An online analysis and optimization method for grid-connected oscillation stability of a wind farm

Abstract

The application discloses a kind of wind farm grid-connected oscillation stability online analysis and optimization method, it is related to wind farm grid-connected technical field, including S1: generating the data set of wind farm grid-connected oscillation stability evaluation;S2: record and mark data set, form input feature set x;S3: according to x, using mode analysis method calculates corresponding eigenvalue, when the real part of eigenvalue is greater than or equal to 0, label value is marked as 0, i.e. there is vibration instability, otherwise, mark 1, label value forms output y, forms vibration stability evaluation model: S4: based on oscillation stability evaluation model LightGBM classification model is built, and evaluation result is output;S6: based on the important feature screening of local interpretable proxy model of generalized additive model to the evaluation result output in step S4;S9: according to the important feature screened in step S6 and the evaluation result in step S4, wind farm grid-connected system scheme optimization can quickly and accurately analyze the vibration stability of wind farm grid-connected.

Description

Technical Field

[0001] This invention relates to the field of wind farm grid connection technology, and in particular to an online analysis and optimization method for the grid-connected oscillation stability of wind farms. Background Technology

[0002] As the scale of wind power generation increases, the system oscillation problems caused by wind power generation are also increasing. In the past decade, there have been hundreds of oscillation events caused by wind farms connecting to the grid worldwide. These events have resulted in many wind turbines being shut down and curtailed due to excessive harmonics, causing great losses to the wind power industry and negatively impacting social and economic benefits. Therefore, how to quickly and accurately assess the oscillation stability of wind farms connected to the grid is a prerequisite for formulating corresponding control strategies, improving oscillation stability, and preventing wind curtailment.

[0003] Existing studies on the interpretability of vibration stability assessment models for wind farms mostly focus on ex-post interpretations unrelated to the model, making it difficult for decision-makers to fully trust the assessment models. Fitting local original models using linear regression cannot satisfy the complex relationships inherent in wind power itself, and the fitting accuracy on complex models is insufficient. As a result, existing studies can only perform local fitting on redundant simple models, such as gradient boosting decision trees (GBDT). Summary of the Invention

[0004] The purpose of this invention is to provide an online analysis and optimization method for the grid-connected oscillation stability of wind farms, so as to solve the problems existing in the prior art and to quickly and accurately analyze the vibration stability of grid-connected wind farms.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] This invention provides an online analysis and optimization method for grid-connected oscillation stability of wind farms, comprising the following steps:

[0007] (I) Offline Training:

[0008] S1: Generate a dataset for assessing the grid-connected oscillation stability of wind farms;

[0009] S2: Record and label the dataset to form the input feature set x;

[0010] S3: Based on the input feature set x, calculate the corresponding eigenvalues ​​using pattern analysis. When the real part of an eigenvalue is greater than or equal to 0, the label value is recorded as 0, indicating vibration instability; otherwise, it is recorded as 1. The label values ​​form the output y, thus forming the vibration stability assessment model: y = F K (x);

[0011] (II) Online Assessment:

[0012] LightGBM is a machine learning method based on learning from historical data and evaluating and analyzing current data online. Using this method, data-driven oscillation stability assessment can be achieved, enabling online prediction of oscillation stability without the need to build a detailed model of the actual system.

[0013] S4: Construct a LightGBM classification model based on the oscillation stability evaluation model in step S3, for online detection and evaluation, and output the evaluation results;

[0014] (III) Decision Optimization

[0015] S6: Based on the evaluation results output in step S4, important features are selected using the locally interpretable surrogate model of the generalized additive model;

[0016] S9: Optimize the wind farm grid connection system scheme based on the key features selected in step S6 and the evaluation results in step S4.

[0017] Preferably, in step S1, various operating scenarios of the wind farm grid-connected system are simulated by setting disturbances, including setting different load fluctuations and different unit output fluctuations, in order to generate a dataset for evaluating the stability of wind farm grid connection oscillations.

[0018] Preferably, in step S2, the dataset is categorized using pattern analysis, with the active power, reactive power, series complement degree, and wind speed on the grid-connected lines of the wind farm as input features. The input feature set x has the form: x = [P1, P2, P3, ..., Q1, Q2, Q3, ..., V w,1 V w,2 V w,3 ,…,C1,C2,C3,],

[0019] Among them, P i and Q i V represents the active power and reactive power on the grid connection line of the i-th wind farm, respectively. wi Let C be the wind speed of the i-th wind farm. i Let be the series compensation degree on the grid connection line of the i-th wind farm.

[0020] Preferably, in step S4, the specific parameters of the LightGBM classification model are set as follows: the number of leaf nodes is set to 31, the learning rate is 0.05, the maximum depth of the tree model is 4, and the remaining parameters are default values.

[0021] Preferably, the method further includes step S5: performance evaluation of the oscillation stability evaluation model, specifically:

[0022] In the oscillation stability assessment of wind farm grid connection, the number of stable and unstable scenarios differs, with unstable scenarios receiving more attention. Using only classification accuracy is insufficient for objectivity; therefore, the following two indicators are used to comprehensively evaluate model performance:

[0023]

[0024]

[0025] In the formula, A accuracy It is the accuracy of the model's predictions, A specificity It is the accuracy of the model in predicting unstable samples, n TP and n FN It is the number of stable samples that are correctly and incorrectly evaluated; n FP and n TN It is the number of unstable samples that are incorrectly and correctly evaluated;

[0026] When the accuracy of the model's predictions and the accuracy of the model's predictions of unstable samples do not meet the standards, the oscillation stability assessment model needs to be modified.

[0027] Preferably, in step S6, the important feature screening specifically includes:

[0028] Fitting a locally interpretable surrogate model for samples of interest using a generalized additive model, specifically including:

[0029] The standard form of the generalized additive model is:

[0030]

[0031] In the formula, x j f is the j-th feature of the sample. j (x j f is a ridge function. i,j (x i ,x j () represents the cross ridge function, where the ridge function is a function that acts on the input features;

[0032] Using binary trees as ridge functions, the features are first processed using binary trees, and then the ridge functions of each feature are linearly combined to obtain the original model f. K The locally interpretable surrogate model for (x) is as follows:

[0033]

[0034] In the formula, z represents the important feature to be focused on, and S is the set of ridge functions corresponding to a single feature;

[0035] Preferably, the method further includes step S7: generating new samples by perturbation near the samples of interest, using the original model to predict label values, thereby obtaining a new dataset, and defining π. x (z) represents the similarity between the samples before and after the perturbation, and the formula is as follows:

[0036]

[0037] In the formula, D(x,z) is the distance formula;

[0038] The objective function for measuring the difference between the locally interpretable surrogate model and the original model is as follows:

[0039] In the formula, f is the original model, g is the interpretable surrogate model, and Ω(g) is the complexity of g.

[0040]

[0041] Preferably, the method further includes step S8: evaluating the performance of the locally interpretable surrogate model using AUC value and accuracy.

[0042] The locally interpretable surrogate model essentially refers to using a simple model with interpretability to fit a part of a complex model. It belongs to the regression model of machine learning, so its performance can be judged by common AUC value and accuracy.

[0043] AUC is the area under the ROC curve for a classification model. The ROC curve is a curve with the TruePositive Rate on the horizontal axis and the False Positive Rate on the vertical axis. The value of AUC ranges from 0 to 1; the closer the value is to 1, the better the classifier's performance.

[0044] Preferably, the method further includes step S10: the optimized wind farm grid connection system scheme in step S9 is used to predict oscillation stability through a locally interpretable model.

[0045] Preferably, the method further includes step S11: verifying the effectiveness of the prediction results of the locally interpretable surrogate model in step S10 through time-domain simulation.

[0046] The present invention achieves the following technical effects compared to the prior art:

[0047] This invention provides an online analysis and optimization method for the oscillation stability of wind farms connected to the grid. It employs a LightGBM model, which boasts fast computation speed and good generalization performance, to predict the oscillation stability risk of wind farms after grid connection. This method is based on historical data and does not require detailed or full-order mathematical models. It enables rapid and accurate analysis of the vibration stability of wind farms connected to the grid. Furthermore, by using a generalized additive model to fit a locally interpretable surrogate model, the results are analyzed on samples of interest to derive auxiliary control strategies. These strategies can be applied online, assisting grid operators in quickly issuing alarms when oscillations occur during actual grid operation and taking emergency optimization control measures to prevent further losses.

[0048] Furthermore, the dataset for assessing the oscillation stability of wind farm grid connection was improved. By setting up disturbances to simulate various operating scenarios of the wind farm grid connection system, the impact of series compensation on oscillation stability on the wind farm grid connection line was taken into account to analyze the system oscillation risk and improve the accuracy of oscillation stability assessment. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 A flowchart illustrating the online analysis and optimization method for grid-connected oscillation stability of wind farms provided by this invention;

[0051] Figure 2 An improved IEEE-39 calculation example diagram for multi-wind farm grid connection in this invention;

[0052] Figure 3 This is a comparison chart of the training time of the LightGBM algorithm of this invention with other algorithms;

[0053] Figure 4 This is a comparison chart of the test times of the LightGBM algorithm of this invention with other algorithms;

[0054] Figure 5 This is a diagram showing the influence of various wind farm characteristics on oscillations, obtained using a locally interpretable surrogate model, as derived in this invention.

[0055] Figure 6 This is a diagram showing the effect of adjusting the output of WT2 on the model prediction results.

[0056] Figure 7 The graph shows the oscillation stability results under two optimized control schemes in the embodiments of the present invention. Detailed Implementation

[0057] 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.

[0058] The purpose of this invention is to provide an online analysis and optimization method for the grid-connected oscillation stability of wind farms, so as to solve the problems existing in the prior art and to quickly and accurately analyze the vibration stability of grid-connected wind farms.

[0059] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0060] This invention provides an online analysis and optimization method for grid-connected oscillation stability of wind farms, such as... Figure 1 As shown, the online analysis and optimization method for grid-connected oscillation stability of wind farms includes the following steps:

[0061] (I) Offline Training:

[0062] S1: By setting up disturbances, various operating scenarios of the wind farm grid-connected system are simulated, including setting different load fluctuations and different unit output fluctuations, to generate a dataset for evaluating the stability of wind farm grid-connected oscillations.

[0063] To demonstrate the effectiveness and superiority of the method described in this embodiment, such as Figure 2 As shown, the improved IEEE-39 example using multiple wind farms connected to the grid includes 10 generators, 39 buses, and 46 lines, with a rated frequency of 60Hz. Node 31 of the generators is a balancing machine. The synchronous generators located at nodes 30, 32, 33, 34, and 35 are replaced with 5 wind farms connected via series compensation (WT1 to WT5, represented by the wind blade shape in the figure).

[0064] Set up the simulation as follows:

[0065] 1) While keeping the total output power constant, adjust the output of each unit to fluctuate randomly between 80% and 120% of the rated value.

[0066] 2) The line load varies randomly within the range of 90% to 110%.

[0067] 3) Set the wind speed of each wind farm to fluctuate randomly from 3 to 15 m / s, and the series compensation degree on the grid connection line to fluctuate randomly from 5% to 50%.

[0068] 4) At t = 0.5s, a fluctuation of 10% of the output is applied to a random unit for a duration of 0.1s, and the simulation duration is 2s.

[0069] 5) Repeat the above process to obtain a total of 33,464 samples, which are then divided into training and test sets in a 3:1 ratio.

[0070] S2: Record and label the dataset to form the input feature set x; specifically, use pattern analysis to classify the dataset, taking the active power, reactive power, series complement degree, and wind speed on the grid-connected lines of the wind farm as input features. The form of the input feature set x is: x = [P1, P2, P3, ..., Q1, Q2, Q3, ..., V w,1 V w,2 V w,3 ,…,C1,C2,C3,],

[0071] Among them, P i and Q i V represents the active power and reactive power on the grid connection line of the i-th wind farm, respectively. wi Let C be the wind speed of the i-th wind farm. i Let be the series compensation degree on the grid connection line of the i-th wind farm;

[0072] S3: Based on the input feature set x, calculate the corresponding eigenvalues ​​using pattern analysis. When the real part of the eigenvalue is greater than or equal to 0, the label value is recorded as 0, indicating vibration instability; otherwise, it is recorded as 1. The label values ​​form the output y, thus obtaining the vibration stability assessment model: y = F K (x);

[0073] (II) Online Assessment:

[0074] LightGBM is a machine learning method based on learning from historical data and evaluating and analyzing current data online. Using this method, data-driven oscillation stability assessment can be achieved, enabling online prediction of oscillation stability without the need to build a detailed model of the actual system.

[0075] S4: Construct a LightGBM classification model based on the oscillation stability evaluation model in step S3, for online detection and evaluation, and output the evaluation results;

[0076] The specific parameters of the LightGBM classification model are set as follows: the number of leaf nodes is set to 31, the learning rate is 0.05, the maximum depth of the tree model is 4, and the other parameters are the default values.

[0077] S5: Performance evaluation of the oscillation stability assessment model, specifically:

[0078] The confusion matrix is ​​used to represent the model performance:

[0079]

[0080] Table 1. Model Classification Confusion Matrix

[0081] In the oscillation stability assessment of wind farm grid connection, the number of stable and unstable scenarios differs, with unstable scenarios receiving more attention. Using only classification accuracy is not objective enough. The following two indicators are used to comprehensively evaluate model performance:

[0082]

[0083]

[0084] In the formula, A accuracy It is the accuracy of the model's predictions, A specificity It is the accuracy of the model in predicting unstable samples, n TP and n FN It is the number of stable samples that are correctly and incorrectly evaluated; n FP and n TN It is the number of unstable samples that are incorrectly and correctly evaluated;

[0085] When the accuracy of the model's predictions and the accuracy of the model's predictions of unstable samples do not meet the standards, the oscillation stability assessment model needs to be modified.

[0086] The algorithm of this invention (lightGBM) is compared with random forest (RF), support vector machine (SVM), gradient boosting tree (GBDT), extreme gradient boosting (XGboost), and convolutional neural network (CNN) algorithms. The parameters of each model were selected based on a trial-and-error method. The number of decision trees in RF was 81, the SVM used a Gaussian kernel function, the maximum number of iterations in GBDT was set to 500, the learning rate was 0.1, the maximum tree depth in XGboost was 6, and the CNN used a 3*3 convolutional kernel. The classification accuracy of each algorithm is shown in Table 2.

[0087]

[0088] Table 2 Classification accuracy of each algorithm

[0089] It can be seen that, on the one hand, using the series complement degree on the grid-connected line of the wind farm as an input feature can significantly improve the accuracy of oscillation stability assessment and the accuracy of assessment of unstable samples; on the other hand, the classification index accuracy of the algorithm of this invention is better than that of other algorithms.

[0090] Furthermore, such as Figure 3 , Figure 4As can be seen from the classification accuracy of various algorithms, using LightGBM not only improves the computational complexity of algorithms such as GBDT, but also significantly improves the model training efficiency and accuracy; the SVM algorithm is suitable for small sample data, although the training time is short, the accuracy does not meet the requirements.

[0091] (III) Decision Optimization

[0092] S6: Based on the evaluation results output in step S4, important features are selected using the locally interpretable surrogate model of the generalized additive model;

[0093] Fitting a locally interpretable surrogate model for samples of interest using a generalized additive model, specifically including:

[0094] The standard form of the generalized additive model is:

[0095]

[0096] In the formula, x j f is the j-th feature of the sample. j (x j f is a ridge function. i,j (x i ,x j () represents the cross ridge function, where the ridge function is a function that acts on the input features;

[0097] Using a binary tree as the ridge function, that is, first processing the features with the binary tree, and then processing each feature...

[0098]

[0099] The original model f is obtained by linear combination of ridge functions. K The locally interpretable surrogate model for (x) is as follows:

[0100] In the formula, z represents the important feature to be focused on, and S is the set of ridge functions corresponding to a single feature;

[0101] S7: First, new samples are generated by perturbation near the samples of interest. The original model is then used to predict the label values, resulting in a new dataset. π is defined as follows: x (z) represents the similarity between the samples before and after the perturbation, and the formula is as follows:

[0102]

[0103] In the formula, D(x,z) is the distance formula;

[0104] The objective function for measuring the difference between the locally interpretable surrogate model and the original model is as follows:

[0105]

[0106] In the formula, f is the original model, g is the interpretable surrogate model, and Ω(g) is the complexity of g.

[0107] For a single sample of interest, a locally interpretable surrogate model is fitted to the original model within its neighborhood. This interpretable surrogate model can identify key features affecting oscillation stability under the current state based on risk analysis results, enabling targeted control and modification of the wind farm and system structure, and the effectiveness of the control measures is measurable.

[0108] S8: Evaluate the performance of the locally interpretable surrogate model using AUC values ​​and accuracy;

[0109] An interpretable surrogate model essentially refers to using a simple model with interpretability to fit a part of a complex model. It belongs to the regression model of machine learning, so its performance can be judged by common AUC values ​​and accuracy.

[0110] AUC is the area under the ROC curve for a classification model. The ROC curve is a curve with the TruePositive Rate on the horizontal axis and the False Positive Rate on the vertical axis. The value of AUC ranges from 0 to 1; the closer the value is to 1, the better the classifier's performance.

[0111] The generalized additive model (M1) of the present invention, which uses a binary tree as a ridge function, is compared with the generalized additive model (M2) which uses a fully connected neural network as a ridge function, and the method of fitting the local original model directly using linear regression.

[0112] Ten datasets, each containing 5000 samples, were generated around the given sample points using ten different random number seeds. These datasets were fitted using M1, M2, and linear regression, and their AUC values ​​and accuracies on the generated datasets are listed in Table 3. The left side of the parentheses represents the AUC value, and the right side represents the accuracy.

[0113]

[0114] Table 3 shows the fitting results of the three models on a dataset randomly generated near the sample of interest.

[0115] It can be seen that, for the complex LightGBM model, the linear regression performs worse overall, while for the algorithm using a fully connected neural network as the ridge function, the binary tree as the ridge function performs better. This is because the algorithm is based on a binary tree, which is more suitable for approximating complex tree models. Therefore, the interpretable surrogate model of this invention performs better.

[0116] S9: Optimize the wind farm grid connection system scheme based on the important features selected in step S6 and the evaluation results in step S4;

[0117] S10: The optimized wind farm grid connection system scheme in step S9 is used to predict oscillation stability through a locally interpretable model;

[0118] S11: Verify the effectiveness of the prediction results of the locally interpretable surrogate model in step S10 through time-domain simulation.

[0119] Based on the fitted model, the sample of interest is analyzed: when the system is predicted to be unstable at a certain operating point, an interpretable model of the stability assessment model is fitted in the neighborhood of the current unstable sample to obtain the importance of each feature of the current operating condition; decision-makers can clearly understand the priority of each feature through the results and make better auxiliary control strategies to suppress oscillations.

[0120] A sample of predicted oscillation instability is randomly selected from all samples and analyzed using a locally interpretable analysis model. Ignoring the unadjustable characteristic wind speed, the importance of the top eight key features affecting oscillation stability is... Figure 5 It is displayed in the middle.

[0121] Depend on Figure 5 It can be seen that the active power and series compensation on the grid-connected lines of the wind farm have a positive impact on the prediction results, that is, as the wind power penetration rate increases or as the series compensation increases, the system tends to be more oscillating and unstable. On the other hand, the reactive power on the grid-connected lines of the wind farm has a negative impact on the prediction results, that is, the increase in reactive power has a positive impact on the oscillation stability of the system. From a physical point of view, in the grid-connected system of the wind farm, the increase in reactive power can regulate the system voltage. When the system voltage is high or low, the regulating effect of reactive power can suppress the system oscillation by bringing the voltage back to a suitable range, thus having a positive impact on the oscillation stability of the system. This shows that the analysis results of this interpretable surrogate model are consistent with objective physical laws. Similarly, the positive and negative correlations between other characteristics and oscillation stability can be obtained through the analysis of this model.

[0122] Furthermore, under the current operating conditions, the influence weight of power output varies among different wind farms. This is related to the different control parameters of each wind farm, as well as the distance of the grid connection lines and the series compensation heterogeneity. Figure 5It can also be seen that the power output of Wind Farm No. 2 and the series compensation degree on its grid connection line have the greatest impact on system oscillation stability. When the data-driven oscillation stability assessment system predicts system oscillation instability online, short-term auxiliary optimization control strategies can consider prioritizing the disconnection of wind turbines at Wind Farm No. 2 and adjusting the series compensation degree on its grid connection line. In the long term, additional damping control elements can be designed at Wind Farm No. 2 to fundamentally suppress oscillations.

[0123] By adjusting the output of wind farm No. 2 using an interpretable surrogate model, the model predicts the probability of system oscillation and instability as follows: Figure 6 As shown, when the output decreases, the probability of the system oscillation and instability predicted by the model also gradually decreases.

[0124] According to the present invention, a strategy for suppressing oscillations is formulated for this embodiment, namely, cutting off 10% of the output of wind farm No. 2 and adjusting the series compensation on the No. 2 grid connection line to 50% of its original value; this is compared with another control strategy (cutting off 10% of the output of wind farms No. 1 and No. 3) and verified by simulation. The results are as follows. Figure 7 .

[0125] Depend on Figure 7 It can be seen that the system exhibits oscillatory divergence under the original operating conditions. If 10% of the output of wind farm No. 2 is cut off, and the series compensation on the No. 2 grid connection line is adjusted to 50% of its original value (Scheme 1), the oscillation decays, and the effect on improving the damping ratio of the system's dominant mode is most significant. If 10% of the output of wind farms No. 1 and No. 3 is cut off (Scheme 2), the system damping is slightly lower, and the system remains in a constant-amplitude oscillation state for a long time. It is evident that, under the same control measures, selecting an effective target will result in better suppression, verifying the correctness of the oscillation analysis system. These findings are consistent with the above analysis results, demonstrating the correctness and feasibility of the oscillation stability analysis and control method of this invention.

[0126] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for online analysis and optimization of grid-connected oscillation stability of wind farms, characterized in that: include: The following steps are required: (a) Offline training: S1: Generate a dataset for assessing the grid-connected oscillation stability of wind farms; S2: Record and label the dataset to form the input feature set x; S3: Based on the input feature set x, the corresponding eigenvalues ​​are calculated using pattern analysis. When the real part of the eigenvalue is greater than or equal to 0, the label value is recorded as 0, indicating vibration instability; otherwise, it is recorded as 1. The label values ​​form the output y, thus forming a vibration stability assessment model. ; (ii) Online assessment: S4: Construct a LightGBM classification model based on the oscillation stability evaluation model in step S3, for online detection and evaluation, and output the evaluation results; (III) Decision Optimization S6: Based on the evaluation results output in step S4, important features are selected using the locally interpretable surrogate model of the generalized additive model; S9: Optimize the wind farm grid connection system scheme based on the important features selected in step S6 and the evaluation results in step S4; In step S2, the dataset is categorized using pattern analysis. The active power, reactive power, series complement degree, and wind speed on the grid-connected lines of the wind farm are used as input features. The input feature set x takes the form: ， in, and The first Active and reactive power on the grid connection lines of each wind farm For the first Wind speed at each wind farm For the first Series compensation degree on the grid connection line of a wind farm; In step S6, the specific steps for selecting important features include: Fitting a locally interpretable surrogate model for samples of interest using a generalized additive model, specifically including: The standard form of the generalized additive model is: In the formula, β0 is the bias term. For the j-th feature of the sample, For ridge functions, The cross ridge function is a function that acts on the input features, and the bias term is the baseline output value of the model when all features are zero or take the baseline value. Using binary trees as ridge functions, the features are first processed using binary trees, and then the ridge functions of each feature are linearly combined to obtain the original model. The locally interpretable proxy model takes the following form: In the formula, μ is the intercept term, that is, when the ridge function of all features... When the value is 0, it is the baseline value of the output of the surrogate model g(z), where z is the important feature to be focused on. A set of ridge functions corresponding to a single feature; It also includes step S7: generating new samples by perturbing the samples of interest, using the original model to predict the label values, thus obtaining a new dataset, and defining... The similarity between samples before and after perturbation is calculated using the following formula: In the formula, Distance formula The objective function for measuring the difference between the locally interpretable surrogate model and the original model is as follows: In the formula, For the original model, For an interpretable proxy model, The time complexity is g.

2. The method for online analysis and optimization of grid-connected oscillation stability of wind farms according to claim 1, characterized in that: In step S1, a dataset for assessing the grid-connected oscillation stability of a wind farm is generated by simulating various operating scenarios of the wind farm grid-connected system through setting up disturbances.

3. The method for online analysis and optimization of grid-connected oscillation stability of wind farms according to claim 1, characterized in that: In step S4, the specific parameters of the LightGBM classification model are set as follows: the number of leaf nodes is set to 31, the learning rate is 0.05, the maximum depth of the tree model is 4, and the remaining parameters are the default values.

4. The method for online analysis and optimization of grid-connected oscillation stability of wind farms according to claim 1, characterized in that: It also includes step S5: performance evaluation of the oscillation stability assessment model, specifically: In the formula, It is the accuracy of the model's predictions. It is the accuracy of the model in predicting unstable samples. and It is the number of stable samples that are correctly and incorrectly evaluated; and It is the number of unstable samples that are incorrectly and correctly evaluated; When the accuracy of the model's predictions and the accuracy of the model's predictions of unstable samples do not meet the standards, the oscillation stability assessment model needs to be modified.

5. The method for online analysis and optimization of grid-connected oscillation stability of wind farms according to claim 1, characterized in that: It also includes step S8: evaluating the performance of the locally interpretable surrogate model using AUC value and accuracy.

6. The method for online analysis and optimization of grid-connected oscillation stability of wind farms according to claim 1, characterized in that: It also includes step S10: the optimized wind farm grid connection system scheme in step S9 is used to predict oscillation stability through a locally interpretable model.

7. The method for online analysis and optimization of grid-connected oscillation stability of wind farms according to claim 6, characterized in that: It also includes step S11: verifying the effectiveness of the prediction results of the locally interpretable surrogate model in step S10 through time-domain simulation.

Images