A power grid broadband oscillation intelligent identification method considering new energy hybrid control mode

Abstract

The application discloses a power grid broadband oscillation intelligent identification method considering a new energy hybrid control mode and belongs to the technical field of smart grids. The method builds a new energy unit hybrid control grid-connected system simulation model containing network type and follow network type control, generates different operation condition simulation data samples and classifies them; uses a support vector machine algorithm to classify oscillation states, quantifies feature contribution degrees in combination with a SHAP algorithm, identifies dominant influence features through K-means clustering, and finally accesses a wide-area monitoring platform to realize differentiated visual risk early warning. The application solves the problems that a traditional method is difficult to adapt to complex scenes of a new power system and the like, realizes broadband oscillation severity evaluation, dominant factor identification and visual display, improves the safe and stable operation capability of a power grid, and has the advantages of strong scene adaptability, high automation level, good interpretability and the like.

Description

Technical Field

[0001] This invention belongs to the field of smart grid technology, and specifically relates to a method for intelligent identification of broadband oscillations in power grids that takes into account the hybrid control mode of new energy sources. Background Technology

[0002] With the continuous advancement of the "dual carbon" target, the penetration rate of new energy sources, represented by wind power and photovoltaics, in the power grid is rapidly increasing, and the power grid is gradually exhibiting new operational characteristics of high proportion of new energy and deep power electronics. However, the large-scale grid connection of new energy sources makes the control mode of the power grid more complex and diverse, and the dynamic characteristics of the system more uncertain. This hybrid control mode, with the coexistence of multiple new control strategies (such as virtual synchronous machine control, droop control, matching control, and grid-following control), makes the power system more prone to broadband oscillation events. In recent years, power grid safety and stability problems caused by broadband oscillations have frequently occurred, seriously affecting the reliable operation of the power grid and even potentially causing large-scale blackouts, which have a serious impact on the social economy. Therefore, accurate prediction and effective risk assessment of oscillation states during power grid operation have become key aspects of current power system dispatch and operation. Using machine learning technology to quickly and accurately identify and predict broadband oscillations, and to identify the dominant factors and key parameters leading to oscillation risks in real time, further realizing proactive and precise control of the power grid's operating state, is of great significance for improving the safe and stable operation capability of the power grid.

[0003] Current traditional oscillation analysis methods for power grids generally focus on low-frequency or subsynchronous oscillations under a single control mode, and are mostly based on classical modeling methods. They only study oscillation modes under specific operating conditions or frequency bands, making it difficult to effectively adapt to the complex scenarios of new power systems with multiple renewable energy sources and multiple control modes. Specifically, traditional methods lack in-depth research on broadband oscillation phenomena occurring under scenarios with a high proportion of renewable energy integration, especially lacking the identification of the dominant factors of oscillation under different control modes and combinations of control parameters. In addition, traditional analysis methods generally suffer from strong uninterpretability of results and a lack of visualization methods, making it difficult to meet the actual needs of current power grid operation safety risk early warning and accurate decision-making. Therefore, it is necessary to design an intelligent identification method based on data-driven and machine learning technologies to address the complex oscillation problem in new hybrid control mode grid connection scenarios. This method can accurately assess the severity of broadband oscillations, accurately identify the dominant factors of oscillations, and assist power system dispatching decisions through differentiated visualization methods. Summary of the Invention

[0004] To address the problem of broadband oscillations in power grids under high-proportion renewable energy grid connection and multiple hybrid control modes, this invention aims to propose an intelligent identification method for broadband oscillations in power grids that considers renewable energy hybrid control modes. This method can automatically identify the severity of system oscillations and the dominant oscillation characteristics, and achieve intelligent monitoring and differentiated visualization of broadband oscillation risks.

[0005] To achieve the above objectives, the technical solution adopted by this invention is as follows: A method for intelligent identification of broadband oscillations in power grids considering a hybrid control mode of new energy sources, comprising the following steps:

[0006] S1: Build a simulation model of a hybrid control grid-connected system for new energy generating units that includes grid-connected control and grid-following control, and define the unit access status and adjustable control parameters.

[0007] S2: By randomly adjusting the connection status and control parameter disturbance method of the unit, simulation data samples of different operating conditions are generated. Through FFT spectrum analysis, the data scenarios are classified into stable scenarios, wideband oscillation scenarios and other scenarios according to the frequency energy distribution and modal characteristics of the system.

[0008] S3: Using the Support Vector Machine (SVM) algorithm with multi-dimensional features, including voltage amplitude, frequency, and power, as input, the oscillation state of different simulation scenarios is classified to distinguish between stable scenarios and broadband oscillation scenarios. S4: The SHAP algorithm is used to analyze the feature contribution of the SVM classification model, quantifying the contribution of different input features to the oscillation classification results and enabling interpretability analysis of the model's results. S5: Using the K-means clustering algorithm with the feature contribution matrix obtained from SHAP analysis as input, unsupervised clustering analysis is performed. By statistically analyzing the average SHAP value of features within each cluster, the dominant influencing features under different broadband oscillation scenarios are identified, clarifying the key factors causing broadband oscillations in the system.

[0009] S6: Integrate the oscillation classification results, SHAP feature contribution analysis results, and the dominant factors of cluster identification into the wide-area monitoring platform, and unify them into the power system wide-area monitoring platform to achieve risk warning through differentiated visualization.

[0010] Furthermore, in step S1, the grid-type control includes virtual synchronous machine control, droop control or matching control, and the grid-type control includes frequency tracking control. The simulation model covers the topology of wind power, photovoltaic energy grid-connected units and common coupling points.

[0011] Furthermore, in step S2, the method for calculating the distribution index of the broadband oscillation scenario is as follows:

[0012] The frequency domain spectral distribution of the simulated time-domain data is obtained by FFT transformation, as shown in the following formula:

[0013]

[0014] Where x[n] is the nth sampling point of the time-domain sequence; X[k] is the spectral amplitude of the kth frequency component; N is the signal length; j is the imaginary unit;

[0015] For a given disturbance condition, the formula for the broadband dominant oscillation distribution index is defined as follows:

[0016]

[0017] Where X(f) represents the spectral amplitude under the disturbance condition, X ss (f) represents the spectral amplitude under steady state, F osc This represents a set of wideband oscillation frequency bands, selecting the frequency bands of interest; F main The set of dominant oscillation frequencies is selected, and the top k frequencies with the largest energy proportion are chosen; the function of argmax is to find all f∈F osc Within the range, the frequency points with the greatest difference; the function of Top-k is to select the top k frequencies with the greatest difference.

[0018] Furthermore, in step S2, the process of generating multi-condition simulation data samples is as follows:

[0019] S21: Randomly set the unit connection status and initialize the control parameters to a reasonable range;

[0020] S22: For each set of access and parameter combinations, run the system time-domain simulation and record the changes of key variables over time;

[0021] S23: Perform frequency domain analysis on the simulation results, and extract the spectral characteristics of voltage or power using techniques such as FFT.

[0022] They are classified according to oscillation mode, energy distribution, and frequency range;

[0023] S24: Based on the dynamic response characteristics of the system, the data samples are divided into stable scenarios, wideband oscillation scenarios, and other types of scenarios.

[0024] Furthermore, the classification decision function of the SVM algorithm described in step S3 is defined by the following formula:

[0025]

[0026] Where, α i For Lagrange multipliers; K(x) i x) is the kernel function; b is the bias.

[0027] The optimal classification hyperplane within this space is represented by the following formula:

[0028] K(x i ,x j )=φ(x i ) T φ(x j )

[0029] Where φ(·) is a nonlinear mapping function,

[0030] The Gaussian radial basis kernel formula is as follows:

[0031] K(x i ,x j )=exp(-γ||x i -x j || 2 )

[0032] By optimizing the classification hyperplane through cross-validation and regularization parameter adjustment, overfitting of the model is avoided. The input features of SVM include node voltage, frequency, PCC point power, and amplitude of the dominant oscillation frequency.

[0033] Furthermore, in step S4, the contribution of each input feature is quantitatively interpreted using the SHAP analysis method, as follows:

[0034] First, perform full combinatorial sampling on all input feature sets N. For each feature i, evaluate the change in the model's predicted output before and after its inclusion, and then calculate the SHAP value φ of feature i. i The principle behind SHAP value calculation is as follows:

[0035]

[0036] Where S is a subset that does not contain feature i; f S (x) represents the model output when only feature S is included; f S∪{i} (x) represents the output when feature i is added to S;

[0037] Then, for multiple samples, the SHAP values ​​are used to statistically determine the feature importance in the following matrix form:

[0038]

[0039] Where, φ i This represents the SHAP value of the j-th feature of the i-th sample.

[0040] Furthermore, in step S5, the feature contribution matrix Φ obtained through SHAP analysis is used as the input data for the K-means algorithm. Each row in the matrix represents the SHAP value of a sample across all features. First, K samples are randomly selected as the initial cluster centers. Then, the distance from each sample point to the cluster center is calculated, and the sample point is assigned to the nearest cluster, as shown in the following formula:

[0041]

[0042] Where, μ k Represents the SHAP value of the j-th feature of the i-th sample; |C k | represents the number of samples in the k-th cluster; by comparing the average SHAP values ​​of each feature, the dominant feature that contributes the most to the oscillation classification result in this scenario is identified, as shown in the following formula:

[0043]

[0044] Where, j * Characteristic numbering for the dominant factor;

[0045] Further statistical analysis was conducted on the frequency of the feature with the largest SHAP value for each sample within this cluster, for cluster C. k For all samples within the range, the indicator function formula is defined as follows:

[0046]

[0047] Then feature j in cluster C k The probability formula for internal features becoming the dominant feature is as follows:

[0048]

[0049] This reflects the dominant distribution of features in actual broadband oscillation samples;

[0050] If the dataset contains label information such as inverter control parameters or grid connection status, a correlation coefficient such as the Pearson correlation coefficient can be used to measure the strength of the association between cluster centers and control parameters, as shown in the following formula:

[0051]

[0052] Where x represents the SHAP value of a dominant feature, and y represents the control parameter.

[0053] Furthermore, in step S6, the differential visualization includes:

[0054] Heat maps and radar charts were used to illustrate the severity of the oscillations and the dominant frequency.

[0055] The relationship between the dominant characteristics and control parameters of clusters is displayed hierarchically by geographical region and time segment;

[0056] Risk level warnings are achieved through color grading, and oscillation mode and key parameter traceability information are superimposed.

[0057] The present invention has the following advantages and beneficial effects:

[0058] This invention first proposes a systematic data-driven simulation and feature sample generation method for complex scenarios involving high proportions of new energy sources and multiple control strategies connected to the grid. By flexibly combining unit access states and parameter disturbances, it can fully reflect the wide-band oscillation characteristics of the actual system across multiple sources, modes, and operating conditions, greatly improving the method's scenario adaptability and engineering application value.

[0059] This invention utilizes advanced machine learning algorithms such as Support Vector Machine (SVM) to automatically and intelligently classify and identify broadband oscillation samples, achieving efficient and accurate determination of system oscillation states. This method overcomes the limitations of traditional rules and empirical thresholds, enabling intelligent classification of large-scale, multi-type samples and significantly improving the automation level of broadband oscillation monitoring.

[0060] This invention innovatively introduces SHAP interpretability analysis and K-means clustering attribution method to quantitatively interpret the decision-making process and dominant relationship of features output by the SVM classification model. Through the SHAP algorithm, the dominant characteristics and key parameters of oscillation risk under different operating conditions can be clearly identified, and the dominant influencing factors under different oscillation mechanisms and scenarios can be automatically summarized by the K-means algorithm. Furthermore, by combining with a wide-area monitoring platform to achieve differentiated and hierarchical visualization of oscillation severity and main causes, the transparency, traceability, and engineering practicality of oscillation monitoring results are greatly enhanced, providing strong decision support for risk warning, differentiated regulation, and precise handling of broadband power grid oscillations. Attached Figure Description

[0061] Figure 1 This is a flowchart of the method of the present invention;

[0062] Figure 2 This is a grid topology diagram for hybrid control mode;

[0063] Figure 3 This is a schematic diagram of the oscillation stability identification process based on SVM;

[0064] Figure 4 This is a diagram illustrating the synergistic effect between the SVM and SHAP algorithms. Detailed Implementation

[0065] Example 1

[0066] like Figure 1 As shown, a method for intelligent identification of broadband oscillations in power grids considering a hybrid control mode of new energy sources is described, with the following steps:

[0067] S1. Build a complex hybrid control grid-connected system simulation model that includes various new energy units such as grid-connected and grid-following control, and define the access status and adjustable control parameters of various units.

[0068] This embodiment considers a complex hybrid control mode system involving multiple grid-connected control units, including grid-connected (such as virtual synchronous machine control VSG, droop control, matching control, etc.) and grid-following control methods, and covers various new energy grid-connected units such as wind power and photovoltaic power. The overall line topology diagram is as follows. Figure 2 The diagram illustrates a typical multi-inverter hybrid control grid-connected system structure. Each renewable energy inverter unit is connected to the main grid via its point of common coupling (PCC). Different inverters employ different control strategies, independently adjusting parameters such as active power (P), reactive power (Q), voltage amplitude (V), and phase angle (θ). Due to the significant differences in dynamic response characteristics between these control strategies, the overall system dynamics are more complex. Especially during large disturbances or operating condition switching, the system may generate multiple oscillation modes at different frequencies. These modes differ significantly in frequency and damping, easily superimposing to form broadband oscillations.

[0069] In a hybrid control grid-connected system, the mechanism of broadband oscillation can be explained using the following formulas and concepts:

[0070]

[0071] Among them, H k (f) is the system transfer function under the k-th control method, U k Y(f) represents the corresponding disturbance input, and Y(f) represents the system output response in the frequency domain.

[0072] S2. By randomly adjusting the connection status and control parameter disturbance methods of the units, a large number of simulation data samples with different operating conditions are generated; then, using FFT spectrum analysis technology, the data scenarios are classified into stable, wideband oscillation and other scenarios according to the frequency energy distribution and modal characteristics of the system.

[0073] During sample generation, each generating unit can freely choose whether to connect to the grid; that is, each unit can be set to "1" (connected) or "0" (disconnected). Different connection combinations simulate various typical and extreme system operation scenarios. After determining each connection combination, the control parameters of each generating unit (such as inertia, damping, PI parameters, etc.) can be further adjusted, and these parameters can be randomly perturbed within a reasonable range. Through parameter perturbation, the impact of changes in equipment parameters and uncertainties in operating conditions during actual operation can be fully simulated, thereby generating more diverse data samples. The specific steps are as follows:

[0074] 1) Randomly set the access status (1 / 0) of each grid-connected unit and initialize all control parameters to a set of values ​​within a reasonable range.

[0075] 2) For each set of access and parameter combinations, run the system time-domain simulation and record the changes of key variables (such as voltage, frequency, output power, PCC point signal, etc.) over time.

[0076] 3) Perform frequency domain analysis on the simulation results, use FFT and other methods to extract the spectral characteristics of voltage or power, and classify them according to oscillation mode, energy distribution and frequency range.

[0077] 4) Based on the dynamic response characteristics of the system, the data samples are divided into stable scenarios (oscillations are suppressed), wideband oscillation scenarios (the system has obvious energy distribution or multimodal coupling in a wide frequency band above 2Hz), and other types of scenarios.

[0078] Through the above process, stable domain datasets, wideband oscillation datasets, and other scenario datasets can be obtained, providing a rich data foundation for subsequent machine learning modeling and the development of intelligent oscillation identification methods.

[0079] To further quantify and differentiate the severity and dominant frequency distribution of system oscillations, this embodiment proposes a wideband Main Oscillation Distribution Index (MODI). This index can not only quantitatively measure the overall deviation of oscillations from the steady state, but also directly show the frequency points where the system's oscillation energy is mainly distributed, thus providing more intuitive and usable information for oscillation mode identification and suppression.

[0080] Specifically, the frequency domain spectral distribution is first obtained by performing an FFT transform on the time-domain data obtained from the simulation, as shown in the following formula:

[0081]

[0082] Where x[n] is the nth sampling point of the time-domain sequence; X[k] is the spectral amplitude (complex number) of the kth frequency component; N is the signal length; and j is the imaginary unit.

[0083] For a given disturbance condition, the broadband dominant oscillation distribution index is defined as:

[0084]

[0085] Where X(f) represents the spectral amplitude under the disturbance condition; X ss (f) represents the spectral amplitude under steady state; F osc This represents a set of wideband oscillation frequency bands (selecting the frequency bands of interest, such as 2–500 Hz); F mainThe set of dominant oscillation frequencies is typically selected from the top k frequencies with the largest energy proportions (e.g., k=3); argmax is used to find all f∈F osc Within the range, the frequency points with the greatest difference; the function of Top-k is to select the top k frequencies with the greatest difference.

[0086] This indicator has the following characteristics:

[0087] 1) Quantitatively reflect the degree of deviation of the overall system oscillation, and ensure the comparability of indicators.

[0088] 2) Highlight the dominant oscillation frequency. By selecting several frequency points with the highest energy, it is convenient to directly display the position and amplitude of the dominant mode under broadband oscillation.

[0089] In summary, MODI can not only serve as a comprehensive evaluation standard for the severity of system oscillations, but also assist in automatically identifying the most dominant oscillation frequency under the current operating conditions, providing theoretical support for subsequent pattern recognition and suppression measures.

[0090] S3. The Support Vector Machine (SVM) algorithm is used to classify the oscillation state of different scenario data obtained from the simulation, taking multi-dimensional feature data such as voltage amplitude, frequency, and power as input, and quickly distinguishing between stable scenarios and wideband oscillation scenarios.

[0091] To achieve efficient and accurate differentiation of large amounts of complex simulation data, this embodiment introduces the Support Vector Machine (SVM) algorithm in the oscillation stability classification module, replacing traditional ensemble learning methods such as random forests. SVM is a typical supervised machine learning method with good generalization ability and high-dimensional data processing capabilities. It can find the optimal hyperplane to distinguish different categories (such as "stable" and "oscillating" states) in the sample space with maximum margin, thereby significantly improving the reliability and robustness of oscillation classification. A flowchart is shown below. Figure 3 As shown:

[0092] The basic idea of ​​SVM classification is to construct a separating hyperplane in a high-dimensional space that maximizes the margin between the two classes of samples. For a given training dataset {(x... i ,y i )}, where x i Let y be the eigenvector. i Let ∈{-1,+1} be the class labels. The classification decision function of SVM can be defined as:

[0093]

[0094] Where, α i Let K(x) be the Lagrange multiplier (weighting coefficient). iLet x be the kernel function (commonly such as linear kernel, radial basis function kernel, etc.), and b be the bias. The introduction of the kernel function enables SVM to achieve nonlinear classification in the original feature space. This involves transforming complex oscillating data to a higher-dimensional space through kernel mapping, and then searching for the optimal classification hyperplane in this space. Formally, this can be represented as:

[0095] K(x i ,x j )=φ(x i ) T φ(x j (6)

[0096] Where φ(·) is a nonlinear mapping function. For a common Gaussian radial basis function (RBF), we have:

[0097] K(x i ,x j )=exp(-γ||x i -x j || 2 (7)

[0098] By appropriately selecting kernel functions and hyperparameters, SVM can effectively adapt to the complex distribution characteristics of broadband oscillating data in the parameter space. Furthermore, to avoid model overfitting, this invention employs cross-validation and regularization parameter adjustment during SVM training, effectively balancing model complexity and classification accuracy.

[0099] In practical oscillation identification applications, the input features of SVM can include multi-dimensional indicators such as the voltage and frequency of each node, the power at the PCC point, and the amplitude of the dominant oscillation frequency. After training, the SVM model can automatically output its stability criteria (such as "wideband oscillation" or "stable operation") for new operating condition samples, realizing automated and intelligent classification of large-scale, multi-source simulation conditions. Overall, the introduction of SVM enhances the applicability and engineering practical value of this invention in practical wideband oscillation monitoring, early warning, and cluster analysis, laying a solid foundation for subsequent feature extraction and dominant factor attribution analysis.

[0100] S4. Use the SHAP algorithm to analyze the feature contribution of the support vector machine classification model in step three, quantify the contribution of different input features to the oscillating classification results, and realize the interpretability analysis of the model's discrimination results.

[0101] This embodiment uses the SHAP (SHapley Additive exPlanations) analysis method to quantify the contribution of each input feature, as follows:

[0102] First, perform full combinatorial sampling on all input feature sets N. For each feature i, evaluate the change in the model's predicted output before and after its inclusion, and then calculate the SHAP value φ of feature i. i The principle behind SHAP value calculation is as follows:

[0103]

[0104] Where S is a subset that does not contain feature i, f S (x) represents the model output when only feature S is included, f S∪{i} (x) represents the output when feature i is added to S. The SHAP value essentially reflects the marginal contribution of feature i to the final discrimination result of the model.

[0105] Then, for multiple samples, the SHAP values ​​can be used to statistically represent the feature importance in the following matrix form:

[0106]

[0107] Where, φ i This represents the SHAP value of the j-th feature of the i-th sample.

[0108] In this embodiment, the above method can be used to clarify the dominant role of different operating conditions and characteristics (such as active power, reactive power, voltage amplitude, dominant frequency, etc.) in oscillation discrimination, realizing interpretable analysis and root cause tracing of broadband oscillations, and providing a data foundation for subsequent parameter optimization and feature selection. The synergistic effect between SVM and SHAP algorithms is as follows: Figure 4 As shown:

[0109] S5. Using the K-means clustering algorithm, unsupervised clustering analysis is performed with the feature contribution matrix obtained from SHAP analysis as input. By statistically analyzing the average SHAP value of features within each cluster, the dominant influencing features under different broadband oscillation scenarios are identified, and the key factors causing broadband oscillations in the system are clarified.

[0110] This embodiment introduces the K-means clustering algorithm to perform unsupervised clustering analysis on the SHAP value matrix, in order to further explore the dominant factors influencing broadband oscillations. The objective function of the K-means clustering algorithm is to minimize the sum of squared Euclidean distances from samples within a cluster to the cluster center, as shown in the following equation:

[0111]

[0112] Where K is the number of clusters, C k Let μ be the sample set of the k-th cluster. k Let x be the center of the k-th cluster, and x be the SHAP value vector of each sample point, i.e., a single sample.

[0113] In practical applications, the feature contribution matrix Φ obtained through SHAP analysis is used as the input data for the K-means algorithm. Each row in the matrix represents the SHAP value of a sample across all features. First, K samples are randomly selected as the initial cluster centers. Then, the distance from each sample point to the cluster center is calculated, and the sample point is assigned to the nearest cluster, as shown in the following formula:

[0114]

[0115] Where, μ k Let |C| represent the SHAP value of the j-th feature of the i-th sample. k | represents the number of samples in the k-th cluster. By comparing the average SHAP values ​​of each feature, the dominant feature that contributes the most to the oscillation classification result in this scenario can be identified, as shown in the following formula:

[0116]

[0117] Where, j * The characteristic number of the dominant factor.

[0118] Secondly, to enhance the robustness of the analysis, we can further statistically analyze the frequency of the feature with the largest SHAP value for each sample within the cluster. For cluster C... k For all samples within the range, the indicator function formula is defined as follows:

[0119]

[0120] Then feature j in cluster C k The probability formula for internal features becoming the dominant feature is as follows:

[0121]

[0122] This frequency statistics can reflect the dominant distribution of features in actual broadband oscillation samples.

[0123] Third, if the dataset contains label information such as inverter control parameters (e.g., inertia, damping coefficient, PI parameters, etc.) or grid connection status, cross-analysis can be used to explore the correlation between the dominant characteristics of each cluster and the system physical parameters. Correlation analysis (e.g., Pearson correlation coefficient) can be used to measure the strength of the correlation between cluster centers and control parameters.

[0124]

[0125] Where x represents the SHAP value of a dominant feature, and y represents the control parameter.

[0126] The final clustering results can be visualized using methods such as principal component analysis projection, radar charts, or heatmaps, allowing maintenance personnel to intuitively grasp the dominant influencing factors and their patterns under various broadband oscillation scenarios. The introduction of the K-means algorithm in this embodiment provides an efficient data analysis tool for the automatic clustering and dominant factor identification of broadband oscillation scenarios. Through clustering, the dominant influencing characteristics under different oscillation mechanisms and scenarios can be automatically summarized, laying a solid foundation for subsequent differentiated monitoring, risk warning, and parameter optimization processes. This process promotes the transformation of broadband oscillation governance from traditional experience-based analysis to intelligent, data-driven approaches, significantly improving the interpretability and decision support capabilities of system operation.

[0127] S6. The above oscillation classification results, SHAP feature contribution analysis results, and dominant factors of clustering identification are uniformly integrated into the power system wide-area monitoring platform. Through real-time and differentiated visualization (such as heat maps, radar charts, etc.), the platform provides an intuitive presentation and graded early warning of broadband oscillation risks, offering precise decision-making basis and proactive control support for power grid operators. Specifically, the oscillation classification results, oscillation severity indicators, and dominant factors of clustering analysis obtained from previous analyses are comprehensively integrated and connected to the power system wide-area monitoring platform. In the specific implementation process, the monitoring results of each region and node are first aggregated into a unified platform database through a data interface, enabling multi-point concurrent monitoring of the entire power grid. The platform can display data in layers according to multiple dimensions such as geographical region, time segment, or oscillation type.

[0128] On the user interface, the system can use various methods such as visualization charts, heatmaps, and color grading to intuitively present key information such as the severity of broadband oscillations (e.g., MODI value), dominant oscillation frequency, and characteristic dominant factors. For example, for areas where broadband oscillations are detected, the platform can intuitively mark their oscillation risk level with different colors such as red, orange, and yellow, and overlay clustering information of dominant oscillation modes and key control parameters to help maintenance personnel quickly identify the source of risk.

[0129] Furthermore, the wide-area monitoring platform supports real-time dynamic updates and historical trend retrospection, enabling automatic tracking and tracing of oscillation events. The platform can also set multi-level alarm thresholds, automatically triggering tiered alerts and sending alarm information to maintenance personnel or the dispatch center via SMS, email, or push notifications. Combined with cluster analysis results, the system can clearly identify the dominant factors causing oscillations and potential key control strategies, supporting maintenance personnel in accurately locating and quickly implementing targeted measures.

[0130] By unifying the management and differentiated display of the above-mentioned oscillation classification, quantitative indicators and main cause analysis results on the same monitoring platform, not only is the visualization capability and intelligence level of system operation improved, but also strong data and technical support is provided for power grid safety prevention and control, intelligent dispatch and operation and maintenance decision-making, significantly enhancing the power grid's proactive defense and response capabilities against high-risk events such as broadband oscillations.

Claims

1. A method for intelligent identification of broadband oscillations in power grids considering a hybrid control mode of new energy sources, characterized in that, Includes the following steps: S1: Build a simulation model of a hybrid control grid-connected system for new energy generating units that includes grid-connected control and grid-following control, and define the unit access status and adjustable control parameters. S2: By randomly adjusting the connection status and control parameter disturbance method of the unit, simulation data samples of different operating conditions are generated. Through FFT spectrum analysis, the data scenarios are classified into stable scenarios, wideband oscillation scenarios and other scenarios according to the frequency energy distribution and modal characteristics of the system. S3: Using the Support Vector Machine (SVM) algorithm with multi-dimensional features, including voltage amplitude, frequency, and power, as input, the oscillation state of different simulation scenarios is classified to distinguish between stable scenarios and broadband oscillation scenarios. S4: The SHAP algorithm is used to analyze the feature contribution of the SVM classification model, quantifying the contribution of different input features to the oscillation classification results and enabling interpretability analysis of the model's results. S5: Using the K-means clustering algorithm with the feature contribution matrix obtained from SHAP analysis as input, unsupervised clustering analysis is performed. By statistically analyzing the average SHAP value of features within each cluster, the dominant influencing features under different broadband oscillation scenarios are identified, clarifying the key factors causing broadband oscillations in the system. S6: Integrate the oscillation classification results, SHAP feature contribution analysis results, and the dominant factors of cluster identification into the wide-area monitoring platform, and unify them into the power system wide-area monitoring platform to achieve risk warning through differentiated visualization.

2. The method for intelligent identification of broadband oscillations in a power grid considering a hybrid control mode of new energy sources, as described in claim 1, is characterized in that... In step S1, the grid-type control includes virtual synchronous machine control, droop control or matching control, and the grid-type control includes frequency tracking control. The simulation model covers the topology of wind power, photovoltaic energy grid-connected units and common coupling points.

3. The method for intelligent identification of broadband oscillations in a power grid considering a hybrid control mode of new energy sources, as described in claim 1, is characterized in that... In step S2, the method for calculating the distribution index of the broadband oscillation scenario is as follows: The frequency domain spectral distribution of the simulated time-domain data is obtained by FFT transformation, as shown in the following formula: ， in, For the time-domain sequence One sampling point; For the first The spectral amplitude of each frequency component; The signal length; The imaginary unit; For a given disturbance condition, the formula for the broadband dominant oscillation distribution index is defined as follows: ， ， in, This indicates the spectral amplitude under disturbance conditions. This represents the spectral amplitude under steady-state conditions. This represents a set of wideband oscillation frequency bands, selecting the frequency bands of interest. The set of dominant oscillation frequencies is selected, with the highest energy percentage among them. The frequency; argmax is used to find all frequencies. The frequency point with the greatest difference within the range; Its function is to select the k frequencies with the greatest differences.

4. The method for intelligent identification of broadband oscillations in a power grid considering a hybrid control mode of new energy sources, as described in claim 3, is characterized in that... In step S2, the process of generating multi-condition simulation data samples is as follows: S21: Randomly set the unit connection status and initialize the control parameters to a reasonable range; S22: For each set of access and parameter combinations, run the system time-domain simulation and record the changes of key variables over time; S23: Perform frequency domain analysis on the simulation results, use FFT to extract the spectral characteristics of voltage or power, and classify them according to oscillation mode, energy distribution and frequency range; S24: Based on the dynamic response characteristics of the system, the data samples are divided into stable scenarios, wideband oscillation scenarios, and other types of scenarios.

5. The method for intelligent identification of broadband oscillations in a power grid considering a hybrid control mode of new energy sources, as described in claim 1, is characterized in that... In step S3, the classification decision function of the support vector machine algorithm is defined by the following formula: ， in, For Lagrange multipliers; For kernel functions; For bias, The optimal classifying hyperplane in the space is represented by the following formula: ， in, It is a nonlinear mapping function. The Gaussian radial basis kernel formula is as follows: ， By optimizing the classification hyperplane through cross-validation and regularization parameter adjustment, overfitting of the model is avoided. The input features of SVM include node voltage, frequency, PCC point power, and amplitude of the dominant oscillation frequency.

6. The method for intelligent identification of broadband oscillations in a power grid considering a hybrid control mode of new energy sources, as described in claim 1, is characterized in that... In step S4, the contribution of each input feature is quantitatively interpreted using the SHAP analysis method, as follows: First, for all input feature sets Perform full combination sampling for each feature The changes in the model's predicted output before and after its inclusion are evaluated, and then the features are calculated. SHAP value The principle behind SHAP value calculation is as follows: ， in, For features not included A subset of; Indicates that only features are included. Model output at that time; Indicates in Adding features based on the existing features Output at time; Then, for multiple samples, the SHAP values ​​are used to statistically determine the feature importance in the following matrix form: ， in, Indicates the first The first sample The SHAP value of each feature.

7. The method for intelligent identification of broadband oscillations in a power grid considering a hybrid control mode of new energy sources, as described in claim 1, is characterized in that... In step S5, the feature contribution matrix obtained through SHAP analysis is... As input data for the K-means algorithm, each row of the matrix contains the SHAP value of a sample across all features. First, K samples are randomly selected as the initial cluster centers. Then, the distance from each sample point to the center of each cluster is calculated, and the sample point is assigned to the nearest cluster, as shown in the following formula: ， in, Indicates the first The first sample The SHAP value of each feature; For the first The number of samples in each cluster; by comparing the average SHAP value of each feature, the dominant feature that contributes the most to the oscillation classification result in this scenario is identified, as shown in the following formula: ， in, Characteristic numbering for the dominant factor; Further analysis was conducted to determine the frequency of the feature with the highest SHAP value for each sample within the cluster. For all samples within the range, the indicator function formula is defined as follows: ， Then features In cluster The probability formula for internal features becoming the dominant feature is as follows: ， This reflects the dominant distribution of features in actual broadband oscillation samples; If the dataset contains label information about inverter control parameters or grid connection status, the Pearson correlation coefficient is used to measure the strength of the association between cluster centers and control parameters, as shown in the following formula: ， in, The SHAP value represents a dominant feature. Represents control parameters.

8. A method for intelligent identification of broadband oscillations in a power grid considering a hybrid control mode of new energy sources, as described in any one of claims 1-7, characterized in that, In step S6, the differential visualization includes: Heat maps and radar charts were used to illustrate the severity of the oscillations and the dominant frequency. The relationship between the dominant characteristics and control parameters of clusters is displayed hierarchically by geographical region and time segment; Risk level warnings are achieved through color grading, and oscillation mode and key parameter traceability information are superimposed.

Images