A Key Factor Screening Method for New Energy Oscillation Bearing Capacity Based on Modal Sensitivity Index and Improved XGBoost

Abstract

This invention belongs to the field of new energy power system technology, specifically a method for screening key factors of new energy oscillation carrying capacity based on modal sensitivity index and improved XGBoost. The method includes establishing a state-space model of a power system containing new energy based on the small-disturbance stability analysis theory of power systems; initially screening four key electrical quantities—active power, reactive power, voltage, and current—through factor analysis to form an initial feature set; using a variational autoencoder to reduce the dimensionality of the initial feature set to obtain a low-dimensional potential representation that retains oscillation characteristics; inputting the low-dimensional potential representation into an improved XGBoost model with an objective function; and combining this with a preset threshold to select a final key feature subset; and constructing a new energy oscillation carrying capacity assessment model based on the final key feature subset to assess the oscillation carrying capacity of a power system containing new energy.

Description

Technical Field

[0001] This invention belongs to the field of new energy power system technology, specifically a method for screening key factors of new energy oscillation bearing capacity based on modal sensitivity index and improved XGBoost. Background Technology

[0002] With the large-scale grid connection of new energy sources such as wind power and photovoltaics through power electronic devices, the power system has encountered a wide-frequency oscillation problem covering the subsynchronous to supersynchronous frequency bands. This problem seriously restricts the absorption capacity of new energy sources. Accurately assessing the system's oscillation bearing capacity (i.e., the maximum power injected by new energy sources that the system can withstand within the stability boundary) has become the key to ensuring the safe operation of the power grid.

[0003] The core challenge in constructing an AI model for assessing oscillation bearing capacity lies in the selection of key features; existing technologies suffer from two major drawbacks:

[0004] First, feature selection often relies on researchers' subjective experience (such as directly using power or voltage data), lacking explanation and support from the system's physical mechanisms, resulting in insufficient rationality in feature sets.

[0005] Secondly, the large scale of the system leads to excessively high feature dimensions, which can easily cause model overfitting and the "curse of dimensionality," reducing the accuracy and reliability of the evaluation results.

[0006] The development of related technologies shows a trend of transitioning from single methods to integrated methods; early research was either based on mechanistic models (such as participation factor analysis) for theoretical derivation, but could not cope with high-dimensional data; or it adopted pure data-driven methods (such as filtering and wrapping feature selection), which could uncover data relationships but lacked physical interpretability.

[0007] Therefore, there is an urgent need for a key feature screening method that combines physical interpretability with efficient data processing capabilities to improve the scientificity and reliability of new energy oscillation bearing capacity assessment. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies and solve the aforementioned technical problems, this invention proposes a method for screening key factors of new energy oscillation bearing capacity based on modal sensitivity index and improved XGBoost.

[0009] The technical solution adopted by this invention to solve its technical problem is as follows: This invention proposes a method for screening key factors of new energy oscillation bearing capacity based on modal sensitivity index and improved XGBoost, including the following steps:

[0010] S1: Based on the small-disturbance stability analysis theory of power systems, establish a state-space model of a power system containing new energy sources;

[0011] S2: Based on the state-space model described in S1, four key electrical quantities—active power, reactive power, voltage, and current—are initially screened out through factor analysis to form an initial feature set.

[0012] S3: The initial feature set described in S2 is reduced in dimensionality using a variational autoencoder to obtain a low-dimensional latent representation that retains oscillatory characteristics;

[0013] S4: The improved XGBoost model, which inputs the low-dimensional latent representation described in S3 into the objective function, is combined with a preset threshold to select the final key feature subset; the improved XGBoost model adds a physical consistency regularization term to the traditional objective function;

[0014] S5: Construct a new energy oscillation bearing capacity assessment model based on the final key feature subset described in S4, and realize the assessment of the oscillation bearing capacity of power systems containing new energy.

[0015] Furthermore, the new energy power system mentioned in S1 includes a wind turbine subsystem and a power system subsystem; when establishing the state space model, the state space equations of the wind turbine subsystem and the power system subsystem are first constructed separately, and then combined with the voltage and current interaction relationship of the wind turbine subsystem and the power system subsystem at the grid connection point to obtain the equations; the oscillation bearing capacity is the maximum new energy injection power that the new energy power system can withstand within the stability boundary.

[0016] Furthermore, the state-space equation of the wind turbine subsystem is based on the infinitesimal increments of the wind turbine state variables relative to their steady-state values. The input vectors are the infinitesimal increments of the voltage and current at the grid connection point relative to their steady-state values, and the output variables are the infinitesimal increments of the active power and reactive power relative to their steady-state values. The state-space equation of the power system subsystem is based on the infinitesimal increments of the power system state variables relative to their steady-state values, and its input vector is consistent with that of the wind turbine subsystem.

[0017] Furthermore, the participation factor analysis described in S2 quantifies the importance of each state variable in a specific oscillation mode by calculating the product of the left and right eigenvectors of the coefficient matrix of the state-space model, and then derives the mapping relationship between the oscillation bearing capacity and the electrical quantities of the system.

[0018] Furthermore, before performing dimensionality reduction on the initial feature set in S3, preprocessing of the initial feature set data is also included; the preprocessing includes removing abnormal data or adjusting electrical quantity data of different magnitudes to the same numerical range.

[0019] Furthermore, the variational autoencoder described in S3 includes an encoder and a decoder; the encoder encodes the initial feature set data into probability distribution parameters of latent variables, and the decoder samples the latent variables from the probability distribution and decodes them back into the original data space; when training the variational autoencoder, the objective is to minimize an optimization objective function that includes KL divergence and mean square error, whereby the KL divergence measures the similarity between the approximate posterior distribution and the prior distribution of the latent variables, and the mean square error measures the difference between the input data and the reconstructed data.

[0020] Furthermore, the traditional XGBoost objective function described in S4 includes a loss function and a regularization term; the loss function measures the difference between the true and predicted values ​​of the samples, and the regularization term controls the model complexity by limiting the number of leaf nodes and weights in the decision tree; the physical consistency regularization term is guided by physical laws by introducing a modal sensitivity index vector, which is calculated based on the state space model and represents the physical importance of each state variable to the oscillation mode.

[0021] Furthermore, the preset threshold mentioned in S4 was determined through multiple experiments. The preset threshold satisfies the requirement of eliminating redundant features in the initial feature set while ensuring evaluation accuracy. During the training of the improved XGBoost model, the weight of the misclassified samples in the previous round is increased, the weight of the correctly classified samples is decreased, and the weak classifier with smaller classification error has a greater weight in the final decision.

[0022] Furthermore, when constructing the evaluation model in S5, the final key feature subset is used as input, and the maximum injection power of new energy that the new energy power system can withstand is used as output. The model is trained using a machine learning model, which is a neural network or a support vector machine.

[0023] Furthermore, the new energy source in the power system containing new energy sources is wind power or photovoltaic power, and it is connected to the power grid through power electronic equipment.

[0024] The beneficial effects of this invention are as follows:

[0025] 1. The new energy oscillation carrying capacity key factor screening method based on modal sensitivity index and improved XGBoost described in this invention, through modal sensitivity index (MSI) analysis and participation factor analysis, provides a rigorous physical mechanism basis for feature screening from the perspective of small disturbance stability theory of power system. It has the advantages of enhancing adaptability to oscillation problems of new energy power system, more accurately reflecting the dynamic response characteristics of state variables at different frequencies, changing the feature selection process from relying on subjective experience to quantitative derivation based on system physical laws, significantly improving the scientificity and interpretability of the method, and laying a reliable theoretical foundation for subsequent artificial intelligence models.

[0026] 2. The new energy oscillation carrying capacity key factor screening method based on modal sensitivity index and improved XGBoost described in this invention achieves nonlinear dimensionality reduction processing of high-dimensional electrical quantity time series data through training optimization of variational autoencoder and "encoder-decoder" structure (with the objective function including KL divergence and mean square error as the training objective). It has the advantages of compressing massive data into low-dimensional potential representations while retaining the core characteristics of broadband oscillation, greatly improving data processing efficiency, effectively enhancing the model's generalization ability by eliminating redundant information, and suppressing the risk of overfitting.

[0027] 3. The method for screening key factors of new energy oscillation carrying capacity based on modal sensitivity index and improved XGBoost described in this invention achieves accurate screening of key factors of new energy oscillation carrying capacity by introducing a physical consistency regularization term into the objective function of the improved XGBoost model and a feature importance evaluation mechanism. It confirms the core position of active power in oscillation carrying capacity assessment from the data level, verifies the rationality of the conclusion from the physical mechanism level, and enables the finally constructed new energy oscillation carrying capacity assessment model to maintain high accuracy while having a solid physical foundation, greatly improving the reliability and engineering practical value of the method. Attached Figure Description

[0028] The invention will now be further described with reference to the accompanying drawings.

[0029] Figure 1 This is a flowchart of the broadband oscillation feature fine screening method based on variational autoencoder and XGBoost of the present invention;

[0030] Figure 2 This is a framework diagram of the wideband oscillation feature fine screening based on variational autoencoder and XGBoost in this invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Example: Combining Figures 1-2 As shown, the implementation process of this invention includes the following:

[0033] Step 1: Using the small-disturbance stability analysis theory of power systems, establish a state-space model of a power system including new energy sources. Based on the modal sensitivity index (MSI) analysis, derive the mapping relationship between oscillation bearing capacity and system electrical quantities from the physical mechanism level, and initially select four key electrical quantities, namely active power, reactive power, voltage and current, as the initial feature set.

[0034] The algorithm principle and specific process are as follows:

[0035] Mechanism-based coarse screening:

[0036] By calculating the sensitivity of system operating parameters (such as active power, reactive power, voltage, and current at the generator outlet) to key oscillation modes, the correlation strength between various electrical quantities and the oscillation source is quantified from a physical mechanism perspective, thereby initially screening out the set of key electrical quantities. To select effective data features, it is necessary to determine which electrical quantities have a close relationship with the oscillation source. First, the range of electrical quantities is roughly determined from a physical mechanism perspective. Then, these are input into XGBoost for influencing factor ranking, selecting the electrical features most relevant to the oscillation source. Finally, these are input into a neural network-based disturbance source localization model to achieve fast and accurate localization.

[0037] Consider a scenario where a direct-drive wind turbine is connected to an AC main grid (equivalent to an infinite system) via an AC power grid. The wind turbine can be considered as a "wind turbine subsystem," while the main grid and other parts of the power system connected to it can be considered as a "power system subsystem."

[0038] Let X w U is a column vector containing all the state variables of the wind turbine. w =[ΔV w ,ΔI w ] T This is the input vector of the "wind turbine subsystem," composed of the voltage and current vectors of the "wind turbine subsystem" and the "power system subsystem" at the grid connection point. V w I w These represent the terminal voltage and current at the wind turbine's grid connection point, respectively; ΔY w =[ΔP w ,ΔQ w ] T For the output variable of the "wind turbine subsystem", P w Q w Let represent the active power and reactive power supplied by the wind turbine to the "power system subsystem" respectively. Based on these definitions, the state-space equation of the "wind turbine subsystem" is expressed as follows: Equation 1:

[0039]

[0040] In the formula, Δ is the infinitesimal increment of the corresponding variable relative to its steady-state value, and A w B represents the state matrix of the "wind turbine subsystem". w C w D w For the coefficients or coefficient matrix of the "wind turbine subsystem";

[0041] Suppose X g Given a column vector containing all state variables of the "power system subsystem", the state-space equation of the "power system subsystem" is as follows:

[0042]

[0043] In the formula, A g B is the state matrix of the "power system subsystem". g C g D g This is the coefficient matrix of the "power system subsystem";

[0044] Combining equations 1 and 2 above, we can obtain the state-space model equation of the entire power system with the direct-drive wind turbine connected to the AC grid, which can be expressed as equation 3 below:

[0045]

[0046] In the formula, A wg The characteristic matrix of the closed-loop system contains information about the system's dynamic characteristics. Based on the eigenvalue analysis method for the stability of the system with small disturbances, a state-space model of the system's state variables is constructed, and the model is linearized at the system's equilibrium state to obtain the small-signal state-space matrix of the system around its equilibrium point.

[0047] Based on the system's coefficient matrix A wg The modal sensitivity index (MSI) can be calculated from the left and right eigenvectors. Based on the traditional participation factor, the MSI incorporates the influence of the rate of change of state variables, more comprehensively reflecting the dynamic response sensitivity of each state variable to a specific oscillation mode of the system. Specifically, the modal sensitivity index includes the participation factor, which is obtained by comparing the contribution of each state variable with the total energy of the mode, thus quantifying the importance of each state variable in a specific oscillation mode. The formula for calculating the modal sensitivity index is as follows:

[0048] Formula 4:

[0049]

[0050] In the formula, MSI kiψ represents the modal sensitivity exponent of the k-th state variable to the i-th oscillation mode; ki and φ ki These are the elements of the left and right eigenvector matrices, respectively (same as the traditional participation factor); λ i γ is the i-th eigenvalue (frequency-related parameter of the oscillation mode); γ is the sensitivity adjustment coefficient, usually taken as 0.1-0.5;

[0051] Compared to traditional participation factors, MSI introduces eigenvalues ​​λ. i The absolute value term enhances the sensitivity to high-frequency oscillation modes. This allows MSI to consider not only the static participation of state variables in oscillation modes but also the dynamic response characteristics of the system at different frequencies, making it particularly suitable for the analysis of broadband oscillation problems in new energy power systems.

[0052] According to Formula 4, if the coefficient matrix A is known, the modal sensitivity index MSI of the system can be calculated. ki Combining formulas 1-4, it can be seen that the coefficient matrix A of the closed-loop system is related to the operating point E of the system; therefore, the relationship between the oscillation source and the electrical quantities of the system can be described by the following formula 5:

[0053] L = f L←e (f e←A (f A←MSI (f MSI←E (E))));

[0054] In the formula, E consists of input variables and output variables, E = [Y w U w ] = [P w Q w V w ,I w ], where P w Q w V w I w These represent the active power, reactive power, voltage, and current at the grid connection point, respectively; L represents the location of the oscillation source; f L←MSI The mapping function between the oscillation source label and the system modal sensitivity index is described. Therefore, the oscillation source of the actual system is closely related to the active power, reactive power, voltage and current at the grid connection point of the wind turbine generator, and these electrical quantities can be obtained through actual measurement. Therefore, when using artificial intelligence methods to locate the disturbance source, the active power, reactive power, voltage and current at the grid connection point of the wind turbine or the generator outlet should be used as a set of electrical quantities to construct a broadband oscillation sample set.

[0055] The second step addresses the issues of high dimensionality and large data volume in the initial feature set by employing a variational autoencoder (VAE) for deep feature extraction and dimensionality reduction. The VAE compresses high-dimensional time-series data into low-dimensional latent representations through an encoder-decoder structure while preserving the data's oscillatory characteristics. The resulting low-dimensional latent representations are then input into an improved XGBoost model with a target function for fine-tuning. A physical consistency regularization mechanism ensures that feature importance assessment conforms to the dynamic physical laws of the system, and a final subset of key features is selected based on threshold values.

[0056] The basic principle of a variational autoencoder is as follows:

[0057] Variational Autoencoders (VAEs) introduce the idea of ​​probabilistic generative models into the autoencoder paradigm, adding a degree of randomness to the encoding process, allowing the model to generate new data points. The core improvement of VAEs is:

[0058] Encoder: Encodes input data into parameters of latent variables (such as mean and variance), which describe the probability distribution of the latent variables, usually assumed to be Gaussian distribution;

[0059] Decoder: Samples latent variables from the probability distribution of latent variables obtained from the encoder, and then decodes these latent variables back to the original data space;

[0060] The encoder and decoder of the VAE are modeled as having parameters. of and p with parameter θ θ The probability distribution models of (x|z) are implemented using neural networks; the input data is transmitted via an encoder. They are encoded as latent variables z, which are typically assumed to follow a standard normal distribution.

[0061] Because the true posterior distribution p θ (x|z) is usually difficult to obtain directly, so VAE uses an approximate distribution. To estimate the posterior distribution p θ (x|z); then, by approximating the distribution Medium sampling to simulate the true posterior distribution p θ The sampling process in (x|z). To achieve an approximate distribution. Compared with the true posterior distribution p θ (x|z) should be as similar as possible;

[0062] Variational autoencoders use Kullback-Leibler (KL) divergence to quantify the similarity between two probability distributions; the smaller the KL divergence value, the more similar the two distributions are. The optimization objective of the neural network aims to minimize the KL divergence between the distribution of the encoded latent variable and the prior distribution, which can be expressed as Equation 6 below:

[0063]

[0064] Where, q φ (z|x) is the variational distribution, used to represent the posterior distribution of the latent variable z given data x; p θ (z|x) is the prior distribution, used to describe prior knowledge of the latent variable z;

[0065] To simplify the model and calculations, the first term can be set as follows: It follows a normal distribution, p θ (z) is a standard normal distribution, i.e., q φ (z|x)~N(μ1,σ1), p θ (z)~N(0,1), where μ and σ are functions parameterized by the encoder network, representing the mean and variance. The KL divergence can then be calculated as shown in Equation 7:

[0066]

[0067] The second term is a performance metric for evaluating the network's reconstruction capabilities, where z is the value obtained from the encoder. get;

[0068] To make the reconstructed data To effectively extract latent representations, the mean square error L between the input and output should be as close as possible to the original input data x. r The minimized objective is incorporated into the total error function, L r It can be represented as Equation 8 below:

[0069]

[0070] The total error function typically consists of two parts: reconstruction error L r And prior knowledge of latent variables (KL divergence). Therefore, the optimization objective function can be expressed as follows: Equation 9:

[0071] L=D KL (q φ (z|x)||p θ (z))+L r ;

[0072] Finally, by minimizing the objective function (Equation 9) using a neural network, VAE can find model parameters that can simultaneously optimize the quality of data reconstruction and the effectiveness of potential representations.

[0073] The improved XGBoost theoretical basis is as follows:

[0074] The XGBoost algorithm combines weak learners in a specific way to form a learner with significantly improved performance. XGBoost increases the weight of samples that were misclassified by the weak classifiers in the previous round, while decreasing the weight of samples that were correctly classified. When combining the weak classifiers, those with smaller classification errors have larger weights in the final decision, meaning they play a greater role in forming the strong classifier. Conversely, weak classifiers with higher error rates are given lower weights, and their influence is reduced accordingly.

[0075] Therefore, the traditional XGBoost model can be defined as follows (Equation 10):

[0076]

[0077] In the formula, F represents the classification and regression tree space, f k As the base learner, f k (x i ) represents the k-th tree pair of sample x i The predicted score It is x i The prediction results. The objective function of the traditional model can be expressed as follows: Equation 11:

[0078]

[0079] The objective function consists of two parts; the first part is the loss function. Used to measure the difference between the true value and the predicted value of the i-th sample;

[0080] The second part is the regularization term of the model, where γ and λ are hyperparameters, and T and w represent the number and weight of leaf nodes, respectively. Larger weights mean stronger constraints. Tree models are inherently prone to overfitting, so the regularization term can prevent overfitting by limiting the complexity of the model and the number of important nodes.

[0081] This invention makes a key improvement to the XGBoost objective function by introducing a third part: a physical consistency regularization term; the improved objective function is shown in Equation 12 below:

[0082]

[0083] Where: λ p R is the physical regularization coefficient, used to control the strength of physical constraints.phys It is a physical consistency regularization term, and its specific form is as follows: Equation 13:

[0084]

[0085] In the formula, The model's predicted values ​​are relative to the input latent features z. i The gradient reflects the sensitivity of a feature to the prediction result; MSI i α is the modal sensitivity exponent vector corresponding to sample i, representing the importance of each state variable to the oscillation mode obtained from the analysis of the physical mechanism; α is the scaling factor used to adjust the magnitude of the physical terms.

[0086] This project, by aligning the data-driven sensitivity of the constrained model with the direction of the physical mechanism importance vector, embeds the physical knowledge of small-disturbance stability analysis of power systems into the machine learning model. This makes feature importance assessment not only dependent on data statistical relationships but also guided by physical laws, thereby selecting key features that are both data-effective and physically interpretable.

[0087] The core idea of ​​XGBoost is that for every new tree added, f t We use this tree to fit the previous tree f. t-1 The residual between the predicted result and the actual value;

[0088] During this process, the loss function value should continuously decrease; if This indicates that the first t-1 trees are paired with sample x. i The prediction results of the first t trees for sample x i The iterative prediction results can be obtained by minimizing the following objective function 14:

[0089]

[0090] A second-order Taylor expansion of equation 12 yields equation 15:

[0091]

[0092] In the formula, g i and h i These are the first and second partial derivatives of the loss function, respectively, and

[0093] Since the term is a constant, the objective function can be further rewritten as Equation 16:

[0094] ;

[0095] Taking the derivative of equation 14 and setting it to 0, we obtain equation 17 as follows:

[0096]

[0097] Will Substituting the objective function, we get the following equation 18:

[0098]

[0099] Equation 16 represents the leaf node weights of the generated t-th tree, which score the structure of the t-th tree. Similar to the impurity in decision tree algorithms, the smaller the value, the better the tree structure. Decision trees are constructed through optimization of the above objective function, resulting in K decision trees. Then, different weights are assigned to each tree to obtain a strong learner.

[0100] The detailed screening framework for broadband oscillation characteristics is as follows:

[0101] When the power system is large and there are many generators, the active power P, reactive power Q, voltage U and current I selected by the factor screening may face two problems when using the gradient boosting machine learning method XGBoost for feature selection: overfitting and excessively long model training time. This may result in the extraction of key electrical features that are not accurate enough, thus affecting the effect of oscillation source localization.

[0102] Therefore, a variational autoencoder, a deep learning method, is first used for data preprocessing. It learns the latent representation of the input data to reduce the dimensionality of high-dimensional data, effectively extracting the most important information and transforming it into a lower-dimensional representation. Then, the obtained latent representation can be input into an improved XGBoost model to evaluate the importance score of each electrical feature, achieving fine-tuning of electrical features. This not only effectively avoids overfitting and reduces model training time but also improves the physical consistency and accuracy of key electrical feature extraction. The entire method framework is as follows: Figure 2 As shown, the specific steps are as follows:

[0103] (1) Collect the active power P, reactive power Q, voltage U and current I electrical quantity data of the system when oscillation occurs, and preprocess the collected data;

[0104] (2) Input the preprocessed data into the variational autoencoder and train the model to learn the latent representation of the data;

[0105] (3) The latent representation obtained from the variational autoencoder is used as input to the improved XGBoost model, and the importance score of each electrical feature is evaluated through its unique physical consistency regularization objective function.

[0106] This invention constructs a new energy oscillation bearing capacity assessment model based on the selected optimal feature subset, achieving rapid and accurate stability assessment; the entire process is interconnected, ensuring both the physical interpretability of feature selection and the high efficiency of data processing.

[0107] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for screening key factors of new energy oscillation bearing capacity based on modal sensitivity index and improved XGBoost, characterized in that: Includes the following steps: S1: Based on the small-disturbance stability analysis theory of power systems, establish a state-space model of a power system containing new energy sources; S2: Based on the state-space model described in S1, four key electrical quantities—active power, reactive power, voltage, and current—are initially screened out through factor analysis to form an initial feature set. S3: The initial feature set described in S2 is reduced in dimensionality using a variational autoencoder to obtain a low-dimensional latent representation that retains oscillatory characteristics; S4: The low-dimensional latent representation described in S3 is input into the improved XGBoost model as the objective function, and a final key feature subset is selected by combining it with a preset threshold. The improved XGBoost model adds a physical consistency regularization term to the traditional objective function. The traditional XGBoost objective function includes a loss function and a regularization term. The loss function measures the difference between the true value and the predicted value of the sample, and the regularization term controls the model complexity by limiting the number of leaf nodes and weights in the decision tree. The physical consistency regularization term is guided by physical laws by introducing a modal sensitivity index vector. The modal sensitivity index vector is calculated based on the state space model and represents the physical importance of each state variable to the oscillation mode. Modal sensitivity index calculation formula: ； in, Indicates the first The state variable is the first... Modal sensitivity index of each oscillation mode; and These are the elements of the left and right eigenvector matrices, respectively; For the first One eigenvalue; The sensitivity adjustment coefficient is used; key improvements are made to the XGBoost objective function, and the improved objective function is as follows: ； in: It is the physical regularization coefficient, used to control the strength of physical constraints; It's a loss function; it's a loss function. As a base learner; It is a physical consistency regularization term, with the following specific form: ； in, It is the model's predicted value relative to the input latent features. The gradient reflects the sensitivity of a feature to the prediction result; It corresponds to the sample The modal sensitivity exponent vector represents the importance of each state variable to the oscillation mode obtained from the analysis of the physical mechanism. It is a scaling factor used to adjust the magnitude of physics terms; S5: Construct a new energy oscillation bearing capacity assessment model based on the final key feature subset described in S4, and realize the assessment of the oscillation bearing capacity of power systems containing new energy.

2. The method for screening key factors of new energy oscillation bearing capacity based on modal sensitivity index and improved XGBoost according to claim 1, characterized in that, The new energy power system mentioned in S1 includes a wind turbine subsystem and a power system subsystem. When establishing the state space model, the state space equations of the wind turbine subsystem and the power system subsystem are first constructed separately, and then combined with the voltage and current interaction relationship of the wind turbine subsystem and the power system subsystem at the grid connection point to obtain the equations. The oscillation bearing capacity is the maximum power of new energy injection that the new energy power system can withstand within the stability boundary.

3. The method for screening key factors of new energy oscillation bearing capacity based on modal sensitivity index and improved XGBoost according to claim 2, characterized in that, The state-space equation of the wind turbine subsystem is based on the infinitesimal increments of the wind turbine state variables relative to their steady-state values. The input vectors are the infinitesimal increments of the voltage and current at the grid connection point relative to their steady-state values, and the output variables are the infinitesimal increments of the active power and reactive power relative to their steady-state values. The state-space equation of the power system subsystem is based on the infinitesimal increments of the power system state variables relative to their steady-state values, and the input vectors are consistent with those of the wind turbine subsystem.

4. The method for screening key factors of new energy oscillation bearing capacity based on modal sensitivity index and improved XGBoost according to claim 1, characterized in that, The participation factor analysis described in S2 quantifies the importance of each state variable in a specific oscillation mode by calculating the product of the left and right eigenvectors of the coefficient matrix of the state-space model, and then derives the mapping relationship between oscillation bearing capacity and system electrical quantities.

5. The method for screening key factors of new energy oscillation bearing capacity based on modal sensitivity index and improved XGBoost according to claim 1, characterized in that, Before performing dimensionality reduction on the initial feature set in S3, preprocessing of the initial feature set data is also included; the preprocessing includes removing abnormal data or adjusting electrical quantity data of different magnitudes to the same numerical range.

6. The method for screening key factors of new energy oscillation bearing capacity based on modal sensitivity index and improved XGBoost according to claim 1, characterized in that, The variational autoencoder described in S3 includes an encoder and a decoder; the encoder encodes the initial feature set data into probability distribution parameters of latent variables, and the decoder samples the latent variables from the probability distribution and decodes them back to the original data space; When training the variational autoencoder, the objective is to minimize an optimization objective function that includes KL divergence and mean squared error. The KL divergence measures the similarity between the approximate posterior distribution and the prior distribution of the latent variables, and the mean squared error measures the difference between the input data and the reconstructed data.

7. The method for screening key factors of new energy oscillation bearing capacity based on modal sensitivity index and improved XGBoost according to claim 1, characterized in that, The preset threshold mentioned in S4 was determined through multiple experiments. The preset threshold satisfies the requirement of eliminating redundant features in the initial feature set while ensuring evaluation accuracy. During the training of the improved XGBoost model, the weight of misclassified samples from the previous round is increased, the weight of correctly classified samples is decreased, and weak classifiers with smaller classification errors have a greater weight in the final decision.

8. The method for screening key factors of new energy oscillation bearing capacity based on modal sensitivity index and improved XGBoost according to claim 1, characterized in that, When constructing the evaluation model in S5, the final key feature subset is used as input, and the maximum injection power of new energy that the new energy power system can withstand is used as output. The model is trained using a machine learning model, which is a neural network or a support vector machine.

9. The method for screening key factors of new energy oscillation bearing capacity based on modal sensitivity index and improved XGBoost according to claim 1, characterized in that, The new energy source in the power system containing new energy sources is wind power or photovoltaic power, and it is connected to the power grid through power electronic equipment.

Images