Method for evaluating dynamic regulation capacity of wind farm based on spatio-temporal correlation feature fusion

Abstract

This invention presents a method for evaluating the dynamic regulation capability of wind farms based on spatiotemporal correlation feature fusion, comprising the following steps: 1) obtaining data from each wind turbine in the wind farm and cleaning abnormal data to obtain normal operation data; 2) constructing a mapping relationship model of wind speed-power-theoretical pitch angle; 3) grouping and equivalencing wind turbines; 4) real-time evaluation of wind farm regulation capability, constructing a wind farm theoretical output model based on a DRN-GRU network. During the real-time evaluation, the theoretical pitch angle is obtained from the wind speed-power-theoretical pitch angle mapping relationship model based on the equivalent wind speed and equivalent power of the wind turbine group. The spatiotemporal correlation feature matrix is ​​then corrected, and the corrected spatiotemporal correlation matrix is ​​input into the wind farm theoretical output model of the DRN-GRU network to obtain the theoretical output of the wind farm. This theoretical output is compared with the current wind farm output to obtain the active power regulation margin of the wind farm. The wind farm regulation capability is then evaluated in real time based on the active power regulation margin.

Description

Technical Field

[0001] This invention belongs to the field of wind power generation, specifically relating to a real-time evaluation method for the dynamic adjustment capability of wind farms based on the fusion of spatiotemporal correlation features. Background Technology

[0002] Wind power is an important component of China's energy structure. In recent years, China's installed wind power capacity has steadily increased. Accurate real-time assessment of wind farm regulation capabilities is the foundation for wind farm frequency and voltage regulation, and is of great significance to the economy, stability, and power generation reliability of wind power systems.

[0003] Unlike conventional wind turbines, the real-time adjustability of wind turbines is difficult to accurately perceive due to constraints such as fluctuations in wind resources within the wind farm, the safety of wind turbine mechanical loads, and differences in wind turbine operating parameters. The real-time adjustment capability of a wind farm with a large number of turbines aggregated is also difficult to accurately assess, mainly due to two aspects:

[0004] 1. Difficulty in calculating dynamic regulation capacity: The single-unit accumulation method requires accurate and complete unit information, which is difficult to implement in engineering; the complex mapping law between the equivalent parameters of the wind turbine group and the aggregated regulation capacity is difficult to determine when using the wind turbine group aggregation equivalent method.

[0005] 2. Difficulty in cluster assessment of wind turbines: Due to the influence of spatial location and temporal factors, the highly differentiated operating status and dispersed distribution characteristics of wind turbines make cluster modeling extremely difficult.

[0006] Real-time regulation capability assessment methods for wind farms involve real-time monitoring of the operating status of a large number of wind turbines and real-time calculation of their regulation capabilities. Existing real-time regulation capability assessments directly evaluate the overall farm capacity, failing to consider the differences between wind turbines due to the influence of micro-topography and micro-meteorology. Real-time assessments using theoretical wind speed-power curves do not account for the temporal dependencies between wind turbines; therefore, the accuracy of real-time assessments needs further improvement. To address this, considering the influence of micro-topography and micro-meteorology, grouping wind farm turbines together, and establishing a theoretical power output model of the wind farm considering the spatiotemporal correlation characteristics between the groups, can yield more reliable results for real-time wind farm regulation capability assessment. Summary of the Invention

[0007] The purpose of this invention is to provide a real-time evaluation method for the dynamic adjustment capability of wind farms based on spatiotemporal correlation feature fusion. This method uses the spatiotemporal correlation feature matrix of wind farms as a foundation and leverages a DRN-GRU network to obtain accurate real-time evaluation results of the wind farm's dynamic adjustment capability.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows:

[0009] A method for evaluating the dynamic regulation capability of wind farms based on spatiotemporal correlation feature fusion, characterized by comprising the following:

[0010] 1) Obtain data from each wind turbine in the wind farm and perform abnormal data cleaning to obtain normal operation data;

[0011] 2) Construct a mapping model of wind speed, power, and theoretical blade pitch angle.

[0012] Based on the normal operation data of the wind turbine in step 1, a mapping relationship between wind speed, power, and pitch angle is established. This mapping relationship is established by a neural network. Its training process takes wind speed and power as input and pitch angle as output to obtain a mapping relationship model of wind speed-power-theoretical pitch angle.

[0013] 3) Equivalent value of wind turbine group

[0014] For the normal operation data in step 1), the sliding time window method is used to perform statistical analysis on the wind turbine data in each time window. The first difference of the wind speed time series is defined as the wind speed change. The relationship between the wind speed change and the wind speed is the time series dependency feature. Based on the time series dependency feature, the K-means clustering method is used to dynamically group the wind farm units into m wind turbine groups. The weighted method is used to calculate the equivalent wind speed, equivalent wind direction and equivalent pitch angle of each wind turbine group.

[0015] 4) Real-time assessment of wind farm regulation capacity

[0016] A sliding time window method is used to determine the benchmark wind turbine group at each moment. The benchmark wind turbine group has the highest proportion of wind turbines and the highest correlation with other wind turbine groups. Based on the Shapley method, the equivalent wind speed weighted correlation coefficient, equivalent pitch angle weighted correlation coefficient, and equivalent wind direction weighted correlation coefficient between the benchmark wind turbine group and other wind turbine groups are obtained. Based on the spatiotemporal correlation between the benchmark wind turbine group and other wind turbine groups, a spatiotemporal correlation feature matrix of the wind farm at that moment is established.

[0017] The spatiotemporal correlation feature matrix has a dimension of m*m. The first element of the matrix is ​​the equivalent wind speed of the benchmark wind turbine group, the diagonal element is the weighted correlation coefficient of the equivalent wind speed between the benchmark wind turbine group and other wind turbine groups, the first column element is the weighted correlation coefficient of the equivalent pitch angle between the benchmark wind turbine group and other wind turbine groups, the first row element is the weighted correlation coefficient of the equivalent wind direction between the benchmark wind turbine group and other wind turbine groups, and the remaining elements are 0.

[0018] A theoretical power output model for wind farms based on DRN-GRU network is constructed. This model takes the spatiotemporal correlation feature matrix of wind farm as input and wind farm power as output. During the operation of wind farm, as the time window moves, a grouping information sequence of {spatiotemporal correlation feature matrix, wind farm power} is generated. This sequence is substituted into the DRN-GRU network for training to obtain the theoretical power output model of wind farms based on DRN-GRU network.

[0019] During the real-time evaluation process, the theoretical pitch angle is obtained from the equivalent wind speed and equivalent power of the wind turbine group by the wind speed-power-theoretical pitch angle mapping model. The weighted correlation of the theoretical pitch angle is calculated to obtain the real-time theoretical equivalent pitch angle weighted correlation coefficient. The information of the equivalent pitch angle weighted correlation coefficient in the spatiotemporal correlation feature matrix is ​​corrected by the theoretical equivalent pitch angle weighted correlation coefficient to obtain the corrected spatiotemporal correlation feature matrix. The corrected spatiotemporal correlation matrix is ​​input into the wind farm theoretical output model of the DRN-GRU network to obtain the wind farm theoretical output. The theoretical output of the wind farm is compared with the current wind farm output to obtain the active power regulation margin of the wind farm. The wind farm regulation capability is evaluated in real time based on the active power regulation margin.

[0020] 2. The process of cleaning the abnormal data is as follows:

[0021] When the wind speed is not greater than the cut-in wind speed V in At that time, the theoretical output of the wind turbine is always equal to 0, in [0, V in Mark negative power values ​​within the interval, and retain non-negative wind power values;

[0022] When the wind speed is greater than the cut-in wind speed V in And not greater than the cut-out wind speed V out Because the wind power fluctuations corresponding to the wind speed at the meteorological tower in the VP scatter plot are large and dispersed, most outlier data points appear at both ends of the probability distribution. First, a nonparametric kernel density estimation method is used to remove the dispersed outlier data points at both ends. Then, the wind speed is divided into n intervals of equal length to obtain the power data set corresponding to each wind speed interval. For the power data in each wind speed interval, a probability distribution function based on kernel density estimation is established. Wind power values ​​whose probability density function value is less than 25% of the maximum value or greater than 75% of the maximum value are considered outlier data. In addition, power data exceeding the rated value and their corresponding wind speeds are also marked as outlier data.

[0023] When the wind speed is greater than the cut-out wind speed V out When the wind turbine is in a stopped state, its theoretical output is always equal to 0, and the wind power in this wind speed range is marked as abnormal data;

[0024] After the above processing, the unmarked anomalies and the retained non-negative values ​​are considered normal operating data.

[0025] The process of wind turbine grouping is as follows:

[0026] For each wind turbine, a time window length of L is set, and each sliding distance is 1 sampling point. Statistical analysis is performed on the wind speed and wind speed change data within each time window. The wind speed is divided into n intervals of equal length to obtain the corresponding wind speed change data set within each wind speed interval. For the wind speed change data under each wind speed interval, a Gaussian distribution function is established to obtain the Gaussian parameters μ and σ corresponding to the current wind speed interval. The wind speed data under each wind speed interval is averaged to obtain the correspondence between the wind speed mean under each wind speed interval and the Gaussian parameters μ and σ. The functional relationship between the wind speed mean and the Gaussian parameters μ and σ is fitted under all wind speed intervals to obtain the time series fluctuation model of the wind speed for each wind turbine, which is used for K-means clustering.

[0027] The number of clusters is set to m. The wind speed value at the midpoint of the interval for each wind turbine is calculated. The wind speed value at the midpoint of the interval for each wind turbine is substituted into the time series fluctuation model of the wind speed of each wind turbine to obtain the Gaussian parameters μ and σ under the wind speed value at the midpoint of the interval. There are n intervals corresponding to n midpoint wind speed values, n Gaussian parameters μ and n Gaussian parameters σ. The Gaussian parameters μ and σ are used as the feature quantities of each wind turbine. The K-means clustering method is used to cluster the wind turbines of the wind farm into m groups with different characteristics based on the minimum Euclidean distance, thus obtaining m wind turbine groups.

[0028] The process of determining the benchmark wind turbine group is as follows: The power correlation coefficients between wind turbine groups are calculated using three correlation description methods: Pearson linear correlation coefficient, Kendall rank correlation coefficient, and mutual information. Then, the three correlation coefficients are weighted using the Shapley value method to obtain the equivalent power weighted correlation coefficient. Next, the mean of the equivalent power weighted correlation coefficients of each wind turbine group with other wind turbine groups, as well as the proportion of each wind turbine group, are calculated to obtain the evaluation coefficient θ for each wind turbine group. The group with the highest evaluation coefficient is selected as the benchmark wind turbine group. The benchmark wind turbine group will change in each time period.

[0029] Compared with existing methods, the advantages of the present invention are as follows:

[0030] (1) This invention combines data cleaning, wind turbine grouping and fusion of spatiotemporal characteristics of wind power output influencing factors to evaluate the real-time adjustment capability of wind farms, and the evaluation results are more accurate.

[0031] (2) Based on the kernel density function, wind speed and power data can be cleaned efficiently, which improves the efficiency of data cleaning.

[0032] (3) The wind farm units are creatively grouped and valued based on the temporal characteristics of wind speed, and the wind turbine groups are weighted by correlation based on the Shapley method. The spatiotemporal correlation characteristics of the wind farm are fully considered, and a neural network evaluation model based on DRN-GRU is constructed, which increases the accuracy of the evaluation results. Attached Figure Description

[0033] Figure 1 Scatter plot of wind speed and power before kernel density cleaning.

[0034] Figure 2 Wind speed-power scatter plot after kernel density cleaning.

[0035] Figure 3 Wind speed time-series fluctuation model diagram for wind turbine No. 1. Detailed Implementation

[0036] The real-time evaluation method for the dynamic adjustment capability of wind farms based on spatiotemporal correlation feature fusion of this invention includes the following:

[0037] 1) Wind turbine anomaly data cleaning based on kernel density function

[0038] When the wind speed is not greater than the cut-in wind speed V in At that time, the theoretical output of the wind turbine is always equal to 0. Therefore, in [0, V] in Marking negative power values ​​within the interval while retaining non-negative wind power values ​​is beneficial for subsequent processing of other abnormal data.

[0039] When the wind speed is greater than the cut-in wind speed V in And not greater than the cut-out wind speed V out Because the wind power fluctuations corresponding to the wind speed of the meteorological tower in the VP scatter plot are large and scattered, and most of the abnormal data points appear at both ends of the probability distribution, the non-parametric kernel density estimation method is first used to remove the scattered abnormal data points at both ends. Then, the wind speed is divided into n intervals of equal length to obtain the power data set corresponding to each wind speed interval. For the power data under each wind speed interval, a probability distribution function based on kernel density estimation is established. Wind power values ​​that are less than 25% of the maximum value of the probability density function or greater than 75% of the maximum value of the probability density function are considered as abnormal data. In addition, power data that are greater than the rated value and their corresponding wind speeds are also marked as abnormal data.

[0040] When the wind speed is greater than the cut-out wind speed V out When the wind turbine is shut down, its theoretical output is always equal to 0, therefore the wind power in this wind speed range is marked as abnormal data.

[0041] The above-mentioned unmarked anomalies and retained non-negative value points are normal operation data.

[0042] 2) Equivalent value of wind turbine group

[0043] For the normal operation data processed in step 1), the sliding time window method is used to perform statistical analysis on the wind turbine data in each time window. Considering the time-series dependence characteristics, a time-series fluctuation model of the wind speed of each wind turbine is established. The K-means clustering method is used to dynamically group the wind farm units, and the weighted method is used to calculate the equivalent wind speed, wind direction and pitch angle of each group.

[0044] Taking wind turbine No. 1 as an example, the first difference of the wind speed time series is defined as the wind speed change, and the dependence of the wind speed change on the wind speed is the time series dependence feature.

[0045] The construction process of the time-series fluctuation model of wind speed is as follows: For each wind turbine, a time window length of L is set, and the sliding distance is 1 sampling point. Statistical analysis is performed on the wind speed and wind speed change data within each time window. The wind speed is divided into n intervals of equal length to obtain the corresponding wind speed change data set within each wind speed interval. For the wind speed change data under each wind speed interval, a Gaussian distribution function is established to obtain the corresponding Gaussian parameters μ and σ under the current wind speed interval. The wind speed data under each wind speed interval is averaged to obtain the correspondence between the wind speed mean and the Gaussian parameters μ and σ under each wind speed interval. The functional relationship between the wind speed mean and the Gaussian parameters μ and σ is fitted under all wind speed intervals (i.e., the relationship between the wind speed mean and the Gaussian function μ, and the relationship between the wind speed mean and the Gaussian function σ) to obtain the time-series fluctuation model of wind speed for each wind turbine (for clustering).

[0046] The number of clusters is set to m. The wind speed value at the midpoint of the interval for each wind turbine is calculated. The wind speed value at the midpoint of the interval for each wind turbine is substituted into the time series fluctuation model of the wind speed of each wind turbine to obtain the Gaussian parameters μ and σ under the wind speed value at the midpoint of the interval. There are n intervals corresponding to n midpoint wind speed values, n Gaussian parameters μ and n Gaussian parameters σ. The Gaussian parameters μ and σ are used as the feature quantities of each wind turbine. The K-means clustering method is used to cluster the wind turbines of the wind farm into m groups with different characteristics based on the minimum Euclidean distance, thus obtaining m wind turbine groups.

[0047] Furthermore, using power as a benchmark, a weighted method is employed to calculate the equivalent wind speed, equivalent wind direction, and equivalent pitch angle for each wind turbine group. Each wind turbine group is then treated as a single unit, denoted as the equivalent wind turbine for each wind turbine group. The data for the wind turbine groups described below are calculated based on the equivalent wind turbine, and the specific calculation process can be implemented using existing technologies.

[0048] 3) Real-time assessment of wind farm regulation capacity

[0049] By employing a sliding time window method, a benchmark wind turbine group is determined at each moment. Based on the spatiotemporal correlation between the benchmark wind turbine group and other wind turbine groups, a spatiotemporal correlation feature matrix of the wind farm at that moment is established. This integrates the factors affecting wind farm power into a whole, ensuring that the factors affecting wind farm power at each moment correspond one-to-one with the theoretical power of the wind farm. Furthermore, a theoretical power output model of the wind farm is established based on the DRN-GRU neural network, and the model is compared with the actual power output of the wind farm to achieve real-time evaluation of the wind farm's regulation capability.

[0050] (1) Correlation evaluation method based on Shapley value

[0051] To improve the accuracy of spatiotemporal correlation description, three correlation evaluation methods, namely Pearson linear correlation coefficient, Kendall rank correlation coefficient, and mutual information, are used to describe the spatiotemporal correlation among aircraft clusters. The weighted correlation coefficient among aircraft clusters is obtained based on the Shapley value method.

[0052] Let the Pearson linear correlation coefficient, Kendall rank correlation coefficient, and mutual information between any two wind turbine groups be ρ, ... P ρ K ρ M There are a total of 7 combinations of the three correlation coefficients, and each combination and its correlation contribution G{} are as follows:

[0053]

[0054] Shapley values ​​G of the three correlation coefficients P G K G M They are respectively:

[0055]

[0056] Then, according to formulas (1) and (2), the weights of the three types of correlation are as follows:

[0057]

[0058] The weighted correlation coefficient between the two wind turbine groups is:

[0059] ρ=w P ·ρ P +w K ×ρ K +w M ·ρ M (4)

[0060] In the formula, ρ P ω is the Pearson linear correlation coefficient among the aircraft clusters. P ρ represents the weights of the Pearson linear correlation coefficient.K Let ω be the Kendall rank correlation coefficient among the clusters. K ρ represents the weights of the Kendall rank correlation coefficient. M For mutual information between machine groups, ω M The weights for mutual information.

[0061] (2) Determine the benchmark wind turbine group:

[0062] The benchmark wind turbine group in a wind farm should have a high proportion of wind turbines and a high correlation with other wind turbine groups. Therefore, the pairwise equivalent power correlation coefficients between each wind turbine group are calculated first. In this embodiment, three correlation description methods—Pearson linear correlation coefficient, Kendall rank correlation coefficient, and mutual information—are applied to calculate the correlation coefficients between wind turbine groups, which can more comprehensively calculate the power correlation between wind turbine groups. Then, the three correlation coefficients are weighted based on the Shapley value method to obtain the equivalent power weighted correlation coefficient. The mean of the equivalent power weighted correlation coefficients between each wind turbine group and other wind turbine groups, as well as the proportion of wind turbines in each wind turbine group, are calculated to obtain the evaluation coefficient θ of each wind turbine group. The wind turbine group with the highest evaluation coefficient is selected as the benchmark wind turbine group. The benchmark wind turbine group will change in each time period.

[0063] Taking Unit 1 and Unit 2 as examples, the three correlation coefficients of the equivalent power of the two groups are the Pearson linear correlation coefficient ρ. Pp Kendall's rank correlation coefficient ρ Kp The correlation coefficient ρ between the two information elements is 1. Mp There are a total of 7 combinations, and the correlation contribution G{} of each combination is as follows:

[0064]

[0065] Then the Shapley value G of the three correlation coefficients Pp G Kp G Mp They are respectively:

[0066]

[0067] The weights w corresponding to the three correlation coefficients are as follows:

[0068]

[0069] The equivalent power weighted correlation coefficient ρ between Unit 1 and Unit 2 is then calculated. p12 for:

[0070] ρ p12 =w Pp ×ρ Pp12 +wKp ×ρ Kp12 +w Mp ·ρ Mp12 (8)

[0071] In equation (8), ρ Pp12 The Pearson linear correlation coefficient is the equivalent power of Unit 1 and Unit 2; ρ Kp12 ρ is the Kendall rank correlation coefficient for the equivalent power of Unit 1 and Unit 2. Mp12 ω represents the mutual information correlation between the equivalent power of Unit 1 and Unit 2. Pp ω represents the weights of the Pearson linear correlation coefficient for equivalent power. Kp ω represents the weights of the Kendall rank correlation coefficient of the equivalent power; Mp The weights are the mutual information of equivalent power.

[0072] Calculate the mean of the equivalent power weighted correlation coefficient between each wind turbine group and other wind turbine groups, as well as the proportion of each wind turbine group; then the evaluation coefficient of each wind turbine group can be obtained. for

[0073]

[0074] In equation (5), C k The percentage of wind turbine units in a group; ρ ave This represents the average power correlation coefficient between each wind turbine group and other wind turbine groups.

[0075] The wind turbine group with the highest evaluation coefficient in the wind farm is selected as the benchmark wind turbine group, and the corresponding equivalent wind turbine is recorded as the benchmark equivalent wind turbine.

[0076] (3) Spatiotemporal correlation feature matrix of wind farm

[0077] According to equation (4), the correlation between the equivalent wind speed, equivalent wind direction, and equivalent pitch angle between the benchmark wind turbine group and other wind turbine groups is calculated using three different correlation coefficients. The weighted correlation coefficients of equivalent wind speed, equivalent wind direction, and equivalent pitch angle are obtained using the Shapley value method. The spatiotemporal correlation feature matrix of the benchmark wind turbine group and other wind turbine groups is constructed according to equation (10) to achieve the fusion of spatiotemporal correlation features of the wind farm. Taking wind turbine group No. 1 as the benchmark wind turbine group as an example, the spatiotemporal correlation feature matrix corresponding to this moment is shown in equation (10).

[0078] Spatiotemporal correlation feature matrix: In the formula, v1 is the equivalent wind speed of the reference wind turbine group when wind turbine group No. 1 is the reference wind turbine group, and ρ dir12 ρ dir13 …ρ dir1mρ is the equivalent wind direction weighted correlation coefficient between the benchmark wind turbine group and other wind turbine groups; pit12 ρ pit13 …ρ pit1m ρ is the weighted correlation coefficient of the equivalent pitch angle between the benchmark wind turbine group and other wind turbine groups; speed12 ρ speed13 …ρ speed1m is the weighted correlation coefficient of equivalent wind speed between the benchmark wind turbine group and other wind turbine groups; m is the number of wind turbine groups in the wind farm.

[0079] The spatiotemporal correlation feature matrix has dimensions m*m. The first element is the equivalent wind speed of the benchmark wind turbine group, the diagonal elements are the weighted correlation coefficients of equivalent wind speeds between the benchmark wind turbine group and other wind turbine groups, the first column is the weighted correlation coefficients of equivalent pitch angles between the benchmark wind turbine group and other wind turbine groups, and the first row is the weighted correlation coefficients of equivalent wind direction between the benchmark wind turbine group and other wind turbine groups. All other elements are 0. Based on the spatiotemporal correlation between the benchmark wind turbine group and other wind turbine groups, the spatiotemporal correlation feature matrix integrates the power-influencing factors in the entire wind farm into a unified whole, ensuring a one-to-one correspondence between the power-influencing factors and the wind farm power at each moment.

[0080] (4) Wind farm theoretical output model based on DRN-GRU

[0081] Deep residual neural networks (DRNs) have a deep ability to recognize images; gated recurrent neural networks (GRUs) can effectively capture the correlation information between long sequences. To this end, a wind farm theoretical output model based on DRN-GRU is constructed. This model takes the spatiotemporal correlation feature matrix of the wind farm as input and the wind farm power as output. During the operation of the wind farm, as the time window moves, a sequence of {spatiotemporal correlation feature matrix, wind farm power} is generated. This sequence is then substituted into the DRN-GRU network for training.

[0082] Based on the normal operating data of the wind turbine after abnormal data cleaning in step 1, a mapping relationship between wind speed, power, and pitch angle is established. This mapping relationship is established by a neural network. Its training process takes wind speed and power as input and pitch angle as output to obtain a mapping relationship model of wind speed-power-theoretical pitch angle.

[0083] During the real-time evaluation process, the theoretical pitch angle is obtained from the wind speed-power-theoretical pitch angle mapping relationship model based on the equivalent wind speed and equivalent power of the wind turbine group, and the weighted correlation coefficient of the theoretical equivalent pitch angle is obtained in real time using equation (4). The information of the weighted correlation coefficient of the equivalent pitch angle in the spatiotemporal correlation feature matrix is ​​corrected by the weighted correlation coefficient of the theoretical equivalent pitch angle, so that the corrected spatiotemporal correlation feature matrix corresponds to the direct correlation coefficient of the pitch angle under the theoretical output state of each group. The theoretical output of the wind farm is obtained through the wind farm theoretical output model of the DRN-GRU network. By comparing it with the current wind farm output, the active power regulation margin of the wind farm is obtained, which is the regulation capability of wind power. A large margin indicates a better regulation capability.

Claims

1. A method for evaluating the dynamic adjustment capability of wind farms based on spatiotemporal correlation feature fusion, characterized in that, Includes the following: 1) Obtain data from each wind turbine in the wind farm and perform abnormal data cleaning to obtain normal operation data; 2) Construct a mapping model of wind speed, power, and theoretical blade pitch angle. Based on the normal operation data of the wind turbine in step 1, a mapping relationship between wind speed, power, and pitch angle is established. This mapping relationship is established by a neural network. Its training process takes wind speed and power as input and pitch angle as output to obtain a mapping relationship model of wind speed-power-theoretical pitch angle. 3) Equivalent value of wind turbine group For the normal operation data in step 1), the sliding time window method is used to perform statistical analysis on the wind turbine data in each time window. The first difference of the wind speed time series is defined as the wind speed change. The relationship between the wind speed change and the wind speed is the time series dependency feature. Based on the time series dependency feature, the K-means clustering method is used to dynamically group the wind farm units into m wind turbine groups. The weighted method is used to calculate the equivalent wind speed, equivalent wind direction and equivalent pitch angle of each wind turbine group. 4) Real-time assessment of wind farm regulation capacity A sliding time window method is used to determine the benchmark wind turbine group at each moment. The benchmark wind turbine group has the highest proportion of wind turbines and the highest correlation with other wind turbine groups. Based on the Shapley method, the equivalent wind speed weighted correlation coefficient, equivalent pitch angle weighted correlation coefficient, and equivalent wind direction weighted correlation coefficient between the benchmark wind turbine group and other wind turbine groups are obtained. Based on the spatiotemporal correlation between the benchmark wind turbine group and other wind turbine groups, the spatiotemporal correlation feature matrix of the wind farm at this moment is established. The spatiotemporal correlation feature matrix has a dimension of m*m. The first element of the matrix is ​​the equivalent wind speed of the benchmark wind turbine group, the diagonal element is the weighted correlation coefficient of the equivalent wind speed between the benchmark wind turbine group and other wind turbine groups, the first column element is the weighted correlation coefficient of the equivalent pitch angle between the benchmark wind turbine group and other wind turbine groups, the first row element is the weighted correlation coefficient of the equivalent wind direction between the benchmark wind turbine group and other wind turbine groups, and the remaining elements are 0. A theoretical power output model for wind farms based on DRN-GRU network is constructed. This model takes the spatiotemporal correlation feature matrix of wind farm as input and wind farm power as output. During the operation of wind farm, as the time window moves, a grouping information sequence of {spatiotemporal correlation feature matrix, wind farm power} is generated. This sequence is substituted into the DRN-GRU network for training to obtain the theoretical power output model of wind farms based on DRN-GRU network. During the real-time evaluation process, the theoretical pitch angle is obtained from the equivalent wind speed and equivalent power of the wind turbine group by the wind speed-power-theoretical pitch angle mapping model. The weighted correlation of the theoretical pitch angle is calculated to obtain the real-time theoretical equivalent pitch angle weighted correlation coefficient. The information of the equivalent pitch angle weighted correlation coefficient in the spatiotemporal correlation feature matrix is ​​corrected by the theoretical equivalent pitch angle weighted correlation coefficient to obtain the corrected spatiotemporal correlation feature matrix. The corrected spatiotemporal correlation matrix is ​​input into the wind farm theoretical output model of the DRN-GRU network to obtain the wind farm theoretical output. The theoretical output of the wind farm is compared with the current wind farm output to obtain the active power regulation margin of the wind farm. The wind farm regulation capability is evaluated in real time based on the active power regulation margin.

2. The method for evaluating the dynamic adjustment capability of wind farms based on spatiotemporal correlation feature fusion according to claim 1, characterized in that, The process of cleaning up the abnormal data is as follows: When the wind speed is not greater than the cut-in wind speed V in At that time, the theoretical output of the wind turbine is always equal to 0, in [0, V in Mark negative power values ​​within the interval, and retain non-negative wind power values; When the wind speed is greater than the cut-in wind speed V in And not greater than the cut-out wind speed V out Because the wind power fluctuations corresponding to the wind speed at the meteorological tower in the VP scatter plot are large and dispersed, most outlier data points appear at both ends of the probability distribution. First, a nonparametric kernel density estimation method is used to remove the dispersed outlier data points at both ends. Then, the wind speed is divided into n intervals of equal length to obtain the power data set corresponding to each wind speed interval. For the power data in each wind speed interval, a probability distribution function based on kernel density estimation is established. Wind power values ​​whose probability density function value is less than 25% of the maximum value or greater than 75% of the maximum value are considered outlier data. In addition, power data exceeding the rated value and their corresponding wind speeds are also marked as outlier data. When the wind speed is greater than the cut-out wind speed V out When the wind turbine is in a stopped state, its theoretical output is always equal to 0, and the wind power in this wind speed range is marked as abnormal data; After the above processing, the unmarked anomalies and the retained non-negative values ​​are considered normal operating data.

3. The method for evaluating the dynamic adjustment capability of wind farms based on spatiotemporal correlation feature fusion according to claim 1, characterized in that, The process of wind turbine grouping is as follows: For each wind turbine, a time window length of L is set, and each sliding distance is 1 sampling point. Statistical analysis is performed on the wind speed and wind speed change data within each time window. The wind speed is divided into n intervals of equal length to obtain the corresponding wind speed change data set within each wind speed interval. For the wind speed change data under each wind speed interval, a Gaussian distribution function is established to obtain the Gaussian parameters μ and σ corresponding to the current wind speed interval. The wind speed data under each wind speed interval is averaged to obtain the correspondence between the wind speed mean under each wind speed interval and the Gaussian parameters μ and σ. The functional relationship between the wind speed mean and the Gaussian parameters μ and σ is fitted under all wind speed intervals to obtain the time series fluctuation model of the wind speed for each wind turbine, which is used for K-means clustering. The number of clusters is set to m. The wind speed value at the midpoint of the interval for each wind turbine is calculated. The wind speed value at the midpoint of the interval for each wind turbine is substituted into the time series fluctuation model of the wind speed of each wind turbine to obtain the Gaussian parameters μ and σ under the wind speed value at the midpoint of the interval. There are n intervals corresponding to n midpoint wind speed values, n Gaussian parameters μ and n Gaussian parameters σ. The Gaussian parameters μ and σ are used as the feature quantities of each wind turbine. The K-means clustering method is used to cluster the wind turbines of the wind farm into m groups with different characteristics based on the minimum Euclidean distance, thus obtaining m wind turbine groups.

4. The method for evaluating the dynamic adjustment capability of wind farms based on spatiotemporal correlation feature fusion according to claim 1, characterized in that, The Shapley method involves using three correlation evaluation methods—Pearson linear correlation coefficient, Kendall rank correlation coefficient, and mutual information—to describe the spatiotemporal correlation between aircraft clusters, and obtaining the weighted correlation coefficient between aircraft clusters based on the Shapley value method. Let the Pearson linear correlation coefficient, Kendall rank correlation coefficient, and mutual information between any two wind turbine groups be ρ, ... P ρ K ρ M There are a total of 7 combinations of the three correlation coefficients, and each combination and its correlation contribution G{} are as follows: Shapley values ​​G of the three correlation coefficients P G K G M They are respectively: Then, according to formulas (1) and (2), the weights of the three types of correlation are as follows: The weighted correlation coefficient between the two wind turbine groups is: p=w P ·r P +w K ·r K +w M ·r M (4) In the formula, ρ P ω is the Pearson linear correlation coefficient among the aircraft clusters. P The weights are the Pearson linear correlation coefficients. ρ K Let ω be the Kendall rank correlation coefficient among the clusters. K ρ represents the weights of the Kendall rank correlation coefficient. M For mutual information between machine groups, ω M The weights for mutual information.

5. The method for evaluating the dynamic adjustment capability of wind farms based on spatiotemporal correlation feature fusion according to claim 1, characterized in that, The process for determining the benchmark wind turbine cluster is as follows: The power correlation coefficients between wind turbine clusters are calculated using three correlation description methods: Pearson linear correlation coefficient, Kendall rank correlation coefficient, and mutual information. Then, the three correlation coefficients are weighted using the Shapley value method to obtain the equivalent power weighted correlation coefficient. Finally, the mean of the equivalent power weighted correlation coefficients of each wind turbine cluster with other wind turbine clusters, along with the proportion of each wind turbine cluster, are calculated to obtain the evaluation coefficient for each wind turbine cluster. The group with the highest evaluation coefficient is selected as the benchmark wind turbine group, and the benchmark wind turbine group will change in each period.

Images