0.75pu for no more than 1s after a fault. The present invention does not consider partitioning for the time being. It is considered that if one of the monitored buses meets the voltage instability criterion, the system is judged to be unstable, and the corresponding category label is 0; otherwise, it is judged to be stable, and the corresponding category label is 1.

[0113] (II) First-stage feature selection based on class separability and feature correlation:

[0114] The first-stage feature selection method under the filtering framework studied in the present invention consists of two parts: class separability criterion and Pearson correlation coefficient.

[0115] The present invention proposes a class separability criterion based on intra-class and inter-class distances that takes into account sample imbalance. The core idea is to evaluate the influence of features in the classification process by the distance between feature samples. That is, the smaller the distance between samples in the same category and the larger the distance between samples in different categories, the better the corresponding feature classification performance.

[0116] The degree of dispersion between samples is reflected by the mean and standard deviation of the distance. Since the physical quantities of the power system are time series data, the dynamic time warping (DTW) algorithm is used here to calculate the distance between sequences. For a sample (v, c)∈(UTS D ,C),UTS D There is a D-dimensional feature space, C = {1,2,…,c} is the class label set. The total number of samples is N, N i Represents the number of samples in the i-th class. For the single feature time series set UTS corresponding to feature k (k∈1,2,…,D) k , calculate UTS k The mean distance between two sequences in the same category i and standard deviation Recalculate UTS k The mean distance between two sequences of different classes (i, j) and standard deviation Each category introduces a weight coefficient ω iand ω j , the weight is inversely proportional to the number of samples in each category, as shown in formula (3). Finally, the score evaluation formula (4) is used as an indicator to evaluate the classification performance of the feature k. The larger the score value, the better the classification performance. By introducing the weight coefficient, the influence of the minority class in calculating the mean and standard deviation can be increased, so that the minority class is paid more attention in calculating the score, reducing the impact of sample imbalance.

[0117]

[0118] The Pearson correlation coefficient is a commonly used tool in statistics to measure the correlation between two features. In the field of time series data analysis, it is mainly used to identify the synchronous change trend of two sets of data. It is widely used because of its convenient calculation and high time efficiency, and can effectively characterize the correlation between data sets.

[0119] Based on the multidimensional time series MTS defined previously, it is expressed as Assume that there are d-th and q-th features of the multidimensional time series, which are UTS d and UTS q The correlation between them can be expressed as:

[0120]

[0121] in, and Represents the characteristics UTS d and characteristic UTS q The average value of:

[0122]

[0123] According to the definition of the Pearson correlation coefficient between time series, combined with the original sample feature set of the power system, the Pearson correlation coefficient matrix between multi-dimensional time series features can be calculated. Combined with the score of the feature classification performance calculated previously, the evaluation criteria for selecting features can be obtained as shown in formula (8):

[0124]

[0125] In formula (13), MTS represents the set of unselected features; S represents the set of selected features; score (UTS d ) represents the feature score corresponding to UTSd; |R(UTS d ,UTS q )| represents the absolute value of the Pearson correlation coefficient between feature UTSd and feature UTSq; |S| represents the number of variables in the selected feature subset.

[0126] According to the evaluation criteria shown in formula (13), the forward method is used to select the features in the first stage. The feature selection process in the first stage is as follows:

[0127] ①. First, set the feature subset to an empty set, and calculate the mean and standard deviation of the distance between different samples in the same class for each feature quantity according to formula (3) and formula (4), and calculate the mean and standard deviation of the distance between different samples in different classes for each feature quantity according to formula (5) and formula (6);

[0128] ②. Calculate the score of each feature according to formula (9) and arrange them in descending order, and add the feature with the highest score to the feature subset;

[0129] ③. Calculate the Pearson correlation coefficient matrix between the feature quantities according to formula (10);

[0130] ④. The remaining feature quantities are calculated according to formula (13) to calculate the ratio of their scores to the average value of the Pearson correlation coefficient between the selected feature quantities, and the feature quantity with the largest result is added to the feature subset;

[0131] ⑤. Set the threshold K and repeat the selection process in step ④. When the number of feature subsets reaches the threshold K, the feature subset data is output to conduct the second stage feature selection study.

[0132] (III) Second stage feature selection based on binary competitive search algorithm:

[0133] The encapsulated feature selection method based on swarm intelligence algorithm has been widely used to solve the optimal feature subset problem, but with the increase of feature dimension, this method often falls into a stagnant state of local optimal solution, affecting the optimization efficiency, which has become an important factor restricting its performance to be further improved. To this end, the present invention uses the more robust and reliable competitive search algorithm (Competitive Search Algorithm, CSA) proposed in the document [1]: 'Xu Y, Liu H, Xie S, et al. Competitive search algorithm: a new method for stochastic optimization [J]. Applied Intelligence, 2022, 52 (11): 12131-1215.' as the search strategy for encapsulated feature selection. CSA is a new swarm intelligence optimization method proposed based on some social activities in human life, such as all-around sports competitions and talent shows. The algorithm has a simple structure, better optimization effect and stronger robustness.

[0134] The specific mathematical model and formula of CSA are as follows:

[0135] The project team will evaluate contestant X according to the scoring criteria. n,d Comprehensive evaluation and ranking are performed, and the players are divided into excellent group and general group. n represents the number of players, and d represents the various indicators for evaluating players in the competition, that is, the variable dimension of the problem that needs to be optimized. Each player has different learning abilities, which will change randomly and are divided into strong learning abilities and normal learning abilities according to the set threshold. In the excellent group, the update formula for players with strong learning abilities is formula (9), and the update formula for players with normal learning abilities is formula (27). In the general group, players with strong learning abilities pay more attention to improving themselves, while players with normal learning abilities hope to break through themselves. The update function of each indicator is shown in formula (28). When the player's learning ability is greater than the threshold, learning is performed according to the indicators of the best player, and the formula is (29). After each round of competition, the total number of players is kept unchanged, some players withdraw, and a corresponding number of new players join, and the learning ability and evaluation indicators of the new players are random.

[0136]

[0137] In equations (26) and (27), S1 and S2 are the search range functions of the strong learning ability contestants and the general learning ability contestants; t is the current iteration number; j is the dimension in which X is located; j = 1, 2, 3, 4, …, d; X i,j is the value of the jth evaluation index of the i-th player, that is, the position information in the j-th dimension; and They represent the upper and lower limits of the function in the j-th dimension search range respectively; L B and U B are all constants; ρ is a value randomly extracted from the matrix [-1,0,1]; A(i) is the learning ability of the current contestant; L1 is the threshold value representing the strength of learning ability in the excellent group, and L1 belongs to the matrix (0,1).

[0138]

[0139] F=P·ο

[0140] In formula (28), F is a negative factor; α is a random number in [-1,1]; Q is a random number in [0,2]; D and L2 are 1×d matrices, but all elements in matrix D are 1, and elements in L2 are randomly distributed with -1 and 1; P is a standard normal distribution with mean 0 and variance 1; ο is a random factor that updates the position of each player and randomly selects [0.1, 0.2, 0.3, 0.4, 0.5] from the matrix.

[0141]

[0142] In formula (29), is the index value of the best player in the jth dimension during the tth iteration; L3 is the reference threshold and belongs to the interval (0,1).

[0143] Feature selection is classified as a combinatorial binary optimization problem, which represents the solution on a binary search space. However, CSA is designed to solve continuous optimization problems, which is not suitable for feature selection problems. To develop a binary version of CSA, the solution should be represented in binary form (0 or 1). Therefore, CSA needs to be modified to meet the requirements.

[0144] Using transfer function is one of the effective methods to convert continuous optimizer into binary optimizer. Compared with other operators, transfer function is cheaper, simpler and faster, so it is easy to implement. Therefore, in the present invention, the continuous CSA is converted into a binary version using S-shaped transfer function. The S-shaped transfer function and the contestant conversion are shown in formula (14) and formula (15) respectively. The converted competitive search algorithm is called binary competitive search algorithm (Binary Competitive Search Algorithm, BCSA).

[0145]

[0146] The Convolution Gated Recurrent Unit (ConvGRU) is an improved and simplified version of the Gated Recurrent Unit (GRU), which retains the original update gate, reset gate and gated recurrent mechanism of GRU. It is mainly used when processing images or spatial data. The dataset can be regarded as having spatial (between features) and temporal dimensions. Therefore, if we want to capture spatial (cross-feature) and temporal dependencies in time series data, in the temporal dimension, the recurrent structure of GRU enables the model to capture information about temporal changes in sequence data; in the spatial dimension, the fully connected layers in GRU are replaced with convolutional layers, which enables ConvGRU to process spatial information at each time step, that is, to learn spatial dependencies between features. The internal unit structure of ConvGRU is as follows: Figure 2 The calculation formula of ConvGRU is shown in equations (16) to (19):

[0147] R t =σ(W xr *X t +W hr *H t-1 ) (16);

[0148] Zt =σ(W xz *X t +W hz *H t-1 ) (17);

[0149]

[0150] The above formula ignores the bias vector. In the above formula, σ and tanh are activation functions; * is the convolution operation; ⊙ is the Hadamard product; R t is the reset gate; Z t is the update gate; X t is the input data at the current moment; Indicates the current candidate value; H t-1 Represents the hidden state at the previous moment.

[0151] Previous studies on stability assessment usually use single-task learning models. However, since transient power angle stability and voltage stability are inseparable in mechanism, it is difficult to share information between tasks in a single-task learning model, resulting in low efficiency in the assessment of multiple stability issues. Multiple single-task learning models not only increase the consumption of storage and computing resources, but also limit the accuracy and efficiency of the assessment. Therefore, the present invention uses a multi-task learning method to simultaneously assess multiple stability issues. Constructing a multi-task transient stability assessment model can jointly improve the learning efficiency and effect of model parameters by optimizing the loss functions of different tasks.

[0152] The present invention defines transient power angle stability assessment as task 1, and transient voltage stability assessment as task 2. The two tasks are linked through multi-task learning to establish a multi-task transient stability integrated assessment model based on convolutional gated recurrent unit ConvGRU, such as Figure 3 As shown in the figure; in order to fully extract the local spatial features of the electrical measurement data and construct the temporal relationship, the convolutional gated recurrent unit ConvGRU network is first used to process the electrical measurement data; then a multi-task learning model is used to input the features extracted by the convolutional gated recurrent unit ConvGRU network into the classifiers corresponding to task 1 and task 2 respectively to obtain the evaluation results, thereby completing the information sharing and joint evaluation between the two tasks 1 and 2.

[0153] The setting of subset evaluation criteria for encapsulated feature selection in the second stage, that is, setting the iterative optimization objective function, mainly considers the feature subset dimension and the classification performance of the classification model on different feature subsets. Considering the sample imbalance problem existing in the power system, that is, the imbalance between the number of samples and the importance, the evaluation criteria are more biased towards minority samples (that is, unstable samples) to reduce the number of missed judgments of the model. Therefore, the recall rate R is defined according to the stable evaluation confusion matrix shown in Table 1. ec , precision Prec and f β Therefore, the objective function is as shown in equations (19) to (24):

[0154] Table 1 Confusion matrix

[0155]

[0156]

[0157] In Table 1, TP and TN are the number of correctly classified stable samples and unstable samples, respectively; FN is the number of stable samples misclassified as unstable samples; FP is the number of unstable samples missed as stable samples;

[0158] F fitness =α(1-f1)+(1-α)f2 (20);

[0159] f1=γf β功角 +(1-γ)f β电压 (twenty one);

[0160]

[0161] Where: R ec represents the “recall rate” of unstable samples; P rec represents the “precision rate” of unstable samples; α and γ are weight coefficients; f β功角 、f β电压 They are the f corresponding to transient power angle stability assessment and transient voltage stability assessment. β Value; f β The value is f β功角 and f β电压 The larger the β is, the greater the weight given to the recall rate is, and the number of missed judgments of the model is minimized. The iterative optimization is performed with the goal of minimizing the fitness function.

[0162] The overall process of the second stage encapsulated feature selection is as follows Figure 4 As shown in the figure, the results of the first stage feature screening are sent to the second stage feature selection, and the position vector of the contestants in each iteration round is updated according to the BCSA algorithm. The feature subset data corresponding to the vector is input into the multi-dimensional time series multi-task classification model based on ConvGRU. The fitness function value corresponding to the feature subset is calculated according to the classification result. The above process is repeated until the number of iterations required is met, and finally the optimal feature subset is output.

[0163] In summary, the overall framework of the power system transient power angle and voltage integrated stability assessment method based on two-stage feature selection proposed in this invention can be expressed as Figure 5 shown.

[0164] The technical effects of the present invention will be further illustrated by examples below:

[0165] Based on the standard New England 10-machine 39-bus system example, the generator at bus No. 34 is replaced with a wind turbine, a doubly fed wind turbine model is used, the generator uses a 4th-order model, all loads use a constant impedance model, and the PSASP software is used for transient stability simulation to obtain experimental samples. In order to make the sample data set more comprehensive and diverse, the specific setting of the fault is as follows: ① Consider wind power penetration of 5%, 10%, and 15%; ② The flow level is 90%, 95%, 100%, 105%, and 110%; ③ The most serious three-phase short circuit fault occurs in the AC line. ④ The fault locations are located at 2%, 50%, and 98% of the AC line respectively; ⑤ In actual situations, the fault removal time is generally greater than or equal to 3 cycles, so the fault duration is set to 3 to 11 cycles.

[0166] The experiment generated 13,365 MTS samples, corresponding to 26,730 classification labels for transient power angle stability and transient voltage stability. Among them, there are 11,291 transient power angle stability samples and 2,074 unstable samples; there are 9,616 transient voltage stability samples and 3,749 unstable samples. The data set is randomly split, and the ratio of training set to test set is 4:1. The training set is used to screen feature subsets, and the test set is used to evaluate the classification performance corresponding to these feature subsets. A feature quantity candidate set containing 205 candidate features is constructed based on the timing characteristics and sampling characteristics.

[0167] To measure the classification performance of the feature quantity, the class separability criterion considering sample imbalance proposed in the present invention is used to calculate the feature scores under the transient power angle stability label and the transient voltage stability label respectively. The larger the score, the better the classification performance. The two scores are then added together to obtain the score of each feature quantity. The feature score distribution after normalization is as follows: Figure 6 As shown, the characteristic quantities with high scores include some bus voltages, active and reactive power output by generators, and active and reactive power of lines.

[0168] The correlation coefficient matrix between the 205-dimensional feature quantities is calculated according to the Pearson correlation coefficient algorithm of the time series. A positive correlation coefficient indicates a positive correlation between the features, a negative correlation indicates a negative correlation between the features, and a zero indicates no correlation. The heat map of the correlation coefficient matrix is ​​as follows: Figure 7 As shown in Figure 2, the feature set is screened according to the evaluation criteria of the first stage feature selection, and the redundancy between features is eliminated while reducing the dimension to obtain the initial screening feature set.

[0169] In order to determine the threshold K for feature selection in the first stage, the influence of different feature numbers on the accuracy of transient stability assessment under the initial screening results is as follows: Figure 8 As shown. Figure 8It can be seen that when the number of features reaches 80, the accuracy of transient power angle stability evaluation has reached 97.28%, and the accuracy of transient voltage stability evaluation has reached 96.35%. As the number of features increases, the change in accuracy does not exceed 0.25%. Therefore, the present invention sets the value of the threshold K to 80, and performs the subsequent second-stage subset optimization based on the first 80 features screened out in the first stage.

[0170] In order to verify whether the swarm intelligence optimization algorithm selected by the present invention has better optimization efficiency and optimization performance in the feature selection process compared with other common swarm intelligence optimization algorithms, it is compared with feature selection methods based on other common algorithms, including: ① Genetic Algorithm (GA); ② Binary Grey Wolf Optimization Algorithm (BGWO); ③ Binary Harris Hawks Optimization (BHHO). The same data set and objective function are used in the verification process. The classifier used is a multi-dimensional time series multi-task classification model based on ConvGRU. Each method will repeat the experiment multiple times to statistically average the results. The experimental results are as follows. Fig. 9 And as shown in Table 2.

[0171] Table 2 Comparison of the effects of different optimization algorithms

[0172]

[0173]

[0174] Depend on Fig. 9 It can be seen that the BCSA algorithm selected by the present invention has stronger optimization ability and better global search ability than other algorithms, and has achieved better fitness value results. From the convergence trend, the convergence curve descent gradient of the algorithm is significantly greater than that of other algorithms, and it still maintains a faster descent speed after 15 iterations. In contrast, other algorithms update very slowly or even fall into local optimality, resulting in stagnation of updates.

[0175] According to Table 2, the BCSA algorithm used in the present invention shows higher accuracy of transient stability assessment, lower subset dimension, and better fitness value in the selection of feature subsets compared with other methods. The method of the present invention reduces the data dimension from 80 dimensions after the initial screening in the first stage to 21 dimensions, which significantly improves the evaluation efficiency of the subsequent transient stability assessment model. Since the GA algorithm evolution process is simpler, subsets with smaller data dimensions can be used earlier in the subset evaluation stage, and the subset evaluation process occupies the main part of the overall optimization time, so the GA algorithm performs better in terms of optimization time. Compared with the GA algorithm, the overall optimization time of the method of the present invention increases by 432s, but avoids premature convergence, and due to the existence of the random elimination mechanism, it avoids falling into the local optimum, so that its final solution obtains a better fitness value.

[0176] In order to verify the effectiveness of the ConvGRU multi-dimensional time series multi-task classification model used in this paper in feature selection and stability assessment, based on the original sample data of 205-dimensional features, the performance of various transient stability assessment models including the gated recurrent unit (GRU), bidirectional long short-term memory (BiLSTM), convolutional neural network (CNN) and K-nearest neighbor (KNN) were compared on the test set. The average results were calculated through multiple experiments. Fig.10 , Fig.11 The comparison results are shown in detail in Table 3. In Table 3, the training time of the model selected by the present invention is used as a benchmark to list the relative multiples of the evaluation time of other models, so as to more intuitively compare the efficiency of each model.

[0177] Table 3 Performance comparison of commonly used stability evaluation models

[0178]

[0179] According to the above Fig.10 , Fig.11Compared with Table 3, it can be clearly seen that ConvGRU and BiLSTM perform well in accuracy compared with other evaluation models. The reason is that the time series memory network has the ability to deeply mine time series information, thereby significantly improving the accuracy of stability evaluation. Although BiLSTM performs well in accuracy, its evaluation time is long, which does not meet the rapidity of subset optimization and online evaluation in encapsulated feature selection. ConvGRU is an improved and simplified version based on BiLSTM and GRU. While maintaining a high evaluation accuracy, it significantly improves the evaluation efficiency, allowing ConvGRU to better meet the speed requirements in the subset optimization process. Although CNN and KNN perform well in evaluation efficiency, they have obvious shortcomings in accuracy. CNN converts the time series matrix into a one-dimensional vector during training. This operation causes the information of the time dimension to be submerged, thereby affecting the accuracy of the evaluation. KNN fails to make full use of the information of the time dimension, resulting in poor classification results. The above results further emphasize the importance of time dimension information in transient stability evaluation. Therefore, the present invention selects ConvGRU as the evaluator in the feature selection process. ConvGRU can effectively extract information in the feature space and analyze the time series information in the data, which not only improves the evaluation accuracy, but also meets the speed requirements in the subset optimization process, providing strong support for accuracy and speed in practical applications.

[0180] The stability evaluation feature subsets screened by the method of the present invention and their corresponding feature scores are arranged in descending order as shown in Table 4.

[0181] Table 4. Second stage feature selection results and their corresponding scores

[0182]

[0183] The results of stability evaluation before and after feature selection are used to verify the effectiveness and necessity of the two-stage feature selection method proposed in this study. The feature sample sets before and after selection are trained using a multi-task transient stability integrated evaluation model based on the convolutional gated recurrent unit ConvGRU. The training results of the model under different feature set dimensions before and after feature selection are listed in Table 5. The training time corresponding to the features selected by the method of the present invention is used as a benchmark, and the relative multiples of the time used for training using the original feature set are listed for intuitive comparison.

[0184] Table 5 Comparison of stable evaluation effects before and after feature selection

[0185]

[0186] The data in Table 5 clearly show the remarkable effect achieved by the feature selection method proposed in the present invention. Through this method, 21 key features were successfully selected from the original feature set, and the feature dimension was compressed by 89.756%, which greatly reduced the model training time by 75.94%. Under the current situation where the complexity of the model is constantly increasing and the amount of sample data is growing sharply, the research of the present invention provides an effective method for simplifying the model training process and improving the computational efficiency. Not only the feature dimension has been significantly compressed, but also the effects of transient power angle stability assessment and transient voltage stability assessment have been improved. The accuracy rate in transient power angle stability assessment has been improved by 1.785%, and the false positive rate and missed rate have been reduced by 1.010% and 4.206% respectively. The accuracy rate in transient voltage stability assessment has been improved by 1.785%, and the false positive rate and missed rate have been reduced by 1.010% and 4.206% respectively. The above results show that the feature selection method of the present invention improves the accuracy and reliability of the multi-task transient stability integrated assessment model while eliminating invalid and redundant features in high-dimensional data.

Claims

1. A power system transient power angle and voltage integrated stability assessment method based on two-stage feature selection, characterized by The following steps are involved: Step 1: Construct a comprehensive feature set to characterize transient power angle stability problems and transient voltage stability problems; Step 2: A two-stage transient stability feature selection method is proposed: in the first stage, the original comprehensive feature set is preliminarily screened based on the class separability score and the Pearson correlation coefficient; in the second stage, the competitive search algorithm CSA is binarized, and a multi-task transient stability integrated evaluation model based on the convolutional gated recurrent unit ConvGRU is embedded in the selection process to further optimize the feature subset; Step 3: Based on the optimal feature subset obtained by the two-stage transient stability feature selection method, the constructed multi-task transient stability integrated evaluation model is used to realize the integrated evaluation of transient power angle stability and transient voltage stability.

2. The method for evaluating the transient power angle and voltage integrated stability of a power system based on two-stage feature selection according to claim 1 is characterized in that In step 1, in the comprehensive feature set, the features related to the generator include G ω , G δ , G △f , G Pm , G P , G Q , respectively represent the angular velocity, power angle, frequency deviation, mechanical power, active power, and reactive power of different generators; The characteristics related to the busbar include B u , B θ , respectively represent the voltage amplitude and phase angle of all buses; Line-related features include C P , C Q , respectively represent the active power and reactive power of all lines.

3. The method for evaluating the transient power angle and voltage integrated stability of a power system based on two-stage feature selection according to claim 2 is characterized in that : Through the transient angle stability index (TASI) η TASI The power angle stability is judged as shown in formula (2): In formula (2), △δ max represents the maximum power angle difference between any two generators during the simulation time; when η TASI >0 is considered stable, and the corresponding category label is 1; otherwise, it is considered unstable, and the corresponding category label is 0.

4. The method for evaluating the transient power angle and voltage integrated stability of a power system based on two-stage feature selection according to claim 3 is characterized in that: The transient voltage stability is judged using practical engineering criteria: the bus node voltage is less than 0.75pu for no more than 1s after a fault. Without considering the partition, if one of the monitored buses meets the voltage instability criterion, the system is judged to be unstable and the corresponding category label is 0; otherwise, it is judged to be stable and the corresponding category label is 1.

5. The method for evaluating the transient power angle and voltage integrated stability of a power system based on two-stage feature selection according to claim 1 is characterized in that: In step 2, since the physical quantity of the power system is time series data, the dynamic time warping (DTW) algorithm is used to calculate the distance between time series; For a time series training set sample (v,c)∈(UTS D ,C), v represents the time series data in the training set, c represents the label corresponding to the sample, UTS D Represents the time series data corresponding to a single feature, UTS D There is a D-dimensional feature space, C = {1, 2, ..., c} is the class label set, c means there are c different categories; For the single feature time series set corresponding to feature k D represents the total number of features; represents the measured value of feature k at the starting time; Then it represents the measured value of feature k at time l; represents the measured value of feature k at the end time; L is the length of the time series data segment; Calculating UTS k The mean distance between two sequences in the same category i and standard deviation As shown in formula (2) and formula (3); In the above formula, N i Represents the number of time series of feature k in class i; and They represent the pth time series and the qth time series of feature k in class i respectively; is the DTW distance between the p-th and q-th time series of feature k in class i; Recalculate UTS k The mean distance between two sequences in different classes (i, j) and standard deviation As shown in formula (4) and formula (5); In the above formula, N i and N j Respectively represent the number of time series of feature k in class i and class j; represents the pth time series of feature k in class i; represents the qth time series of feature k in class j; is the DTW distance between the p-th and q-th time series of feature k in different classes (i, j); Each category introduces a weight coefficient ω i and ω j , the weight of each category is inversely proportional to the number of samples in each category, as shown in equations (7) and (8); In the above formula, N is the total number of samples; N i Represents the number of samples in the i-th class; N j represents the number of samples in the jth class; The score evaluation formula (9) is used as an indicator to evaluate the classification performance of feature k. The larger the score value, the better the classification performance. In formula (9), score(k) represents the score corresponding to feature k; ω i and ω j Represent the weights of class i and class j respectively; Represents the mean distance between sequences of feature k in the same category i; Represents the mean distance between sequences of feature k in the same class j; Represents the mean distance between sequences of feature k in different classes (i, j); Represents the standard deviation of the distance between sequences in the same category i for feature k; It represents the mean standard deviation of the distance between sequences of feature k in the same class j; Represents the standard deviation of the distance between feature k sequences in different classes (i, j).

6. The method for evaluating the transient power angle and voltage integrated stability of a power system based on two-stage feature selection according to claim 1 is characterized in that: Based on multidimensional time series MTS, UTS d represents the time series data corresponding to the dth feature, d∈{1,2,…,D}, D represents the total number of features; L is the length of the time series data segment; Then it represents the measured value of feature d at time l; Assume that there are d-th and q-th features of the multidimensional time series MTS, which are represents the measured value of feature d at time l, and L is the length of the time series data segment; represents the measured value of feature q at time l, and L is the length of the time series data segment; UTS d and UTS q The correlation between them is expressed as: In formula (10): R(UTS d ,UTS q ) indicates feature UTS d and characteristic UTS q The correlation between represents the measured value of feature d at time l; represents the measured value of feature q at time l; L is the length of the time series data segment; and Represents the characteristics UTS d and characteristic UTS q The average value of: In formula (11) and formula (12): represents the measured value of feature d at time l; represents the measured value of feature q at time l; L is the length of the time series data segment; Combining the calculated score of the feature k classification performance and the Pearson correlation coefficient between the features, the evaluation criteria for selecting features can be obtained as shown in formula (13): In formula (13), MTS represents the set of unselected features; S represents the set of selected features; score (UTS d ) indicates UTS d The corresponding feature score; |R(UTS d ,UTS q )|Indicates feature UTS d and characteristic UTS q The absolute value of the Pearson correlation coefficient between |S| represents the number of variables in the selected feature subset; UTS d Indicates the dth feature; UTS q represents the qth feature.

7. The method for evaluating the transient power angle and voltage integrated stability of a power system based on two-stage feature selection according to claim 6 is characterized in that: In step 2, according to the evaluation criteria shown in formula (13), the forward method is used to perform the first stage feature selection, and the first stage feature selection process is as follows: ①. First, set the feature subset to an empty set, and calculate the mean and standard deviation of the distance between different samples in the same class for each feature quantity according to formula (3) and formula (4), and calculate the mean and standard deviation of the distance between different samples in different classes for each feature quantity according to formula (5) and formula (6); ②. Calculate the score of each feature according to formula (9) and arrange them in descending order, and add the feature with the highest score to the feature subset; ③. Calculate the Pearson correlation coefficient matrix between the feature quantities according to formula (10); ④. The remaining feature quantities are calculated according to formula (13) to calculate the ratio of their scores to the average value of the Pearson correlation coefficient between the selected feature quantities, and the feature quantity with the largest result is added to the feature subset; ⑤. Set the threshold K and repeat the selection process in step ④. When the number of feature subsets reaches the threshold K, the feature subset data is output to conduct the second stage feature selection study.

8. The method for evaluating the transient power angle and voltage integrated stability of a power system based on two-stage feature selection according to claim 7 is characterized in that: The continuous CSA is converted into a binary version using an S-shaped transfer function. The S-shaped transfer function and the contestant conversion are shown in formula (14) and formula (15) respectively. The converted competitive search algorithm is called a binary competitive search algorithm (BCSA). In formula (14), e is the base of natural logarithm; represents the value of the jth evaluation index of the i-th player in the t+1th iteration, that is, the position information in the j-th dimension; The continuous variable Convert to probability representation of binary variables; In formula (15), is a continuous variable Convert to probability representation of binary variables; It represents the value of the jth evaluation index of the ith player in the t+1th iteration after binary conversion, that is, the position information in the jth dimension; rand(0,1) represents a random number that follows a uniform distribution in the interval [0,1].

9. The method for evaluating the transient power angle and voltage integrated stability of a power system based on two-stage feature selection according to claim 8 is characterized in that: In the step 2, a multi-task transient stability integrated evaluation model based on a convolutional gated recurrent unit ConvGRU is constructed, transient power angle stability evaluation is defined as task 1, transient voltage stability evaluation is defined as task 2, and the two tasks are linked through multi-task learning. In order to fully extract the local spatial features of the electrical measurement data and construct a temporal relationship, the convolutional gated recurrent unit ConvGRU network is used to process the electrical measurement data; the features extracted by the convolutional gated recurrent unit ConvGRU network are input into the classifiers corresponding to task 1 and task 2 to obtain the evaluation results, thereby completing the information sharing between the two tasks 1 and 2 and the joint transient stability evaluation.

10. The method for evaluating the transient power angle and voltage integrated stability of a power system based on two-stage feature selection according to claim 9 is characterized in that: In step 2, the setting of the subset evaluation criteria for the second stage feature selection, that is, setting the iterative optimization objective function, is shown in formulas (20) to (25): F fitness =α(1-f1)+(1-α)f2 (20); f1=γf β功角 +(1-γ)f β电压 (21); Where: F fitness represents the total fitness value; f1 represents the fitness value corresponding to the transient stability integrated assessment result; f2 represents the fitness value corresponding to the feature subset dimension; |S| represents the dimension of the selected feature subset; |X| represents the total dimension of the feature set; TP and TN are the number of correctly classified stable samples and unstable samples, respectively; FN is the number of stable samples misclassified as unstable samples; FP is the number of unstable samples missed as stable samples; R ec represents the "recall rate" of unstable samples; P rec represents the "precision rate" of unstable samples; α, γ, β are weight coefficients; f β功角 、f β电压 They are the f corresponding to transient power angle stability assessment and transient voltage stability assessment. β Value; f β The value is a comprehensive consideration of recall and precision. The larger β is, the greater the weight given to recall. The process of feature selection in the second stage is as follows: ①. The results of the first stage of preliminary feature screening are sent to the BCSA feature selection algorithm part of the second stage of feature selection; ②. Update the position vector of the contestant in each iteration according to the BCSA algorithm, and input the feature subset data corresponding to the position into the multi-task transient stability integrated evaluation model based on the convolutional gated recurrent unit ConvGRU; ③. According to the classification results of the multi-task transient stability integrated evaluation model based on the convolutional gated recurrent unit ConvGRU, the fitness function value corresponding to the feature subset is calculated using equations (20) to (25), and the iterative optimization is performed with the goal of minimizing the fitness value shown in equation (20); ④. Repeat the above steps ② and ③ until the required number of iterations is met, and finally output the optimal feature subset.

Citation Information

Cited By

  • Wiring terminal voltage detection method, device and system

    CN119861316A

  • Transient voltage stability control method for multi-wind-power-plant grid-connected system during power grid fault duration

    CN120546031A

  • Characteristic distance measurement method and system for dimension reduction visualization of power system data

    CN121705756A